Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining

Lecture Notes in Social Networks

Nitin Agarwal · Nima Dokoohaki  Serpil Tokdemir Editors

Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining

Lecture Notes in Social Networks Series editors Reda Alhajj, University of Calgary, Calgary, AB, Canada Uwe Glässer, Simon Fraser University, Burnaby, BC, Canada Huan Liu, Arizona State University, Tempe, AZ, USA Rafael Wittek, University of Groningen, Groningen, The Netherlands Daniel Zeng, University of Arizona, Tucson, AZ, USA Advisory Board Charu C. Aggarwal, Yorktown Heights, NY, USA Patricia L. Brantingham, Simon Fraser University, Burnaby, BC, Canada Thilo Gross, University of Bristol, Bristol, UK Jiawei Han, University of Illinois at Urbana-Champaign, Urbana, IL, USA Raúl Manásevich, University of Chile, Santiago, Chile Anthony J. Masys, University of Leicester, Ottawa, ON, Canada Carlo Morselli, University of Montreal, Montreal, QC, Canada

More information about this series at http://www.springer.com/series/8768

Nitin Agarwal • Nima Dokoohaki • Serpil Tokdemir Editors

Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining

123

Editors Nitin Agarwal Collaboratorium for Social Media and Online Behavioral Studies (COSMOS) Information Science Department University of Arkansas at Little Rock Little Rock, Arkansas, USA

Nima Dokoohaki Intellectera AB Stockholm, Sweden

Serpil Tokdemir Collaboratorium for Social Media and Online Behavioral Studies (COSMOS) University of Arkansas at Little Rock Little Rock, Arkansas, USA

ISSN 2190-5428 ISSN 2190-5436 (electronic) Lecture Notes in Social Networks ISBN 978-3-319-94104-2 ISBN 978-3-319-94105-9 (eBook) https://doi.org/10.1007/978-3-319-94105-9 Library of Congress Control Number: 2018952350 © Springer International Publishing AG, part of Springer Nature 2019 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Acknowledgments

This effort is funded in part by the U.S. National Science Foundation (IIS-1636933, ACI-1429160, and IIS-1110868), U.S. Office of Naval Research (N00014-10-1-0091, N00014-14-1-0489, N00014-15-P-1187, N00014-16-1-2016, N00014-16-1-2412, N00014-17-1-2605, and N00014-17-1-2675), U.S. Air Force Research Lab, U.S. Army Research Office (W911NF-16-1-0189), U.S. Defense Advanced Research Projects Agency (W31P4Q-17-C-0059), and the Jerry L. Maulden/Entergy Endowment at the University of Arkansas at Little Rock. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the funding organizations. The researchers gratefully acknowledge the support.

v

Contents

Part I Emerging Social Issues Intent Mining for the Good, Bad, and Ugly Use of Social Web: Concepts, Methods, and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hemant Purohit and Rahul Pandey

3

Bot-ivistm: Assessing Information Manipulation in Social Media Using Network Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Matthew C. Benigni, Kenneth Joseph, and Kathleen M. Carley

19

Studying Fake News via Network Analysis: Detection and Mitigation . . . . Kai Shu, H. Russell Bernard, and Huan Liu

43

Predictive Analysis on Twitter: Techniques and Applications . . . . . . . . . . . . . . Ugur Kursuncu, Manas Gaur, Usha Lokala, Krishnaprasad Thirunarayan, Amit Sheth, and I. Budak Arpinar

67

Part II Fundamental Research Problems Using Subgraph Distributions for Characterizing Networks and Fitting Random Graph Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Benjamin Cabrera Testing Assessment of Group Collaborations in OSNs. . . . . . . . . . . . . . . . . . . . . . . 131 Izzat Alsmadi and Mohammad Al-Abdullah Dynamics of Overlapping Community Structures with Application to Expert Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Mohsen Shahriari, Ralf Klamma, and Matthias Jarke On Dynamic Topic Models for Mining Social Media. . . . . . . . . . . . . . . . . . . . . . . . . 209 Shatha Jaradat and Mihhail Matskin

vii

viii

Contents

Part III Broader Challenges and Impacts Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 Soon Jye Kho, Swati Padhee, Goonmeet Bajaj, Krishnaprasad Thirunarayan, and Amit Sheth Privacy in Human Computation: User Awareness Study, Implications for Existing Platforms, Recommendations, and Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 Mirela Riveni, Christiaan Hillen and Schahram Dustdar Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

About the Editors

Nitin Agarwal is a Distinguished Professor and Maulden-Entergy Endowed Chair of Information Science at the University of Arkansas at Little Rock. He is also the Director of the Collaboratorium for Social Media and Online Behavioral Studies (COSMOS). His research interests include social computing, deviant behavior modeling, mis/disinformation dissemination, computational propaganda analysis, group dynamics, social-cyber forensics, data mining, artificial intelligence, and privacy. His research has been supported by the U.S. National Science Foundation (NSF), the Army Research Office (ARO), the Office of Naval Research (ONR), the Air Force Research Laboratory (AFRL), the Defense Advanced Research Projects Agency (DARPA), and the Department of Homeland Security (DHS) with a total funding of over $10 million. He is a fellow of the prestigious International Academy, Research and Industry Association (IARIA). Dr. Agarwal received his doctorate at the Arizona State University in 2009 with outstanding dissertation recognition and was recognized as top 20 in their 20s by Arkansas Business. He has published over 100 peer-reviewed articles with several best paper awards and has been recognized as an expert in social, cultural, and behavioral modeling by several international news media organizations. Nima Dokoohaki is a senior data scientist. He is currently affiliated with Intellectera, a data science research and development company where together with cofounders he develops and delivers solutions for consumer behavior modeling and analytics. In addition, he maintains collaboration with a research group at Software and Computer Systems department of Royal Institute of Technology (KTH) as an external advisor. His research interests include trust and privacy, applied machine learning, social computing, and recommendation systems. He received his Ph.D. in information and communications technology (ICT) in 2013. The main theme of his research was how to understand and leverage the notion of social trust so online service providers can deliver more transparent and privacy-preserving analytical services to their end users. His research has been backed by European projects funded from EU FP7 and Horizon 2020 framework programs, as well as distinguished public funding organizations including Swedish Research Council ix

x

About the Editors

and Vinnova. In 2014, he received a distinguished fellowship from the European Research Consortium for Informatics and Mathematics (ERCIM). He has published over 30 peer-reviewed articles. In addition to two best paper awards, he has been interviewed for his visible research, and his lecture has been broadcasted on Swedish public television. An ACM professional member, he is a certified reviewer for prestigious Knowledge and Information Systems (KAIS) as well as occasional reviewer for recognized international venues and journals. Serpil Tokdemir is a research project analyst at the Office of Medicaid Inspector General (OMIG), Little Rock, Arkansas, USA. Dr. Tokdemir has a joint affiliation with the Collaboratorium for Social Media and Online Behavioral Studies (COSMOS) at UALR as research associate. Her work involves extracting raw data from Fraud and Abuse Detection System (FADS), cluster analysis, anomaly/outlier detection, predictive analysis and decision support systems, data visualization, content mining, and network analysis. Dr. Tokdemir obtained her Ph.D. from UALR in 2015 with support from U.S. National Science Foundation (NSF). Bringing together the computational modeling and social science theories, her dissertation explored the role of social media in coordinating online collective action in the context of Saudi Arabian Women’s campaigns for right to gender equality. She has published several articles in this domain and won the most published student distinction by Engineering and Information Technology college at UALR. She obtained her Bachelor of Science in Computer Science from Marmara University, Istanbul, Turkey, in 2003. She completed her Master of Science (MS) in Computer Science from Georgia State University in 2006, Atlanta, Georgia, USA.

Part I

Emerging Social Issues

Intent Mining for the Good, Bad, and Ugly Use of Social Web: Concepts, Methods, and Challenges Hemant Purohit and Rahul Pandey

Abstract The social web has empowered us to easily share information, express opinions, and engage in discussions on events around the world. While users of social media platforms often offer help and emotional support to others (the good), they also spam (the bad) and harass others as well as even manipulate others via fake news (the ugly). In order to both leverage the positive effects and mitigate the negative effects of using social media, intent mining provides a computational approach to proactively analyze social media data. This chapter introduces an intent taxonomy of social media usage with examples and describes methods and future challenges to mine the intentional uses of social media.

1 Introduction The rapid adoption of social media has made the activity on online social networks (OSNs) an integral part of our daily lives. As per Pew Research Center survey,1 nearly seven in every ten people in the USA use some type of OSNs (as of January 2018). The trend for the adoption of OSNs is not limited to the USA alone but worldwide, as evident from more than two billion monthly active users on Facebook across the world. The large scale of such digital connectivity comes with a medium to share information rapidly and interact with others virtually anywhere and anytime. Thus, OSNs facilitate an opportune playground for the users with varied intent (the purpose for an action), from helping others during disasters [36] to harassing and hate speech conversations [31] as well as manipulation with fake news

1 http://www.pewinternet.org/fact-sheet/social-media/.

H. Purohit () · R. Pandey Department of Information Sciences and Technology, George Mason University, Fairfax, VA, USA e-mail: [email protected]; [email protected] © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_1

3

4

H. Purohit and R. Pandey

Fig. 1 A spectrum to demonstrate the variety of user intents existing on social media for the diverse uses from the social good to ugly

[45] and bots [16]. The good, bad, and ugly uses of OSNs have a profound impact on the evolution of our society. In fact, the gush2 of fury and vitriol on the OSN companies in recent times for their inability to control the spread of disinformation [5, 28, 43] motivates the need to better understand the diverse uses of OSNs during real-world events. We have seen many examples of the diverse usage of OSNs in the last decade, such as for coordination and self-organization during different types of social activism. For instance, #BlackLivesMatter [11] and #OccupyWallStreet [10] for justice and inequality, #Metoo [54] and #ILookLikeAnEngineer [26] for eradicating workplace harassment and stereotypes, etc. Likewise, OSNs have provided a valuable information exchange platform to support and rebuild communities after catastrophic natural disasters such as #HurricaneSandy [36] and #HaitiEarthquake [30] as well as enable community healing after man-made disasters such as mass shootings [19] and terror attacks [21]. Unfortunately, OSNs have also facilitated the amplification of malicious agenda such as to harass and bully others especially youth [8, 34], to spread disinformation for alternative narratives, as well as to manipulate public opinions via fake news during elections [2, 45]. One approach to understanding the nature and motives of information sharing on OSNs is to analyze the potential intent types (c.f. Fig. 1) associated with the OSN user interactions. Recognizing user intent helps collect evidences for interpreting and predicting potential actions and consequences—analogous to the problem of plan recognition in Artificial Intelligence [47]. We can model intent based on the content of the message shared, activity logs of the user sharing the message, and the link structure in OSNs that support the information flow of the message. Given the volume, variety, and velocity of information flowing on OSNs, computational approaches of intent mining provide a promising direction to help study the varied types of intent at large scale.

2 https://journalistsresource.org/studies/society/internet/fake-news-conspiracy-theories-

journalism-research.

Intent Mining on Social Media

5

Rest of the chapter provides an extensive overview of concepts, methods, and challenges in mining intent. In particular, Sect. 2 describes related concepts for a taxonomy of intent types, Sect. 3 provides an overview of different methods to process content, user, and network structure data for modeling intent, where the data may exist in different modalities—text, images, and videos. Finally, Sect. 4 describes the challenges in mining intent for future research directions to lead the society towards the good use of OSNs.

2 Concepts This section describes first the concept of intent from multidisciplinary perspective, followed by describing the taxonomy of intent types for the diverse uses of OSNs, on a spectrum of positive to negative effects as shown in Fig. 1.

2.1 Intent: Multidisciplinary Perspective Intent in the simplest form can be defined as a purpose for an action. In a more in-depth form, one can understand the broader view of intent from the concept of “intentional stance” proposed by the well-known philosopher and cognitive scientist Daniel Dennett [13]. Intentional stance is the highest abstract level of strategies for predicting and thereby, explaining and understanding the behavior of an entity (e.g., OSN user). Likewise, in Artificial Intelligence research community, the intent recognition problem has been studied for understanding the behavior of agents in the context of goal and plan recognition [47]. The power to recognize the plans and goals of other agents enables effective reasoning about the actions of the agents. In our context of the different uses of OSNs, a user can express desires and beliefs for certain intentionality in either message content or through his interactions and activities on OSNs. Therefore, a variety of factors can affect an individual’s expression of intentionality through different information modalities. For example, “I wanna give #blood today to help the victims #sandy” shows the intent to donate blood for the desire to help and for the belief of resource scarcity to treat victims in the aftermath of Hurricane Sandy [36]. Intent can be expressed both explicitly and implicitly in a given content. Table 1 shows examples of messages with different intent types.

2.2 Intent Taxonomy Given the diverse uses of OSNs and the endless possibilities of actions, the variety of intent behind the actions would be vast. Therefore, an approach to better understand

6

H. Purohit and R. Pandey

Table 1 Modified examples (for anonymity) of OSN messages from past events, with intent expressed in the textual content Social media message M1. I want to send some clothes for hurricane relief #sandy M2. I support what u said about shooting here in Florida, Ill stand with u at any time. I am a retired teacher M3. You can find the @user student Developer Pack here: URL M4. it’s better to start 3rd-world war instead of letting Russia & assad commit #HolocaustAleppo M5. hi @user, we sincerely apologize for your inconvenience, in order to regain access to your account, please visit: URL M6. One of the suspects (according to BPD) is Sunil Tripathi. The missing Brown student NEWS reported on in March URL M7. you’re a despicable whore M8. We are a people whose true lives begin after their death. #hijrah #jihad #shahadah M9. DONT EVEN ASK EM WHO DEY WIT JUS BLOW EM FACES M10. there’s a new drink called Sandy, it is a watered down Manhatten M11. No luck needed to #SAVE up to 60% off! Visit URL details of #vacation package M12. white women have lied about rape against black men for generations M13. There’s no New Clinton, never has been. Shes same rape defending, racist, homophobic liar shes been for 70 yrs

Intent [implication] Offering Help [community rebuilding and trust for collective action] Emotional Supporting [community healing for psychological support] Expertise Sharing [improving learning from experiences of peers and mentors] Propagandizing [group-specific beliefs leading to echo chambers] Deceiving [financial frauds and stealth of personal information] Rumoring [creating uncertainty in situational awareness for poor decision support] Harassing [affecting mental health and physical well-being] Manipulating [shifting public attitude towards radicalized outfits] Bullying [threatening and creating fear and insecurity in the society] Joking [creating junk for some sections of the community] Marketing [spamming in the information ecosystem] Accusing [giving an alternative, supporting narrative to stereotypical groups] Sensationalizing [diverting from key issues and politicizing environment]

Intent can also be expressed using other information modalities (image, audio, or video); Fig. 2 shows an image example

the diverse uses and organize the associated intent for actions, we can consider a spectrum representation for the OSN uses with positive to negative effects. We also create a taxonomy of intent types as shown in Fig. 1. On the left side of the spectrum, social good uses of OSNs lead to the positive, enlightening effects of inspiration, cooperation, and trust in our society and strengthen the value of social networking in our lives. On the other hand, as we move towards the right end, social bad and ugly uses start to lead the negative effects of creating distrust, radicalization, and fear in our society. The social bad and ugly uses discredit and ruin the social networking values in our lives.

Intent Mining on Social Media

7

Fig. 2 Fake image shared during Hurricane Sandy 2012 (https://mashable.com/2012/10/29/fakehurricane-sandy-photos/#Wc2mpf4QXgqV)

The proposed spectrum of Fig. 1 is flexible to extend the OSN uses as well as the intent taxonomy with different interpretations in the future. We broadly define five types of intent: (a) intent for the social good use, (b) intent for the social bad use, (c) intent for the social ugly use, (d) intent for the mixed social good and bad uses, and (e) intent for the mixed social bad and ugly uses. The proposed intent types are described in the following with real examples of OSN messages in Table 1. (a) Intent for the Social Good Use People have good intentions and attitudes who believe in the social welfare and who would come forward to help others in the times of needs. OSNs facilitate a medium for such users to not only assist during disasters to provide emotional support and donations, but also, in general, offer help with the expertise to educate, inform, and caution-advice others. Illustrative intents in this category are: – Help Offering: to express assistance to people in need of a resource or service. For example, message M1 in Table 1 shows a user offering clothing donations for disaster relief during Hurricane Sandy [36]. Likewise, users also offer to help with resources often, like blood donation [37]. – Emotional Supporting: to express care and sympathy for someone affected by an event. For example, message M2 in Table 1 shows support for the affected community of a mass shooting event. OSNs have played such roles in supporting a community for psychological well-being and caring of the affected people from depression and trauma [19].

8

H. Purohit and R. Pandey

– Expertise Sharing: to suggest or give advice to an information seeker based on expertise. For example, message M3 in Table 1 shows the answer to a user with a query to seek resources. OSNs provide hashtag and replybased affordances for conversation chains, to allow expertise and knowledge sharing. (b) Intent for the Social Bad Use OSN users are not just humans but also social bots, who often participate for different motives in the conversations on social media. Both these types of users have contributed to create propaganda and spread the spamming content extensively in the recent years. Illustrative intents in this category are: – Propagandizing: to create certain perception or belief towards an agenda of an organization or a group. For example, message M4 in Table 1 shows a strong justification for the government policies and attempts to convince the audience to believe in them [27]. – Deceiving: to spread spam or malicious content for a financial fraud or the purposeful misleading. For example, message M5 in Table 1 shows a clickbait and a potential scam for attracting readers to malicious sites related to buying some products and then stealing personal and financial information [25]. – Rumoring: to share unverified information aligned with emotions of someone that creates uncertainty. For example, message M6 in Table 1 shows a rumor indicating an emotionally charged message during Boston bombing and drawing everyone’s attention to a misguided fact [46]. (c) Intent for the Social Ugly Use Unfortunately, OSNs have become an avenue for conspiracy theories in recent years, where fake user accounts incite social tensions and radicalize others. Furthermore OSNs provide a medium to easily connect and converse with anyone that is abused (especially among youth) to engage in the online harassment and bullying, with the strong mental health implications. Illustrative intents in this category are: – Harassing: to cause emotional distress to someone by insults, misogyny, or hateful messages and trolling for publicly shaming someone. For example, message M7 in Table 1 shows a sender harassing a receiver, which can lead to both mental and physical harm to the receiver [14]. – Manipulating: to purposefully divert a discourse to radicalize as well as politically or socially divide people. For example, message M8 in Table 1 shows how a potential member of a terror group can influence others and boost their recruitment drives [17]. – Bullying: to threaten or intimidate for creating a fear among a recipient. For example, message M9 in Table 1 shows a message of a gang member involved in illegal activities who threatens the rival gang, creating a fear in the social environment of the local region [3].

Intent Mining on Social Media

9

(d) Intent for the Mixed Social Good and Bad Uses OSN users come from all sections of our society and their participation motives can range from personal to commercial usage. In this case, not all members of the society would benefit from all the activities of such users (e.g., a repetitive irrelevant advertisement) and therefore, the OSN use can be considered as mixed. Illustrative intents in this category are: – Joking: to ridicule for fun or make a mockery of some event, object, or person. For example, message M10 in Table 1 shows a user making fun of Hurricane Sandy that may be amusing to some but contributes to the information overload on others, such as emergency services who would be working hard to monitor OSN streams for situational awareness [36]. – Marketing: to promote and advertise a product or service for selling. For example, message M11 in Table 1 shows a brand user creating a marketing pitch to attract more buyers that may be useful to some users who are looking to buy a travel package but a spam for those who are not traveling [12]. (e) Intent for the Mixed Social Bad and Ugly Uses Users on OSN platforms may hold specific beliefs and may be associated with specific ideological identities such as political, religious, and social activist groups. Thus, their propaganda activities on OSNs can be motivated to meet the purpose of those belief and ideologies, however giving rise to echo chambers, which are the drivers of conspiracies. Illustrative intents in this category are: – Accusing: to accuse someone and doubt publicly for creating an alternative reality. For example, message M12 in Table 1 shows a user trying to develop a narrative by accusing a female rape victim publicly and, thus, trying to undermine the key social issue of rape myths [4, 39]. – Sensationalizing: to provoke the audience to divert to an issue for frightening and politicizing the environment. For example, message M13 in Table 1 shows how a social issue can be mixed with a political context and divert the focus in a conversation away from the social issue (i.e., against rape) [40].

3 Methods This section presents different types of methods to mine intent types described in the previous section. Early research in online intent mining was focused on search engines, questionanswering and product review forums, ad recommendation systems as well as spam detectors in information networks. For search systems, the key challenge was to understand information seeking intent of users in the queries on search engines using logs and give the relevant results to the users. Although, user query intent covers only a few categories of the broad variety of intents possible for uses of OSNs. In particular, query intent can be navigational, informational, or transactional

10

H. Purohit and R. Pandey

information to meet a user’s information requirement [23], but the intent types in a social environment relate to communication and engagement with others in a conversation for different purposes, such as offering help or manipulating others. For question-answering and product forums, the possible intent types are centered around information seeking and knowledge sharing. For the ad recommendation systems, the commercial intent of buying and selling are priorities. For spam detection in networks, researchers focus on modeling patterns of malicious behavior but there are other types of intent possible for OSNs. Additionally, researchers have investigated intent across different modalities of information than the textual content, such as fake images [20] for rumors (see Fig. 2). Literature shows intent modeling in OSNs based on content of a message, user profile activities over time, and the network of user interactions as well as structural links of friendship and trust. We describe the methods under three major categories of content-based, user-based, and network-based approaches.

3.1 Content-Based Intent Mining This type of methods solve the problem of inferring intent from a given instance of a message content shared on an OSN. Inferring intent from content is challenging due to possibilities of multiple natural language interpretations in a given text message. Therefore, to make the intent mining problem computationally tractable, prior research primarily exploited the text classification problem format [38]. Although it is different from the well-studied text analytics tasks of topic classification (focused on the subject matter) as well as opinionated text classification of sentiment or emotion (focused on the current state of affairs). For instance, in a message “people in #yeg feeling helpless about #yycflood and wanting to help, go donate blood,” the task of topic classification focuses on the medical resource “blood,” the task of sentiment and emotion classification is focused on the negative feeling expressed for being helpless. In contrast, intent classification concerns the author’s intended future action, i.e., “wanting to help/donate.” Therefore, the choice of feature representation is different across the tasks (e.g., adjectives are considered important for capturing sentiment and emotion, and likewise, verbs are important for indicating intent or action). Given the complexity to understand intent from natural language, researchers have explored various classifier designs using both rule-based systems and machine learning techniques. Rule-based approaches are appropriate for small-scale data while for the largescale data with intent labels, machine learning approaches can be leveraged. We summarize few approaches from the literature for brevity. Among rule-based classification approaches, Ramanand et al. [41] created rules for transactional (buying–selling) wishes in the product review text (e.g., “{window of size 3} ”) and Purohit et al. [37] created rules for help-seeking and offering behavior during disasters (e.g., “ (Pronoun except you = yes) ∧ (need/want = yes) ∧ (Adjective = yes/no) ∧

Intent Mining on Social Media

11

(Thing = yes)” for seeking help about a “Thing” such as food). Among machine learning approaches, we can develop a classifier for detecting credible information messages to undermine potential rumor intent, such as Castillo et al. [7] proposed a classification method using the diverse features from message content, posting and re-tweeting behavior of users, and from citations to external sources. The key challenge of classification methods is to design good features that can efficiently capture the intent representation. Hollerit et al. [22] created a binary classifier for buying–selling posts on Twitter by exploring n-grams and POS tags-based features, and Carlos and Yalamanchi [6] proposed a supervised learning classifier for commercial intent based on features grounded in speech act theory. Purohit et al. [36, 38] proposed pattern-aided supervised classification approaches to identify the intent of help-seeking or offering during disasters, by combining the features from a bag-of-tokens model with patterns extracted from a variety of declarative and psycholinguistic knowledge sources. Likewise, Nazer et al. [33] proposed a system for identifying help-seeking request intent during disasters by combining content-based and context-based features such as the device type of a message source and location. While creating an exhaustive set of user-defined features from the user-generated content of social media can be challenging, researchers also explored deriving some valuable data-driven features for better generalization. Wang et al. [50] proposed a semi-supervised learning approach using the link prediction task in a graph of the tweet and intent-specific keyword nodes, in order to categorize intent tweets into different categories of general interests such as food and drink, travel, and goods and services. Given the possible lack of sufficient labeled data in an application domain, one can also use the transfer learning paradigm. Among such approaches, Chen et al. [9] built a combined classifier based on two classifiers trained on different source and target domains, in order to identify cross-domain intentional posts of commercial value (buying/selling) in discussion forums. Likewise, Ding et al. [15] proposed a convolutional neural network-based method for identifying user consumption intent for product recommendations, by transferring the mid-level sentence representation learned from one domain to another by adding an adaptation layer. Pedrood and Purohit [35] proposed sparse coding-based feature representation for efficient transfer learning to detect intent of help-seeking or offering in the future disaster event by exploiting data of historic disaster events.

3.2 User Profile-Based Intent Mining This type of methods solve the problem of inferring the intent of a user by exploiting the patterns of activities or messages of the user in his historic profile data. It is similar to the idea of personalized recommender systems, which exploit all the historical data of a user to create his interest profile. The primary focus of these types of methods for OSNs is to model malicious user behavior such as spamming behavior to identify spammer networks or specific orientation towards

12

H. Purohit and R. Pandey

some beliefs. We explain a few approaches for brevity. The majority of such methods for learning and modeling the user behavior from historic data rely on machine learning techniques given the possibility to leverage large-scale historic data. Intent mining literature has different methods from supervised to unsupervised learning for modeling user behavior, by leveraging features of all modalities such as text and images as well as temporal patterns of user activities. For instance, Jin et al. [25] created a detection system for users with malicious intent (spamming) by using both image and textual content features from the historic user profiles as well as the social network features. Lee et al. [29] created a supervised classifier to identify malicious content polluters using a diverse set of features from historic profile data including demographics, social network structure, the content of messages, as well as temporal behavior patterns in the activity. Among unsupervised learning approaches, Mukherjee et al. [32] proposed a method to exploit observed behavioral footprints of fake reviewers using a Bayesian framework. Furthermore, Ferrara et al. [16] review different methods for social bot detection using both feature-based and graph-based and crowdsourcing-based approaches. Beyond the bot users, human users are also involved in the social bad and ugly uses of OSNs, such as with the intents of bullying and threatening others. Squicciarini et al. [44] proposed an approach to study both the detection of cyberbullies and the identification of the pairwise interactions between OSN users, who contributed in spreading the bullying intent. Salawu et al. [42] provide an extensive survey of the state of the art cyberbullying detection approaches. Likewise, Balasuriya et al. [3] studied the problem of detecting gang member profiles on Twitter that often share messages with the threatening intent, by proposing a method of supervised classification with diverse features of tweet text, profile information, usage pattern of emoji symbols, as well as additional information from the descriptions and comments on the external links of YouTube videos. On the other side of the OSN use spectrum, we can also model the user behavior in general for understanding the intent of non-malicious kind. For instance, Tomlinson et al. [48] proposed a method to detect a user’s long-term intent and analyzing differences across cultures in expressing intent. Authors captured the latent cultural dimensions via the Singular Vector Decomposition technique. Such methods can be valuable for large-scale studies to assist multidisciplinary research at the intersection of social, humanities, and computing sciences.

3.3 Network-Based Intent Mining Methods in this category focus on inferring the intent of a user by exploiting a given network structure of social relationships of the user in an OSN. The patterns of network structure can inform the membership to spam communities as well as information propagation cascades with the distinctive signatures of fake or rumor spreading intent. The network-based approaches have an advantage of being

Intent Mining on Social Media

13

language independent of the content, although they have to deal with a challenge of acquiring the network structure for any data modeling. Social network analysis methods are valuable for extracting the structural patterns. We summarize some of these approaches next. The malicious users whether social bots or spammers or even radicalized group users often form community structures in the network, for sharing content and giving others a deceiving perception of general users. For instance, Ghosh et al. [18] investigated Twitter network for link farming—an approach to acquire a large number of follower links—by studying nearly 40,000 spammer accounts suspended by Twitter. Their analysis showed that the link farming is very common, where a majority of the links are acquired from a small fraction of Twitter users that are themselves seeking links. Likewise, a study conducted by Al-khateeb et al. [1] discovered cyber propaganda campaigns against NATO’s Trident Juncture Exercise 2015 using social network analysis. There are also approaches for combining both features of network structure and content or user interaction patterns. Yu et al. [53] proposed a subgroup detection method to identify deceptive groups from their conversations, by combining linguistic signals in the content of interactions and signed network analysis for dynamic clustering. Among the approaches to model information propagation for identifying the intent of users, Starbird [45] studied the network generated from the common URL domains in the potentially malicious user messages on Twitter, which contained alternative narratives about mass shooting events and discovered the patterns of different domains and how they connect to each other. A model proposed by Wu and Liu [52] for the propagation of messages in OSNs infers embeddings of users with network structures as well as represents and classifies propagation pathways of a malicious intent message. Jiang et al. [24] provide an extensive survey of the approaches for malicious intent behavior detection across the categories of traditional spam, fake reviews, social spam, and link farming. On the other side of the spectrum in Fig. 1 for positively using OSNs also, researchers have designed network-based approaches to glean intent of social good to help others. Welser et al. [51] identified key roles of Wikipedia editors such as substantive experts and vandal fighters by extracting patterns from edit histories as well as egocentric networks of users. Likewise, Tyshchuk et al. [49] presented a methodology that combined natural language processing and social network analysis to construct a network of actionable messages, for discovering communities and extracting leaders with a social good intent to help. In summary, the approaches described above provide an overview of how one can study a variety of intent types in OSN uses by leveraging the message content, user profile history, and the social network structure.

14

H. Purohit and R. Pandey

4 Challenges and Future Research Directions The use of OSNs in the future is going to be dependent on how OSN providers address the concerns of intent related to the social bad and ugly uses, which have given the perception that social networks are broken.3 It is an open question— how we can create OSN platform affordances that would help manage both accountability of user activities and verification for trusted user networks, while discouraging the actors with social bad intents.4 Similarly, it will be very important to boost the OSN uses with social good intents, such that we can still preserve some level of trust for OSN uses in the society. We describe some of these challenges in the following that future researchers can build on: – Profiling anonymous identities. The cases for bullying and harassing intents often include harassers with anonymous profiles. The impact of such virtual anonymity leads to a lack of accountability and trust, due to the abuse of the medium of information sharing on OSNs. The anonymous users can spread information with malicious agenda but still remain unaccountable for the consequential effects. We need to address the challenge of understanding content and interaction patterns of such anonymous profiles for designing efficient user profiling methods. – Transforming social bots. It is not clear how many users of OSNs are actually human users versus social bots, some of those present a threat to the information ecosystem of our society. While existing methods of bot detection provide some capability at scale to detect the bots, it is not clear beyond suspending them if we could alternatively transform the behavior of these bots. For example, teaching the intent behavior of social good as opposed to social bad (e.g., as observed in 2016, for the Microsoft chatbot5 ) could present an interesting opportunity to the human-in-the-loop Artificial Intelligence research. – Fixing erroneous spreading of malicious intent. Sometimes the OSN users rapidly spread unverified, fake information due to emotional provocation such as after looking at an image of a disaster-affected site, although without a malicious goal. In this case, even if the user would like to change his course of action, the current OSN affordances only allow deletion of content for that individual user but the effect on the network is not handled effectively. Future research can also investigate this challenge of how to fix the issue of controlling message propagation. – Hybrid information filtering. OSNs have been criticized lately to control what information a user can see, based on their content filtering and ranking algorithms. It leads to the formation of echo chambers with negative consequences.

3 https://www.technologyreview.com/s/610152/social-networks-are-broken-this-man-wants-to-

fix-them. 4 https://datasociety.net/output/dead-reckoning/. 5 http://www.bbc.com/news/technology-35890188.

Intent Mining on Social Media

15

There is a need for fairness and diversity in representation of information shown to a user such that the resulting content covers the varied intents of a story. It should further de-prioritize strongly subjective content and also provide an opportunity to the user to change the prioritization. To conclude, this chapter presented a detailed overview of different uses of OSNs on the spectrum of social good to social ugly and also introduced an intent taxonomy. It further described intent mining methods and future challenges, which can help discover the varied types of intent behind the uses of OSNs. Acknowledgements The authors thank Professor Amit Sheth at Kno.e.sis Center, Wright State University for valuable feedback and US National Science Foundation (NSF) for partially supporting this research on intent mining through grant award IIS-1657379. Opinions in this chapter are those of the authors and do not necessarily represent the official position or policies of the NSF.

References 1. Al-khateeb, S., Hussain, M.N., Agarwal, N.: Social cyber forensics approach to study twitter’s and blogs’ influence on propaganda campaigns. In: International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, pp. 108–113. Springer, Berlin (2017) 2. Allcott, H., Gentzkow, M.: Social media and fake news in the 2016 election. J. Econ. Perspect. 31(2), 211–36 (2017) 3. Balasuriya, L., Wijeratne, S., Doran, D., Sheth, A.: Finding street gang members on twitter. In: 2016 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 685–692. IEEE, New York (2016) 4. Boux, H.J., Daum, C.W.: At the intersection of social media and rape culture: how Facebook postings, texting and other personal communications challenge the real rape myth in the criminal justice system. Univ. Illinois J. Law Technol. Policy 1, 149 (2015) 5. Brown, J.: Is social media bad for you? the evidence and the unknowns (2018). http://www.bbc. com/future/story/20180104-is-social-media-bad-for-you-the-evidence-and-the-unknowns 6. Carlos, C.S., Yalamanchi, M.: Intention analysis for sales, marketing and customer service. In: Proceedings of COLING 2012: Demonstration Papers, pp. 33–40 (2012) 7. Castillo, C., Mendoza, M., Poblete, B.: Information credibility on twitter. In: Proceedings of the 20th International Conference on World Wide Web, pp. 675–684. ACM, New York (2011) 8. Chatzakou, D., Kourtellis, N., Blackburn, J., De Cristofaro, E., Stringhini, G., Vakali, A.: Mean birds: detecting aggression and bullying on twitter. In: Proceedings of the 2017 ACM on Web Science Conference, pp. 13–22. ACM, New York (2017) 9. Chen, Z., Liu, B., Hsu, M., Castellanos, M., Ghosh, R.: Identifying intention posts in discussion forums. In: Proceedings of the 2013 conference of the North American chapter of the association for computational linguistics: human language technologies, pp. 1041–1050 (2013) 10. Conover, M.D., Ferrara, E., Menczer, F., Flammini, A.: The digital evolution of occupy wall street. PLos One 8(5), e64679 (2013) 11. De Choudhury, M., Jhaver, S., Sugar, B., Weber, I.: Social media participation in an activist movement for racial equality. In: ICWSM, pp. 92–101 (2016) 12. De Vries, L., Gensler, S., Leeflang, P.S.: Popularity of brand posts on brand fan pages: an investigation of the effects of social media marketing. J. Interact. Mark. 26(2), 83–91 (2012) 13. Dennett, D.C.: The Intentional Stance. MIT Press, Cambridge (1989)

16

H. Purohit and R. Pandey

14. Dinakar, K., Jones, B., Havasi, C., Lieberman, H., Picard, R.: Common sense reasoning for detection, prevention, and mitigation of cyberbullying. ACM Trans. Interact. Intell. Syst. 2(3), 18 (2012) 15. Ding, X., Liu, T., Duan, J., Nie, J.Y.: Mining user consumption intention from social media using domain adaptive convolutional neural network. In: AAAI vol. 15, 2389–2395 (2015) 16. Ferrara, E., Varol, O., Davis, C., Menczer, F., Flammini, A.: The rise of social bots. Commun. ACM 59(7), 96–104 (2016). http://doi.acm.org/10.1145/2818717 17. Gates, S., Podder, S.: Social media, recruitment, allegiance and the Islamic state. Perspect. Terrorism 9(4), 107–116 (2015) 18. Ghosh, S., Viswanath, B., Kooti, F., Sharma, N.K., Korlam, G., Benevenuto, F., Ganguly, N., Gummadi, K.P.: Understanding and combating link farming in the twitter social network. In: Proceedings of the 21st International Conference on World Wide Web, pp. 61–70. ACM, New York (2012) 19. Glasgow, K., Vitak, J., Tausczik, Y., Fink, C.: Grieving in the 21st century: Social media’s role in facilitating supportive exchanges following community-level traumatic events. In: Proceedings of the 7th 2016 International Conference on Social Media & Society, p. 4. ACM, New York (2016) 20. Gupta, A., Lamba, H., Kumaraguru, P., Joshi, A.: Faking sandy: characterizing and identifying fake images on twitter during hurricane sandy. In: Proceedings of the 22nd International Conference on World Wide Web. pp. 729–736. ACM, New York (2013) 21. He, X., Lin, Y.R.: Measuring and monitoring collective attention during shocking events. EPJ Data Sci. 6(1), 30 (2017) 22. Hollerit, B., Kröll, M., Strohmaier, M.: Towards linking buyers and sellers: detecting commercial intent on twitter. In: Proceedings of the 22nd International Conference on World Wide Web, pp. 629–632. ACM, New York (2013) 23. Jansen, B.J., Booth, D.L., Spink, A.: Determining the informational, navigational, and transactional intent of web queries. Inf. Process. Manag. 44(3), 1251–1266 (2008) 24. Jiang, M., Cui, P., Faloutsos, C.: Suspicious behavior detection: current trends and future directions. IEEE Intell. Syst. 31(1), 31–39 (2016) 25. Jin, X., Lin, C., Luo, J., Han, J.: A data mining-based spam detection system for social media networks. Proc. VLDB Endowment 4(12), 1458–1461 (2011) 26. Johri, A., Karbasian, H., Malik, A., Handa, R., Purohit, H.: How diverse users and activities trigger connective action via social media: lessons from the twitter hashtag campaign# ilooklikeanengineer. In: Proceedings of the 51st Hawaii International Conference on System Sciences (2018) 27. Kavanaugh, A.L., Fox, E.A., Sheetz, S.D., Yang, S., Li, L.T., Shoemaker, D.J., Natsev, A., Xie, L.: Social media use by government: from the routine to the critical. Gov. Inf. Q. 29(4), 480–491 (2012) 28. Lazer, D.M.J., Baum, M.A., Benkler, Y., Berinsky, A.J., Greenhill, K.M., Menczer, F., Metzger, M.J., Nyhan, B., Pennycook, G., Rothschild, D., Schudson, M., Sloman, S.A., Sunstein, C.R., Thorson, E.A., Watts, D.J., Zittrain, J.L.: The science of fake news. Science 359(6380), 1094– 1096 (2018). http://science.sciencemag.org/content/359/6380/1094 29. Lee, K., Eoff, B.D., Caverlee, J.: Seven months with the devils: A long-term study of content polluters on twitter. In: ICWSM (2011) 30. Meier, P.: Digital Humanitarians: How Big Data is Changing the Face of Humanitarian Response. CRC Press, Boca Raton (2015) 31. Mondal, M., Silva, L.A., Benevenuto, F.: A measurement study of hate speech in social media. In: Proceedings of the 28th ACM Conference on Hypertext and Social Media, pp. 85–94. ACM, New York (2017) 32. Mukherjee, A., Kumar, A., Liu, B., Wang, J., Hsu, M., Castellanos, M., Ghosh, R.: Spotting opinion spammers using behavioral footprints. In: Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 632–640. ACM, New York (2013)

Intent Mining on Social Media

17

33. Nazer, T.H., Morstatter, F., Dani, H., Liu, H.: Finding requests in social media for disaster relief. In: 2016 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 1410–1413. IEEE, New York (2016) 34. O’Keeffe, G.S., Clarke-Pearson, K., et al.: The impact of social media on children, adolescents, and families. Pediatrics 127(4), 800–804 (2011) 35. Pedrood, B., Purohit, H.: Mining help intent on twitter during disasters via transfer learning with sparse coding. In: Proceedings of the 11th International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation (2018, in press). http://ist.gmu.edu/~hpurohit/informatics-lab/papers/sbp18transferlearning-camera-ready-FINAL.pdf 36. Purohit, H., Castillo, C., Diaz, F., Sheth, A., Meier, P.: Emergency-relief coordination on social media: Automatically matching resource requests and offers. First Monday 19(1) (2013). http:// firstmonday.org/ojs/index.php/fm/issue/view/408 37. Purohit, H., Hampton, A., Bhatt, S., Shalin, V.L., Sheth, A.P., Flach, J.M.: Identifying seekers and suppliers in social media communities to support crisis coordination. Comput. Supported Coop. Work 23(4–6), 513–545 (2014) 38. Purohit, H., Dong, G., Shalin, V., Thirunarayan, K., Sheth, A.: Intent classification of shorttext on social media. In: IEEE International Conference on Smart City/SocialCom/SustainCom (SmartCity), 2015, pp. 222–228. IEEE, New York (2015) 39. Purohit, H., Banerjee, T., Hampton, A., Shalin, V.L., Bhandutia, N., Sheth, A.: Gender-based violence in 140 characters or fewer: a# bigdata case study of twitter. First Monday 21(1) (2016). http://firstmonday.org/ojs/index.php/fm/issue/view/408 40. Purohit, H., Stabile, B., Grant, A., Pandey, R.: Modeling policy-relevant intent related to gender violence myths on social media using social construction theory. In: International Conference on Computational Social Science (IC2S2) (2018, forthcoming). http://ist.gmu.edu/~hpurohit/ informatics-lab/papers/modeling-gbv-policy-intent-ic2s218.pdf 41. Ramanand, J., Bhavsar, K., Pedanekar, N.: Wishful thinking: finding suggestions and ‘buy’ wishes from product reviews. In: Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pp. 54–61. Association for Computational Linguistics, Los Angeles (2010) 42. Salawu, S., He, Y., Lumsden, J.: Approaches to automated detection of cyberbullying: a survey. IEEE Trans. Affect. Comput. (2017). https://doi.org/10.1109/TAFFC.2017.2761757, http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8063898&isnumber=5520654 43. Solon, O.: Tim Berners-Lee: we must regulate tech firms to prevent ‘weaponised’ web (2018). https://www.theguardian.com/technology/2018/mar/11/tim-berners-lee-tech-companiesregulations?CMP=Share_iOSApp_Other 44. Squicciarini, A., Rajtmajer, S., Liu, Y., Griffin, C.: Identification and characterization of cyberbullying dynamics in an online social network. In: 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 280–285. IEEE, New York (2015) 45. Starbird, K.: Examining the alternative media ecosystem through the production of alternative narratives of mass shooting events on twitter. In: ICWSM, pp. 230–239 (2017) 46. Starbird, K., Spiro, E., Edwards, I., Zhou, K., Maddock, J., Narasimhan, S.: Could this be true?: I think so! expressed uncertainty in online rumoring. In: Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pp. 360–371. ACM, New York (2016) 47. Sukthankar, G., Geib, C., Bui, H.H., Pynadath, D., Goldman, R.P.: Plan, activity, and intent recognition: theory and practice. In: Newnes (2014) 48. Tomlinson, M., Bracewell, D., Krug, W.: Capturing cultural differences in expressions of intentions. In: Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pp. 48–57 (2014) 49. Tyshchuk, Y., Li, H., Ji, H., Wallace, W.A.: Evolution of communities on twitter and the role of their leaders during emergencies. In: Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, pp. 727–733. ACM, New York (2013)

18

H. Purohit and R. Pandey

50. Wang, J., Cong, G., Zhao, W.X., Li, X.: Mining user intents in twitter: a semi-supervised approach to inferring intent categories for tweets. In: AAAI, pp. 318–324 (2015) 51. Welser, H.T., Cosley, D., Kossinets, G., Lin, A., Dokshin, F., Gay, G., Smith, M.: Finding social roles in Wikipedia. In: Proceedings of the 2011 iConference, pp. 122–129. ACM, New York (2011) 52. Wu, L., Liu, H.: Tracing fake-news footprints: Characterizing social media messages by how they propagate. In: Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM’18, pp. 637–645. ACM, New York (2018). http://doi.acm.org/10. 1145/3159652.3159677 53. Yu, D., Tyshchuk, Y., Ji, H., Wallace, W.: Detecting deceptive groups using conversations and network analysis. In: Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, vol. 1, pp. 857–866 (2015) 54. Zarkov, D., Davis, K.: Ambiguities and dilemmas around #metoo: #forhow long and #whereto? Eur. J. Women’s Stud. 25(1), 3–9 (2018). https://doi.org/10.1177/1350506817749436

Bot-ivistm: Assessing Information Manipulation in Social Media Using Network Analytics Matthew C. Benigni, Kenneth Joseph, and Kathleen M. Carley

Abstract Social influence bot networks are used to effect discussions in social media. While traditional social network methods have been used in assessing social media data, they are insufficient to identify and characterize social influence bots, the networks in which they reside and their behavior. However, these bots can be identified, their prevalence assessed, and their impact on groups assessed using high dimensional network analytics. This is illustrated using data from three different activist communities on Twitter—the “alt-right,” ISIS sympathizers in the Syrian revolution, and activists of the Euromaidan movement. We observe a new kind of behavior that social influence bots engage in—repetitive @mentions of each other. This behavior is used to manipulate complex network metrics, artificially inflating the influence of particular users and specific agendas. We show that this bot behavior can affect network measures by as much as 60% for accounts that are promoted by these bots. This requires a new method to differentiate “promoted accounts” from actual influencers. We present this method. We also present a method to identify social influence bot “sub-communities.” We show how an array of sub-communities across our datasets are used to promote different agendas, from more traditional foci (e.g., influence marketing) to more nefarious goals (e.g., promoting particular political ideologies).

M. C. Benigni · K. M. Carley () Institute for Software Research, Carnegie Mellon University, Pittsburgh, PA, USA e-mail: [email protected] K. Joseph Computer Science and Engineering, SUNY Buffalo, Buffalo, NY, USA e-mail: [email protected] © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_2

19

20

M. C. Benigni et al.

1 Introduction How can individuals and groups be manipulated in social media? What messaging strategies can be used to shape behavior and alter opinions? In general, opinions and behaviors are a function of social influence [1]. That is, whom you know impacts not just what you know, but also your opinions and behavior. Consequently, group members, particularly those in tightly knit groups can come to share the same opinions and behaviors. Herein, we examine this process in social media. We identify a new type of bot that operates through social influence, the social influence bot network (SIBN). We then demonstrate the information strategies employed by such SIBNs to manipulate individuals and groups in social media. Initially used to spread spam [2] and malware [3], a substantial literature now documents the use of bots on Twitter to influence global politics [4–7]. Many of the original bots discovered were individual bots acting in isolation to effect social goals. These bots or “social bots” [7, 8] have been used to shape discussions during political revolutions [9, 10] and in recruiting and propaganda efforts for terrorist groups [11, 12]. Social bots, and individual bots, were also pervasive and highly active in online conversations during the 2016 US presidential election [13]. More recently, we find evidence of concerted coordinated effort using networks of bots. Thus, we deviate from the existing literature on social bot networks in that we study the creation and use of a new form of social bot—the social influence bot network (SIBN). Most prior work has assessed the network structure of bots on the Twitter follower network [14] or on the directed @mention network [8]. In the latter case, the focus is generally on how bots mention real users in order to gain attention. SIBNs use @mentions in this way and in doing so change the effective influence of those users whom they @mention. Importantly, SIBNs also use @mentions to manipulate the Twitter social network by @mentioning each other. In other words, they operate by altering the social network structure, and so impact who has social influence on whom. An example of the way social influence bots in our data use @mentions is shown in Fig. 1. The tweet in the figure was sent by a bot in our Syrian revolution dataset and contains only a string of mentions to nine other similarly named accounts. Shortly after, the bot sending this tweet was itself mentioned in similarly structured tweets by the other bots mentioned in Fig. 1. The sole purpose of these tweets is to artificially manipulate the reciprocal @mention, or co-mention graph, creating networks of bots with strong ties in the @mention network. More specifically, this kind of behavior produces a mention core of bots—a sub-community of social influence bots displaying “core-like” behavior [15] that have anomalously strong connections in the co-mention network. To distinguish this particular form of social bot network, we will refer to those social bot networks that have a mention core as social influence bot networks (SIBNs). Social influence bots can impact the Twitter ecosystem on at least three levels. First, they can be used at the “content-level” to rapidly spread specific tweets and/or particular URLs and make them appear artificially popular [8]. Second, they can be

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

21

Fig. 1 Depicts core-bot behavior in the Firibinome social influence bot network

used at the “user-level” to artificially inflate the importance of themselves and/or target users as measured by standard metrics of influence, like a user’s number of followers [13]. Finally, social influence bots can be used at the “community-level,” forging fake communities or embedding themselves into communities of real users to promote particular ideologies [9]. In other words, they can and do manipulate what is being said, who is communicating to whom, and the relative influence of particular actors. These means of manipulation are mutually reinforcing, and thus, are often used simultaneously as part of a social influence information campaign. Social influence [1] is the process through which the opinions and behaviors of actors are influenced by the opinions and behaviors of those with whom they have a structural relation. As the network of ties among actors change, so too does who has influence, the rate at which opinions and behaviors change, the rate at which the group reaches consensus, etc. Ties among actors, however, are continually constructed as actors learn new information, share opinions, and so forth [16, 17]. The result is the mutual development of who is tied to or has relations with whom and who shares what beliefs or opinions with whom. In general the more social relations or ties within a group, the higher the density, the faster the social influence process. At its extreme, if there are only a few connections, but it is still a connected network, if all the actors have the same information, same beliefs, the network will

22

M. C. Benigni et al.

quickly become fully connected. The social network and the knowledge network connecting actors are co-evolutionary. These high dimensional networks where actors are connected by social ties and knowledge/opinion ties are topic groups. In Twitter to find topic groups we use as the social network the mentions network (the network formed by all instances of who mentions or retweets or replies to whom) and as the knowledge network the shared hashtag network (the network formed by all instances of who shared the same hashtags and note that it could be same concepts in the tweet content). Using the IVCC process [18], which has now been realized in ORA-PRO [19], these two networks are used to find a set of topic groups. As the connections in both the social network and the shared knowledge network increase the topic group begins to grow into an echo-chamber. In other words, echo-chambers are a set of actors who are densely connected in the social network and in the shared knowledge network; hence, in Twitter this would be a set of users who are highly connected in the mentions network and in the shared hashtag network. In the extreme, a topic group formed of a set of actors who all connect to each other and all share the same knowledge is a pure echo-chamber. That is, a pure echo-chamber is a completely connected subgraph in both the social network and the shared knowledge network. As topic groups become more echo-chamber like in nature emotions can escalate, language variation can decrease, and ideas will flow faster. The more echo-chamber like a topic group is the more prone it is to groupthink. Pure echo-chambers can be very influential as all members spread the same message, and it is difficult for those outside the group to get their message into the group. Social influence bot networks (SIBNs) influence topic groups and have as their core an echo-chamber. An SIBN is composed of a network of core and non-core bots, such that the core bots form an echo-chamber. SIBNs connect into topic groups and alter the makeup of these groups. Due to this constructural process, one of the most powerful mechanisms for increasing the manipulative power of social influence bots is the construction of network ties within topic groups. SIBNs, through the use of @mentions, are altering the social networks of groups on Twitter and so impacting who is influencing whom. While these social influence bot networks (social botnets or SIBNs [7, 9])1 are in some ways easier to detect than isolated accounts, they are also much more able to achieve desired manipulations of the Twitter ecosystem. In addition to allowing faster spread of content [8], creating a network of social ties among bots can inflate the importance of bots in network metrics like degree and betweenness centrality [22]. The present work presents an exploration of the impact of SIBNs in three very different political discussions—“alt-right” themed discussions of the 2016 US presidential elections, the Syrian revolution, and the Euromaidan movement in the Ukraine. For each topic, we collect a different dataset via snowball sampling of the follower network. We seed each snowball sample with a set of core members

1 Social

bot networks are also referred to as Sybils in the computer security literature [20, 21].

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

23

of these discussion communities, allowing us the ability to explore the social structure of the users and SIBNs associated with those networks and influencing user behavior. Within each of these discussion communities there are topic groups. Associated with these topic groups are one or more SIBNs. Despite the distinct contexts in which these three communities are set, we observe high levels of similarity in how SIBNs within them impact user-level metrics and community structure. Specifically, we observe that: • SIBNs operate as echo-chambers. Each SIBN is a set of bots that @mention other bots forming a fully connected social network, and which tweets about the same topics in similar ways. By mentioning each other, the SIBN forms a dense core thus giving weight to the importance of the individual bots in the SIBN in terms of message prioritization. • SIBNs manipulate the apparent influentialness of key actors. SIBNs are used to greatly inflate complex influence metrics on the social network induced by reciprocal mentioning behavior on Twitter. We find that some accounts drop in influence by as much as 60% in important network measures like coreness [15] after removing bot-like behavior from our data. • SIBNs manipulate what information is flowing to achieve various goals. Each SIBN appears to have a distinct goal. These included SIBNs with more traditional foci (e.g., explicit influence marketing) as well as those aimed towards more nefarious goals (e.g., promoting particular political ideologies). • SIBNs build network ties among users so as to better manipulate those users. Bot creators within our communities used directed social engineering to accomplish particular goals. This included, for example, a dataset sharing lewd images of women to attract young men interspersed with calls to violence in our Euromaidan data. The obvious questions that remain, of course, are why bot creators manipulate the co-mention graph in this way and how SIBNs are used in our datasets. With respect to the “why” of SIBNs, the creation of a mention core in the comention graph allows the SIBN to confound radial-volume centrality measures like PageRank [23] and coreness [15] that are supposedly more robust to spamlike behavior. Because of this, owners of SIBNs can inflate the centrality of promoted accounts, which are fully or semi-automated accounts that are mentioned by the core bots in the SIBN and that attempt to influence a specific online community of interest. These promoted users thus have influenced level metrics that distort our understanding of true community influencers. While we find that true community influencers, individuals with real influence on the non-automated discussion community of interest, tend to be mentioned frequently by SIBNs as well, they can be differentiated from promoted users by the fact that their ranking on radial-volume does not depend significantly on bot activity. Assume that Twitter users are more likely to have recommended to them to follow more central actors, to follow topics mentioned by more central actors, and are more likely to have messages from such actors prioritized in their lists. If this is indeed the case, and it appears to be so, then promoting accounts, and promoting certain messages,

24

M. C. Benigni et al.

will alter who is connected to whom in the social network and who shares what information in the knowledge network. With respect to how SIBNs are used, a graph-level perspective is required. In other words, the utility of SIBNs is best understood by first uncovering the different SIBNs that exist in our dataset, and then considering how each may be used in different ways to manipulate influence across particular sets of users and themes using various social engineering strategies. To uncover the different SIBNs within our dataset, we develop a methodology based on dense subgraph detection [24]. We use this method to survey multiple different SIBNs in our datasets, each of which have unique goals.

2 Related Work 2.1 Overview For an expansive overview of social bots, we point the reader to [7]. Here, we restrict ourselves to a discussion only of recent work related to the study of social bot networks and automated manipulation of influence metrics. Boshmaf et al. define a social bot network as a set of social bots with three components: the botherder who controls the social bots, the social bots that carry out tasks assigned by the botherder, and a Command and Control channel used to facilitate task assignment [25, 26]. In order to construct social bot networks, as opposed to just a series of isolated bots, botherders must engage in some form of link creation between bots and link farming to external accounts. Ghosh et al. [14] define link farming as the process by which “users, especially spammers, try to acquire large numbers of follower links in the social network.” They find, somewhat surprisingly, that bots do not necessarily create mutual following ties only with other bots, but also with other “real” Twitter users with low thresholds for reciprocity. Here, we discuss how bots attempt to use @mentions in similar ways. However, our focus is also on how links between the bots themselves corrupt influence metrics. In turn, several works have shown the ease with which a social bot network can artificially inflate such metrics, from obvious ones like number of followers [26] to more opaque metrics like Klout scores [3, 8, 27]. This research shows that these more opaque metrics can be influenced by both following and @-mentioning behavior. Herein, we show one way in which these opaque metrics might be impacted by @mentioning behavior among bots themselves, which can significantly impact important social network metrics in non-obvious ways.

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

25

2.2 Defending Against Social Bot Networks The most common approach to defending against social bot networks is to identify individual bot accounts using machine-learning methods [28–30]. As noted by Ferrara et al. [7], the best performing of such methods tend to take into account a variety of behavioral features of accounts. Such methods are quite fragile. In contrast, we detect SIBNs using only social graph features. This effort aligns more closely to a host of Sybil-detection algorithms in the literature (for a nice review, see [20]). Our method differs from most prior work; however, in that we focus on reciprocal mentioning behavior, rather than following [14] or friend (on Facebook) [26] ties. While “real” users have been observed to reciprocate friendships or following ties with social influence bots (leading to difficulties in Sybil detection [7]), the reciprocal @mention network has generally been found to signal strong social relationships [31, 32]. Perhaps more importantly, such directed communication is unlikely to occur in reciprocal formats between non-automated and automated accounts. Moving beyond bot detection at the account level is important, however, because an increasing amount of automated, malicious activity happens within “cyborg” accounts [7]. These cyborg accounts, which blend human behavior with automated behavior produced by Twitter applications the user has subscribed to, are extremely difficult to detect using existing bot detection methods. Cresci et al. [33] find that existing methods perform at near-chance levels on classifying these accounts that blend automated and (at least seemingly) non-automated behavior. These findings are worrying for two reasons. First, as it becomes increasingly difficult to determine whether or not an account is controlled by a human, analysts and scholars will have an increasingly difficult time studying socio-theoretic models of human behavior [34, 35]. Second, these results suggest it is increasingly difficult for “real” Twitter users to tell the different between bots and non-bots. This leads to an increasingly level of social interaction between “real” users and at least semiautomated accounts. The present work therefore focuses on automated behavior at the level of the edge (interaction), rather than the node (the individual account). We identify anomalous edge behavior and then assess how this behavior impacts user influence metrics and leverage it to identify SIBNs. The prevalence of cyborgs also implies that differentiating “bots” from “humans” is a potentially misleading problem [7]. Scholars have thus also began to focus on developing methods to address potential impacts of automated behaviors on social media ecosystems without having to determine which accounts are bots [14, 36, 37]. These tolerance-based [20] defense schemes rely on tactics like down-weighting influence scores for accounts that are “bot-like” and assessing the percentage of “bot-like” activity around a certain topic. Our approach to differentiating community influencers from promoted accounts is an example of a tolerance-based approach.

26

M. C. Benigni et al.

2.3 The Impact of Social Bot Networks A growing literature details the way in which social bot networks are used for promotion of political agendas of varying forms [9, 10, 38–40]. Abokhodair et al. [9] describe the 35-week lifespan of a social bot network on Twitter designed to express opinion, testimony of ongoing events, and engage in preliminary conversation associated with the Syrian revolution. Similar techniques have been observed in the case of ISIS’ online video dissemination as well [12]. The present work complements these efforts by showing how SIBNs can conflate not only topical discussions but also social network metrics, i.e., they impact both the social and the knowledge network. This is demonstrated using the evidence of highly related behavior in three different datasets. Scholars have also observed that social bot networks were responsible for a significant amount of content produced around conspiracy theories central to the 2016 US presidential election [40]. Nied et al. observed in this context that in a cluster of bot-like accounts, “[u]nexpectedly, 62% of all mentions were between users of which neither followed the other.” Our work sheds significant insight into this behavior. This recent work exploring the impact of social bot networks falls, however, on a backdrop of a large literature analyzing Twitter data without considering the impact of bots or only considering bots superficially using heuristics. The present work shows that in addition to follower networks [26], directed mention networks [41], and topical themes [10] being susceptible to social bot network behavior, so too is the network of reciprocal mentioning behavior.

3 Data The present work studies three distinct datasets. Each has been collected using snowball sampling [42]. In a snowball sample, a set of individuals are chosen as “seed agents,” and individuals with social ties to one or more seed agents are added to the sample. This technique can be iterated in a series of steps or “hops.” On each “hop,” the set of accounts sampled on the prior hop are used as seed for collection in the next hop. Thus, a “1-hop” snowball sample collects the social ties of the seed agents; a “2-hop” sample collects the social ties of the seed agents’ social ties (“friends of friends”), and so on. For each of our datasets, we seed searches with a set of users that are known members of a community of interest. We then collect additional users based on a snowball sample using relationships defined by a datasetspecific social graph. For each user added to the snowball sample, we collect up to the last 3200 tweets sent by the user, as allowed by the Twitter API. This method of collecting data is different from most prior work on the influence of bots on geopolitical discussions [9, 40], which focuses collection around tweets mentioning one or more terms of interest. It is potentially this form of sampling,

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

27

rather than a focus on collecting tweets, that allows us to more easily observe SIBN behavior. The first dataset we collect is the alt-right community (ALT16) dataset. In October 2016, we seeded a one-hop snowball sample of 2482 users who each followed five influential Twitter users associated with the Alt-Right political movement: Richard Spencer, Jared Taylor, American Renaissance, Milo Yiannopoulos, and Pax Dickinson. The search resulted in 106K users and 268 million tweets. The second dataset is the Euromaidan community (EUR17) dataset. The Euromaidan revolution occurred as a wave of demonstrations starting in Ukraine in November 2013 and resulted in the removal of Ukrainian President Viktor Yanukovych from power. In an attempt to study messaging themes used within the Euromaidan movement we conducted a one-hop snowball sample of starting from 1209 and following their ties in the directed @mention network (e.g., everyone any of the seed users have mentioned) from March 2014 to September 2017. The 1209 seed users were collected via a combination of automated search and manual annotation from subject matter experts [43]. The snowball search resulted in 92,706 Twitter users and 212 million tweets. The final dataset is the Syrian revolution (SYR15) dataset. With research objectives similar to our Euromaidan Study, we conducted a one-hop snowball sample of the friend graph for 13,949 accounts that had been identified as ISISsupporting in prior work [18]. We then removed all nodes with degree less than 2 in the following graph. The search resulted in 87,046 Twitter users and 179 million tweets.

4 Observing Anomalous @Mentioning Behavior Much like [9], our datasets were originally collected for the study of online political activism but were found to be deficient for these purposes until bot activities within it were better understood. Here, we briefly detail how the anomalous @mention behavior we observe was discovered and how it motivated the proposed methodology. Because we assume readers will be most familiar with the context and users central to the ALT16 dataset, we use it as a running example throughout this section. As noted above, prior work has suggested that co-mentioning behavior—when two users reciprocate @mentions of each other—is a strong signal of social relationships between two Twitter users [31, 32]. Correspondingly, we began our analysis of these two datasets by constructing the directed mention graph. Mathematically, the weighted mention graph M(V, E) is a weighted directed graph with vertices V: {v1 , . . . ,vn } consisting of the users returned from our search and edges Em : {eM,1 , . . . ,eM,m } defined as the number of unique tweets where useri mentions userj . We then constructed the primary graph of interest, the reciprocal or co-mention graph, R(V, E). R has the same vertices as M but had undirected

28 Table 1 Top users in the ALT16 dataset co-mention graph using weighted degree centrality

M. C. Benigni et al. Rank 1 2 3 4 5 6 7 8 9 10

Top ALT16 users nayami_rescue Socialfave sundoghigh TheMisterFavor SarCatStyX saravastiares KoichicCheryl realDonaldTrump Easy_Branches lupash7

edges with weights set to the minimum number of times useri mentions userj , or vice versa. As we were interested in understanding who the influential users in the comention graph, we began by considering weighted degree centrality of the nodes in this network. Weighted degree centrality, which simply measures the total weight of each user’s edges in the co-mention graph, is one of the easiest and most common network metrics to compute. Table 1 presents the top 10 users in the ALT16 dataset according to weighted degree centrality. These users were not necessarily those one would expect to be central in a snowball sample seeded with highly politically motivated accounts. Upon closer inspection, we found some of these users to have sent several tweets containing only a string of @mentions to other accounts, e.g., “@user1 @user2 @user3 @user4 Hey!”. Looking at the profiles of @user1, @user2, and @user3, we further observed that some of these accounts were cyborgs leveraging the same kind of automated behavior. Others were clearly bot accounts, sending only these @mention tweets, with no attempt to appear human or perform human-like behaviors. As it is well known that bots can influence simple network metrics like degree centrality, results in Table 1 were discouraging but not surprising. We then turned to more complex network metrics that are better able to withstand spam and bot-like behavior. Specifically, we consider PageRank [23] and coreness [15], two common, more complex metrics for measuring influence in complex networks. Both metrics are drawn from the family of radial-volume centralities [44], which attempt to quantify an account (or more generally, a network vertex) x’s influence based on the influence of the other vertices to which x is linked. These measures can be viewed on a continuum based on the length of walk considered in the neighborhood of a given vertex. Coreness [45] is calculated based on the concept of K-shells within a graph. A K-shell is defined as the maximal subgraph of a given graph G, where all vertices are of degree greater than or equal to k. A vertex’s coreness, Ks , indicates the greatest value k for with the node remains in the corresponding K-shell. In addition to coreness, we also look at PageRank, or eigenvector centrality, which is roughly defined as “accounts who are popular with other accounts who are popular” [46].

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

29

Table 2 Top users in the ALT16 dataset co-mention graph using PageRank (left) and coreness (right) Rank 1 2 3 4 5 6 7 8 9 10

Top ALT16 users (PageRank) realDonaldTrump HillaryClinton YouTube POTUS FoxNews CNN nytimes timkaine NASA wikileaks

Top ALT16 users (Coreness) 2020sahara JuliaZek Jerz_Gal DilrubaLees 7artistai nayami_rescue wanderingstarz 1 JulezPooh Dollhouse Edward733

Table 2 presents the top 10 nodes on PageRank and coreness for the ALT16 dataset. We see that results from the PageRank algorithm pass tests of face validity, while this is still not true of coreness. Upon further investigation of these coreness metrics, we again found the anomalous @mention behavior we observed to confound measures of degree also impacted measures of more complex measures like coreness. Below, we will show that even for PageRank, designed specifically for spam-like behavior, rank ordering of nodes below the top 10 can be significantly impacted by this behavior. SIBNs thus inhibited our ability to analyze influential users in the ALT16 dataset, as well as the other two datasets we do not discuss here. We were left with two questions for which methods needed to be introduced. First, we wanted to be able to “tolerate” [20] the influence of automated @mentioning behavior on our analysis of influential users in our datasets. In the terminology used in the present work, we wanted to be able to differentiate between promoted accounts and actual community influencers. Second, we were interested in better understanding the structure of SIBNs in our dataset. In particular, we wanted to understand if many SIBNs existed within each dataset, and if so, what their utility was. To do so, we needed to find a means of detecting mention cores, the foundation of SIBNs, in our data. The following section introduces the straightforward, scalable methods we adopt to address these issues.

4.1 Methodology 4.1.1

Differentiating Promoted Accounts from Community Influencers

We use a straightforward procedure to differentiate between promoted users and community influencers. Our strategy is based on the fact that if we are only concerned with identifying any core bot, any bot that is likely to be part of a mention

30

M. C. Benigni et al.

core, rather than distinct SIBNs, we can simply use a measure of @mention activity. Correspondingly, we first define γ p as the mention per tweet ratio for each user that is greater than a percentage p of users within a given dataset. For example, γ 0.995 would represent the mentions per tweet ratio that is greater than the mentions per tweet ratio for 99.5% of the users in a given dataset. By setting p to a large number, we can isolate core bots of any SIBNs in our dataset, as they are the only accounts to produce only tweets that had only @mentions. Upon removing these accounts and recomputing network metrics like coreness and PageRank, we can then differentiate promoted users from community influencers by how the network metrics of these accounts change. Metrics of promoted users should suffer significantly when core bots are removed. In contrast, true community influencers should not be impacted by the removal of core-bot behavior.

4.1.2

Detecting SIBNs

Because mention cores, the base of an SIBN, create highly dense communities, dense subgraph detection offers a logical means to detect them. Dense subgraph detection allows an analyst to extract only subgraphs of a network in which all nodes within that subgraph have an anomalously high density of (weighted) interactions among them. Dense subgraph detection can therefore be preferable when a complete clustering of the data is not desired, but only dense substructure is of interest [24, 47]. This is precisely the case in our setting we would like to extract out only the dense mention cores of SIBNs. Our method for detecting dense subgraphs follows the work of [24], with one difference. Specifically, Chen and Saad [24] define the density of a subgraph as an unweighted link count. Because of the repetitive use of @mentioning by mention cores, we instead choose to define subgraph density by summing link weights instead of link counts. Algorithm 1 Find core botnet members of an SIBN Input: Given a large sparse, weighted, reciprocal mention graph R, and density threshold dmin Output: Set D:d1 (v,e), . . . ,dn (v,e) such that each subgraph contains a subset of users exhibiting behavior that meets the definition of a SIBN core bot member. 1. Compute Matrix CR as defined in (2) 2. Sort the largest t non-zero entries of CR in ascending order, where t = nz(A). Denote Q the sorted array. 3. Construct the hierarchy T according to the sorted vertex pairs designated by Q. 4. Extract Subgraphs of where dG ≥ dmin 5. Manually inspect subgraphs for SIBN like behavior While we leave a detailed discussion of the model to [24], we provide a summary of our approach here and in Algorithm 1. Just as presented in [24], we search for

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

31

dense subgraphs (mention cores) by constructing AR , a weighted adjacency matrix of R. Let us define CR , the cosine matrix, as:    AG (:, i) , AR :, j   (1) CR i, j =       A (:, i)  A :, j  We then set t = 2 × |ER | and sort the largest t non-zero entries of CR in ascending order. We denote this sorted array as Q and construct a hierarchy T based on sorted vertex pairs based on the measure CR . We then extract the largest distinct subgraphs from T that meet a minimum size, smin and density threshold, dmin . We have set smin to 30 and dmin to 0.75 in this work based on manual inspection. In sum, the algorithm we use, defined here and in Algorithm 1, takes as input a weighted co-mention graph for a particular dataset and produces a set of SIBNs. Each SIBN has at least smin accounts (here, 30) and a weighted density of dmin (here, 0.75). Note that because the method we propose is unsupervised, it is possible that it captures human communities as well as SIBNs. In practice, we find that the method returns almost exclusively subgraphs of bots, and that where this is not the case we can easily differentiate SIBNs from “real” sub-communities.

5 Results 5.1 Differentiating Promoted Accounts from Community Influencers Figure 2 depicts the effect of core-bot activity on non-core-bot accounts in the ALT16 dataset with respect to coreness (top panel) and PageRank (bottom panel). The figure shows that the removal of core bots significantly worsens (increases) the ranking on these metrics for a large set of promoted accounts. For PageRank, this removal significantly improves (decreases) the rankings for a smaller set of community influencers. For coreness, we simply see no change in these rankings. Examples of highly promoted accounts (blue labeled points) and community influencers (red labeled points) are shown in Fig. 2. To perform the analysis for Fig. 2, we use γ 0.995 = 5.94 to select 507 core bots. For each panel, in Fig. 2, the x-axis depicts the (log-scaled) number of times a given user is mentioned by these 507 accounts. The y-axis depicts change in centrality with respect to these two metrics for all non-core-bot accounts when core-bot edges are removed. Larger positive values with respect to the y-axis therefore indicate inflated metrics based on core-bot activity. These are promoted accounts. Large negative values on the bottom panel imply that PageRank improves once core bots are removed. These accounts, although highly mentioned by core bots, actually increase in centrality when core bots are removed. These users are community influencers.

32

M. C. Benigni et al.

Fig. 2 Depicts the effect of core-bot activity on non-core-bot accounts with respect to coreness (top panel) and PageRank (bottom panel) in ALT16 dataset. In each panel the x-axis depicts the number of times a given user is mentioned by core bots in log-scale. The y-axis depicts change in coreness and PageRank in the top and bottom panels, respectively, when core-bot edges are removed

To further characterize promoted users and community influencers across our datasets, Tables 3 and 4 depict top community influencers and promoted accounts in the ALT16 and EUR17 datasets, respectively. To determine this set of users, we first sample the set of accounts that represent the top 0.99 percentile of users mentioned by core bots. Promoted users in Table 3 are then the users in this collection who are most negatively impacted by the removal of core bots. In contrast, the community influencers are those that most benefit on these metrics from the removal of bots.

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . . Table 3 Top 0.99 percentile of users mentioned by core bots in the ALT16 dataset

Rank 1 2 3 4 5 6 7 8 9 10

Promoted accounts PollaPrenadora MktgSciences miarianmoreno monicasloves saravastiares Alicelovelb webcamfamosas NudeArt6969 jimkoz69.jim verovvp

33

Community influencers realDonaldTrump HillaryClinton POTUS YouTube FoxNews CNN nytimes mitchellvii seanhannity Cernovich

The left column depicts promoted users whose PageRank decreases the most when core-bot activity is removed. The right column depicts the top 10 users whose PageRank increases most when core-bot activity is removed Table 4 Top 0.99 percentile of users mentioned by core bots in the EUR17 dataset

Rank 1 2 3 4 5 6 7 8 9 10

Promoted accounts dilruba_lees PollaPrenadora goodenough03 NudeArt6969 saravastiares Isobg69 monicasloves patdefranchis V_Samokhova lenlekk

Community influencers YouTube rianru Pravitelstvo_RF MID_RF KremlinRussia history_RF mod_russia zvezdanews kpru wordpressdotcom

The left column depicts promoted users whose PageRank decreases the most when core-bot activity is removed. The right column depicts the top 10 users whose PageRank increases most when core-bot activity is removed

The top community influencers in Tables 3 and 4 consist of accounts one would expect to be influential within each datasets centered around far right political and the Euromaidan movement, respectively. In contrast, manual inspection of promoted accounts highlights a variety of promotional but not necessarily relevant interests. One obvious promotional interest is pornography, highlighted by promoted users 1, 5, 7, 8, and 10 in Table 3, and nearly all promoted users in Table 4. Interestingly, although datasets were seeded from an entirely different set of users, we see that promoted accounts 1, 4, 5, and 8 appear in both datasets. Promoted accounts 2, 6, and 9 in Table 3 all provide links to third party community management applications that facilitate core-bot activity on behalf of “cyborg” subscribers. Finally, promoted accounts 3 and 4 appear to be bloggers who subscribe to some type of community

34

M. C. Benigni et al.

management application as their timelines appear to contain posts containing human language, while others contain strings of mentions. Although similar behavior is observed in the SYR15 dataset, it is not at the same scale as what we observe in the EUR17 and ALT16 datasets. We expect this is due in part to the conservative choice of p used here. Consequently, the proposed method may be improved by leveraging a more automated mechanism of selecting the p parameter of γ . Regardless, however, results show that the methods we develop do a good job of differentiating between community influencers and promoted accounts. After doing so, we found that promoted accounts served a variety of different purposes, most frequently pornography or subscriptions to Twitter applications that promote cyborg behavior. In the following section, we delve deeper into the different SIBNs we found in each dataset and how they help to understand these various purposes for promoted accounts.

5.2 Detected SIBNs We ran the method and found 11, 4, and 1 distinct SIBNs in the ALT16, EUR17, and SYR15 datasets, respectively. We will now discuss how these networks function and how they can be used for a variety of promotional ends. To illustrate how SIBNs function, we use an SIBN from the SYR15 dataset. We call this SIBN “Firibinome” because of its easily identifiable naming convention, as is illustrated in one of its core-bot tweets depicted in Fig. 1. Firibinome is another politically motivated SIBN—each of the core-bot accounts shares the same profile image, a flag associated with Jabhat al-Nusra the predominant al-Qaeda affiliate in Syria. Figure 3 depicts the directed mention graph of all accounts mentioned by the Firibinome SIBN. Nodes represent users within the neighborhood of the SIBN and are colored based on coreness within the co-mention graph; they are sized based on in-degree coreness which emphasizes users who are mentioned by people who are highly mentioned. Mention core edges are depicted in red. The promoted user is highlighted in blue and represents a charitable foundation for Syrian children alleged to be a revenue generating scam for Jabhat al-Nusra. The hashtags in posts retweeted by the Firibinome SIBN are summarized in Fig. 4 and are clearly consistent with Jabhat al-Nusra’s interests in the region. Without analyzing SIBNs at the graph level, one would fail to identify how the mention cores network structure influences more sophisticated measures of influence. While SIBNs all seem to serve the purpose of inflating the importance of particular users2 it is important to understand that within each dataset we find

2 And,

although we have not discussed it here, likely content as well.

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

35

Fig. 3 Depicts the Firibinome SIBN. Nodes represent users within the neighborhood of the SIBN and are colored based on coreness within the co-mention graph and sized based on in-degree coreness in the mention graph. Mention core edges are depicted in red. The promoted user is highlighted in blue and example community influencers in red. Examples of community influencers are in red and represent pro-Nusra propaganda as well as an Islamic Scholar Hani al Sibai a wellknown al-Qaeda supporter

multiple distinct SIBNs that have different goals with respect to which users and themes they are interested in promoting. For example, the 507 ALT16 core bots form 11 distinct SIBNs when we run dense subgraph detection. As depicted in Fig. 7, many of these SIBNs have identifiable and differentiable promotional agendas. Figure 7 summarizes the most dense, distinct subgraphs consisting of between 30 and 200 users within the Alt-Right Twitter Search and found with the dense subgraph detection method described above. The x-axis connotes weighted subgraph density, while the yaxis connotes size in terms of users. Black circles in Fig. 7 display collections of users that we have manually identified as clearly SIBNs. We identify each group’s promotional objective through manual inspection by summarizing the type of content retweeted. The two subgraphs showing highest density refer to third party applications designed to increase the social influence of subscribers and profile descriptions of core bots state these objectives explicitly. Religious and politically

36

M. C. Benigni et al.

Fig. 4 Depicts translated hashtags retweeted by core bots within the Firibinome SIBN

focused cores like the Evangelical, #opisis, and #BoycottIsrael can clearly be identified by summarizing retweeted hashtags in a manner similar to those depicted in Fig. 4. In each of these cases, the communities could be formed by real users via third party applications, which will be discussed later in this section. SIBNs can also be observed to leverage particular content to achieve very targeted objectives. For example, an SIBN in the Euromaidan dataset was found that consists of 14 Twitter accounts designed to bridge two distinct online communities. In the Spring of 2015, each tweet from the core of this SIBN shared nude images of women and strings of mentions that pointed to accounts that either shared similar images or inflammatory news updates covering Ukrainian government corruption and Russian occupation of Crimea. The most frequent hashtags shared by the botnet are translated to English and depicted in Fig. 5. Black terms highlight those hashtags associated with news and propaganda. The community structure of this SIBN is depicted in Fig. 6 and is parameterized identically to Fig. 3. The density of red edges bridging the two communities is highlighted in the figure, and when the botnet’s combined betweenness centrality makes it the most powerful bridge in the entire EUR17 dataset. Although the objective of this SIBN is unknown, such behavior would be consistent with campaigns designed to expose young men to recruiting propaganda.

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

37

Fig. 5 Depicts translated hashtags retweeted by core bots within the Euromaidan Image Sharing SIBN. Terms in black highlight content associated with Euromaidan propaganda and Russian occupation of Crimea. Terms in grey are predominantly associated with the sharing of pornographic pictures

In some cases, SIBNs are the product of third party applications advertised as “influence enhancers” or “community managers.” Four examples, #InfluenceMarketers, Evangelical, #opisis, and #followback, are depicted in Fig. 7. In each example, these communities consist of accounts that appear to be real users who retweet and post content that is core-bot-like with respect to mentions. In each case the post source annotated in the tweet json indicates that they were posted from third party applications like Follow Friday Assistant or commune.it which can generate core-bot behavior. For example, Follow Friday Assistant generates and post lists of accounts a user has interacted with in the past week, day, or month, and will split long lists into multiple tweets. In these cases, core-bot accounts are “cyborgs” or bot-assisted-humans [48]. The content associated with each SIBN summarized in Fig. 7 suggests these third party applications are leveraged for politically motivated reasons. These applications allow cyborg users to post with high volume and generate large online activist communities. In fact, one could argue

38

M. C. Benigni et al.

Fig. 6 Depicts the Euromaidan Image Sharing SIBN. Nodes represent users within the neighborhood of the SIBN and are colored based on coreness within the co-mention graph and sized based on in-degree coreness in the mention graph. Mention core edges are depicted in red. This SIBN appears to be designed to first build a community through image sharing and then to expose young men in that community to recruiting propaganda

that these third party applications slower the technical threshold for users to apply directed social engineering to accomplish their marketing objectives. This activity makes understanding online influence increasingly complex.

6 Conclusion In this paper, we have studied the problem of automated social engineering in Twitter. We have introduced a specific class of social bot networks, social influence bot networks (SIBNs). SIBNs manipulate topic groups by creating an artificial corelike group in Twitter mention and co-mention graphs, promote particular users and cites, build connections among actors, and alter messages. We employed novel network techniques using meta-networks with multiple types of nodes to identify these SIBNs and to understand their usage. Specifically with respect to SIBNs we show:

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

39

Fig. 7 Summarizes the most dense, distinct subgraphs consisting of between 30 and 200 users within the Alt-Right Twitter Search. The x-axis connotes weighted subgraph density, while the y-axis connotes size in terms of users. Black circles are SIBNs, but many of the subgraphs showing high graph density exhibit core-bot-like behavior and identifiable promotional agendas

• SIBNs are pervasive in a large number of domains and are influencing many different topic-groups. • SIBNs are used to greatly inflate complex influence metrics on the social network induced by reciprocal mentioning behavior on Twitter. • SIBNs formed multiple sub-communities of bots or cyborgs within each dataset, each with distinct intentions. These included SIBNs with more traditional foci (e.g., explicit influence marketing) as well as those aimed towards more nefarious goals (e.g., promoting particular political ideologies). • Bot creators and cyborg users within our communities used directed social engineering to accomplish particular goals. This work is limited in that we have not attempted to study these SIBNs influence with respect to following ties or content diffusion. For example, similar methods could be used to generate trending hashtags or disseminate URLs. Furthermore, this work implies the need for a detailed study of third party applications and their use for social engineering. SIBNs have at their core an automated echo-chamber. This social structure raises interesting ethical questions. In effect, the methods described herein can be used to promote a perception of popularity or validity that does not necessarily reflect real user opinion making it increasingly difficult to determine trustworthiness. Moreover, an industry designed to market within this online ecosystem has emerged leveraging these behaviors and others with great effect. As Twitter has become one of the world’s largest publication platforms the implications of providing an API that facilitates such activity merits future research.

40

M. C. Benigni et al.

Bots, cyborgs, and the special form of these as SIBNs are pervasive. They appear to operate in part by exploiting the social media architecture’s recommendation systems and the way in which we as humans learn, recall, and make sense of information. The SIBNs are particularly powerful at manipulating groups in social media because they engage in strategic information maneuvers that create change at two levels—the social network and the shared knowledge network. No doubt, the specific form these bots will take will evolve, as will the form of these bots on other social media platforms. However, we anticipate that those bots and cyborgs that are most effective will be those that like SIBNs manipulate human activity in terms of both who is talking to whom and who shares what information or opinions.

References 1. Friedkin, N.E.: A Structural Theory of Social Influence, vol. 13. Cambridge University Press, Cambridge (2006) 2. Yardi, S., Romero, D., Schoenebeck, G., Boyd, D.: Detecting spam in a twitter network. First Monday 15, 1 (2009). http://firstmonday.org/ojs/index.php/fm/article/view/2793/ 2431?utm_source=twitterfeed&utm_medium=twitter 3. Zhang, J., Zhang, R., Zhang, Y., Yan, G.: On the impact of social botnets for spam distribution and digital-influence manipulation. In: 2013 IEEE Conference on Communications and Network Security (CNS), pp. 46–54. IEEE (2013). http://ieeexplore.ieee.org/xpls/ abs_all.jsp?arnumber=6682691 4. Ratkiewicz, J., Conover, M., Meiss, M., Gonçalves, B., Flammini, A., Menczer, F.: Detecting and tracking political abuse in social media. ICWSM. 11, 297–304 (2011). http:// www.aaai.org/ocs/index.php/ICWSM/ICWSM11/paper/viewFile/2850/3274/ 5. Ratkiewicz, J., Conover, M., Meiss, M., Bruno, G., Patil, S., Flammini, A., Menczer, F.: Truthy: mapping the spread of astroturf in microblog streams. In: Proceedings of the 20th International Conference Companion on World Wide Web, pp. 249–252. ACM, New York (2011). http:// dl.acm.org/citation.cfm?id=1963301 6. Woolley, S.C.: Automating power: social bot interference in global politics. First Monday 21, 4 (2016). http://journals.uic.edu/ojs/index.php/fm/article/view/6161 7. Ferrara, E., Varol, O., Davis, C., Menczer, F., Flammini, A.: The rise of social bots. Commun. ACM. 59(7), 96–104 (2016). http://dl.acm.org/citation.cfm?id=2818717 8. Zhang, J., Zhang, R., Zhang, Y., Yan, G.: The rise of social botnets: attacks and countermeasures. IEEE Trans. Dependable Secure Comput. 99, 1 (2016). https://doi.org/10.1109/TDSC.2016.2641441 9. Abokhodair, N., Yoo, D., McDonald, D.W.: Dissecting a social botnet: growth, content and influence in Twitter. In: Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing, pp. 839–851. ACM, New York (2015). https://doi.org/10.1145/2675133.2675208 10. Wei, W., Joseph, K., Liu, H., Carley, K.M.: The fragility of Twitter social networks against suspended users. In: Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, pp. 9–16. ACM, New York (2015) 11. Berger, J.M., Morgan, J.: The ISIS Twitter Census: defining and describing the population of ISIS supporters on Twitter. The Brookings Project on US Relations with the Islamic World 3, 20 (2015) 12. Al-khateeb, S., Agarwal, N.: Examining botnet behaviors for propaganda dissemination: a case study of ISIL’s beheading videos-based propaganda. In: 2015 IEEE International Conference on Data Mining Workshop (ICDMW) (2015-11), pp. 51–57 (2015). https://doi.org/10.1109/ICDMW.2015.413

Bot-ivistm: Assessing Information Manipulation in Social Media Using. . .

41

13. Bessi, A., Ferrara, E.: Social bots distort the 2016 U.S. Presidential election online discussion. First Monday 21, 11 (2016). http://firstmonday.org/ojs/index.php/fm/article/view/7090 14. Ghosh, S., Viswanath, B., Kooti, F., Sharma, N.K., Korlam, G., Benevenuto, F., Ganguly, N., Gummadi, K.P.: Understanding and combating link farming in the twitter social network. In: Proceedings of the 21st International Conference on World Wide Web, pp. 61–70. ACM, New York (2012). http://dl.acm.org/citation.cfm?id=2187846 15. Liu, Y., Tang, M., Zhou, T., Do, Y.: Core-like groups result in invalidation of identifying superspreader by k-shell decomposition. arXiv preprint arXiv:1409.5187 (2014) 16. Carley, K.M.: Group stability: a socio-cognitive approach. Adv. Group Process. 7(1), 44 (1990) 17. Carley, K.M., Martin, M.K., Hirshman, B.R.: The etiology of social change. Top. Cogn. Sci. 1(4), 621–650 (2009) 18. Benigni, M., Joseph, K., Carley, K.M.: Online extremism and the communities that sustain it: detecting the ISIS supporting community on Twitter. PLoS One. 12(12), e0181405 (2017) 19. Carley, K.M.: ORA: a toolkit for dynamic network analysis and visualization. In: Alhajj, R., Rokne, J. (eds.) Encyclopedia of Social Network Analysis and Mining. Springer, New York (2017). https://doi.org/10.1007/978-1-4614-7163-9_309-1 20. Viswanath, B., Mondal, M., Clement, A., Druschel, P., Gummadi, K.P., Mislove, A., Post, A.: Exploring the design space of social network-based Sybil defenses. In: 2012 Fourth International Conference on Communication Systems and Networks (COMSNETS 2012), pp. 1–8 (2012). https://doi.org/10.1109/COMSNETS.2012.6151333 21. Yang, Z., Wilson, C., Wang, X., Gao, T., Zhao, B.Y., Dai, Y.: Uncovering social network sybils in the wild. ACM Trans. Knowl. Discov. Data (TKDD). 8(1), 2 (2014). https://doi.org/10.1145/2556609 22. Messias, J., Schmidt, L., Oliveira, R., Benevenuto, F.: You followed my bot! Transforming robots into influential users in Twitter. First Monday 18, 7 (2013) 23. Page, L., Brin, S., Motwani, R., Winograd, T.: The PageRank Citation Ranking: Bringing Order to the Web. Technical Report. Stanford InfoLab. 2 (1999) 24. Chen, J., Saad, Y.: Dense subgraph extraction with application to community detection. IEEE Trans. Knowl. Data Eng. 24(7), 1216–1230 (2012). https://doi.org/10.1109/TKDE.2010.271 25. Boshmaf, Y., Muslukhov, I., Beznosov, K., Ripeanu, M.: The social bot network: when bots socialize for fame and money. In: Proceedings of the 27th Annual Computer Security Applications Conference (2011) (ACSAC ’11), pp. 93–102. ACM, New York (2011). https://doi.org/10.1145/2076732.2076746 26. Boshmaf, Y., Muslukhov, I., Beznosov, K., Ripeanu, M.: Design and analysis of a social botnet. Comput. Netw. 57(2), 556–578 (2013) 27. Freitas, C., Benevenuto, F., Ghosh, S., Veloso, A.: Reverse engineering social bot infiltration strategies in Twitter. In: Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, pp. 25–32. ACM, New York (2015). http:/ /dl.acm.org/citation.cfm?id=2809292 28. Zhang, C.M., Paxson, V.: Detecting and analyzing automated activity on Twitter. In: Passive and Active Measurement, pp. 102–111. Springer, Berlin, Heidelberg (2011). https://doi.org/10.1007/978-3-642-19260-9_11 29. Wu, L., Hu, X., Morstatter, F., Liu, H.: Adaptive Spammer Detection with Sparse Group Modeling. ICWSM (2017) 30. Ferrara, E., Varol, O., Menczer, F., Flammini, A.: Detection of promoted social media campaigns. In: Tenth International AAAI Conference on Web and Social Media (2016) 31. Romero, D.M., Tan, C., Kleinberg, J.: On the interplay between social and topical structure. In: Proceedings of the 7th International AAAI Conference on Weblogs and Social Media (ICWSM) (2013) 32. Joseph, K., Carley, K.M.: Culture, Networks, Twitter and Foursquare: Testing a Model of Cultural Conversion with Social Media Data (2015) 33. Cresci, S., Di Pietro, R., Petrocchi, M., Spognardi, A., Tesconi, M.: The paradigmshift of social spambots: evidence, theories, and tools for the arms race. arXiv preprint arXiv:1701.03017 (2017). https://arxiv.org/abs/1701.03017

42

M. C. Benigni et al.

34. Ruths, D., Pfeffer, J.: Social media for large studies of behavior. Science. 346(6213), 1063– 1064 (2014). https://doi.org/10.1126/science.346.6213.1063 35. Tufekci, Z.: Big questions for social media big data: representativeness, validity and other methodological pitfalls. In: ICWSM ’14: Proceedings of the 8th International AAAI Conference on Weblogs and Social Media (2014). http://arxiv.org/abs/1403.7400 36. Viswanath, B., Bashir, M.A., Zafar, M.B., Bouget, S., Guha, S., Gummadi, K.P., Kate, A., Mislove, A.: Strength in numbers: robust tamper detection in crowd computations. In: Proceedings of the 2015 ACM on Conference on Online Social Networks (COSN ’15), pp. 113–124. ACM, New York (2015). https://doi.org/10.1145/2817946.2817964 37. Kakhki, A.M., Kliman-Silver, C., Mislove, A.: Iolaus: securing online content rating systems. In: Proceedings of the 22nd International Conference on World Wide Web (WWW ’13), pp. 919–930. ACM, New York (2013). https://doi.org/10.1145/2488388.2488468 38. Gupta, A., Lamba, H., Kumaraguru, P., Joshi, A.: Faking sandy: characterizing and identifying fake images on twitter during hurricane sandy. In: Proceedings of the 22nd International Conference on World Wide Web, pp. 729–736. ACM, New York (2013). http://dl.acm.org/ citation.cfm?id=2488033 39. Conover, M., Ratkiewicz, J., Francisco, M.R., Gonçalves, B., Menczer, F., Flammini, A.: Political polarization on twitter. ICWSM. 133, 89–96 (2011). http://www.aaai.org/ocs/index.php/ICWSM/ICWSM11/paper/viewFile/2847/3275.pdf 40. Conrad Nied, A., Stewart, L., Spiro, E., Starbird, K.: Alternative narratives of crisis events: communities and social botnets engaged on social media. In: Companion of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing, pp. 263–266. ACM, New York (2017). http://dl.acm.org/citation.cfm?id=3026307 41. Zhang, X., Li, Z., Zhu, S., Liang, W.: Detecting spam and promoting campaigns in Twitter. ACM Trans. Web. 10(1), 4:1–4:28 (2016). https://doi.org/10.1145/2846102 42. Goodman, L.A.: Snowball sampling. Ann. Math. Stat. 32(1), 148–170 (1961). https://doi.org/10.1214/aoms/1177705148 43. Benigni, M.: Detection and analysis of online extremist communities. Ph.D. Thesis, School of Computer Science, Carnegie Mellon University (2016) 44. Bonacich, P.: Power and centrality: a family of measures. Am. J. Sociol. 92(5), 1170–1182 (1987) 45. Kitsak, M., Gallos, L.K., Havlin, S., Liljeros, F., Muchnik, L., Eugene Stanley, H., Makse, H.A.: Identification of influential spreaders in complex networks. Nat. Phys. 6(11), 888–893 (2010) 46. Kwak, H., Lee, C., Park, H., Moon, S.: What is Twitter, a social network or a news media. In: Proceedings of the 19th International Conference on World Wide Web, pp. 591–600. ACM, New York (2010) 47. Kumar, R., Raghavan, P., Rajagopalan, S., Tomkins, A.: Trawling the Web for emerging cybercommunities. Comput. Netw. 31(11), 1481–1493 (1999) 48. Chu, Z., Gianvecchio, S., Wang, H., Jajodia, S.: Who is tweeting on Twitter: human, bot, or cyborg? In: Proceedings of the 26th Annual Computer Security Applications Conference, pp. 21–30. ACM, New York (2010). http://dl.acm.org/citation.cfm?id=1920265

Studying Fake News via Network Analysis: Detection and Mitigation Kai Shu, H. Russell Bernard, and Huan Liu

Abstract Social media is becoming increasingly popular for news consumption due to its easy access, fast dissemination, and low cost. However, social media also enables the wide propagation of “fake news,” i.e., news with intentionally false information. Fake news on social media can have significant negative societal effects. Identifying and mitigating fake news also presents unique challenges. To tackle these challenges, many existing research efforts exploit various features of the data, including network features. In essence, a news dissemination ecosystem involves three dimensions on social media, i.e., a content dimension, a social dimension, and a temporal dimension. In this chapter, we will review network properties for studying fake news, introduce popular network types, and propose how these networks can be used to detect and mitigate fake news on social media.

1 Introduction Social media has become an important means of large-scale information sharing and communication in all occupations, including marketing, journalism, public relations, and more [35]. The reasons for this change in consumption behaviors are clear: (1) it is often faster and cheaper to consume news on social media compared to news on traditional media, such as newspapers or television; and (2) it is easier to share, comment on, and discuss the news with friends or other readers on social media. However, the low cost, easy access, and rapid dissemination of information of social media draws a large audience and enables the wide propagation of “fake news”, i.e., news with intentionally false information. Fake news on social media

K. Shu () · H. Liu Computer Science and Engineering, Arizona State University, Tempe, AZ, USA e-mail: [email protected]; [email protected] H. R. Bernard Institute for Social Science Research, Arizona State University, Tempe, AZ, USA e-mail: [email protected] © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_3

43

44

K. Shu et al.

is growing quickly in volume and can have negative societal impacts. First, people may accept deliberate lies as truths [17]; second, fake news can change the way people respond to legitimate news; and finally, the prevalence of fake news has the potential to break the trustworthiness of the entire news ecosystem. In this chapter, we discuss recent advancements—based on a network perspective—for the detection and mitigation of fake news. Fake news on social media presents unique challenges. First, fake news is intentionally written to mislead readers, which makes it nontrivial to detect simply based on content. Second, social media data is large scale, multimodal, mostly user generated, sometimes anonymous and noisy. Third, the consumers of social media come from different backgrounds, have disparate preferences or needs, and use social media for varied purposes. Finally, the low cost of creating social media accounts makes it easy to create malicious accounts, such as social bots, cyborg users, and trolls, all of which can become powerful sources for proliferation of fake news. The news dissemination ecosystem on social media involves three dimensions (Fig. 1), a content dimension (“What”), a social dimension (“Who”), and a temporal dimension (“When”). The content dimension describes the correlation among news pieces, social media posts, comments, etc. The social dimension involves the relations among publishers, news spreaders, and consumers. The temporal dimension illustrates the evolution of users’ publishing and posting behaviors over time. As we will show, we can use these relations to detect and mitigate the effects of fake news. Detection of fake news can be formalized as a classification task that requires feature extraction and model construction. Recent advancements of network representation learning, such as network embedding and deep neural networks, allow us to better capture the features of news from auxiliary information such as friendship network, temporal user engagements, and interaction networks. In addition, knowledge networks as auxiliary information can help evaluate the veracity of news through network-matching operations such as path finding and flow optimization. For mitigation, the aim is to proactively block target users or start a mitigating

Fig. 1 The information dimensions of the news dissemination ecosystem

Studying Fake News via Network Analysis: Detection and Mitigation

45

campaign at an early stage. We will show that network diffusion models can be applied to trace the provenance nodes and provenance paths of fake news. In addition, the impact of fake news can be assessed and mitigated through network estimation and network influence minimization strategies. We begin, then, with an introduction to network properties.

2 Network Properties In this section, we outline the potential role of network properties for the study of fake news. First, users form groups with like-minded people, resulting in what are widely known as echo chambers. Second, individual users play different roles in the dissemination of fake news. Third, social media platforms allow users to personalize how information is presented to them, thus isolating users from information outside their personalized filter bubbles. Finally, highly active malicious user accounts become powerful sources and proliferators of fake news.

2.1 Echo Chambers The process of seeking and consuming information on social media is becoming less mediated. Users on social media tend to follow like-minded people and thus receive news that promotes their preferred, existing narratives. This may increase social polarization, resulting in an echo chamber effect [2]. The echo chamber effect facilitates the process by which people consume and believe fake news based on the following psychological factors [17]: (1) social credibility, which means that people are more likely to perceive a source as credible if others perceive it as such, especially when there is not enough information available to assess the truthfulness of that source; and (2) frequency heuristic, which means that consumers may naturally favor information they hear frequently, even if it is fake news. In echo chambers, users share and consume the same information, which creates segmented and polarized communities.

2.2 Individual Users During the fake news dissemination process, individual users play different roles. For example: (1) persuaders spread fake news with supporting opinions to persuade and influence others to believe it; (2) gullible users are credulous and easily persuaded to believe fake news; and (3) clarifiers propose skeptical and opposing viewpoints to clarify fake news. Social identity theory [31] suggests that social acceptance and affirmation is essential to a person’s identity and self-esteem,

46

K. Shu et al.

making persuaders likely to choose “socially safe” options when consuming and disseminating news information. They follow the norms established in the community even if the news being shared is fake news. The cascade of fake news is driven not only by influential persuaders but also by a critical mass of easily influenced individuals [7], i.e., gullible users. Gullibility is a different concept from trust. In psychological theory [20], general trust is defined as the default expectations of other people’s trustworthiness. High trusters are individuals who assume that people are trustworthy unless proven otherwise. Gullibility, on the other hand, is insensitivity to information revealing untrustworthiness. Reducing the diffusion of fake news to gullible users is critical to mitigating fake news. Clarifiers can spread opposing opinions against fake news and avoid one-sided viewpoints. Clarifiers can also spread true news which can: (1) immunize users against changing beliefs before they are affected by fake news; and (2) further propagate and spread true news to other users.

2.3 Filter Bubbles A filter bubble is an intellectual isolation that occurs when social media websites use algorithms to personalize the information a user would want to see [16]. The algorithms make assumptions about user preferences based on the user’s historical data, such as former click behavior, browsing history, search history, and location. Given these assumptions, the website is more likely to present information that will support the user’s past online activities. A filter bubble can reduce connections with contradicting viewpoints, causing the user to become intellectually isolated. A filter bubble will amplify the individual psychological challenges to dispelling fake news. These challenges include: (1) Naïve Realism [33]: consumers tend to believe that their perceptions of reality are the only accurate views, while others who disagree are regarded as uninformed, irrational, or biased; and (2) Confirmation Bias [15]: consumers prefer to receive information that confirms their existing views.

2.4 Malicious Accounts Social media users can be malicious, and some malicious users may not even be real humans. Malicious accounts that can amplify the spread of fake news include social bots, trolls, and cyborg users. Social bots are social media accounts that are controlled by a computer algorithm. The algorithm automatically produces content and interacts with humans (or other bot users) on social media. Social bots can be malicious entities designed specifically for manipulating and spreading fake news on social media. Trolls are real human users who aim to disrupt online communities and provoke consumers to an emotional response. Trolls enable the easy dissemination of fake news among otherwise normal online communities.

Studying Fake News via Network Analysis: Detection and Mitigation

47

Finally, cyborg users can spread fake news in a way that blends automated activities with human input. Cyborg accounts are usually registered by a human as a disguise for automated programs that are set to perform activities on social media. The easy switch between humans and bots offers the cyborg users unique opportunities to spread fake news.

3 Network Types In this section, we introduce several network structures that are commonly used to detect and mitigate fake news. Then, following the three dimensions of the news dissemination ecosystem outlined above, we illustrate how homogeneous and heterogeneous networks can be built within a specific dimension and across dimensions.

3.1 Homogeneous Networks Homogeneous networks have the same node and link types. As shown in Fig. 2, we introduce three types of homogeneous networks: friendship networks, diffusion networks, and credibility networks. Each of these types is potentially useful in detecting and mitigating fake news. Friendship Networks A user’s friendship network in the social layer can be represented as a directed graph GF = (U, EF ), where U and EF are the node and edge sets, respectively. A node u ∈ U represents a user, and (u1 , u2 ) ∈ E represents whether a social relation exists. Homophily theory [13] suggests that users tend to form relationships with like-minded friends, rather than with users who have opposing preferences and

Fig. 2 Homogeneous networks. Three types of homogeneous networks are illustrated: (a) friendship network, (b) diffusion network, and (c) credibility network. Node u indicates a user, and s represents a social media post

48

K. Shu et al.

interests. Likewise, social influence theory [12] predicts that users are more likely to share similar latent interests towards news pieces. Thus, the friendship network provides the structure to understand the set of social relationships among users. The friendship network is the basic route for news spreading and can reveal community information. Diffusion Networks A diffusion network in the social layer can be represented as a directed graph GD = (U, ED , p, t), where U and E are the node and edge sets, respectively. A node u ∈ U represents an entity, which can publish, receive, and propagate information at time ti ∈ t. A directed edge, (u1 → u2 ) ∈ ED , between nodes u1 , u2 ∈ U represents the direction of information propagation. Each directed edge, (u1 → u2 ) ∈ ED , between nodes u1 , u2 ∈ U represents the direction of information propagation. Each directed edge (u1 → u2 ) is assumed to be associated with an information propagation probability, p(u1 → u2 ) ∈ [0, 1]. The diffusion network is important for learning about representations of the structure and temporal patterns to help identify fake news. By discovering the sources of fake news and the spreading paths among the users, we can also better mitigate fake news problem. Credibility Networks A credibility network [9] can be represented as an undirected graph GC = (V , EC , s), where V denotes the set of social media posts with corresponding credibility scores that are related to original news pieces, and the edges E denote the link type, such as supporting and opposing, between two nodes, and c(v1 , v2 ) indicates the conflicting degree of credibility values of node v1 and v2 . Users express their viewpoints towards original news pieces through social media posts. In these posts, they can either express the same viewpoints (which mutually support each other), or conflicting viewpoints (which may reduce their credibility scores). By modeling these relationships, the credibility network can be used to evaluate the overall truthfulness of news by leveraging the credibility scores of each social media post relevant to the news.

3.2 Heterogeneous Networks Heterogeneous networks have a different set of node and link types. The advantages of heterogeneous networks are the abilities to represent and encode information and relationships from different perspectives. During the news dissemination process, different types of entities are involved, including users, the social media posts, the actual news, etc. Figure 3 shows the common types of heterogeneous networks for analyzing fake news: knowledge networks, stance networks, and interaction networks. Knowledge Networks A knowledge network (i.e., knowledge graph) GK = (I, EI , R) is constructed with the nodes I representing the knowledge entities,

Studying Fake News via Network Analysis: Detection and Mitigation

49

Fig. 3 Heterogeneous networks. Three types of heterogeneous networks are illustrated: (a) knowledge network, (b) stance network, and (c) interaction network. Node o indicates a knowledge entity, v represents a news item, and p means a news publisher

edges EI indicating the relation between them, R is the relation sets, and g : E → R is the function labeling each edge with a semantic predicate. The knowledge network integrates linked open data, such as DBdata and Google Relation Extraction Corpus (GREC), as a heterogeneous network topology. Factchecking using a knowledge graph checks whether the claims in news content can be inferred from existing facts in the knowledge networks [6, 23]. Stance Networks A heterogeneous stance network can be represented as a heterogeneous network GS = ({U, S, V }, ES ), where the nodes can be users, news items, and social media posts; and the edges ES denote the link type between two nodes, such as posting between users and posts, and stance between posts and news items. The stances are treated as important signals and can be aggregated to infer the news veracity. Stances (or viewpoints) indicate the users’ opinions towards the news, such as supporting, opposing, etc. Typically, fake news pieces will provoke tremendous controversial views among social media users, in which denying and questioning stances are found to play a crucial role in signaling claims as being fake. Interaction Networks An interaction network GI = ({P , U, V }, EI ) consists of nodes representing publishers, users, news, and the edges EI indicating the interactions among them. For example, edge (p → v) demonstrates that publisher p publishes news item v, and (v → u) represents that news v is spread by user u. The interaction networks can represent the correlations among different types of entities, such as publisher, news, and social media post, during the news dissemination process [26]. The characteristics of publishers and users, and the publisher–news and news–users interactions have potential to differentiate fake news.

50

K. Shu et al.

4 Fake News Detection Fake news detection evaluates the truth value of a news piece, which can be formalized as a classification problem. The common procedure is feature extraction and model construction. In feature extraction, we capture the differentiable characteristics of news pieces to construct effective representations; based on these representations, we can construct various models to learn and transform the features into a predicted label. To this end, we introduce how features and models can be extracted and constructed in different types of networks.

4.1 Interaction Network Embedding Interaction networks describe the relationships among different entities such as publishers, news pieces, and users. Given the interaction networks, the goal is to embed the different types of entities into the same latent space, by modeling the interactions among them. We can leverage the resultant feature representations of news to perform fake news detection. News Embedding We can use news content to find clues to differentiate fake news and true news. Using nonnegative matrix factorization (NMF), we can attempt to project the document–word matrix to a joint latent semantic factor space with low dimensionality, such that the document–word relations are modeled as the inner n×t product in the space. Specifically, giving the news–word matrix X ∈ R+ , NMF n×d t×d methods try to find two nonnegative matrices D ∈ R+ and V ∈ R+ by solving the following optimization problem, min  X − DVT 2F

D,V≥0

(1)

where d is the dimension of the latent topic space. In addition, D and V are the nonnegative matrices indicating low-dimensional representations of news and words. User Embedding On social media, people tend to form relationships with likeminded friends, rather than with users who have opposing preferences and interests [28]. Thus, connected users are more likely to share similar latent interests in news pieces. To obtain a standardized representation, we use nonnegative matrix factorization to learn the user’s latent representations (we will introduce other methods in Sect. 4.3). Specifically, giving user–user adjacency matrix A ∈ {0, 1}m×m , we learn nonnegative matrix U ∈ Rm×d by solving the following + optimization problem, min Y  (A − UTUT )2F

U,T≥0

(2)

Studying Fake News via Network Analysis: Detection and Mitigation

51

d×d where U is the user latent matrix, T ∈ R+ is the user–user correlation matrix, and Y ∈ Rm×m controls the contribution of A. Since only positive samples are given in A, we can first set Y = sign(A), then perform negative sampling and generate the same number of unobserved links and set weights as 0.

User–News Embedding The user–news interactions can be modeled by considering the relationships between user attributes and the level of veracity of news items. Intuitively, users with low credibilities are more likely to spread fake news, while users with high-credibility scores are less likely to spread fake news. Each user has a credibility score that we can infer using his/her published posts [1], and we use c = {c1 , c2 , . . . , cm } to denote the credibility score vector, where a larger ci ∈ [0, 1] indicates that user ui has a higher credibility. The user–news engaging matrix is represented as W ∈ {0, 1}m×n , where Wij = 1 indicates that user ui has engaged in the spreading process of the news piece vj ; otherwise, Wij = 0. The user–news embedding objective function is shown as follows: min

 1 + yLj ||Ui − Dj ||22 Wij ci 1 − 2 i=1 j =1



r m  

Truenews r m  



1 + yLj + Wij (1 − ci ) 2 i=1 j =1



(3)

||Ui − Dj ||22

Fakenews

where yL ∈ Rr×1 is the label vector of all partially labeled news. The objective considers two situations: (1) for true news, i.e., yLj = −1, which ensures that the distance between latent features of high-credibility users and that of true news is small; and (2) for fake news, i.e., yLj = 1, which ensures that the distance between the latent features of low-credibility users and that of true news is small. Publisher–News Embedding The publisher–news interactions can be modeled by incorporating the characteristics of the publisher and news veracity values. Fake news is often written to convey opinions or claims that support the partisan bias of the news publisher. Publishers with a high degree of political bias are more likely to publish fake news [26]. Thus, a useful news representation should be good at predicting the partisan bias score of its publisher. The partisan bias scores are collected from fact-checking websites and can be represented as a vector o. We can utilize publisher partisan labels vector o ∈ Rl×1 and publisher–news matrix B ∈ Rl×n to optimize the news feature representation learning as follows: ¯ min  BDQ − o22

(4)

where the latent feature of news publisher can be represented by the features of all ¯ B¯ is the normalized user–news publishing relation the news it published, i.e., BD.

52

matrix, i.e., B¯ kj =

K. Shu et al. B n kj

j =1 Bkj

. Q ∈ Rd×1 is the weighting matrix that maps news

publishers’ latent features to corresponding partisan label vector o. The finalized model combines all previous components into a coherent model. In this way, we can obtain the latent representations of news items D and of users U through the network embedding procedure, which can be utilized to perform fake news classification tasks.

4.2 Temporal Diffusion Representation The news diffusion process involves abundant temporal user engagements on social media [21, 25, 34]. The social news engagements can be defined as a set of tuples E = ei to represent the process of how news items spread over time among m users in U = {u1 , u2 , . . . , um }. Each engagement ei = {ui , ti , si } represents that a user ui spreads news article at time ti by posting si . As shown in Fig. 4, the information diffusion network consists of two major parts of knowledge: (1) temporal user engagements and (2) a friendship network. For example, a diffusion path between two users ui and uj exists if and only if (1) uj follows ui ; and (2) uj posts about a given news only after ui does so. The goal of learning temporal representations is to capture the user’s pattern of temporal engagements with a news article vj . Recent advances in the study of deep neural networks, such as recurrent neural networks (RNN), have shown promising performance for learning representations. RNN are powerful structures that allow the use of loops within the neural network to model sequential data. Given the diffusion network GD , the key procedure is to construct meaningful features xi for each engagement ei . The features can generally be extracted from the contents of

Fig. 4 A diffusion network consists of temporal user engagements and a friendship network

Studying Fake News via Network Analysis: Detection and Mitigation

53

Fig. 5 RNN framework for learning news temporal representations

si and the attributes of ui . For example, xi consists of the following components: xi = (η, Δt, xui , xsi ). The first two variables, η and Δt, represent the number of total user engagements through time t and the time difference between engagements, respectively. These variables capture the general measure of frequency and time interval distribution of user engagements of the news piece vj . For the content features of users posts, the xsi can be extracted from handcrafted linguistic features, such as n-gram features, or by using word embedding methods such as doc2vec [11] or GloVe [18]. We can extract the features of users xui by performing a singular value decomposition of the user–news interaction matrix W ∈ {0, 1}m×n , where Wij = 1 indicate that user ui has engaged in the process of spreading the news piece vj ; otherwise, Wij = 0. The RNN framework for learning news temporal representations is demonstrated in Fig. 5. Since xi includes features that come from different information space, such as temporal and content features, so we do not suggest incorporating xi into RNN as the raw input. Thus, we can add a fully connected embedding layer to convert the raw input xi into a standardized input features x˜ i , in which the parameters are shared among all raw input features xi , i = 1, . . . , m. Thus, the RNN takes a sequence x˜ 1 , x˜ 2 , . . . , x˜ m as input. At each time-step i, the output of previous step hi−1 , and the next feature input x˜ i are used to update the hidden state hi . The hidden states hi is the feature representation of the sequence up to time i for the input engagement sequence. Thus, the hidden states of final step hm is passed through a fully connected layer to learn the resultant news representation, defined as vj = tanh(Wr hm + br ), where Wr is the weight matrix and br is a bias vector. Thus, we can use vj to perform fake news detection and related tasks [21].

4.3 Friendship Network Embedding News temporal representations can capture the evolving patterns of news spreading sequences. However, we lose the direct dependencies of users, which plays an important role in fake news diffusion. The fact that users are likely to form echo chambers strengthens our need to model user social representations and to explore its added value for a fake news study. Essentially, given the friendship network GF , we want to learn latent representations of users while preserving the structural properties of the network, including first-order and higher-order structure, such as second-order structure and community structure. For example, Deepwalk [19] can

54

K. Shu et al.

preserve the neighborhood structure of nodes by modeling a stream of random walks. In addition, LINE [30] can preserve both first-order and second-order proximities. Specifically, we can measure the first-order proximity by the joint probability distribution between the user ui and uj , p1 (ui , uj ) =

1 1 + exp(−ui T uj )

(5)

where ui (uj ) is the social representation of user ui (uj ). We can model the secondorder proximity by the probability of the context user uj being generated by the user ui , as follows: exp(uj T ui ) p2 (uj |ui ) = |V | T k=1 exp(uk ui )

(6)

where |V | is the number of nodes or “contexts” for user ui . This conditional distribution implies that users with similar distributions over the contexts are similar to each other. The learning objective is to minimize the KL-divergence of the two distributions and empirical distributions, respectively. Network communities may actually be the more important structural dimension because fake news spreaders are likely to form polarized groups [25]. This requires the representation learning methods to be able to model community structures. For example, a community-preserving node representation learning method, modularized nonnegative matrix factorization (MNMF), is proposed [32]. The overall objective is defined as follows: min  S − MUT 2F + αH − UCT 2F − βtr(HT BH)



M,U,H,C≥0







Proximity Mapping Community Mapping Modularity Modeling

(7)

s.t. tr(H H) = m T

and comprises three major parts: proximity mapping, community mapping, and modularity modeling. In proximity mapping, S ∈ Rm×m is the user similarity matrix constructed from the user adjacency matrix (first-order proximity) and neighborhood similarity matrix (second-order proximity), and M ∈ Rm×k and U ∈ Rm×k are the basis matrix and user representations. For community mapping, H ∈ Rm×l is the user–community indicator matrix that we optimize to be reconstructed by the product of the user latent matrix U and the community latent matrix C ∈ Rl×m . For modularity modeling, it represents the objective to maximize the modularity function [14], and B ∈ Rm×m is the modularity matrix.

Studying Fake News via Network Analysis: Detection and Mitigation

4.3.1

55

Credibility Network Propagation

The basic assumption is that the credibility of a given news event is highly related to the credibilities of its relevant social media posts [9]. To classify whether a news item is true or fake, we can collect all relevant social media posts. Then, we can evaluate the news veracity score by averaging the credibility scores of all the posts. Given the credibility network GC for specific news pieces, the goal is to optimize the credibility values of each node (i.e., social media post), and infer the credibility value of corresponding news items [9]. In the credibility network GC , there are (1) a post credibility vector T = {C(s1 ), C(s2 ), . . . , C(sn )} with C(si ) denoting the credibility value of post si ; and (2) a matrix W ∈ Rn×n , where Wij = f (si , sj ) which denotes the viewpoint correlations between post si and sj , that is, whether the two posts take supporting or opposing positions. Network Initialization Network initialization consists of two parts: node initialization and link initialization. First, we obtain the initial credibility score vector of nodes T0 from pre-trained classifiers with features extracted from external training data. The link is defined by mining the viewpoint relations, which are the relations between each pair of viewpoint such as contradicting or the same. The basic idea is that posts with the same viewpoints form supporting relations which raise their credibilities, and posts with contradicting viewpoints form opposing relations which weaken their credibilities. Specifically, a social media post si is modeled as a multinomial distribution θi over K topics, and a topic k is modeled as a multinomial distribution ψtk over L viewpoints. The probability of a post st over topic k along with L viewpoints is denoted as pik = θi × ψik . The distance between two posts st and sj is measured by using the Jensen–Shannon distance: Dis(si , sj ) = DJ S (pik ||pj k ). The supporting or opposing relation indicator is determined as follows: it is assumed that one post contains a major topic-viewpoint, which can be defined as the largest proportion of pi k. If the major topic-viewpoints of two posts si and sj are clustered together (they take the same viewpoint), then they are mutually supporting; otherwise, they are mutually opposing. The similarity/dissimilarity measure of two posts is defined as: f (si , sj ) =

(−1)a DJ S (pik ||pj k ) + 1

(8)

where a is the link type indicator, and if a = 0, then si and sj take the same viewpoint; otherwise, a = 1. Network Optimization Posts with supporting relations should have similar credibility values; posts with opposing relations should have opposing credibility values. Therefore, the objective can be defined as a network optimization problem as below:

56

K. Shu et al.

⎛ Q(T) =μ

n 

⎞2

C(sj ) ⎟ ⎜ C(si ) |Wij | ⎝  − bij  ⎠ ¯ Dii ¯ jj D i,j =1

(9)

+ (1 − μ)T − T0 2  ¯ is a diagonal matrix with D ¯ ii = where D k |Wik | and bij = 1, if Wij ≥ 0; otherwise, bij = 0. The first component is the smoothness constraint that guarantees the two assumptions of supporting and opposing relations; the second component is the fitting constraint to ensure variables not change too much from their initial values; and μ is the regularization parameter to trade off two constraints. Then, the credibility propagation on the proposed network GC is formulated as the minimization of this loss function: T∗ = arg min Q(T)

(10)

T

The optimum solution can be solved by updating T in an iterative manner through ¯ −1/2 WD ¯ −1/2 . the transition function T(t) = μHT(t − 1) + (1 − μT0 ), where H = D As the iteration converges, each post receives a final credibility value, and the average of them is served as the final credibility evaluation result for the news.

4.4 Knowledge Network Matching In this section, we focus on exploiting knowledge networks to detect fake news. Knowledge networks are used as an auxiliary source to fact-check news claims. The goal is to match news claims with the facts represented in knowledge networks.

4.4.1

Path Finding

Fake news spreads false claims in news content, so a natural means of detecting fake news is to check the truthfulness of major claims in the news article. Factchecking methods use external sources such as knowledge networks, to assess the truthfulness of information. Specifically, a news claim can be checked automatically by finding the matching path to knowledge networks. A claim in news content can be represented by a subject–predicate–object triple c = (s, p, o), where the subject entity s is related to the object entity o by the predicate relation p. We can find all the paths that start with s and end with o, and then evaluate these paths to estimate the truth value of the claim. This set of paths, also named knowledge stream [24], are denoted as P(s, o). Intuitively, if the paths involve more specific entities, then the claim is more likely to be true. Thus, we can define a “specificity” measure S(Ps,o ) as follows:

Studying Fake News via Network Analysis: Detection and Mitigation

S(Ps,o ) =

1+

1 n−1 i=2

log d(oi )

57

(11)

where d(oi ) is the degree of entity oi , i.e., the number of paths that entity o participates. One approach is to optimize a path evaluation function: τ (c) = max W(Ps,o ), which maps the set of possible paths connecting s and o (i.e., Ps,o ) to a truth value τ . If s is already present in the knowledge network, it can assign maximum truth value 1; otherwise, the objective function will be optimized to find the shortest path between s and o.

4.4.2

Flow Optimization

We can assume that each edge of the network is associated with two quantities: a capacity to carry knowledge related to (s, p, o) across its two endpoints, and a cost of usage. The capacity can be computed using S(Ps,o ), and the cost of an edge in knowledge is defined as ce = log d(oi ). The goal is to identify the set of paths responsible for the maximum flow of knowledge between s and o at the minimum cost. The maximum knowledge a path Ps,o can carry is the minimum knowledge of its edges, also called its bottleneck B(Ps,o ). Thus, the objective can be defined as a minimum cost maximum flow problem, τ (e) =



B(Ps,o ) · S(Ps,o )

(12)

Ps,o ∈Ps,o

where B(Ps,o ) is denoted as a minimization form: B(Ps,o ) = min{xe | ∈ Ps,o }, with xe indicating the residual capacity of edge x in a residual network [24]. Discussion The knowledge network itself can be incomplete and noisy. For example, the entities in fake news claims may not correspond to any path exactly or may match multiple entities in the knowledge network. In this case, only performing path finding and flow optimization is not enough to obtain a good result to assess the truth value. Therefore, additional tasks (e.g., entity resolution, and link prediction) need to be considered in order to reconstruct the knowledge network and to facilitate its capability. Entity resolution is the process of finding related entities and creating links among them. Link prediction predicts the unseen links and relations among the entities.

4.4.3

Stance Network Aggregation

We can present the stance of users’ posts either explicitly or implicitly. Explicit stances are direct expressions of emotion or opinion, such as Facebook’s “like” actions. Implicit stances can be automatically extracted from social media posts.

58

K. Shu et al.

Consider the scenario where the stances are explicitly expressed in “like” actions on social media. Given the stance network GS = {U ∪ S ∪ V , ES }, the first step is to construct a bipartite graph (U ∪ V , L), where L is the set of likes actions. The idea is that user express “like” actions due to both the user reputations and news qualities. The users and news items can be characterized by the Beta distributions Beta(αi , βi ) and Beta(αj , βj ), respectively. The distribution of a user Beta(αi , βi ) represents the reputation or reliability of user ui , and the distribution of a new piece Beta(αj , βj ) represents the veracity of news vj . The expectation values of the Beta i distribution are used to estimate the degree of user reputation (pi = αiα+β ) or new i veracity (pj =

αj αj +βj

). To predict whether a piece of news is fake or not, the linear α −β

transformation of pj is computed: qj = 2pj − 1 = αjj +βjj , where a positive value indicates true news; otherwise, it is fake news. The model is trained in a semi-supervised manner. Let the training set consists of two subsets VF , VT ⊆ V for labeled fake and true news, and Φi = {ui |(ui , vj ) ∈ L} and Φj = {vj |(ui , vj ) ∈ L}. The labels are set as qj = −1 for all vj ∈ IF , and qj = 1 for all vj ∈ IT , and qj = 0 for unlabeled news pieces. The parameter optimization of user ui is performed iteratively by following updating functions: αi =Δα +



qi

qi >0,i∈Φi

βi =Δβ −



qi

(13)

qi 0,j ∈Φj

βj =Δβ −



qj

(14)

qj 0) can only see user profile and activities. (2) The system can start with an initial setting of two options: (a) Open-privacy: where all users with an edge-value > 0 can see user profile and all posted activities. As an alternative, and for manual consideration, user can temporarily block some of their friends from seeing their profiles and activities. This static decision will keep those “friend” values as zeros, which will make them “temporary unfriends”, (b) Closed-privacy in which no user can see any activity unless explicitly added by the user. This decision is memorized by the system for future activities and will only be nullified by second explicit decisions. (3) As friends interact with user activities, their edge-trust values are gradually increased. (This applies for the closed-privacy model, as open-model assumes all friends can see all activities initially.) (4) Eventually, the privacy level is decided by the level of interaction from those friends. This can keep increasing or decreasing with time based on their future interactions. This means that future privacy level is decided by current interaction level with the user.

6.2 Friends’ Recommendation System One of the main functionalities in OSNs is to promote expanding the usage of their social networks through the promotion of more interaction from users with network applications or with each other using the network tools. Most of the algorithms to recommend new friends use similar attributes related to the person country, profession, interests, common friends, etc. Our goal is to use earlier proposed interactive-based reputation system to provide friends’ recommendations based on individual strengths of relations with friends and friends interactions with user activities. Our study of relations in social networks showed that the majority of friends in any user in OSNs are silent or low profile friends who either rarely or

Testing Assessment of Group Collaborations in OSNs

145

occasionally interact with user activities. For example, in Facebook the three main friends’ interactive actions are: likes, comments, and wall tags. Wall tags seem to be the highest indicator of relationship-strength specially as it requires a special permission by node owner. Commenting on photos, videos, posts, etc. is the middle level activity where a number of friends will make comments periodically. The volume of such comments can be another indicator of the closeness or the weight of the relationship. The amount of likes seems to be the largest and the lowest friendship-indicator. We surveyed a large number of users in Facebook and noticed that in many cases, 50% of friends or more have only (like) actions with many of their friends. Users in OSNs rarely take the decision to “unfriend” a friend although they may have friends in their lists without any interactions. There are several possible common features that can be extracted based on studying users activities in OSNs in addition to their friends interactions with those activities. These include for example how much those users are involved in the social network. Despite many common features, if two users are different in the volume of involvement in the social network they may not be considered with similar profiles. Our model includes also the level of interactions with groups, public pages, etc. Unlike classical friend matching algorithms that may try to use similar groups or research interests, we want to further evaluate the users’ involvements in those groups or public pages. Users can register in many groups or public websites while they may rarely interact with their activities. Two users that heavily interact with a certain group or a public website should be seen as good candidate friends even if they have many different attributes.

6.3 A Recommendation System for Accepting Invitations from Strangers This application for our reputation system is related to invitations from “strangers” so in the OSN if you receive an invitation to connect or anything else from a stranger, our reputation system will try to help you make the decision whether to accept or reject such invitation. The model is interaction based which means that this is a stranger to you, and never interact with, so how can you decide if they are good or bad, how about consult people who interact with this person.

6.4 Professional Referrals In OSNs referrals and users recommendations specially about commercial products are seen to many businesses as necessary tools for marketing and the assessment of users’ perceptions of commercial products. In addition to recommendations on

146

I. Alsmadi and M. Al-Abdullah

Fig. 1 Professional referral architecture

Referral/ Recommondation

Recommonder

Recomndation Context

Recommonded

products, books, etc. recommendations can be used in seeking professional or expert referrals on services (e.g., a physician, consultant, educator, clinic, restaurant, brand, etc.). Users may trust their friends and their recommendations. Some professional related OSNs such as LinkedIn allow users to seek recommendations from their friends. Interactions between friends can help us distinguish the nature of the relation between those friends (e.g., personal or professional). Such classification should be also considered when considering referral recommendations from friends. For example, a professional recommendation from a personal friend can fall within the category of conflict of interests. We assume here that OSN is capable to determine the nature of the relation that exists between those two friends (i.e., personal or professional). This requires modeling referral decisions based on four elements: (1) recommended, (2) recommender, (3) the nature of their relation in addition to (4) the context in which the recommended person is seeking the recommendation for. Figure 1 showed the general professional recommendation components. We showed the recommender in the middle as recommender needs to know all components in the system. In addition, the recommendation should not only include information about the recommended person and the context but also about the recommender.

6.5 Credit Reporting Several references indicated that credit reporting agencies (e.g., Experian and Equifax and TransUnion) are already using or aggregating information from OSNs. Our proposed reputation rank for users in OSNs can be used directly as part of credit or background reporting. Our goal is to allow users to transparently monitor and be able to improve this value that can possibly impact decisions made to give those users credits, or any other possible decisions. In classical credit and background reporting, an individual score can be negatively impacted by incidents that users are

Testing Assessment of Group Collaborations in OSNs

147

not aware of or that they have no power to change them. Reputation score is not an alternative to credit reporting that involve user financial decisions and activities but may only complement it or show a new or different dimension related to the user e-present or e-behavior in particular. Users can improve their reputation score through their friends’ selections and through the kind of activities they create. Reports showed that some reporting agencies tried to use illegal tools to access private conversations between users in OSNs. All reporting organizations depend on the aggregation of a large amount of data about users from different sources (e.g., utility bills, cell phones, government agencies, OSNs, etc.). In many cases, they may pay to acquire such data or may acquire it illegally. The availability of such wealthy data publicly through OSNs encourage those data aggregations to use them as one of their major sources of information. Our reputation model proposes an alternative where users and OSNs can control and manage the process. Our model proposes a statistical approach that does not require looking into the content of either public or private conversation. Based on the goal or usage of credit reporting, the nature of extracted information can vary. In general, the reputation rank itself can be used similar to the credit score where for example, the user will qualify for a rent application if their credit score is more than a score A. OSNs should be able to generate those models. Although some information used in the processing maybe private, however, final report should only include general statistics that does not violate users’ privacy.

6.6 Spam Detection in OSNs Spam is very popular in emails where users frequently receive unsolicited emails sent for a large number of email users largely for marketing or financial purposes. Spam problem starts also growing rapidly in OSNs. In some cases, spammers register in publicly open pages to keep posting their spam messages. In other cases, they may seek friendships to many users to be able to spread their marketing messages in those users’ activities. Many OSN users may accept friendship requests from anonymous users, intentionally or unintentionally. In addition, rarely OSN users screen their friendship lists for friends to verify whether they want to keep or remove some names. Their accounts can be possibly exposed to security problems or at least spammers. Based on our model, we proposed to alert users periodically (e.g., a monthly report) about users who rarely interact with their activities. Users may screen those names and make decisions whether they want to keep or remove those names. This can be a convenient alternative to manual investigation specially for users who have a large number of friends. This is also an alternative to contentbased spam detection which may produce a possible number of false negative or positive elimination of friends. Not all low-interactive friends should be considered as spam and system should be able to make that distinction.

148

I. Alsmadi and M. Al-Abdullah

6.7 Assessments of Networks Evolution The formation of groups in OSNs is dynamic and can change from one network to another. Information or knowledge extraction related to groups’ formations, interactions, interests, etc. can be related to many research fields. However, there are several challenges that face such knowledge extraction including the large amount of data and the need to have a focused context upon which search can be concentrated. Instrumentation or automation for such process can then be possible. Through this paper, we showed examples of how to measure dynamically the strength of relations among members in OSN agile groups. Those groups or cliques are not formally formed or defined.

6.8 Positive or Negative Interactions?! Our original reputation model is based on statistical data that does not require the analysis of text in posts, comments, etc. As such interactions are considered regardless of the nature of such interaction (e.g., whether it is a positive or negative interaction). Nonetheless, we proposed using semantic analysis to analyze public posts and information to see if friends comments with each other are positive or negative. Ultimately, friendship weighted edges can vary from −1 to 1 rather from zero to 1. For example, a friend can be heavily interacting with a user, but based on sentimental analysis, such interactions are always negative and hence the overall weight value of the relation should be negative.

6.9 Context-Driven Automatic Privacy Assessment We propose an automatic context-driven privacy system that can alert users, based on their posts or activities on who can or should see their activities. There are two modes in which this can operate: A manual mode in which the activity author can decide on the level of privacy of the activity. The privacy assessment system can then decide who can or cannot view, comment on, or propagate the activity based on the users’ friendship classification in author profile. In the second mode, which is automated, the privacy assessment system can decide the activity confidentiality classification. Sentimental analysis methods can be used to make a final judgment on the activity to be one class label from several alternatives (e.g., normal and classified or confidential). Based on these initial confidentiality classifications, the second step will be similar to the manual mode where different friends will be assigned different visibility or access levels. There are two components in this objective and each one will have its own expected outcome: • An automated system for information classification: In sentimental analysis, or opinion mining, analysts are interested to discover overall people

Testing Assessment of Group Collaborations in OSNs

149

rating/evaluation (e.g., positive or negative) regarding a political or social event, product, etc. In this project, we are interested to use similar models, but for information classification, where the system output for each message/activity is its information classification level (e.g., secret, confidential, public trust, and unclassified). Automatic Information classification in the public Internet is considered a complex and time-consuming task where application dominates the balance between performance and accuracy. • Privacy-based user-connections’ classification system: Another expected output from this objective is users’ connections’ classification based on privacy or information classification. Typically in OSNs, user connections are all classified as “friends,” “followers,” “following,” etc. (based on the OSN). Some OSN can also have classifications such as family and close friends. One of the show cases we were inspired by to develop an autonomous privacy assessment model is related to information leakage in OSNs from military families. The show case indicates that while military personnel may receive the right security training, information leakage may come from their spouses or families. As one proposed solution, we suggested to enable users to classify their friends among different categories including, family and work categories. This can help them eventually to provide different access or visibility levels on their posts according to this classification. Our goal is to develop an automatic privacy assessment system for users that help OSN users to make reasonable privacy assessment decisions for their activities and profiles with the following constraints and guidelines: • Each node in the OSN will have relations with all other nodes in the social network with a weighted value that varies between (zero; not friends) up to 1 (very close friend). The initial assumption is that only users who are friends (e.g., weighted edge is not zero) can only see user profile and activities. • The system can start with an initial setting of two options: – Open-privacy: where all users with an edge-value of more than zero can see user profile and all posted activities. As an alternative, and for manual consideration, user can temporarily block some of their friends from seeing their profiles and activities. This static decision will keep those friend values as zeros, which will make them temporary unfriends. – Closed-privacy in which no user can see any activity unless explicitly added by the user. This decision is memorized by the system for future activities and will only be nullified by a second explicit decision. • As friends interact with user activities, their edge-trust values are gradually increased. (This applies for the closed-privacy model, as open-model assumes all friends can see all activities initially.) • Eventually, the privacy level is decided by the level of interaction from those friends. This can keep increasing or decreasing with time based on their future interactions. This means that future privacy level is decided by current interaction level with the user.

150

I. Alsmadi and M. Al-Abdullah

In order to develop context-driven or sensitive privacy assessment system, following major tasks will be accomplished: • Information sensitive sentimental analysis model: This is the first major task in this system. We have to conduct first an extensive study to enable classifying certain text (e.g., post, tweet, etc.) based on information classification. If content is not information sensitive, default decision can be made (e.g., make it visible to public, friend, or default). To the best of our knowledge, no existing dictionaries (public) are available that can classify the sentiment of words based on information sensitivity. We are planning to create information sensitive dictionaries, based on standard government or military classification. • Privacy recommendation system: Most OSNs enable users to make their privacy and security settings based on different levels of details. Those settings are made on the user level. Our proposed system extends this to the activity level as we acknowledge the complexity of making decisions at the user level. It is inconvenient for users to make fine-grained privacy assessment per each activity they create and hence users tend to ignore such options. The goal of our system is to assess users in this process rather than controlling their decisions. Users can still have the option to override system decisions (if they decided to). • Access and visibility control: Our proposed system extends options in OSNs to control visibility of users’ created activities. Those visibilities represent the action points in which privacy recommendations are enforced. • Intelligent data integration module. Any IS system that will take decisions automatically on users behalf should be equipped with methods to collect and process historical data. Such historical data (related to users and systems information sensitivity and content visibility decisions) will be used for making future decisions. Such data can be also used for system, or third party information intelligent purposes. • A major task, in terms of time, in this goal is data collection and analysis for (1) information classification and (2) for privacy assessment. We started collecting data, based on our model from Facebook and Twitter. We need to extend the collection and analysis process to collect a “significantly” large enough dataset.

7 Conclusion In an earlier paper, [1] we proposed a reputation score model in OSNs. The goal is to promote information accuracy and enhance knowledge extraction from those OSNs. We showed in this paper, several applications of how to use this reputation model. We developed a dataset from Twitter and make it public as a prototype for this model and its applications. With the huge amount of information exchanged through OSNs, it became necessary for those networks to promote information quality specially as information in those networks is an integral part of Internet body of knowledge in general.

Testing Assessment of Group Collaborations in OSNs

151

References 1. Alsmadi, I., Xu, D., Cho, J.: Interaction-based reputation model in online social networks. In: Proceedings of the Second International Conference on Information Systems Security and Privacy (ICISSP2016) (2016) 2. Altman, I.: The Environment and Social Behavior: Privacy, Personal Space, Territory, Crowding. Brooks/Cole Publishing, Monterey (1975) 3. Backstrom, L., Sun, E., Marlow, C.: Find me if you can: improving geographical prediction with social and spatial proximity. In: Proceedings of the 19th International Conference on World Wide Web, pp. 61–70. ACM, New York (2010) 4. Carminati, B., Ferrari, E., Perego, A.: Enforcing access control in web-based social networks. ACM Trans. Inf. Syst. Secur. (TISSEC) 13(1), 6 (2009) 5. Castelfranchi, C., Falcone, R.: Trust is more than subjective probability: mental components and sources of trust. In: Proceedings of the 33rd Hawaii International conference on System Science, Maui, HI (2000) 6. Cho, J.-H., Alsmadi, I., Xu, D.: Privacy and social capital in online social networks. In: IEEE Global Communications Conference (GLOBECOM), pp. 1–7. IEEE, Piscataway (2016) 7. Dey, R., Tang, C., Ross, K., Saxena, N.: Estimating age privacy leakage in online social networks. In: INFOCOM’12 (2012) 8. Dhekane, R., Vibber, B.: Talash: friend finding in federated social networks. In: LDOW (2011) 9. Dwyer, C., Hiltz, S., Passerini, K.: Trust and privacy concern within social networking sites: a comparison of Facebook and MySpace. In: AMCIS 2007 Proceedings, p. 339 (2007) 10. Fong, P.W.L., Anwar, M., Zhao, Z.: A privacy preservation model for Facebook-style social network systems. In: Computer Security–ESORICS 2009, pp. 303–320. Springer, Heidelberg (2009) 11. Gilbert, E., Karahalios, K.: Predicting tie strength with social media. In: Proceedings of the 27th International Conference on Human Factors in Computing Systems (CHI’09), NY, pp. 211–220 (2009) 12. Grabner-Kräuter, S., Bitter, S.: Trust in online social networks: a multifaceted perspective. In: Forum for Social Economics, 1–21 (2013) 13. Guo, L., Yang, S., Wang, J., Zhou, J.: Trust model based on similarity measure of vectors in P2P networks. In: Grid and Cooperative Computing-GCC 2005, pp. 836–847. Springer, Berlin (2005) 14. Guo, L., Zhang, C., Fang, Y.: Trust-based privacy-preserving friend recommendation scheme for online social networks. IEEE Trans. Dependable Secure Comput. 12(4), 413–427 (2015) 15. Hardin, R.: Trust and Trustworthiness. Russell Sage Foundation, New York (2002) 16. Himel, D., Ali, M., Hashem, T.: User interaction based community detection in online social networks. In: The 19th International Conference on Database Systems for Advanced Applications (DASFAA), Bali (2014) 17. Hogg, T., Adamic, L.: Enhancing reputation mechanisms via online social networks. In: EC 2004, New York (2004) 18. Hussain, F.K., Chang, E.: An overview of the interpretations of trust and reputation. In: 3rd Advanced International Conference on Telecommunications (AICT 2007) (2007) 19. Jøsang, A., Pope, S.: Semantic constraints for trust transitivity. In: Proceedings of Asia-Pacific Conference of Conceptual Modeling, vol. 3 (2005) 20. Kim, T.H.-J., Yamada, A., Gligor, V., Hong, J., Perrig, A.: Relationgram: tie-strength visualization for user-controlled online identity authentication. In: Proceedings of the International Conference on Financial Cryptography and Data Security, pp. 69–77. Springer, Berlin (2013) 21. Korzaan, M.L., Boswell, K.T.: The influence of personality traits and information privacy concerns on behavioral intentions. J. Comput. Inf. Syst. 48(4), 15–24 (2008) 22. Kresnova, H., Spiekermann, S., Koroleva, K., Hildebrand, T.: Online social networks: why we disclose. J. Inf. Technol. 25(2), 109–125 (2010)

152

I. Alsmadi and M. Al-Abdullah

23. Kumar, R., Novak, J., Tomkins, A.: Structure and evolution of online social networks. In: Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 611–617 (2006) 24. Leskovec, J., Backstrom, L., Kumar, R., Tomkins, A.: Microscopic evolution of social networks. In: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2008) 25. Lewis, K., Kaufman, J., Christakis, N.: The taste for privacy: an analysis of college student privacy settings in an online social network. J. Comput. Mediat. Commun. 14, 79–100 (2008) 26. Mayer, R.C., Davis, J.H., Schoorman, D.F.: An integrative model of organizational trust. Acad. Manag. Rev. 20(3), 709–734 (1995) 27. McKnight, H., Chudhury, V., Kacmar, C.: Developing and validating trust measures for ecommerce: an integrative typology. J. Inf. Syst. Res. 13(3), 334–359 (2002) 28. Mesch, G.S.: Is online trust and trust in social institutions associated with online disclosure of identifiable information online? Comput. Hum. Behav. 28, 1471–1477 (2012) 29. Piccoli, G., Ives, B.: Trust and the unintended effects of behavior control in virtual teams. MIS Q. 27(3), 365–395 (2003) 30. Robins, G., Pattison, P., Kalish, Y., Lusher, D.: An introduction to exponential random graph (p*) models for social networks. Soc. Netw. 29(2), 173–191 (2007) 31. Smith, H.J., Millberg, S.J., Burke, S.J.: Information privacy: measuring individuals’ concerns about organizational practices. MIS Q. 20(2), 167–196 (1996) 32. Squicciarini, A.C., Xu, H., Zhang, X.L.: CoPE: enabling collaborative privacy management in online social networks. J. Am. Soc. Inf. Sci. Technol. 62(3), 521–534 (2011). Wiley Online Library 33. Toivonen, S., Lenzini, G., Uusitalo, I.: Context-aware trust evaluation functions for dynamic re-configurable systems. In: Proceedings of the Workshop on Models of Trust for the Web (MTW’06), Edinburgh (2006) 34. Van De Garde-Perik, E., Markopoulos, P., De Ruyter, B., Eggen, B., Ijsselsteijn, W.: Investigating privacy attitudes and behavior in relation to personalization. Soc. Sci. Comput. Rev. 26(1), 20–43 (2008) 35. Von Arb, M., Bader, M., Kuhn, M., Wattenhofer, R.: Veneta: serverless friend-of-friend detection in mobile social networking. In: Networking and Communications, 2008. WIMOB’08. IEEE International Conference on Wireless and Mobile Computing, pp. 184–189. IEEE, Piscataway (2008) 36. Weber, R.: Internet of things-new security and privacy challenges. J. Comput. Law Secur. Rev. 25(6), 522–527 (2009) 37. Wishart, R., Robinson, R., Indulska, J., Josang, A.: SuperstringRep: reputation-enhanced service discovery. In: Proceedings of the 28th Australasian Conference on Computer Science (ACSC’05), pp. 49–57. Australian Computer Society, Darlinghurst (2005) 38. Xiang, R., Neville, J., Rogati, M.: Modeling relationship strength in social networks. In: Proceedings of the 19th International Conference of World Wide Web (WWW’10), pp. 981– 990. ACM Press, New York (2010) 39. Young, A.L., Quan-Haase, A.: Information revelation and internet privacy concerns on social network sites: a case study of Facebook. In: Proceedings of the International Conference on Communities and Technology, pp. 265–274 (2006) 40. Yu, B., Singh, M.P.: A social mechanism for reputation management in electronic communities. In: Proceedings of the 4th International Workshop on Cooperative Information Agents (CIA ’00), pp. 154–165. Springer, Berlin (2000) 41. Zhou, B., Pei, J.: Preserving privacy in social networks against neighborhood attacks. In: IEEE 24th International Conference on Data Engineering, 2008. ICDE 2008, pp. 506–515. IEEE, Piscataway (2008)

Dynamics of Overlapping Community Structures with Application to Expert Identification Mohsen Shahriari, Ralf Klamma, and Matthias Jarke

Abstract Social media are connected to the daily life of people. While participating in a particular social platform, people belong to structures named communities. Detection of communities and analysis of their life cycles have been a topic of interest in the social network analysis research. There have been several methods—sometimes rather complex—that approach this area from different aspects. However, these methods may fail in real applications such as recommender systems while they may lack a proper social dynamic behind their formulations. Furthermore, our little wisdom regarding the community formation and evolution in dynamical networks persuade us to answer research questions like which structural properties of communities are significant for their fate and how they are related with community detection algorithms. Keeping in mind the deficiencies, we consider a multifaceted approach for the analysis of overlapping communities. Overlapping community structures are suitable indicators as for a real analysis in this domain. As such, we propose a two-phase algorithm based on two significant rather simple social dynamics named Disassortative degree Mixing and Information Diffusion—this algorithm is called DMID. In the first phase, DMID detects influential nodes with the help of degree information of nodes and a steady-state random-walk. In the second phase, remainder nodes are assigned to communities by an information diffusion process. Afterward, we opted a real-world domain example, i.e., expert identification, as a suitable evaluation scenario. To this end, we combined community structures with original ranking algorithms and employed them to return the relevant list of experts. Furthermore, we analyzed our algorithms by considering their structural properties. Keeping a dynamical perspective on the problem, we constructed a classifier by using logistic regression classification model and combining community structures and suitable community-related features. Results indicate that DMID competitively wins in

M. Shahriari () · R. Klamma · M. Jarke RWTH Aachen University, Aachen, Germany e-mail: [email protected]; [email protected]; [email protected] © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_7

153

154

M. Shahriari et al.

several cases over the baselines. In the community prediction problem, it could achieve the highest and the most reliable predictions. Furthermore, we ranked significant features for community evolution prediction based on their importance— the size of a community is important for such an investigation. We explained our detailed findings and studies in the result section. Furthermore, we implemented all our algorithms as a generalized and extensible Web service. This research paves the way for people who require either to detect communities in their network or to predict their future.

1 Introduction We frequently encounter networks in our daily life. Our life challenges strongly connect to complex networks and their interactions—using public transportation systems to commute, getting acquainted with people in a sports club, or collaborating with colleagues in an office—constitute interactions of such a network. All these communications and collaborations can be represented by graphs, i.e., networks [28, 54]. With the flourish of advanced computational technologies and data science field, we have further been swamped with dynamical networks, in general, and online social networks, in specific. Since WWW establishment in 1990s, Internet has been widely used in political movements and protests, and conflicts nourishments and immense online polling advantages. Generally speaking, fit is possible to map individuals and their communications to nodes and links of a graph, respectively. Popular social networks like Facebook, Google+, LinkedIn, and Twitter have tremendous number of visitors and active users every day. Shrinking diameter, temporarily, small-world-ness, and motifs are among essential properties of social networks [15, 41, 48]. People join and leave social networks regularly and this causes temporal dynamics. The substantial involvement of online users and their instant communications spawn feelingly small networks with shrinking diameter. Often, we would like to recognize people in these platforms that probe accommodations, submit a paper to the same conference, or look for jobs. Although there exist differences between social networks, e.g., topics on LinkedIn and Twitter are more profession related in comparison to Facebook and Google+, overlapping community structure is a common property in all of them [42, 43, 56, 79]. Not only detection of community structures is essential but also their analysis should be comprehensive and manifold. One should consider them from different aspects. The first aspect of communities is their dynamism. Communities encounter various evolution stages such as birth, death, growth, shrink, split, and merge. In other words, two communities may merge to a bigger community in size or one may split into several smaller communities. One day, a community opens its eyes to the world, and another may die [5, 8, 13, 36, 53]. In addition to evolution, detection of communities is still a challenging problem since 2004.

Dynamics of Overlapping Community Structures with Application to Expert. . .

155

Many research works have approached the topic from different aspects and improved both the complexity and precision of the algorithms. Local or global, static or dynamic, structural or contextual are among properties of available overlapping community detection (OCD) algorithms. However, these algorithms are sometimes complicated and no one has proved their suitability for particular applications. To put this another way, simple algorithms with reliable social dynamics may be more effective than others. In this regard, algorithms in this area have only been tested statistically on various datasets but not on specific domains. Correspondingly, the influence of overlapping community structures on expert finding is not evident. First, we may require to define who an expert is and what expertise is. In this regard, experts possess a higher level of knowledge in particular domains and are adept to perform a specific task. Similar to real world, experts have higher levels of expertise in comparison to other users in the virtual world. Often, they qualify to judge success of an approach, strategy, or an activity [60, 77]. For instance, they may help novices to accomplish a task in a question and answer forum via answering askers’ questions. In general, expert identification is a significant task because experts significantly affect on learning rate of novices and amateurs when they share their knowledge. Moreover, expertise of a person enhances the trustworthiness in resources, specifically, informal learning environments require the experience of such people. In fact, expertise contributes in engaging new connections and interactions among professionals and increases trust. Expert identification and OCD algorithms are branches which need to be considered together. In specific domains such as learning environments, overlapping members can be recommended to experts to expand the borders of communities and enhance the information flow. In other areas like open-source developer networks, we may require to distinguish the core and periphery members and to identify people with particular expertise regarding requirements [37, 68]. In this regard, we have proposed a two-phase OCD algorithm based on two social dynamics. The first phenomenon is disassortative degree mixing—an indicator showing dissimilarity among neighbors of nodes. To clarify it, let us consider the example of the relation among professors and students in a university or the potential connections among high-level with middle- or lower-level people in societies. In the first phase of the algorithm—which is named DMID, disassortative degree mixing alongside the degree of nodes are combined to compute influential nodes. This algorithm detects as well the hierarchical structure of the networks, i.e., local and global leaders. The first phase of the algorithm survives without requiring any input parameter, and it identifies the number of communities that equals to the number of leaders. The second social dynamic involves diffusion of opinions. It is possible to consider several strategies to simulate information diffusion that one of them is the network coordination game. Considering binary states for opinions like A and B and equal resistance threshold for the nodes contribute to agree to a simple diffusion strategy. A simple example of opinion formation and informed agents can be a new mobile brand. People seeing your brand new mobile might imitate your propensity especially if you have high charismatic character (being an influential member or leader). Thus, people around identified leaders in the first stage may accept the

156

M. Shahriari et al.

opinions of the leaders based on their dependency on these set of influential nodes. Membership magnitude of common members of communities, i.e., leaders can be computed based on the dependence and the time that the synchronization happens with the leaders. To evaluate DMID, we extend classical HITS and PageRank algorithms to obtain community-aware HITS and PageRank algorithms to benefit community structures. However, before manipulating the ranking formulas, we may need to ask ourselves why communities can be performant in this regard? To respond to this question, we know that nodes inside communities can more reliably declare opinions about other nodes as long as they reside in the same communities. In other words, nodes placed under the auspice of the same communities may genuinely know each other than those living in separate communities. In HITS, we play with hubs and authority vectors, on the contrary, we manipulate the random-walker in the PageRank algorithm to activate the effects of community structures. Accepting HITS and PageRank as the baselines and applying these community-aware ranking methods indicated the excellent achievement of community structures. Ranking algorithms based on DMID are quite successful and generate the relevant list of experts among others. Results characterize the suitability of DMID comparing other algorithms correctly. We compared DMID with SLPA, CLiZZ, SSK, and Link Communities (Link) from the literature. Results in real-world networks indicate traces of better modularity values and time complexity in many cases. Additionally, synthetic network generators are also very popular for evaluation which we employ for the assessments. Expert identification outcomes also corroborate that community structures affirmatively improve the result set for the queries—mean reciprocal rank (MRR) and mean average precision (MAP) confirm it. Detecting hierarchical structures, suitable for temporal settings, and showing its merit in real domain applications such as expert identification verify the appropriateness of DMID. In addition to detection of community structures, their effect on the prediction task is not well investigated. Overlapping communities are connected to the prediction problem; unfortunately, almost all the research works do not approach the problem from a holistic perspective. To further shed light on this issue, it is not still determined which features are significant in the prediction of overlapping communities, and how they connect with the detection of community structures. To address this issue, we applied the proposed OCD algorithm together with other baselines on dynamic networks including Email, DBLP, and Facebook datasets. We contrasted the properties of some of these algorithms against each other regarding number and size of communities and frequencies of overlapping members. At this juncture, DMID detects an almost smaller number of communities, in fact, bigger size communities with the higher overlapping degree. We also mapped the communities over time to perform a temporal analysis through Group Evolution Discovery (GED) technique. Next, we identified events happening to each community in which they include dissolve, merge, split, and survive. We applied logistic regression classifier together with a couple of community- and node-level features. Results indicate that size of communities is a distinctive indicator to predict the fate of communities. Besides, DMID generates the highest community prediction accuracies among its competitors.

Dynamics of Overlapping Community Structures with Application to Expert. . .

157

Regarding the framework, current implementations are based on various tools which make them scattered to the social network analysis research community. These challenges inspired us to initiate research on the detection, evolution analysis, and applications of overlapping communities. We develop a generalized RESTful framework for our purpose and expose it to the research communities. It is opensource, and other algorithms can be integrated with the current platform. We also innovate simple but effective social dynamics such as DMID to the case of OCD. To summarize, we mention our contributions as follows: • We proposed a two-phase algorithm based on social dynamics of disassortative degree mixing and opinion diffusion; we compared this algorithm with several baselines. Results indicate the suitability of DMID compared to other algorithms. This algorithm and part of findings were published in World Wide Web conference, companion volume [66]. • The proposed algorithm including baselines, evaluation measures, and synthetic networks is implemented as RESTful Web services and can be widely employed and extended by the researcher community. This framework is open-source, and further algorithms can be integrated with it. The initial framework was published at the proceedings of i-KNOW conference [67]. • The problem of community structure detection, community structure application, evolution analysis, and prediction is investigated with a holistic approach. A novel community detection approach is proposed, then together with baselines, they are applied in expert finding domain. Moreover, we further employed them to perform a community prediction model. Part of community prediction and expert identification results were published at World Wide Web companion volume and proceedings of i-KNOW conference [68, 69]. • Results of the community prediction indicate that the accuracy of community structure prediction depends on the used algorithms and the intended event. Altogether, size of a community is revealed to be a significant feature for community analysis. • Community-aware ranking algorithms are proposed based on HITS and PageRank algorithms. For the first time, we connected overlapping communities to expert identification task. Results indicate that proper detection of overlapping community structures can improve the returned list of relevant items. The rest of this article is organized as follows. In Sect. 2, we first describe the related work regarding the detection of overlapping communities. Afterward, we discussed approaches to identify experts. Next, methods of mapping communities over time and analytic approaches and results in community research domains are explored. Furthermore, we introduced available tools to investigate graphs and communities. In Sect. 3, we discussed the anatomy of the DMID algorithm. Afterward, community-aware ranking algorithms are explained. Next, the method of mapping communities over time, features, and the classifier for community prediction are elaborated. In Sect. 4, we mentioned baselines to evaluate the OCD and community-aware ranking algorithms. Subsequently, in Sect. 5, we organized the rendering of the whole framework as an integrated and extendible RESTful Web

158

M. Shahriari et al.

service. In Sects. 7 and 8, we define the evaluation protocol and describe the results, respectively. In this regard, results include evaluation of OCD approaches, expert finding methods, and prediction values. Eventually, we discuss the strong and weak points of the article in Sect. 9 and conclude in Sect. 10.

2 Related Work In this section, we demonstrate the related work. First, the related work regarding the detection of overlapping communities is explained. Afterward, we discuss real applications of community structures such as expert identification. Moreover, we converse about the related work regarding analysis and prediction of overlapping communities. Finally, we review available tools and services to detect community structures.

2.1 Overlapping Community Detection Numerous disjoint community detection algorithms have been proposed [19, 44]; however, they are not proper to identify communities in real-world networks. While studying the literature on community detection algorithms, we notice several different categorizations. In a global perspective, we may designate the approaches to three classes of static or dynamic, local or global, and structural or content-based algorithms. In addition to the above categories, algorithms in this area can be nominated to categories such as clique percolation, line graphs or link communities, local optimization or leader-based methods, random-walks, or agent-based algorithms. Global methods apply universal metrics to identify communities [19]. These metrics can be, for example, modularity, and can be optimized through a global optimization approach [34]. Local methods of community detection consider local information of the network and thus approaches using random-walk processes, cliques, influential nodes, or leaders can be classified as local methods [38, 40, 71, 72, 74]. Furthermore, considering other sources of information may provide more realistic data to identify communities. In this regard, content-based and attribute-based approaches take attributes of nodes and edges besides to the connection information [62, 80]. In addition to locality and relationship information, some algorithms are suitable for temporal environments. Complex networks are dynamic; however, static community detection algorithms can also be applied to each snapshot of the graph separately. When network properties are temporal, then static solutions may not be suitable and stable, and thus adaptive methods of community detection emerged. Adaptive methods approach the issues encountered in static methods, and they behave more stable [3, 18]. In the following, we briefly explain each of these categories.

Dynamics of Overlapping Community Structures with Application to Expert. . .

2.1.1

159

Leader-Based Techniques

Considering neighboring of nodes to identify communities is informative while nodes’ proximities may not be as dynamic as the global structure of the network. Often, leader-based approaches employ local dynamics to identify several groups of nodes as influential members of the communities. Afterward, membership of other nodes to the communities is computed based on some specific metrics [72]. Influential nodes can be identified based on different approaches such as random-walk, degree, and even considering the influence range of other members. Computing the membership of nodes to communities can also be calculated via the random-walk, cascading behaviors, and shortest paths [1, 32, 42]. In these approaches, leaders are located on top of the communities and in many of these approaches, they lead to hierarchical communities. In the baseline section, we demonstrate these methods in more detail.

2.1.2

Agent-Based Approaches: Random-Walks and Label Updates

One subclass of OCD approaches is the agent-based methods. In this category, algorithms based on the random-walk are noticeable. In the random-walk-based methods, the surfer starts a walk until infinity. It selects the probability of its outgoing path based on a transition matrix. Many of the algorithms employ the steady state of the random-walker and the spectrum of its walk to identify the overlapping communities. For instance, in this category of the approaches they address every node of the network by unfolding a community around it through a random-walk process; then community structures around nodes can be combined [14, 35]. Other methods may consider walks with limited length or constrain for the walks [74]. For instance, Jin et al. applied a limited-length random-walk and found the walks of length n2 and then identified paths of length k. In this method, edges overlapping in the paths gain weights which determine the communities [45]. Other techniques impose constraints on the walk such as visiting active nodes (nodes with the higher degree than average degree of the network) and nodes with a predefined property named associate degree [27]. Another approach combines modularity with the random-walk and considers the difference between random-walk value in the network and its corresponding value in the null model [4]. Aston and Hu discovered communities based on hierarchical clustering; therefore, it needs to compute distances. Node–node and community–node distances are computed based on random-walks to define the distances [55]. Finally, other ideas such as applying coding techniques are incorporated with optimization techniques or including the history of previous states besides current states are among them [61]. Agent-based techniques can also behave based on updating labels of the nodes [31]. For instance, SLPA is an agent-based technique that applies memories for each node [78]. Labels can take several different states and be updated based on some predetermined rules. Other agent-based techniques consider utility functions to join, leave, and switch communities. Communities are related based on the Nash equilibrium and agent’s

160

M. Shahriari et al.

gain, and loss functions are defined for them. Usually, in these methods, each node in the social network may be considered as a separate agent, deciding on its community [2, 3, 16].

2.1.3

Clique Percolation Approaches

Clique percolation methods work based on finding the k-clique groups. k-cliques are subgraphs that are fully connected. The algorithm identifies these set of kcliques and maps each of these cliques to a new node. These subgraphs can overlap, and a set of cliques are mapped into a community while a series of k neighboring cliques are identified. In this regard, sharing of the k − 1 member can be considered as neighboring cliques. There are two problems ascribed to clique percolation techniques that not only they lack the time efficiency but also they merely may be suitable for dense graphs [38, 40, 71]. There have been several variations to clique-based methods. For instance, EAGLE combines agglomerative techniques with maximal cliques and considers maximal cliques instead of nodes to form hierarchical clustering. It neglects all maximal cliques with threshold smaller than a certain k value [71]. Moreover, Bi-clique communities consider the clique percolation problem in bipartite networks. It defines the concept of Ka,b cliques and considers ka cliques in one set and Kb in the other set [40]. In another work, clique optimization is employed to identify granular overlaps. Clique optimization as a fine-grained approach is employed to detect the nodes that are connected to distinct communities and are highly connected to each community [75]. Additionally, finding cliques are attributed to communicability and binary matrices. The adjacency matrix is transformed to a communicability graph which can be performed by walks of length k between every two nodes. Afterward, by converting the communicability graph to the binary matrix and constituting the similarity matrix of cliques, cliques are merged to form communities [23]. As the efficiency of cliques has been an issue in some applications such as prediction of protein domains, researchers proposed other extensions. For instance, SECOM eliminates all the edges with weights smaller than a predefined threshold then the algorithm starts with cliques of size k and uses projection [26].

2.1.4

Line Graphs and Link Communities

The idea behind this category of approaches is to find the overlapping communities on links of a graph rather than nodes. In other words, instead of using partition methods for OCD, these methods convert the original graphs to line graphs and then partition the nodes of the new graph. In this regard, there are two ways, which map the original graph to a line graph: in the first method, a vertex of degree k is mapped into k(k − 1)/2 edges of the line graph, so the random-walk will be a link–link random-walk. The other method converts the graph to an affiliation graph in which edges of the original graph are in one side, and the vertices are on the

Dynamics of Overlapping Community Structures with Application to Expert. . .

161

other side [25]. In another method, Ahn et al. applied the edge similarity to form the hierarchical communities [1]. Furthermore, link communities have been considered with other methods such as cliques and random-walks [24, 25].

2.2 Expert Identification Community structures have been applied to various domains such as sign and link prediction, misbehavior detection, and routing [2, 49, 65, 70]. Another informative application domain is to identify communities of people for the expert finding. Peer production systems such as question and answer forums require experts who possess higher levels of knowledge in particular domains. There have been several different approaches accosting the challenge. Bozzon et al. found the corresponding profiles of workers in Facebook, Twitter, and LinkedIn and looked for experts in them. To achieve that, they identified some people and asked them questions about their expertise with values ranging from 1 to 7. These people were from different disciplines like computer science, music, and so on. They employed information retrieval techniques and came to conclusions like information by others on the profiles of people are informative, and Twitter appeared to be the best social network for expert matching [11]. In a similar work, Bozzon et al. targeted the IBM company and extracted corresponding persons and their information on LinkedIn and Facebook. These works propose analysis towards a correspondence of companies and their professional social networks. Similar to social networks, the expert finding has been investigated in DBLP networks. Deng et al. cordially applied statistical graphical models like generative probabilistic models. In addition to the statistical model, weighted language model together with citation of papers are considered for finding experts. In the same article, a topic-based model is investigated in which documents are converted to topics, and the relation between the topics and the queries is found [22]. Reichling et al. employed an approach by combining self-reported directory information with keyword mining from users’ files and folders. Their recommender system has been used in the real-world organization of NIA [58]. Macdonal and Ounis considered the expert search task as a voting problem where documents vote for the candidates with relevant expertise [47]. Data fusion techniques are also applied and evaluated by utilizing different document weighting models [7]. The actual evaluation is conducted within the context of expert search task of the TREC enterprise track. Researchers have addressed question and answer forums with two main categories of algorithms. In some of the approaches, models and methods are not directly applied to find experts but rather to find similar and related items. One of such directions is feed distillation in which related blogs to a query are retrieved. In the first approach, resources associated with specific queries are identified, and contribution of people in those resources is calculated. Structural-based ranking algorithms and expert identification based on HITS and PageRank reside in this category. In other words, graph induced from the query is employed to apply

162

M. Shahriari et al.

classical or tuned HITS and PageRank algorithms on them. Zhou et al. applied Latent Dirichlet Allocation (LDA) to extract topics and contextual information from generated askers and answers activities. They devised a tuned PageRank algorithm to take advantage of topical similarity in addition to structural information [82]. Yang et al. challenged the previous works on topic and expertise modeling. They considered tags of users while posting questions as for their topic interests and similarities. Moreover, they employed generated votes of users as traces of their expertise. Two information sources, tags and votes, are leveraged to be consumed for expert finding [81]. Zhu et al. innovated a new approach for computation of relevant topics and categories. They further created a ranking method employing two information resources of target and related categories [83, 84]. Bouguessa et al. did not appreciate methods generating a ranked list of nodes while they had problems identifying the threshold in different categories in Yahoo! Answers. Thus, they leveraged authoritative and non-authoritative authority values with the gamma distribution. The number of components and the parameters of the model are estimated via Bayesian Information Criterion and EM algorithm [10] . Our proposed community-aware ranking algorithms belong to these categories and effect of community structures are assimilated with ranking algorithms.

2.3 Community Evolution and Prediction To analyze the evolution of communities over time, one may first need to find a way to map them over time. Afterward, we may need to decide how we can predict their life.

2.3.1

Community Mapping over Time

Tracking the evolution of (overlapping) communities is significant for their analysis and extraction of dynamics over time. The main category of approaches investigate the inclusion of one community in another, or they track the influential members of communities. The first category is matching information from one community in one snapshot with all the communities in the next snapshot. Ma and Huang proposed community update and tracking (CUT) algorithm which maps the graph information to a bipartite graph. The authors take the idea of connected components named cliques (here 3 cliques) on one side of the bipartite graph and the connections among these cliques on the other side. The algorithm tracks node and edge adding/removal by tracking these clique updates [46]. Hopcroft et al. defined a tracking method by the repeated creation of agglomerative clustering and considering a match value ranging between 0 and 1 [33]. Greene et al. innovated a simple but rather efficient method which employs aggregate information from other communities. They define a front set and compare the identified community (Cta with front

Dynamics of Overlapping Community Structures with Application to Expert. . .

163

ta ∩Fi | (Fi ) based on Jaccard similarity (sim(Cta , Fi ) = |C |Cta ∪Fi | )) [30]. Savic et al. devised and implemented a Java code to detect community evolution life cycles including birth, death, contraction, growth, split, and stability in Apache Ant class collaboration network. They consider a similarity measure and specify several rules for communities that place transitions in predefined life cycles [63]. Cuzzocrea and Folino track the flow of members through a confusion matrix, and they define internal and external transitions for communities. External transitions leverage the relationships of the community with other communities, and internal transitions handle transformations happening inside the community. They further apply similarity and overlap measures to find sharing of clusters [20]. Correspondingly, Palla et al. consider clusters of two consecutive time slots ti and ti+t as one cluster and apply the clique percolation methods. This method can identify, merge, grow, and unchanged events which are achieved by considering an overlap coefficient [53]. CommTracker follows influential nodes to discover how their properties change and how they are updated. This algorithm is suitable for large-scale networks and does not require any input parameter [76]. Similarly, Chen et al. considered community representatives to track the community dynamics. They defined communities as cliques and community representative as nodes with the minimum number of appearances in other communities. They bring several themes and constitute a decision tree for describing the events happening to community representatives [17]. HOCTracker is a node-based processing approach that maintains intermediate transition through a Log list. The Log-based approach contributes to avoiding the requirements to define overlapping coefficient and similarity measures and enhances the performance of community comparisons [8]. Finally, the GED method [13] is a widely used approach that is applied in the literature and this work. Community dynamics are handled through community inclusion factor. This metric quantitatively and qualitatively measures the overlaps among two consecutive clusters. We further demonstrate it in the method section.

2.3.2

Community Evolution Analysis and Prediction

Community evolution analysis and prediction have been investigated to a certain extent. Community prediction methods apply either parameter-free or parameterbased approaches for prediction of communities. Parameter-based approaches learn the parameters of the model by employing supervised learning algorithms and apply train and test sets. Takaffoli et al. applied a disjoint community detection method and mapped the communities over time. They predicted events such as survive, merge, split with employing several static and dynamic features [73]. However, their method did not consider overlapping communities and balancing the distribution of the classes. Similarly, Brodka et al. applied GED technique for community evolution tracking. Moreover, they employed features such as group size and event types in previous time steps [12]. Sekara et al. proposed a framework for the changes happening to the community. In fact, they shed lights on high

164

M. Shahriari et al.

stability of core gatherings and predictability through cores. They further discussed the concept of soft and hard boundaries, the context of meeting or community and recreational gathering versus work gatherings [64]. Palla et al. mapped communities over time and indicated that the stability of large groups depends highly on their adaptability to change the group composition. In contrast, small groups retain their core members for stability [53]. Goldberg et al. developed a scalable algorithm to predict the lifespan of communities. They mention strength axioms for the evolution of communities such as monotonicity and extension. Furthermore, size and intensity of communities have high predictive powers [29]. Backstrom et al. studied a vast spectrum of properties including topological and communal features to investigate joining of members to communities. By applying decision trees they indicated that probability of joining to a community depends not only on the number of common neighbors in the community but also on the composition of the internal connections inside a community. Furthermore, they identified larger growth rates of communities in a given period of time [5]. Kairam et al. performed a similar experiment and considered the diffusion growth as the dominant growth in NING social network [36]. Baek et al. used content similarity of bloggers and communities to show that content may create communities. They indicated that not only content is an important factor in community formation but also it does expedite the evolutions [6].

2.4 Overlapping Community Detection Tools and Web Services In this subsection, we give an overview of the tools and services regarding graph and community analysis. Although there is quite a wide range of tools, however, users (researchers) still encounter challenges while using them. For instance, some of these tools do not provide a suitable graphical interface for the person working with them; however, the research community has a few time to do reverse engineering to understand the code. Secondly, the tools do not provide complete functionalities expected from a community analytic framework. In other words, sometimes there is no evaluation measure or synthetic generator to support the users. Or even preprocessing and post-processing functionalities are missing. Additionally, most of the tools are not Web-based and need to be deployed via a suitable Integrated Development Environment (IDE) and sometimes several libraries may be missing. Furthermore, most of the tools are not open-source and are not possible to be easily extended by the research community. With the reasoning mentioned above, it may be lighting to introduce some of these tools. GANXiS1 is programmed in Java and is an extended version for SLPA. JUNG provides Java implementation of some graph theory, social network analysis, and data mining techniques and it also includes undirected and unweighed implementation of Link Communities.2

1 https://sites.google.com/site/communitydetectionslpa/. 2 http://jung.sourceforge.net/.

Dynamics of Overlapping Community Structures with Application to Expert. . .

165

Moreover, pylouvain-igraph renders a python implementation of multiple versions of Louvain method.3 Community detection modularity suit4 is based on C++ and R and contains three community detection algorithms based on modularity measure. It further contains boot-strapping facilities to test cluster robustness. Additionally, CFinder is a Java-based implementation of clique percolation method. The software is suitable for dense graphs and provides suitable visualization and computation of graphs and covers through proper diagrams.5 Furthermore, InfoMap has a corresponding Java implementation named Map Equation which provides dynamic visualization of algorithm execution and the community structure.6 Finally, Jmod7 is an open-source tool implemented in Java that can be integrated into thirdparty software applications. It comprises several community detection algorithms with the possibility of parsing the networks in different formats. Similar to our framework, it contains benchmark graphs and the LFR method for generation of synthetic networks. Other tools such as Apache Commons Graph,8 JGraphX,9 graph stream10 , and JGraphT11 provide functionalities regarding handling and processing of graphs.

2.5 Related Work Summary Figure 1 shows an overview of the related work section. Each paper is assigned to its category and contribution of each article with community domain is depicted. One can figure out that the community literature is related to several disciplines in social network analysis. From one side, OCD is an exciting area which has been under much investigation, and four categories of approaches including leader-based, agent-based, clique percolation, and line graphs have been proposed. Moreover, community mapping and community analysis and prediction are two areas that several research papers have been published in that regard. Matching techniques and methods based on influential nodes are among the approaches that belong to community mapping. As for the community prediction, we mainly depicted the parameter-free, and parameter-based approaches in this figure. We also gathered the literature regarding currently available community analytic tools in one node, we as

3 https://launchpad.net/pylouvain-igraph. 4 https://sourceforge.net/projects/cdmsuite/. 5 https://sourceforge.net/projects/cdmsuite/. 6 http://www.mapequation.org/index.html. 7 http://tschaffter.ch/projects/jmod/. 8 http://commons.apache.org/sandbox/commons-graph/project-reports.html. 9 https://github.com/jgraph/jgraphx. 10 http://graphstream-project.org/doc/Tutorials/Getting-Started/. 11 http://jgrapht.org/.

166

M. Shahriari et al.

Fig. 1 An overview of the related work regarding overlapping communities. Communities are related to different disciplines in social network analysis such as detection, mapping, evolution, prediction, tools, and supporting applications

well plotted some essential works for expert identification. Overall, one can see that communities are connected with various fields in the area of social network analysis. Of course, the connection is more than what we observe now, but this figure gives the reader a rough idea of the necessary disciplines.

3 Methodology 3.1 Problem Formulation Now, we formally define the problem and explain the necessary terms and variables. If we need to consider and analyze a social network, we can model it as a graph. This graph is denoted by G(V , E) in which V is the set of nodes and E is the set

Dynamics of Overlapping Community Structures with Application to Expert. . .

167

of edges. In other words, V is a set that comprises V = {V1 , V2 , . . . , VN } and E contains E = {E1 , E2 , . . . , Em }, respectively. Each connection Ei is constituted between two arbitrary nodes Vi and Vj . Neighbors of node i are denoted by Nie(i) and degree of node i is denoted by |Nie(i)|. In unsigned social networks, all the connections are positive, and the adjacency matrix can take either 0 or 1, i.e., Aij = 0, 1. In the community detection problem, we aim to discover connected components that are densely connected internally and sparsely connected to each other. Different algorithms result in various community resolution levels. If one applies OCD algorithms on the graph of a social network, then we obtain overlapping communities. We can denote the communities by C = {C1 , C2 , . . . , CL } that L is the number of found communities and Ci ∩Cj = ∅. While networks evolve, we can integrate time symbols with the notation sets. For instance, the graph and covers can be denoted as Gt (V t , E t ) and V t = {V1t , V2t , . . . , VNt }. Furthermore, in expert finding problem, it is ideal to figure out people with higher levels of knowledge. Although expertise usually depends on different structural and contextual factors, we may categorize the vertices based on their ranking. Hence, in expert finding problem, we are interested to reliably rank the nodes that possess the higher level of expertise in comparison to other users. Last but not least, the problem still connects with community evolution, and we apply tracking and the OCD algorithms to the community evolution prediction problem; therefore, we simply define it here. Given two snapshots t and t + 1, we would like to match the communities; in other words, given Cit , we intend to figure out its fate. Research has mainly identified evolutions such as birth (Cit+1 is only observed at time t + 1), death (Cit is not observed in snapshot t + 1), grow and shrink (Ci is observed at snapshots t and t + 1 but with different sizes), split and merge (sometimes a community at time t split into several communities at time t + 1; similarly several communities may be merged and form a bigger but single community at time t + 1). In the community evolution problem, we are interested to apply either supervised or unsupervised learning algorithms to reliably predict what happens to a certain community Ci . In other words, what event is the major undergoing event of a community. The prerequisite for community evolution prediction is that we reach a common understanding of what a community is and how to map the communities over the period of time. Table 1 shows the list of symbols that we shall use to describe the algorithms in the method section.

3.2 Overlapping Community Detection In this subsection, we introduce the algorithm based on DMID.

168

M. Shahriari et al.

Table 1 An overview of the symbols used for description of OCD algorithms Symbol deg(i) AS TAS DV LV AFD RT RG RGA Follower(i) GL M Mil Successor(i) ti LLD(i) PayOff(i) k μ Exi α Cr ai hi β PRi Outdegintra (j ) Outdegextra (j ) N

3.2.1

Description Degree of node i Assortative matrix Row-normalized disassortative transition matrix Disassortative vector Vector containing relative leadership values Average follower degree Resistance threshold Received gain Resistance threshold for behavior A List of followers of node i Number of leaders Membership matrix membership dependence of node i on community l The set of successors of node i The time point node i changes its behavior Local leadership value of node i Payoff that node i receives Average degree in LFR networks The faction of edges sharing with other communities (mixing parameter) Expertise value of node i Adjusts the effect of inside and outside connections Sample cover identified by any OCD algorithm Authority value of node i Hub value of node i Damping factor PageRank value of node i Nodes in which node j refers to them and in the same community as node j is Nodes in which node j refers to them but in different community from node j is Number of nodes in the network

Disassortative Degree Mixing and Information Diffusion

DMID works based on two simple social properties named DMID. The algorithm assumes that communities, often, are constituted around high degree vertices; we consider these nodes as leaders or influential components. One may consider simple degree, closeness, and betweenness centralities or rank values as leadership values; however, these metrics do not take into account disassortative degree mixing which is a common property of social networks. In other words, influential nodes not only possess high structural degree level but also tend to own disassortative degree mixing with their neighbors. Disassortative degree mixing is a homophilic measure which indicates any dissimilarity; for simplicity, we confine the problem to structural differences. For instance, high degree nodes tend to connect with

Dynamics of Overlapping Community Structures with Application to Expert. . .

169

low degree nodes and vice versa. In the first phase, DMID finds hubs with high disassortative degrees. We begin the first phase by defining an N ∗ N assortative matrix named AS that can be computed as follows: ⎧ ⎨|Nei(i)| − |Nei(j )| , if j ∈ Nei(i) , (1) ASij = ⎩0 , otherwise   where |Nei(i)| is the degree of node i and |Nei(i)| − |Nei(j )| is the absolute value. Now, we have a matrix that contains the degree difference corresponding to nodes of this matrix. If we apply a random-walker on the corresponding row-normalized transition matrix, paths can be considered as disassortative paths. In other words, the walker tends to flow in directions with higher degree differences. The rownormalized disassortative matrix can be computed based on the following: ASij TASij = |N | . k=1 ASik

(2)

Afterward, a Disassortative Vector (DV) is considered to hold the disassortative value of each node. We initialize DV with |N1 | , and we update it with the help of disassortative transition matrix TAS; therefore, the update is based on: DVt = DVt−1 × TAS,

(3)

after enough iterations, the process converges, and we obtain the disassortative value of each node. To be considered as a leader, the simple degree of nodes needs to be combined in addition to homophile. The leadership value of node i (LV(i)) can be calculated as follows: LV(i) = DV(i) × |Nei(i)|.

(4)

So far, each node i can be represented by its relative leadership value LV(i). To further proceed towards final leaders, vertices need to decide regarding local leadership locally. In other words, for each node i, a local leader needs to be selected as follows: LV(i) > LV(j ) ∀j ∈ Nei(i).

(5)

As the formula represents, node j is the follower of node i; therefore, a forest is condensed out of the above process. Leafs of the forest are not good candidates to be considered as global leaders but the local leaders, by comparing the number of followers of local leaders to the average number of followers in the forest. The average number of followers (ANF) is defined as follows:  ANF =

i∈LL

  Follower(i) |LL|

,

(6)

170

M. Shahriari et al.

here LL is the local leader set and Follower(i) is the set of followers of node i (nodes that have lower leadership degree in comparison to node i and are connected to it). At last, global leaders are extracted from local leaders based on the following formula:   Follower(i) > ANF.

(7)

Now, |GL| (number of leaders) has been identified which corresponds to the number of communities; however, the dependency of nodes to communities is missing. To compute membership degree of vertices to covers, we simply apply an information diffusion strategy with profitability gain for each member. We can consider different strategies and behaviors in the network, and some nodes share common behaviors. In this regard, nodes’ opinions and beliefs are influenced by the innovation of their neighboring vertices. To simulate opinion formation and information diffusion, we may consider some resistance threshold (RT) and received gains (RG). For node i, the RG value for behavior A is aggregated based on the following:   {j ∈ Nei(i) : j has behavior A   RGA (i) = . Nei(i)

(8)

For each node i, if the RG value is greater than the RT value, then it will accept the new behavior; otherwise, it will resist with its opinion. To further clarify our membership calculation strategy, consider a node i with three neighbors and belief A. If two of the neighbors believe in behavior A and one of them believe in behavior B, then the RG(i) = 0.33. If the RT value for node i is 0.5, then it will resist on its current belief. Similarly in our simulations, we consider a unique behavior (B) for all of the nodes. All of a sudden, one of those detected global leaders changes its behavior to a new behavior A. Neighboring nodes are influenced and the new behavior A is cascaded through the network. To simplify the second phase, we consider equal RT values for all of the nodes. For each of global leaders, it is needed to initiate such a changeover in behaviors. We start by RT = 1 and reduce it gradually—repeating the cascading process until all the nodes are member of at least one community. In this information diffusion phase, the sooner a node adopts the new behavior the stronger is its dependence on the corresponding leader and community. In our simulations, we consider the following formula to calculate soft degree membership value of node i to the leader (actually community) l: Mil =

1 , ti2

where Mil is the membership dependence of node i on community l.

(9)

Dynamics of Overlapping Community Structures with Application to Expert. . .

171

Time Complexity The time complexity of the first phase of the algorithm is O(M) (M is the number of edges). In other words, the disassortative matrix is calculated with O(M); the random-walk process can be simulated in O(M). Moreover, the leadership vector is computed in O(N ) and local leaders are identified in O(M). Because all of the steps in the first phase need to be performed sequentially, the total time complexity of this phase is estimated to be O(M). To analyze time complexity of the second phase of the algorithm, we consider communities with the same size N O( |L| ). In the worst case, one node may adopt the behavior among the neighbors and ¯ then the time complexity can be O( N 2 d). ¯ In other if average degree of nodes is d, |L|

N words, the second phase can be estimated as O(M |L| ). Finally, the time complexity

N N of the whole algorithm can be deduced as max(O(M),O(M |L| )), which is O(M |L| ) in the worst case.

DMID Weighted and Directed The previous formula set up is for the undirected and weighted graphs. We further developed the directed and weighted version of the DMID. In the first phase, the disassortative matrix and the transition matrix are calculated based on nodes’ weighted in-degrees instead of node degrees. Moreover, the payoff definition slightly changes by: RGA (i) =

|{j ∈ Successor(i) : j has behavior A}| , |Successor(i)|

(10)

that Successor(i) determines the set of successors of node i.

3.3 Community-Aware Ranking Algorithms in Expert Finding To identify experts, community-aware ranking methods based on classical HITS and PageRank are proposed. Community-aware concept values more to the shared knowledge among people in the same community than to the people in different communities. In other words, people working in the same cluster are more probable to have more commitment towards their shared interests and innovations.

3.3.1

Overlapping Community-Aware HITS

Similar to the classical HITS algorithm, we contemplate about applying hubs and authority vectors. Hubs are vectors pointing to some other nodes, on the contrary, authorities receive ranks. We initialize the hub and authority (Ex) vectors with |N1 | and update them as follows:

172

M. Shahriari et al.



ai = α ×

j,i∈Cr &j ∈Ej i



hi = α ×



hj + (1 − α) × aj + (1 − α) ×

j,i∈Cr &j ∈Eij

3.3.2

hj ,

j,i ∈C / r &j ∈Ej i



(11) aj ,

j,i ∈C / r &j ∈Eij

Overlapping Community-Aware PageRank

Similar to PageRank algorithm, we perform a random-walk on the communityaugmented graph. If we consider the expertise value of a node i as Exi , this vector is initialized by |N1 | . It will repeatedly be updated until convergence based on the following formula: ⎛ 

⎜ Exi = β × ⎝α ×

j,i∈Cr &i∈Nei(j )

 +(1 − β) ∗

1 N

⎛ ⎝

⎞



Exj ⎠ + (1 − α) × Nei+ jin j,i ∈C /

r &j ∈Neii

⎛ ⎝

Exj Nei+ jout

⎞⎞ ⎠⎟ ⎠

,

(12)

here α determines to what extent communities affect the random-walk. Walks will be diverted towards inside and on the boards of communities when α is more significant than 0.5. β is the teleporting parameter to avoid dead ends. Like to the previous case, any OCD algorithm can be employed—DMID is the main one.

3.4 Evolution Analysis and Prediction of Overlapping Community Structures To analyze the evolution of communities over time, one needs to map them over sequential time slots. In other words, the significant events happening to communities need to be detected. There have been several different techniques to achieve this, but here we start by GED technique.

3.4.1

Community Evolution Prediction

To predict the evolution of communities, we need to map the communities over time and build several prediction models. Any supervised prediction task requires features and labels to train the data. As for the features, static and temporal community-level features may be extracted; however, communities need to be detected and mapped over time. To clarify the issue, if communities are mapped,

Dynamics of Overlapping Community Structures with Application to Expert. . .

173

then we can extract the events that communities undergo. For instance, two communities Ci and Cj in snapshot t can be merged into one single community in snapshot t + 1, therefore having features of the community Ci , its corresponding label can be analytically extracted. To be able to build a community prediction model, we may need such mapping and event extraction phase, for which there have been several different proposed approaches, but we prefer using a method named GED. Regarding the prediction phase, we employed the logistic regression classifier. Besides, we introduce the list of features in Table 2. Features in Table 2 can be classified into three categories. Node-level features such as leader ratio, leader degree, closeness, and eigenvector centralities are related to the leaders of a community. Several community-level features are also considered such as size ratio, density, cohesion, clustering coefficient, average assortative degree mixing, closeness, degree, and eigenvector centralities of communities. These features indicate how good the general structure of a community looks like. Other node- and community-level features are taken into account as temporal values. In other words, change in the features mentioned above in consecutive time slots t + 1 and t is taken into account for temporal features as for the classifier. Events that happen to communities are merge, split, survive, and dissolve. We consider a binary classification problem that constitute polar labels of {survive, not-survive}, {merge, not-merge}, {dissolve, not-dissolve}, and {split, notsplit}. For instance, not-survive consists of merge, split, and dissolve. We apply the logistic regression to perform the classification. Logistic regression extends linear regression that fits a line to data. It works based on logistic function and computes some parameters to construct the line going through the data [9]. Gradient descent or any other optimization techniques can be employed to minimize the error while using the following hypothesis function: hθ (x) =

1 1 + e−θ

T (x)

,

(13)

θ and x are the parameters and feature vectors.

3.4.2

Group Evolution Discovery

GED technique considers the consecutive snapshot of communities and takes their representative graphs. In other words, inclusion of a community in another community is a significant decision rule to extract events like survive, merge, split, and dissolve. The inclusion can be based on quantitative and qualitative metrics. Quantitative metrics are only about the percentage of overlapping nodes in two consecutive communities. On the contrary, qualitative inclusion metric may take into account other information such as rank and position of nodes. If we indicate the overlaps of communities Cit and Cjt+1 with OLCij , then their inclusion can be computed as follows:

174

M. Shahriari et al.

Table 2 List of features employed in the community evolution prediction task Features Leader ratio

Definition

Description Ratio of leaders

Value range (0,1]

Leaders average degree

(0,1]

Leaders average closeness

(0,1]

Leaders average eigenvector

(0,1]

Average community size

(0,1]

Density of the community

(0,1]

How cohesive is the community

(0,∞]

Counting number of triads

(0,1]

Dissimilarities among node Rank

[−1,1]

Simple degree measure

(0,1]

Centrality based on closeness

(0,1]

Centrality based on eigenvectors

(0,1]

t #leaderst+1 c #leadersc t+1 t Sizec − Sizec Dent+1 − Dentc c t+1 COc − COtc CCt+1 − CCtc c

Change in leader ratio Change in size Change in density Change in cohesion Change in clustering coefficient

[0,1] [0,1] [0,1] [0,1] [0,1]

D t+1 − D t

Change in degree

(0,1]

Cct+1 − Cct

Change in closeness centrality

(0,1]

Ect+1 − Ect

Change in eigenvector centrality

(0,1]

Survivetc

Survive as previous event?

True,false

mergetc

Merge as previous event?

True, false

splittc dissolvetc

Split as previous event? Dissolve as previous event

True, false True, false

#leaders N 

Leader average degree

deg(i)

i∈leaders

N 

C(i)

Leader average closeness

i∈leaders

Leader eigen centrality

i∈leaders

Size ratio

SizePi = " Ni 2|EiP | DenPi = P P

Density

N 

E(i) N |V P |

#

|Vi |(|Vi |−1)

2|EiP | |ViP |(|ViP |−1)  |EiP | |ViP |(|Vi |−|ViP |)

Cohesion

COPi =

Clustering coefficient

CC(V ) =

Assortative degree mixing Degree centrality

ρ =1−

Closeness centrality Eigen vector centrality δ leader ratio δ size ratio δ density δ cohesion δ clustering coefficient δ degree centrality δ closeness centrality δ eigen vector centrality δ previous survive δ previous merge δ previous split δ previous dissolve

C(u) =

v=u

E(u) =

1 λ

Δ T riples



 6 d2 V (V 2 −1)

deg(u) n−1

D(u) =



sh(u,v)

n−1



N Au,v E(v)

v=1

These features include node- and community-level ones that each of them can be regarding the current snapshot or the previous time slot. δ also shows the change in two consecutive values

Dynamics of Overlapping Community Structures with Application to Expert. . .

 OLCij =

(|Cit

∩ Cjt+1 | |Cit |

×

vi ∈Cit ∩C t+1 &vi ∈V



vi ∈Cit

175

(NR(Cit (vi )))

NR(Cit (vi ))

,

(14)

where NR is a representation of a node that can be its centrality, degree, or position of the node. In simulations, we consider NR as the centrality of a node; 0 ≤ OL ≤ 1. We apply some threshold values to assign the communities to certain events. Two main threshold values that can be selected are α and β values. By considering different threshold values, an event decision tree is constructed. For instance, if OL ≥ α and OL ≥ β, then it will be a survive event.

4 Baseline Approaches 4.1 Overlapping Community Detection Baselines 4.1.1

Speaker Listener Label Propagation

Speaker Listener Label Propagation (SLPA) is a stochastic based algorithm working based on label propagations on nodes [78]. Raghavan et al. flipped the original idea of this algorithm as Label Propagation Algorithm (LPA) [57]. LPA considers individual labels, in which it updates the labeling behavior based on the majority votes of its neighbors. In contrast, SLPA considers multiple labels which are updated based on speaking and listening rules of the algorithm. A node conserves several labels that are updated based on received signaling rules. This information can be, for instance, a received random label opted based on its frequency. Majority occurrence of labels among neighbors can accept this signal. At last, the algorithm post-processes the labels to form overlapping communities. The information propagation process is imitated based on communicating ability. 4.1.2

Stanoev, Smilkov, and Kocarev (SSK)

Stanoev et al. proposed a two-phase algorithm based on influence dynamics and membership computations. In the first phase, a random-walk is employed to calculate the local and global influence matrices. SSK assumes that relationships of nodes and their influences weigh more than considering the direct connection. In other words, proxies among nodes are better established while there exist triangles among nodes. Via adjacency matrix and triangle occurrences among nodes, influence matrix is constituted. This matrix is further applied to achieve the local and global influential nodes—correspondingly the hierarchy of the network. In other words, they calculate the transitive link matrix as follows:  tlj i = tlj i + tlvjki , (15) k

176

M. Shahriari et al.

where tlvjki = min Aki , Aj k is the transitive link weight for the edge (i, j ) which goes through k. The corresponding transition matrix for doing the random-walk can be obtained by row normalizing the tl matrix. After doing the random-walk, the most influencing neighbors of node i are identified based on (Ninfluential = j |Tj i = maxk Tki ()) that T shows the computed link weight transition matrix. By comparing the influencing neighbors with their neighbors’ influences, leaders can be detected. Afterward, membership of nodes to set of leaders can be identified by considering weighted average membership of neighbors. The updating rule for membership computation is as follows: Mi (t + 1) =

n 

(16)

Aij Mj (t),

j =1

where Aij is the row-normalized adjacency matrix. The hierarchical and decentralized working behaviors are among properties that SSK possesses.

4.1.3

CLiZZ

This algorithm comprises two main steps. One includes identifying leader nodes, and the other contains computing the membership of nodes to communities. It computes influence range of members based on shortest distance to identify leader nodes. It determines the mutual effects of nodes towards each other based on the following formula: n 

LVi = j =1 ;

e−

dij δ

(17)

,

3δ dij ≤= √  2

where dij is the shortest path from node i to node j , LSi indicates leadership value of node i within a range of  √3δ . An influential node has strong linkage with other 2 nodes of the network. Afterward, membership values of nodes to influential nodes need to be computed. To do this, a random-walk process with the help of initial stationary membership values is employed. The membership vector is indicated by M and each entry of this vector is updated based on the following: Mi (t + 1) =

1+

1 n

j =1 Aij

" Mi (t) +

n 

# Aij Mj (t) .

(18)

j =1

The algorithm needs to determine the δ based on the topological entropy of nodes. CLiZZ is suitable for directed and weighted networks.

Dynamics of Overlapping Community Structures with Application to Expert. . .

4.1.4

177

Link Communities

Instead of considering nodes, Link Communities maximizes the density of links inside communities and minimizes the density of links between communities. Similarity index helps in computing the similarity of edges. Hierarchical clustering together with similarity between edges build a dendrogram. By cutting this dendrogram at different levels, overlapping communities appear. The similarity between links can be calculated as follows:   Nei(i) ∩ Nei(j ) . (19) S(eik , ej k ) =  Nei(i) ∪ Nei(j ) Edges eik and ej k assumed to be similar when nodes i and j have an approximately high number of shared neighbors. After constituting the link dendrogram, we require to cut it at a suitable point. Partition Density (PD) is applied to find a suitable point to cut the dendrogram. PD can be defined as follows: PD =

|| #Eα − (#Nα − 1) 2  , #Eα M (#Nα − 2)(#Nα − 1)

(20)

α=1

where  is the number of link communities, #Eα is the amount of edges, and #Nα is the amount of nodes incident to an edge of link community Cα .

4.1.5

Merging of Overlapping Communities

This algorithm starts by assigning each node in one community. The communities of nodes are then merged and expanded based on a predefined fitness function. The fitness function f can be defined as follows: f (C, α) = $

C kinside +1 C kinside

C + koutside

%β ,

(21)

where C is a community and β is a parameter to identify community resolution, kin is the internal degree, and kout is the external degree of the intermediate community C. Each of communities with minimum size is visited and checked to see whether the fitness value is increased. Hence, the change in the fitness function can be calculated as follows: % $ $ % C∪{i} C log kinside + 1 − log kinside +1 % $ % $ Δf (C, i) = . (22) C∪{i}) G − log ktotal log ktotal

178

M. Shahriari et al.

After investigating the nodes, those with highest Δf will be added to the community. The process is continued until no communities can be further expanded. Meanwhile, communities are checked to observe whether there are duplicate ones with the same number of nodes.

4.1.6

AFOCS

Nguyen et al. proposed a two-phase algorithm that adaptively discovers overlapping communities [50]. In the first phase, they employ a density function to identify the cluster of nodes. For the next snapshots, they propose scenarios, which tackle adding/removing of nodes and edges. The density function (γ ), which is employed for cover evaluation, is defined as follows: |C in | γ =  , |C| 2

(23)

where C ∈ V . Moreover, they define a threshold to identify the internal density of connections.

4.2 Expert Ranking Baselines As to compare our expert ranking results including MRR and MAP values, we apply classical ranking algorithms including HITS and PageRank.

4.2.1

HITS

HITS is a ranking algorithm working based on two vectors called hub and authority. Hubs generate outlinks to other members, while authorities receive incoming connections. It is one of the classical algorithms used for ranking of the nodes. The updating vectors are shown in the following: ai =



hj

j ∈Ej i

hi =



(24) aj ,

j ∈Eij

that ai and hi are authority and hub vectors. We initialize these vectors with random values.

Dynamics of Overlapping Community Structures with Application to Expert. . .

4.2.2

179

PageRank

PageRank is another structure-based classical ranking algorithm proposed in the beginning of 1990s. Since then, it has been employed in certain domains and applications. If we indicate the PageRank value of a node by PRi , then it can be updated as follows: PRi = β ×

 j ∈Ej i

PRj 1 + (1 − β) × , outdeg(j ) n

(25)

where β is the damping factor controlling walks which are throttled in one part of the network.

5 Overlapping Community Detection as a Web Service There are several problems with the existing community detection tools. Most of OCD algorithms are based on different implementation tools that are not suitable for researchers looking forward to finding the communities of their data quickly. Moreover, after applying the algorithms, evaluation challenges, especially visualization of the small covers remain unsolved. Last but not least, most of the frameworks do not change obviously after some years and face a halt in their way of development. All the abovementioned challenges inspired us to create and develop the algorithms of our research work as Web services. By Web service, it is meant that it can be accessible from outside and quickly exposed to people need to find useful information in their data. Hence, REpresentational State Transfer (REST) as an architectural style, to develop HTTP-based and stateless communications, has been employed for this purpose. The developed Web service supports a RESTful interface. REST defines a couple of constraints that services can work properly on the Web. A RESTful service employs a stateless communication protocol such as Hypertext Transfer Protocol (HTTP) in which a request is independent of others [59]. Methods such as GET, POST, PUT, and DELETE are employed and each resource, here like an algorithm or a metric, can be denoted by Uniform Resource Identifier (URI). As for the programming language, we selected Java because it is well supported with graph processing libraries and it is also quite fast in comparison to other programming languages. The object-oriented capability of Java also enables us to design a flexible system which can be directly integrated with other software packages and to be extended by other developers. Different libraries including Java Persistent API, Apache Derby, yFiles, and la4j have been used to facilitate the implementation of our framework. la4j provides necessary linear algebra functionalities with high performance. Apache derby and Java Persistent API are, respectively, employed for database storage and data persistence. Finally, yFiles has been employed to envision visualizations. Furthermore, all the micro-services are developed by a

180

M. Shahriari et al.

Web service - las2pper

OCD Service • Un signed & Signed & Content-Enriched OCD algorithms • Synthetic Network Generators • Evaluation Metrics

Web Application web Client

Network Data

Viewer Service Visualization of Graphs & Covers Overlapping Community Detection

Fig. 2 It consists of two services: OCD and viewer service. OCD service is responsible for detecting overlapping communities in different types of networks such as (un)signed and content-enriched networks. Furthermore, it includes preprocessing, postprocessing, evaluation, and benchmark functionalities. Viewer service visualizes the found covers and networks

federated peer-to-peer framework named las2peer. las2peer is based on the Java which handles all the requirements to process HTTP request outward/inward the services and as well does mapping of the HTTP request to Java methods internally. In addition to the technical functionalities of the Web service, community-related requirements are indeed met during the implementation process. Several OCD algorithms with different levels of precision and runtime have been integrated with the framework. Additionally, some microservices are embedded for evaluation and visualization functionalities. Evaluation metrics including Normalized Mutual Information (NMI), Modularity, runtime, and different structural formats of reading the graphs and writing of covers are included. A general view of the architecture is indicated in Fig. 2. Of course, there is the possibility of uploading large graphs and receiving the response through Java clients, but online users (mainly researchers) may need to find the communities of data with the help of this platform. The platform is named

Dynamics of Overlapping Community Structures with Application to Expert. . .

181

WebOCD which is deployed online for data analytic purposes.12 The code for the WebOCD framework is hosted in the code repository GitHub.13

6 Datasets In this section, we describe structural properties of datasets14 used in the experiments.

6.1 Datasets for Community Analysis We have employed datasets that are mainly available online.15,16 These datasets belong to different domains; for instance, they belong to areas including social, communication, email, and technical network structures. Information about the number of nodes, edges, and type of these networks is available in Table 4.

6.2 Expert Ranking Question–Answer Forums Stack exchange17 is a question and answer forum dedicated to discussing issues in specific domains. The topics range from social, political, technical, and health issues. Physical fitness and computer science forums are among two topics that people share their opinions and innovations. In the health-care forum, people ask for health-related topics. In contrast, computer science forum is targeted for technical and theoretical issues in computer science domain. These two different contexts might cause differences in research analysis. In addition to stack exchange forum, we received a dataset from Nature18 This data contains topics and discussions regarding the wild life in Estonia. Information about these datasets including the number of posts, users, questions, answers, and the time period is indicated in Table 3.

12 http://dbis.rwth-aachen.de/acis/apps/ocd/login.html. 13 https://github.com/rwth-acis/REST-OCD-Services. 14 All

of the datasets are available online except Nature that Tallinn University gave us the data.

15 https://snap.stanford.edu/data/. 16 http://konect.uni-koblenz.de/. 17 http://stackexchange.com/sites. 18 http://www.looduskalender.ee/forum/.

182

M. Shahriari et al.

Table 3 Number of posts, users, questions, answers, and the time period of stack exchange and Nature forums Forum Physical fitness Computer science Nature

Posts 11,522 21,731 162,325

Users 7567 22264 1370

Questions 3602 8955 227

Answers 7735 12,415 162,098

Time period 03.01.2011–03.05.2014 25.11.2008–08.03.2015 01.01.2013–10.02.2015

7 Evaluation Protocol and Metrics We need to define an evaluation protocol and required metrics to evaluate the algorithms and methodology. We evaluate community detection algorithms with statistical and knowledge-driven measures. Statistical measures are informative for the scenarios for which ground truth information is not available. Example of such a situation is real-world networks that we may have confusion about the resolution of communities. Among statistical measures, modularity has been applied to evaluate community detection algorithms. In contrast, knowledge-driven measures such as NMI are suitable when true cover knowledge is accessible. While generating synthetic networks based on complex dynamics and stochastic processes, community information is not unknown anymore.

7.1 Modularity The original modularity formula proposed by Newman is not suitable for overlapping communities, hence a version suitable for both directed graphs and overlapping communities is applied [51]. The implemented modularity considers a simple random graph named as the null model. The null model contributes to figure out how the cover is modular. This random structure has the same number of nodes, edges, in-degree, and out-degree distribution as the original graph. A belonging factor is defined as follows to compute to what extent an edge belongs to a community: γ(i,j ),l = f (Mi,l , Mj,l ),

(26)

where γ is the belonging factor, f is a function computing the factor, and M is the membership matrix. f (Mi,l , Mj,l ) = Mi,l Mj,l is suggested in the original paper. Probability that a node belongs to a community is assumed to be stochastically independent from the probabilities of other nodes. Similar to the edge belonging factor, one may proceed by considering the belonging probability of nodes to a + certain community. Hence, we define γ(i,j ),l as the belonging factor of an edge (i, j ) which starts from node i and belongs to community Cl :  j ∈V Mi,l Mj,l + , (27) γ(i,j ),l = |V |

DBLP 1959 16,354 Co-Authorship

Dolphins 62 318 Social

Email 1133 10,902 Social

Facebook 4039 176,468 Social

Hamsterster 2000 32,196 Social

This table includes number of nodes, edges, and type of each network

Graph Nodes Edges Type

Internet 6474 25,144 Technological

Jazz 198 5484 Social

Table 4 This table shows real-world networks that are employed to compare OCD algorithms Power Grid 4941 13,188 Technological

Sawmill 36 124 Social

Sawmill Strike 24 76 Social

Zachary 34 156 Social

Dynamics of Overlapping Community Structures with Application to Expert. . . 183

184

M. Shahriari et al.

− similarly γ(i,j ),l indicates the belonging factor of an edge (i, j ) referring to node j which is defined as follows:  − i∈V Mi,l Mj,l = γ(i,j , (28) ),l |V |

finally to compute the modularity, we proceed by subtracting the probabilities related to original graph and the null model as follows: Q=

|CL |  1   (γ(i,j ),l Aij − γ(i,j ),l ), m

(29)

l=1 i,j ∈V

 where γ(i,j ),l is the belonging probability regarding the null model.

7.2 Normalized Mutual Information NMI is intended to calculate the quality of overlapping building blocks [21]. If we consider that each node i is a member of a community Cl , then the entry ij in the membership matrix M can be considered as a random variable with the probability distribution as follows: P (Xil = 1) =

#|Cl | , P (Xil = 0) = 1 − P (Xil = 1), |V |

(30)

where #|V|C|l | is defined as the number of nodes in community Cl . The same probabilistic relationship is assumed to be hold for the ground truth cover. The uncertainty whether i belongs to what community can be considered by conditional entropy as follows: H (Xk |Yl ) = H (Xk , Yl ) − H (Yl ),

(31)

where it only depends on distributions of P (Y ) and P (Xk , Y − l). Observing the vector Yl then the entropy of Xk given the entire Y is defined as follows: H (Xk |Y ) =

min

l∈{1,..., |CL |}

H (Xk |Yl ),

(32)

if we normalize and average over the communities, it can be written as follows: Hnorm (X|Y ) =

1  H (Xk |Y ) , |CL | H (Xk ) k

(33)

Dynamics of Overlapping Community Structures with Application to Expert. . .

185

similarly one can compute Hnorm (Y |X) and proceed to compute the total NMI value as follows: 1 NMI(X|Y ) = 1 − [Hnorm (X|Y ) + Hnorm (Y |X)]. 2

(34)

The actual NMI values range between 0 and 1. 1 indicates the highest and the best match. To apply the measure, synthetic generators like LFR networks can be useful. It is possible to employ runtime to evaluate the speed of algorithms. As classical ranking algorithms are extended to community-based ranking algorithms for expert identification, we may require assessing their goodness. Expert identification, as a subclass of information retrieval approaches, seeks to retrieve the list of experts. Thus, metrics such as MAP and MRR from information retrieval are suitable. In the Community Evolution Prediction (CEP) problem, we deal with a prediction task, hence classification metrics such as precision, MAE, and RMSE can be employed.

7.3 LFR Synthetic Networks Synthetic networks, like computer-generated complex graphs, consider several properties from the real world. These parameters identify the number of nodes, size of communities that take values from a power-law degree distribution. Other parameters include mixing parameter μ, the average degree of nodes, minimum and maximum size of a community, number of overlapping nodes, and the number of communities a node may participate. After identifying the size of the network and number of overlapping communities, we can identify the minimum and maximum degree of nodes. Nodes can participate in connections crossing outside of a community that can be determined by μ [39].

7.4 Precision Precision can play different roles in different contexts. First, precision is employed in information retrieval problems such as (expert) ranking. In the other role, precision is informative in classification context where true/false positive rates are required. 7.4.1

Precision in Information Retrieval Context

A set of queries are considered to compute the MAP in information retrieval systems. MAP is the average precision score of each query that can be computed as follows:  q∈Q AP(q) MAP = , (35) |Q|

186

M. Shahriari et al.

where AP is the average precision for the results returned by each query which can be calculated as follows:  P @k(q) . (36) AP = k∈K |K| K determines the cut-off to ignore documents ranked lower than the threshold value. P @k computes the fraction of relevant results in the top k [11].

7.4.2

Precision in Classification Context

While having a classification problem, known classes are binary or multi-target cases. The true positive and false positive rates need to be calculated to evaluate such supervised learning algorithms. Positive and negative are signs of the classifier’s expectations or better to be named predictions [9]. In contrast, true and false prove to be correspondent with ground truth labels. If one denotes the true positive rates with (TP) and false positive rates with (FP), then precision can be calculated as follows: Precision =

TP . TP + FP

(37)

In a binary classification problem while having two distinct labels or classes, the Prediction Accuracy (PA) may be a better measure indicating the goodness of the prediction as follows: PA =

TP + TN . TP + TN + FP + FN

(38)

7.5 Mean Reciprocal Rank Another metric which is applied to evaluate the results of expert ranking lists is MRR which can be stochastically defined as follows:  MRR =

1 q∈Q rankq

|Q|

.

(39)

where rankq is the rank position of the first relevant document for the query q [11].

Dynamics of Overlapping Community Structures with Application to Expert. . .

187

8 Results In this section, we discuss the evaluation results. First, we describe the modularity and runtime of the proposed DMID algorithm with the baselines. Moreover, we illustrate the results of DMID ranking algorithm for the expert identification domain. Then, we apply DMID and other baselines to the community evolution prediction problem.

8.1 Overlapping Community Detection 8.1.1

Results on Synthetic Networks

We employ both real-world and synthetic networks to compare different OCD algorithms. Synthetic networks can be generated based on different parameters. In general, average node degrees, mixing parameter, and number of overlapping nodes may challenge the algorithms in the evaluation. In this case, DMID algorithm is compared with a couple of algorithms including SLPA, SSK, CLiZZ, and Link Community (Link). Figure 3 indicates a clear picture regarding the comparison of synthetic networks based on different parameters. k indicates the average node degree and μ is a sign of mixing parameters that intertwine communities. As we can observe k = 12 and μ = 0.1 indicate a network with a rather low node average degree and low mixing parameter. When the percentage of overlapping nodes are small from 0 to 10%, SLPA has superior NMI value in comparison to others. It is mainly followed by SSK, merging of overlapping communities (MONC), DMID, Link, and CLiZZ. However, as the number of overlapping nodes increases, the performance of SLPA slumps over DMID, Link, and CLiZZ. CLiZZ has the worst NMI value when the number of overlapping nodes increases in networks with assigned parameters. DMID also has almost a stable NMI value (around 0.2) as the number of overlapping nodes increases. Best cases are MONC followed by Link when overlap enhances which results in NMI value of around 0.3. One can observe a bit different results when average node degree increases and is set to k = 24. Now, MONC is quite improved to 0.5 when overlapping percentage increases to 40%. Similarly, at the beginning in case of zero overlapping percentage, SLPA and SSK have the best NMI performance in comparison to others. SLPA performance decreases but this time is a bit better than SSK, CLiZZ, and DMID. Again, DMID has a stable performance of around 0.2 for all of the overlapping percentages for these parameter settings. For instance, if we look at 40% overlapping, MONC (0.49), SLPA (0.24), Link(0.22), DMID (0.21), SSK (0.15), and CLiZZ (0.07), respectively, obtain NMI values. Furthermore, in the cases when mixing parameter increases to μ = 0.3 the same pattern is kept for both k = 12 and k = 24.

188

M. Shahriari et al.

Fig. 3 This figure indicates the runtime of the different algorithms on LRF synthetic networks. The upper plots show networks with k = 12 and the lower plots indicate networks with k = 24. The left and right plots show networks with μ = 0.1 and μ = 0.3, respectively

Figure 4 shows the runtime in seconds for different algorithms and various parameter settings. The first interesting point is the high runtime of Link which could not be plotted in the figures. Moreover, SSK and MONC approximately had worse runtime performance in comparison to other algorithms. As for all of the parameter settings and the number of overlapping percentage, SLPA, CLiZZ, and DMID finish in less than 10 s. SLPA wins with a bit better performance, and it is followed by DMID and CLiZZ, respectively.

8.1.2

Results on Real-World Networks

To compare the appropriateness of different algorithms, they are applied to a wide range of datasets available online. We inserted the results of the modularity and the CPU times in Tables 5 and 6. In Table 5, we can observe that DMID

Dynamics of Overlapping Community Structures with Application to Expert. . .

189

Fig. 4 This figure shows the NMI values of different OCD algorithms on LFR synthetic networks. The upper plots show networks with k = 12 and the lower plots indicate networks with k = 24. The left and right plots show networks with μ = 0.1 and μ = 0.3, respectively

has approximately satisfactory modularity performance in comparison to other algorithms. As we can observe, SLPA wins the first rank in seven cases including DBLP (0.7115), Dolphins (0.7457), Email (0.6730), Hamsterster (0.5703), Internet (0.5410), and Zachary (0.6993). Moreover, it obtains the second rank in two cases including Sawmill (0.6812) and Facebook (0.9318). Moreover, DMID receives the first rank in two cases including Sawmill (0.6854) and Sawmill Strike (0.7523) and obtains the second rank for two cases in Zachary (0.6914) and Internet (0.5229). MONC is neither among the first nor the second rank algorithms, and CLiZZ only reaches the second rank for Power Grid (0.6810). Furthermore, SSK results in the first rank in two cases of Power Grid (0.7014) and Facebook (0.9407), and it obtains the second rank for JAZZ (0.4135) and Sawmill Strike (0.7029). Finally,

190

M. Shahriari et al.

Link achieves the second rank in two cases of Email (0.4410) and Hamsterster (0.5110). On the other hand, regarding the runtime of these algorithms, SLPA has the best runtime in almost all of the datasets except in very much small datasets that the difference may be negligible. Link and SSK have the worst performance among the other algorithms. Moreover, DMID surpasses the others in many cases including Dolphins and Email. However, CLiZZ gets the superior performance in DBLP, Facebook, Hamsterster, Internet, and JAZZ. Finally, MONC wins only in Power Grid network.

8.2 Expert Finding 8.2.1

Mean Average Precision and Mean Reciprocal Rank Values

We applied Overlapping Community-Aware HITS DMID (OCAHD), Overlapping Community-Aware HITS SLPA (OCAHS), Overlapping Community-Aware PageRank DMID (OCAPD), Overlapping Community-Aware PageRank SLPA (OCAPS), HITS, and PageRank (PR) on three question and answer forums including Fitness, Nature, and computer science forums. We employ MRR and MAP together with correlation values among the algorithms to evaluate the runs. Figure 5 shows the MRR and MAP values for approximately a 2-year period from 2013 to 2015. Community-aware ranking algorithms outperform for the two measures in comparison to its counterparts. Regarding MRR values, from 3.1.2013 to 7.1.2014, OCAHD has the best performance in comparison to others. After 7.1.2014, OCAHD and OCAHS have the superior performance regarding MAP values. It shows that community-aware ranking algorithms outperform others. As for MRR, the situation is a bit different and OCAHS and HITS care about the position of the relevant item. This dominance is kept for approximately all of the time slots. One interesting point about the results is that at the beginning nearly all of the algorithms face the cold start problem and are unable to generate accurate results. To continue regarding the MRR and MAP values, one may look at Fig. 6. The performance of the algorithms is quite similar in this dataset. All the algorithms start from approximately low MAP values around 30% and reach MAP values of more than 67%. Respectively, OCAPD, OCAPS, PR, OCAHD, HITS, and OCAHS obtain low precision levels. After 6.1.2012, they all reach more than 0.9. MAP and MRR values on Fitness forum also confirm the superiority of overlapping communityaware algorithms. If we look at Fig. 7, we can obtain a similar understanding of the positive performance of community information on expert identification task. Regarding MRR, OCAPD and OCAPS win over the other algorithms. For instance, at 6.1.2014, OCAPD and OCAPS obtain 93.83 and 92.49, respectively, which are better than PR (91.39), HITS (90.65), OCAHD (89.33), and OCAHS (87.74). Similarly, overlapping community-aware ranking algorithms yield better results in comparison to classical HITS and PR values regarding MRR measure.

Graph DMID CLiZZ MONC SLPA SSK Link

DBLP 0.4203 0.4281 0.2102 0.7115 0.4855 0.3497

Dolphins 0.5140 0.5218 0.3218 0.7457 0.5506 0.3599

Email 0.3457 0.2483 0.0886 0.6730 0.4283 0.4410

Facebook 0.8257 0.8820 0.0000 0.9318 0.9407 NaN

Hamsterster 0.2801 0.3683 0.1727 0.5703 0.3665 0.5110

Internet 0.5229 0.2253 NaN 0.5410 0.4065 0.0808

Table 5 Modularity values for different OCD algorithms on real-world datasets Jazz 0.3461 0.0000 0.2014 0.7532 0.4135 0.3386

Power grid 0.6433 0.6810 0.3985 0.5345 0.7014 0.0902

Swamill 0.6854 0.4550 0.3638 0.6812 0.5882 0.3407

Swamill Strike 0.7523 0.0000 0.4957 0.6925 0.7029 0.3680

Zachary 0.6914 0.5929 0.3321 0.6993 0.5929 0.2370

Dynamics of Overlapping Community Structures with Application to Expert. . . 191

Graph DMID CLiZZ MONC SLPA SSK Link

DBLP 110.77 84.91 159.08 2.89 3911 12,501

Dolphins 0.380 0.418 0.441 0.657 0.360 0.545

Email 5.51 7.59 109.01 2.82 129.98 3074

Facebook 3102 528.90 3808 82.09 3528 NaN

Hamsterster 734.19 29.55 87.38 5.04 7406 133,532

Table 6 Runtime of OCD algorithms on real-world networks Internet 2523 2484 NaN 5.09 6546 48,736

Jazz 0.203 0.140 0.296 2.29 0.656 304.33

Grid 3040 400,037 62.41 42.54 204,797 5083

Swamill 0.040 0.108 0.020 0.129 0.236 0.065

Swamill Strike 0.028 0.011 0.009 0.018 0.029 0.014

Zachary 0.018 0.011 0.074 0.073 0.023 0.045

192 M. Shahriari et al.

Dynamics of Overlapping Community Structures with Application to Expert. . .

Fig. 5 This figure shows the MAP and MRR values for Nature forum

Fig. 6 This figure indicates the MAP and MRR values for fitness forum

Fig. 7 This figure shows the MAP and MRR values for computer science forum

193

194

M. Shahriari et al.

Table 7 Pearson correlation among PageRank, HITS, and community-aware HITS and PageRank algorithms on computer science forum dataset Dataset PR CA-PR-DMID CA-PR-SLPA HITS CA-HITS-DMID CA-HITS-SLPA

8.2.2

PR 1.000 0.999 0.999 0.964 0.965 0.963

CA-PR-DMID 0.999 1.000 0.999 0.964 0.965 0.963

CA-PR-SLPA 0.999 0.999 1.000 0.964 0.965 0.963

HITS 0.964 0.964 0.964 1.000 0.999 0.999

CA-HITSDMID 0.965 0.965 0.965 0.999 1.000 0.999

CA-HITSSLPA 0.963 0.963 0.963 0.999 0.999 1.000

Spearman Correlation Values

To figure out how similar the original HITS and PageRank algorithms are correlated with community-aware ranking algorithms, we calculated the Pearson correlation of these algorithms. Table 7 shows the correlation values for computer science forum. The bold values indicate higher values of correlation. As we can observe the PR, OCAPD and OCAPS have quite high similarity values (≈ 0.999). Classical HITS and HITS algorithms combined with the community have quite high correlation values among themselves. On the contrary, HITS-related algorithms have the lower correlation with PageRank-related algorithms, for instance, OCAHD-PR (0.965), OCAHD-OCAPD (0.965). Table 8 also indicates very similar results to computer science forum with a bit lower cross-correlation among algorithms OCAHDOCAPS (0.898), OCAHS-OCAPS (0.897), OCAHD-OCAPD (0.904), OCAHD-PR (0.902), OCAHS-PR (0.901), OCAHS-OCAPD (0.901), and so on. The correlation results are a bit different for Nature forum as we can see in Table 9, in which crosscorrelation values among HITS-related and PageRank-related algorithms slump. It is easily observable that PageRank-related algorithms and HITS-related algorithms have very much similar correlation among themselves. Moreover, HITS-PR (0.543), HITS-OCAPD (0.521), HITS-OCAPS (0.519), OCAHD-PR (0.558), OCAHDOCAPD (0.537), OCAHD-OCAPS (0.534), OCAHS-PR (0.535), OCAHS-OCAPD (0.512), and OCAHS-OCAPS (0.513) obtain correlation values of ≈ 0.5 which is much lower than the other two datasets. This observation indicates that similarity of algorithm behavior depends on the nature and structure of the dataset.

8.3 Community Evolution Prediction 8.3.1

Community Properties

Before we demonstrate the results regarding the significant features of community evolution prediction and prediction accuracies, we investigate structural properties of the algorithms. The number of overlapping nodes, average community sizes,

Dynamics of Overlapping Community Structures with Application to Expert. . .

195

Table 8 Pearson correlation among PageRank, HITS, and community-aware HITS and PageRank algorithms on Fitness forum dataset Dataset PR CA-PR-DMID CA-PR-SLPA HITS CA-HITS-DMID CA-HITS-SLPA

PR 1.000 0.999 0.998 0.901 0.902 0.901

CA-PR-DMID 0.999 1.000 0.996 0.901 0.904 0.901

CA-PR-SLPA 0.998 0.996 1.00 0.897 0.898 0.897

HITS 0.901 0.901 0.897 1.000 0.999 0.997

CA-HITSDMID 0.902 0.904 0.898 0.999 1.000 0.996

CA-HITSSLPA 0.900 0.901 0.897 0.997 0.996 1.000

Table 9 Pearson correlation among PageRank, HITS, and community-aware HITS and PageRank algorithms on Nature forum dataset Dataset PR CA-PR-DMID CA-PR-SLPA HITS CA-HITS-DMID CA-HITS-SLPA

PR 1.000 0.997 0.990 0.543 0.558 0.535

CA-PR-DMID 0.997 1.000 0.993 0.521 0.537 0.512

CA-PR-SLPA 0.990 0.993 1.0 0.519 0.534 0.513

HITS 0.543 0.521 0.519 1.000 0.996 0.994

CA-HITSDMID 0.558 0.537 0.534 0.996 1.000 0.990

CA-HITSSLPA 0.535 0.512 0.513 0.994 0.990 1.000

and number of communities are among the explored properties; Table 10 shows the information. We can figure out that the number of overlapping nodes for DMID algorithm is approximately high in comparison to SLPA and AFOCS. SLPA detects less overlapping nodes but more than AFOCS. For instance, let us consider three datasets of Power, NetScience, and PolBlogs, DMID has the highest overlapping percentage of 0.96, 0.55, and 0.839, respectively. AFOCS detects 0.076, 0.1, and 0, respectively, and SLPA identifies 0.404, 0.16, and 0.002, respectively. Regarding the number of detected communities, different algorithms show different resolution levels. The number of communities identified by AFOCS is way more than DMID and SLPA. SLPA generates yet more communities in comparison to DMID. Similarly, regarding Power, NetScience and PolBlogs, 4256, 658, and 823 are detected by AFOCS. Moreover, SLPA detects 737, 407, and 5; in contrast, DMID detects 569, 61, and 15. When there are a low number of communities with high overlapping percentage, we may deduce that size of communities should be large. In this regard, DMID has the biggest community sizes of more than 500 for bigger datasets which are followed by SLPA and AFOCS. In this regard, AFOCS communities are limited to the size of around 3–5 with an exception regarding PolBlogs with the average number of 148 members.

196

M. Shahriari et al.

Table 10 Number of communities, number of overlapping nodes and average size of communities for three algorithms including AFOCS, DMID and SLPA are computed on a couple of real-world networks SLPA C Karate Club 4 Dolphins 3 PolBlogs 5 NetScience 407 Power 737 CA-GrQc 954 p2p-Gnutella08 230

SLPA AvgSize 10.5 216 245.4 4.1 9.9 6.9 18.24

SLPA Ovl 0.23 0.048 0.002 0.1656 0.40376 0.24475 0.50436

AFOCS C 0 34 823 658 4256 2683 5274

AFOCS AvgSize 0 3.3 148 4.2 2.1 3.9 2.39

AFOCS Ovl 0 0.008 0 0.1 0.076 0.038 0.04

DMID C 2 6 15 61 569 263 614

DMID AvgSize 15 25.2 1026.2 177.31 462.8 654.7 573.4

DMID Ovl 0.43 0.65 0.839 0.55 0.96 0.8051 0.6803

One may observe the structural property differences of these three algorithms

8.3.2

Significance of Features

To compare the significance of applied features for prediction of each OCD algorithm, we plotted their HeatMap in Fig. 8. A HeatMAP figure best describes the significance of various features. The features are plotted in bold blue when the feature is important for the prediction task. The first finding is the significance of size ratio feature in all of the datasets and with the contribution of all algorithms. Although different algorithms lead into approximately different communities and different predictions, size of a community has been shown to be a significant and prevailing factor to predict various events happening to a community. Regarding DMID algorithm, in survive degree centrality, in merge temporal features such as previous merge or previous split, in split degree centrality and leader degree centrality and in dissolve density and eigenvector centralities are more important. In Enron dataset, we can observe that leader closeness centrality is more important in survive, eigenvector centrality and previous split are important for merge, previous merge for split and finally previous split, cohesion and change in cohesion can be important for dissolve. Finally, leader eigenvector centrality, change in size ratio and change in density are important for the survive event. Regarding merge leader ratio, regarding split leader closeness centrality and delta leader ratio are important. Moreover, as for dissolve, previous merge is important. If we take a look at other algorithms and datasets, we can observe a different pattern for significant features. This observation may be because of the inherent properties of the algorithms which cause different levels of importance of features and events. For instance, with SLPA and survive event in Facebook, we can observe the change in clustering coefficient as an important feature, but for AFOCS and Facebook, the previous status of the community is important. One challenging point that comes to mind is whether for each dataset, each separate event, and each algorithm, we need to employ the methodology mentioned above and look at the result? The answer to this question can be manifold. Yes, we need to do that because correlating the dynamics of the algorithm with the property of the event is not easy and thus this may depend on

Dynamics of Overlapping Community Structures with Application to Expert. . .

197

Fig. 8 Comparison of important features in Facebook, Enron and DBLP datasets through SLPA, DMID and AFOCS algorithms

the context of the social network and the properties of the algorithm. From another aspect, communities need to be defined explicitly, and we may need to consider context and content of communities. Adding contextual and content-based features may help to figure out more stable and consistent community properties; which we put it for our future work. The third aspect regarding the question mentioned above is the application of other statistical prediction methods. Perhaps using logistic regression and a huge number of features may not lead into variant results.

198

8.3.3

M. Shahriari et al.

Community Evolution Prediction Accuracies

To observe how different feature sets behave in community evolution problem, we applied different categories of features including structural, temporal, and selective features for each dataset and each algorithm. By structural features, we apply node- and community-level features. By temporal, we employ temporal nodeand community-level features. Selective features are the best from temporal and structural feature categories that are selected and applied to give the prediction accuracy results. We can see the results in Table 11. The first interesting observation is that selective features give better prediction accuracies in almost all of the cases. There is yet an exception for DMID-Enron and the survive event. There may exist other exceptions, but in the majority of the cases, selective features yield the best performance. If we consider the selective features of the algorithms, datasets, and the events, the comparison may become easier. In Facebook dataset and the survive event, AFOCS generates the best prediction accuracy of 82.35 which is higher than 75 and 78.57 of SLPA and DMID, respectively. Regarding dissolve and Facebook, DMID gives the best result with the prediction accuracy of 78.57. This can be due to big communities identified by DMID. Finally, regarding split, the highest prediction accuracy (77.94) belongs to SLPA. If we look at Enron dataset, DMID gives the best prediction accuracy result (88.71) that is better than SLPA (76.92) and AFOCS (82.14) for the survive event. Regarding dissolve, SLPA generates the best prediction accuracy of 93.75. Regarding merge and split, we can figure out that DMID wins with 95.59 and 88.71 on Enron dataset, respectively. As for the DBLP dataset and the survive event again DMID wins with the value of 82.14. Results indicate that, regardless of the datasets, selective features based on DMID algorithm achieve the highest prediction accuracy for the survive event. Regarding dissolve and DBLP, we can observe the DMID superiority in prediction task. Although Enron dataset indicated SLPA as the best candidate for dissolve but DMID gives better results for Facebook and DBLP. Regarding merge event, DMID obtains the best prediction value of 66.76 which is higher than 64.78 (SLPA) and 63.62 (AFOCS). Except for the Enron dataset, DMID as well leads to the best prediction accuracy results for the merge event. Finally, as for split, DMID again takes the lead and wins with obtaining the value of 74.29 over SLPA (65.56) and AFOCS (67.78). Except for the Facebook dataset, DMID as well gets the highest prediction accuracy for the split event. In summary, although in some cases SLPA is also good, DMID generates the highest prediction accuracies in all of the datasets and over all the events. This issue may be because of big communities generated by DMID. Prediction of what happens to a small number of big communities might be an easier task while AFOCS never could reach good results in comparison to DMID and SLPA. This aspect might be due to a large number of communities and may be difficult to predict what happens to them. Altogether, the importance of events was quite different and somehow unstable. However, the dominance regarding prediction accuracies was stable and was achieved mostly by DMID. One question which may come to

Dynamics of Overlapping Community Structures with Application to Expert. . .

199

Table 11 Prediction accuracy results of SLPA, AFOCS, and DMID on three datasets including Facebook, Enron, and DBLP SLPA-Facebook-Structural SLPA-Facebook-Temporal SLPA-Facebook-Selective DMID-Facebook-Structural DMID-Facebook-Temporal DMID-Facebook-Selective AFOCS-Facebook-Structural AFOCS-Facebook-Temporal AFOCS-Facebook-Selective SLPA-Enron-Structural SLPA-Enron-Temporal SLPA-Enron-Selective DMID-Enron-Structural DMID-Enron-Temporal DMID-Enron-Selective AFOCS-Enron-Structural AFOCS-Enron-Temporal AFOCS-Enron-Selective SLPA-DBLP-Structural SLPA-DBLP-Temporal SLPA-DBLPA-Selective DMID-DBLP-Structural DMID-DBLP-Temporal DMID-DBLP-Selective AFOCS-DBLP-Structural AFOCS-DBLP-Temporal AFOCS-DBLP-Selective

Survive 58.33 50 75 58.93 51.79 78.57 75.49 80.39 82.35 71.15 65.39 76.92 91.94 88.71 88.71 77.38 77.38 82.14 53.92 60.78 67.16 62.5 62.5 82.14 54.62 55.65 55.99

Dissolve 72.99 70.3 72.82 67.7 66.4 78.57 61.52 62.27 63.28 43.75 68.75 93.75 79.55 72.73 88.64 62.05 56.25 79.17 53.92 61.77 67.2 62.23 65.69 66.42 55.07 56.09 55.99

Merge 66.67 69.38 67.83 67.35 65.82 67.22 64.5 63.54 65.02 65.15 56.06 78.79 89.71 85.29 95.59 56.14 70.17 74.56 58.98 62.85 64.78 63.35 64.49 66.76 58.38 60.88 63.62

Split 66.18 61.77 77.94 70.62 53.75 63.75 60.09 63.78 66.03 50 50 83.33 91.93 88.71 88.71 62.5 64.58 81.25 62.78 63.61 65.56 65.71 67.86 74.29 62.03 64.04 67.78

These prediction accuracies are shown for four events comprising survive, dissolve, merge, and split

mind is what needs to be done if a new network is given. In others words, which algorithm needs to be chosen and which set of features one may use. Can these set of features give rise to reliable results when applying to real-world scenarios? To answer this research question, we may need more context-dependent and contentbased prediction models that consider contextual dynamics and processes. Although the structural features regarding an algorithm could give satisfactory prediction accuracies by considering train and test classes, one may need to identify robust and rather reliable community borders. To put this another way, we may require scrutinizing the collective and contextualized behavior of members in a community.

200

M. Shahriari et al.

8.4 REST Overlapping Community Detection Service To gain a better perspective regarding the OCD Web service, they are evaluated by online users. In other words, several master and Ph.D. students studying in the field of computer science were invited to do tests with the WebOCD service. The main elements and functionalities in WebOCD were fairly explained to the participants. They were supposed to import some graphs, run some algorithms, compute modularity and NMI measures, and observe the visualization. Some Webbased questioners are devised and handed out online among participants to ask for their feedback. We consider questions with possible answers of a 5-point scale ranging from strongly agree to strongly disagree. Moreover, they were required to accomplish the task within 20 min in which all of the participants succeeded. It is worth noticing that the survey attendees did not receive any remunerations for all their efforts except some chocolates. The questions including the mean, median, and standard deviation are shown in Table 12. As we can notice, the WebOCD framework received approximately satisfactory feedback. Almost all of the questions obtained the mean value more than four. We may notice only one question that received a three value. It is about the error handling of the framework. The participants commented out at the end that they neither agree nor disagree with it, or they did not face any problem which happens when using the service. Other aspects of the design of the framework also obtained an acceptable assessment. It was very nice to acquire a positive response regarding using the framework for future.

9 Discussion and Future Work As you now read the article, thousands of people check their Facebook profiles and hundreds join new communities, and dozens may leave their communities. A plethora of user activities has made data science a significant tool to analyze and emend social context. Communities, as the mesoscopic building blocks of social media, are required to be identified and analyzed. Analysis of communities brings many benefits from different aspects. Knowing the communities of people, buying a certain brand, may provide opportunities to introduce other products of the same brand for the companies. In learning environments, knowing the group of amateur learners can better contribute them to catch their learning materials. Maybe, they are offered new learning resources. Besides, boundary spanners can expedite the flow of information. In open-source developer communities, end users can be further supported with similar requirements which may be interesting for them. Keeping in mind the increasing value of community analytic in online media, in this paper, we proposed a holistic approach for detection and analysis of overlapping communities. We combined dynamics extracted from social theories; DMID. The algorithm has several informative properties; suitable for hierarchy detection, can

Dynamics of Overlapping Community Structures with Application to Expert. . .

201

Table 12 This table indicates results of survey by participants which includes mean, median, and standard deviation of the feedback values Questions The client has an intuitive design is easy to understand The client is easy to use The client responded to my actions in the way I would expect I was able to execute the given tasks without having problems When I made a mistake, a useful error message was shown It is easy to add new graphs to the system It is easy to add new covers to the system The graph visualization gives a good overview over the graph structure The cover visualization gives a good overview over the communities The visualizations are easy to handle (i.e., to zoom and to move) The system is useful for handling OCD algorithms The system is useful for handling OCD benchmarks The system is useful for handling OCD metrics The system responded to my actions with a convenient speed The system works well

Mean 4.33

Median 4.5

Standard deviation 0.67

4.17 5.0

4.0 5.0

0.17 0.0

4.83

5.0

0.17

3.0

3.0

0.0

4.67

5.0

2.67

4.83

5.0

0.17

4.83

5.0

0.17

4.33

4.0

0.27

4.5

4.5

0.3

4.67

5.0

0.27

4.5

4.5

0.3

4.67

5.0

0.27

4.67

5.0

0.27

4.5

4.5

3.0

be implemented parallel, reliable results for community prediction and identifying fuzzy and overlapping community structures. Moreover, it is competitive concerning statistical and knowledge-driven measures. The first problem of this article is concerned with the ability of the algorithm to leverage time adaptively. The algorithm can be employed in a separate snapshot of the network; however, this may endanger both the running performance and the stability of the algorithm. We intend to tackle this issue and propose some adaptive leadership update for the first phase and some sort of efficient forward or back propagation dynamic for the second phase. Moreover, communities are mainly formed by the force of context dimensions which are hidden with the structural analysis methods. We further plan to equip our method with contextual information such as enriched properties of users and connections. Secondly, DMID currently encounters problems while the networks are large scale in the range of millions of users and items which may put a heavy challenge and thus we may be required to implement it in parallel or improve the runtime.

202

M. Shahriari et al.

Additionally, with the help of expert identification task, it is possible to assess the goodness of community structures statistically. However, user satisfaction in an (expert) recommender system is way more complicated. Although the ranking algorithms bridged a connection between communities and experts and the community structure improved the MRR and MAP values, we need to observe if users really will be satisfied by those experts. In fact, experts might be contacted via the users, but returning top list can be overwhelming at some point. Considering task management and automatic extraction of expert properties might be required to be recognized for the future work. Regarding the community prediction perspective, we could not observe stable results regarding the use of different algorithms. The only feature that could be observed for all the algorithms and all the events was the size of communities. This issue prompts us to think of more stable approaches to measure the predicting properties of community features. Applying contextual features besides to the structural features can better predict the community structures. Furthermore, we are interested in figuring out how the communities are connected together via overlapping members and how these community relations evolve. Although we evaluated the algorithms with modularity measure, our offside experiments showed that assigning a single node to a community may generate as well a high value of modularity. Hence, we may ruminate about the versatility of evaluation metrics. Last but not least, the invention of GPS functionalities and location-based social networks provide the location of users and more enriched sources of information about individuals. Tracking the individuals and mapping their activities and intentions may even help us to mine their future intentions and thus even better predict the trajectories of community paths. One interesting direction may be to figure out the correlation between the contextual data generation of users and their community structures. In other words, there may exist meaningful patterns between the location of users and their community belonging. Altogether combining the location of users and their communities can even enhance the goodness of recommender systems.

10 Conclusion Tremendous production of data by social organizations has put a big challenge to data science field. Nowadays, instead of individuals, communities and their reputations gain more attention to people worldwide. While you are hearing about a great invention or an excellent software development, the first thing that pops up in mind is the concept of groups or communities; what group did it? With the already increased branch of community mining, many challenges exist, regarding the detection and prediction of communities, that give researchers enough motivations to initiate further work in this direction. Common definitions and suitable algorithms for OCD are necessary. Besides, questions like how to prolong longevity of certain communities and how to manipulate other communities to even die sooner are not clear in certain domains. To accost the problem equipped with simplicity and suitable social dynamic, we propose a two-phase leader-based approach, which

Dynamics of Overlapping Community Structures with Application to Expert. . .

203

owns DMID as its basic dynamic. We also connected the community detection problem with ranking algorithms to categorize the experts in question and answer forums. The idea behind the combination is that people inside a shared community are more familiar with each other than those that are not in the same community. Next, we applied logistic regression as the classifier with community and leader features to predict the events including survive, split, merge, and dissolve. OCD algorithms were tested on real-world and synthetic networks. Experimental results indicate the goodness of DMID algorithm in comparison to its competitors. Moreover, precise and relevant list of experts were improved by exploiting the community structure information. Finally, our experiments reveal features suitable for each event happening to the community. Also, prediction accuracy to predict the events of the community demonstrated that DMID could achieve higher accuracies in comparison to its competitors. Acknowledgements The work has received funding from the European Commission’s FP7 IP Learning Layers under grant agreement no 318209. We would also like to extend our thanks to Sebastian Krott, Sathvik Parekodi, Stephen Gunashekar, and Marven von Domarus, who helped us to accomplish this research work.

References 1. Ahn, Y.-Y., Bagrow, J.P., Lehmann, S.: Link communities reveal multiscale complexity in networks. Nature 466(7307), 761–764 (2010). http://dx.doi.org/10.1038/nature09182. http:// www.nature.com/nature/journal/v466/n7307/full/nature09182.html 2. Alvari, H., Hashemi, S., Hamzeh, A.: Detecting overlapping communities in social networks by game theory and structural equivalence concept. In: Proceedings of the Third International Conference on Artificial Intelligence and Computational Intelligence, pp. 620–630. Springer, New York (2011). ISBN 978-3-642-23886-4. http://dl.acm.org/citation.cfm?id=2045820. 2045910 3. Alvari, H., Hajibagheri, A., Sukthankar, G.: Community detection in dynamic social networks: a game-theoretic approach. In: 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining: ASONAM 2014, pp. 101–107 (2014). ISBN 978-14799-5877-1 4. Aston, N., Hu, W.: Community detection in dynamic social networks. Commun. Netw. 6, 124–136 (2014). ISSN 1949-2421. http://dx.doi.org/10.4236/cn.2014.62015 5. Backstrom, L., Huttenlocher, D., Kleinberg, J.M., Lan, X.: Group formation in large social networks: membership, growth, and evolution. In: Ungar, L., Craven, M., Gunopulos, D., Eliassi-Rad, T. (eds.) Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD ’06, pp. 44–54. ACM Press, New York (2006). ISBN 1595933395. http://doi.acm.org/10.1145/1150402.1150412 6. Baek, S.C., Kang, S., Noh, H, Kim, S.W.: Contents-based analysis of community formation and evolution in blogspace. In: 2009 IEEE 25th International Conference on Data Engineering, pp. 1607–1610 (2009). http://dx.doi.org/10.1109/ICDE.2009.159. 7. Balog, K., de Rijke, M., Weerkamp, W.: Bloggers as experts: feed distillation using expert retrieval models. In: Proceedings of the 31st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 753–754 (2008) 8. Bhat, S., Abulaish, M.: HOCTracker: tracking the evolution of hierarchical and overlapping communities in dynamic social networks. IEEE Trans. Knowl. Data Eng. 27, 1019 (2014) ISSN 1041–4347. http://dx.doi.org/10.1109/TKDE.2014.2349918

204

M. Shahriari et al.

9. Bishop, C.M.: Pattern Recognition and Machine Learning (Information Science and Statistics). Springer, New York, Secaucus, NJ (2006). ISBN 0387310738 10. Bouguessa, M., Dumoulin, B., Wang, S.: Identifying authoritative actors in question-answering forums: the case of Yahoo! Answers. In: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’08, pp. 866–874, New York, NY. ACM, New York (2008). ISBN 978-1-60558-193-4. http://doi.acm.org/10.1145/1401890. 1401994 11. Bozzon, A., Brambilla, M., Ceri, S., Silvestri, M., Vesci, G.: Choosing the right crowd: expert finding in social networks. In: Paton, N.W., Guerrini, G. (eds.) The 16th International Conference on Extending Database Technology, pp. 637–648, New York. ACM, New York (2013). ISBN 978-1-4503-1597-5. http://dx.doi.org/10.1145/2452376.2452451 12. Bródka, P., Kazienko, P., Kołoszczyk, B.: Predicting group evolution in the social network. In: Aberer, K., Flache, A., Jager, W., Liu, L., Tang, J., Guéret, C. (eds.) Social Informatics. Lecture Notes in Computer Science, vol. 7710, pp. 54–67. Springer, Berlin (2012). ISBN 978-3-642-35385-7. http://dx.doi.org/10.1007/978-3-642-35386-4_5 13. Bródka, P., Saganowski, S., Kazienko, P.: GED: the method for group evolution discovery in social networks. Soc. Netw. Anal. Min. 3(1), 1–14 (2013). ISSN 1869-5450. http://dx.doi.org/ 10.1007/s13278-012-0058-8 14. Cai, B., Wang, H., Zheng, H.: An improved random walk based clustering algorithm for community detection in complex networks. In: 2011 IEEE International Conference on Systems, Man and Cybernetics - SMC, pp. 2162–2167 (2011) 15. Cazabet, R., Amblard, F., Hanachi, C.: Detection of overlapping communities in dynamical social networks. In: 2010 IEEE Second International Conference on Social Computing (SocialCom), pp. 309–314 (2010). http://dx.doi.org/10.1109/SocialCom.2010.51 16. Chen, W., Liu, Z., Sun, X., Wang, Y.: A game-theoretic framework to identify overlapping communities in social networks. Data Min. Knowl. Disc. 21(2), 224–240 (2010) 17. Chen, Z., Wilson, K.A., Jin, Y., Hendrix, W., Samatova, N.F.: Detecting and tracking community dynamics in evolutionary networks. In: 2010 IEEE International Conference on Data Mining Workshops (ICDMW), pp. 318–327 (2010). http://dx.doi.org/10.1109/ICDMW. 2010.32 18. Chen, Y., Vikas, K., Rahul, U.: Detecting overlapping temporal community structure in timeevolving networks. eprint: arXiv:1303.7226, abs/1303.7226 (2013). http://dblp.uni-trier.de/ rec/bib/journals/corr/abs-1303-7226 19. Clauset, A., Newman, M.E.J., Moore, C.: Finding community structure in very large networks. Phys. Rev. E 70(6), 066111 (2004) 20. Cuzzocrea, A., Folino, F.: Community evolution detection in time-evolving information networks: proceedings of the joint EDBT/ICDT 2013 workshops. In: Proceedings of the Joint EDBT/ICDT 2013 Workshops (2013). http://doi.acm.org/10.1145/2457317.2457333 21. Danon, L., Duch, J., Diaz-Guilera, A., Arenas, A.: Comparing community structure identification. J. Stat. Mech Theory Exp. P09008 (2005). ISSN 1742-5468. http://dx.doi.org/10.1088/ 1742-5468/2005/09/P09008. http://arxiv.org/abs/cond-mat/0505245 22. Deng, H., King, I., Lyu, M.R.: Formal models for expert finding on DBLP bibliography data. In: Eight IEEE International Conference on Data Mining: ICDM’08, pp. 163–172 (2008) 23. Estrada, E., Hatano, N.: Communicability graph and community structures in complex networks. Appl. Math. Comput. 214(2),500–511 (2009). http://dx.doi.org/10.1016/j.amc.2009. 04.024 24. Evans, T.S.: Clique graphs and overlapping communities. J. Stat. Mech Theory Exp. 2010(12), (2010). ISSN 1742-5468. http://dx.doi.org/10.1088/1742-5468/2010/12/P12037 25. Evans, T.S., Lambiotte, R.: Line graphs of weighted networks for overlapping communities. Eur. Phys. J. B 77(2), 265–272 (2010). http://dx.doi.org/10.1140/epjb/e2010-00261-8 26. Fan, M., Wong, K.-C., Ryu, T., Ravasi, T., Gao, X.: Secom: a novel hash seed and community detection based-approach for genome-scale protein domain identification. PLoS one 7(6) (2012). http://dx.doi.org/10.1371/journal.pone.0039475

Dynamics of Overlapping Community Structures with Application to Expert. . .

205

27. Fu, X.H., Wang, C., Wang, Z.: Threshold random walkers for community structure detection in complex networks. J. Softw. 8(2), 286–295 (2013) 28. Girvan, M., Newman, M.E.J.: Community structure in social and biological networks. Proc. Natl. Acad. Sci. 99(12),7821–7826 (2002). http://dx.doi.org/10.1073/pnas.122653799 29. Goldberg, M., Magdon-Ismail, M., Nambirajan, S., Thompson, J.: Tracking and predicting evolution of social communities, pp. 780–783 (2011). http://dx.doi.org/10.1109/PASSAT/ SocialCom.2011.102. 30. Greene, D., Doyle, D., Cunningham, P.: Tracking the evolution of communities in dynamic social networks. In: 2012 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, pp. 176–183 (2010) 31. Gregory, S.: Finding overlapping communities in networks by label propagation. New J. Phys. 12(10), 1–21 (2010). http://dx.doi.org/10.1088/1367-2630/12/10/103018. 32. Havemann, F., Heinz, M., Struck, A., Gläser, J.: Identification of overlapping communities and their hierarchy by locally calculating community-changing resolution levels. J. Stat. Mech Theory Exp. (2011). ISSN 1742-5468. http://dx.doi.org/10.1088/1742-5468/2011/01/P01023 http://arxiv.org/abs/1012.1269 33. Hopcroft, J., Khan, O., Kulis, B., Selman, B.: Tracking evolving communities in large linked networks: proceedings of the national academy of sciences (2004). http://www.pnas.org/cgi/ content/full/101/suppl_1/5249 34. Jiang, J.Q., McQuay, L.J.: Modularity functions maximization with nonnegative relaxation facilitates community detection in networks. Physica A 391(3), 854–865 (2012). http://dx.doi. org/10.1016/j.physa.2011.08.043 35. Jin, D., Yang, B., Baquero, C., Liu, D., He, D., Liu, J.: Markov random walk under constraint for discovering overlapping communities in complex networks. CoRR, abs/1303.5675 (2013). http://doi.acm.org/10.1088/1742-5468/2011/05/P05031 36. Kairam, S.R., Wang, D.J., Leskovec, J.: The life and death of online groups: predicting group growth and longevity. In: Adar, E. (ed.) Proceedings of the Fifth ACM International Conference on Web Search and Data Mining, New York, NY, pp. 673–682. ACM, New York (2012). ISBN 1450307477. http://doi.acm.org/10.1145/2124295.2124374 37. Klamma, R.: Community learning analytics - challenges and opportunities. In: Wang, J.-F., Lau, R.W.H. (eds.) Advances in Web-Based Learning: ICWL 2013. Lecture Notes in Computer Science, vol. 8167, pp. 284–293. Springer, Berlin (2013) ISBN 978-3-642-41174-8. http://dx. doi.org/10.1007/978-3-642-41175-5_29 38. Kumpula, J.M., Kivela, M., Kaski, K., Saramaki, J.: Sequential algorithm for fast clique percolation. Phys. Rev. E 78(2), 026109 (2008). http://dx.doi.org/10.1103/PhysRevE.78. 026109 39. Lancichinetti, A., Fortunato, S.: Community detection algorithms: a comparative analysis. Phys. Rev. E 80(5) (2009). http://dx.doi.org/10.1103/PhysRevE.80.056117 40. Lehmann, S., Schwartz, M., Hansen, L.K.: Biclique communities. Phys. Rev. 78(1) (2008). http://dx.doi.org/10.1103/PhysRevE.78.016108 41. Leskovec, J., Kleinberg, J.M., Faloutsos, C.: Graph evolution: densification and shrinking diameters. ACM Trans. Knowl. Discov. Data 1(1) (2007). ISSN 1556-4681. http://dx.doi.org/ 10.1145/1217299.1217301 42. Li, H.J., Zhang, J., Liu, Z.P., Chen, L., Zhang, X.S.: Identifying overlapping communities in social networks using multi-scale local information expansion. Eur. Phys. J. B 85(6), (2012). http://dx.doi.org/10.1140/epjb/e2012-30015-5. 43. Li, H.-J., Wang, Y., Wu, L.-Y., Liu, Z.-P., Chen, L., Zhang, X.-S.: Community structure detection based on Potts model and network’s spectral characterization. EPL 97(4), 48005 (2012). http://dx.doi.org/10.1209/0295-5075/97/48005 44. Lin, Z., Jiang, X., Zhao, B., Qu, Y.: A simple but fast agglomerative strategy to detect community structure based on vertex closeness. Int. J. Innov. Comput. Inf. Control 7(9), 5163–5174 (2011)

206

M. Shahriari et al.

45. Liu, X., Zhou, Y., Hu, C., Guan, X., Leng, J.: Detecting community structure for undirected big graphs based on random walks. In: Proceedings of the 23rd International Conference on World Wide Web: Proceedings of the 23rd WWW Conference, WWW ’14, pp. 1151–1156, Republic and Canton of Geneva, 2014. International World Wide Web Conferences Steering Committee. ISBN 978-1-4503-2744-2. http://dx.doi.org/10.1145/2567948.2580060 46. Ma, H.S., Huang, J.W.: CUT: community update and tracking in dynamic social networks. In: Proceedings of the 7th Workshop on Social Network Mining and Analysis (2013). http://dl. acm.org/citation.cfm?id=2501026 47. Macdonald, C., Ounis, I.: Voting for candidates: adapting data fusion techniques for an expert search task. In: Proceedings of the 15th ACM International Conference on Information and Knowledge Management, pp. 387–396 (2006) 48. Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N., Chklovskii, D., Alon, U.: Network motifs: simple building blocks of complex networks. Science 298(5594), 824–827 (2002). http://dx. doi.org/10.1126/science.298.5594.824. 49. Moradi, F., Olovsson, T., Tsigas, P.: Overlapping communities for identifying misbehavior in network communications. In: Tseng, V.S., Ho, T., Zhou, Z.-H., Chen, A.L.P., Kao, H.-Y. (eds.) Advances in Knowledge Discovery and Data Mining. Lecture Notes in Computer Science, vol. 8443, pp. 398–409. Springer International Publishing, Berlin (2014). ISBN 978-3-319-066073. http://dx.doi.org/10.1007/978-3-319-06608-0_33 50. Nguyen, N.P., Dinh, T.N., Tokala, S., Thai, M.T.: Overlapping communities in dynamic networks: their detection and mobile applications. MobiCom, pp. 85–96 (2011). http://doi. acm.org/10.1145/2030613.2030624 51. Nicosia, V., Mangioni, G., Carchiolo, V., Malgeri, M.: Extending the definition of modularity to directed graphs with overlapping communities. J. Stat. Mech Theory Exp. 2009(03), P03024 (2009). ISSN 1742-5468. http://dx.doi.org/10.1088/1742-5468/2009/03/P03024 52. Palla, G., Derényi, I., Farkas, I., Vicsek, T.: Uncovering the overlapping community structure of complex networks in nature and society. Nature 435(7043), 814–818 (2005). http://dx.doi. org/10.1038/nature03607. 53. Palla, G., Barabási, A.-L., Vicsek, T.: Community dynamics in social networks. Fluctuation Noise Lett. 07(03), L273–L287 (2007). http://dx.doi.org/10.1142/S0219477507003933 54. Palla, G., Pollner, P., Barabási, A.-L., Vicsek, T.: Social group dynamics in networks. In: Groß, T., Sayama, H. (eds.) Adaptive Networks: NECSI, New England Complex Systems Institute Book Series, pp. 11–38. Springer, Berlin (2009). ISBN 978-3-642-01283-9. http://dx.doi.org/ 10.1007/978-3-642-01284-6_2 55. Pons, P., Latapy, M.: Computing communities in large networks using random walks. J. Graph Algorithms Appl. 10,284–293 (2004) 56. Psorakis, I., Roberts, S., Ebden, M., Sheldon, B.: Overlapping community detection using Bayesian non-negative matrix factorization. Phys. Rev. E 83(6), 066144 (2011). http://dx.doi. org/10.1103/PhysRevE.83.066114 57. Raghavan, U.N., Albert, R., Kumara, S.: Near linear time algorithm to detect community structures in large-scale networks. Phys. Rev. E 76, 036106 (2007). http://dx.doi.org/10.1103/ PhysRevE.76.036106 58. Reichling, T., Wulf, V.: Expert recommender systems in practice: Evaluating semi-automatic profile generation. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 59–68 (2009). ISBN 978-1-60558-246-7. http://doi.acm.org/10.1145/1518701. 1518712 59. Richardson, L., Ruby, S.: RESTful web services. O’Reilly, Sebastopol, CA, USA, 1 edition, 2007. ISBN 9780596529260. http://proquestcombo.safaribooksonline.com/9780596529260 60. Rizova, P.: Are you networked for successful innovation? MIT Sloan Manag. Rev. 47(3), 49–55 (2006) 61. Rosvall, M., Bergstrom, C.T.: Maps of random walks on complex networks reveal community structure. Proc. Natl. Acad. Sci. 105(4), 1118–1123 (2008). http://dx.doi.org/10.1073/pnas. 0706851105

Dynamics of Overlapping Community Structures with Application to Expert. . .

207

62. Ruan, Y., Fuhry, D., Parthasarathy, S.: Efficient community detection in large networks using content and links. CoRR, abs/1212.0146 (2012) 63. Savic, M., Radovanovic, M., Lvanovic, M.: Community detection and analysis of community evolution in Apache ant class collaboration networks. In: Proceeding of the Fifth Balkan Conference in Informatics, pp. 229–234 (2012). http://doi.acm.org/10.1145/2371316.2371361 64. Sekara, V., Stopczynski, A., Lehmann, S.: The fundamental structures of dynamic social networks. Proc. Natl. Acad. Sci. USA 36(113), 77–82 (2016) http://dx.doi.org/10.1073/pnas. 1602803113. 65. Shahriari, M., Klamma, R.: Signed social networks: Link prediction and overlapping community detection. In: The 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), Paris (2015). http://dx.doi.org/10.1145/ 2808797.2810250 66. Shahriari, M., Krott, S., Klamma, R.: Disassortative degree mixing and information diffusion for overlapping community detection in social networks (DMID). In: Proceedings of the 24th International Conference on World Wide Web, WWW ’15 Companion, pp. 1369–1374 (2015). ISBN 978-1-4503-3473-0. http://dx.doi.org/10.1145/2740908.2741696 67. Shahriari, M., Krott, S., Klamma, R.: WebOCD: a restful web-based overlapping community detection framework. In: Proceedings of the 15th International Conference on Knowledge Technologies and Data-driven Business, i-KNOW ’15. ACM, New York (2015). ISBN 978-14503-3721-2 68. Shahriari, M., Parekodi, S., Klamma, R.: Community-aware ranking algorithms for expert identification in question-answer forums. In: Proceedings of the 15th International Conference on Knowledge Technologies and Data-driven Business, i-KNOW ’15, pp. 8:1–8:8. ACM, New York (2015). ISBN 978-1-4503-3721-2. http://doi.acm.org/10.1145/2809563.2809592 69. Shahriari, M., Gunashekar, S., von Domarus, M., Klamma, R.: Predictive analysis of temporal and overlapping community structures in social media. In: Proceedings of the 25th International Conference Companion on World Wide Web, WWW ’16 Companion, pages 855–860, Republic and Canton of Geneva, Switzerland, 2016. International World Wide Web Conferences Steering Committee. ISBN 978-1-4503-4144-8. http://dx.doi.org/10.1145/ 2872518.2889292 70. Shahriary, S.R., Shahriari, M., Noor, R.M.D.: A community-based approach for link prediction in signed social networks. Sci. Program. J. (2014). http://downloads.hindawi.com/journals/sp/ aip/602690.pdf 71. Shen, H., Cheng, X., Cai, K., Hu, M.-B.: Detect overlapping and hierarchical community structure in networks. Physica A 388(8), 1706–1712 (2009). http://dx.doi.org/10.1016/j.physa. 2008.12.021. 72. Stanoev, A., Smilkov, D., Kocarev, L.: Identifying communities by influence dynamics in social networks. Phys. Rev. 84(4), (2011) http://dx.doi.org/10.1103/PhysRevE.84.046102 73. Takaffoli, M., Rabbany, R., Zaiane, O.R.: Community evolution prediction in dynamic social networks. In: 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining: ASONAM 2014, pp. 9–16 (2014). ISBN 978-1-4799-5877-1. http://dx. doi.org/10.1109/ASONAM.2014.6921553 74. Torghabeh, R.P., Narayana, S.P.: Community detection using slow mixing Markov models. CoRR, abs/1510.02583 (2015). http://arxiv.org/abs/1510.02583 75. Wang, Q., Fleury, E.: Fuzziness and overlapping communities in large-scale networks. J. Universal Comput. Sci. 18(4), 457–486 (2012) 76. Wang, Y., Wu, B., Pei, X.: Commtracker: a core-based algorithm of tracking community evolution. J. Front. Comput. Sci. Technol. 3,282–292 (2009). http://dx.doi.org/10.3778/j.issn. 1673-9418.2009.03.006. http://fcst.ceaj.org/EN/abstract/article_105.shtml. 77. Worsley, M., Blikstein, P.: What’s an expert? using learning analytics to identify emergent markers of expertise through automated speech, sentiment and sketch analysis. In: Pechenizkiy, M., Calders, T., Conati, C., Ventura, S., Romero, C., Stamper, J. (eds.) 4th International Conference on Educational Data Mining, pp. 235–240 (2011). ISBN 978-90-386-2537-9

208

M. Shahriari et al.

78. Xie, J., Szymanski, B.K., Liu, X.: SLPA: uncovering overlapping communities in social networks via a speaker-listener interaction dynamic process. CoRR. abs/1109.5720 (2011) 79. Xie, J., Kelley, S., Szymanski, B.K.: Overlapping community detection in networks. ACM Comput. Surv. 45(4) (2013). ISSN 0360-0300. http://dx.doi.org/10. 1145/2501654.2501657. http://dl.acm.org/citation.cfm?id=2501654.2501657&coll=DL&dl= GUIDE&CFID=280963611&CFTOKEN=70191141 80. Yang, J., McAuley, J., Leskovec, J.: Community detection in networks with node attributes. In: 2013 IEEE 13th International Conference on Data Mining, pp. 1151–1156 (2013) 81. Yang, L., Qiu, M., Gottipati, S., Zhu, F., Jiang, J., Sun, H., Chen, Z.: CQARank: jointly model topics and expertise in community question answering. In: Proceedings of the 22nd International Conference on Information & Knowledge Management, pp. 99–108 (2013). http://doi.acm.org/10.1145/2505515.2505720 82. Zhou, G., Lai, S., Liu, K., Zhao, J.: Topic-sensitive probabilistic model for expert finding in question answer communities. In: Proceedings of the 21st ACM International Conference on Information and Knowledge Management, CIKM ’12, pp. 1662–1666, New York, NY. ACM, New York (2012). ISBN 978-1-4503-1156-4. http://doi.acm.org/10.1145/2396761.2398493 83. Zhu, H., Cao, H., Xiong, H., Chen, E., Tian, J.: Towards expert finding by leveraging relevant categories in authority ranking. In: Proceedings of the 20th ACM International Conference on Information and Knowledge Management, CIKM ’11, pp. 2221–2224, New York, NY. ACM, New York (2011). ISBN 978-1-4503-0717-8. http://doi.acm.org/10.1145/2063576.2063931 84. Zhu, H., Chen, E., Xiong, H., Cao, H., Tian, J.: Ranking user authority with relevant knowledge categories for expert finding. World Wide Web J. 17(5), 1081–1107 (2014) ISSN 1386-145X. http://dx.doi.org/10.1007/s11280-013-0217-5

On Dynamic Topic Models for Mining Social Media Shatha Jaradat and Mihhail Matskin

Abstract Analyzing media in real time is of great importance with social media platforms at the epicenter of crunching, digesting, and disseminating content to individuals connected to these platforms. Within this context, topic models, specially latent Dirichlet allocation (LDA), have gained strong momentum due to their scalability, inference power, and their compact semantics. Although, state-ofthe-art topic models come short in handling streaming large chunks of data arriving dynamically onto the platform, thus hindering their quality of interpretation as well as their adaptability to information overload. In this manuscript (Jaradat et al. OLLDA: a supervised and dynamic topic mining framework in twitter. In: 2015 IEEE international conference on data mining workshop (ICDMW), November 2015. IEEE, Piscataway, pp. 1354–1359), we evaluate a labeled and online extension to LDA (OLLDA), which incorporates supervision through external labeling and capability of quickly digesting real-time updates thus making it more adaptive to Twitter and platforms alike. Our proposed extension has capability of handling large quantities of newly arrived documents in a stream, and at the same time, is capable of achieving high topic inference quality given the short and often sloppy text of tweets. Our approach mainly uses an approximate inference technique based on variational inference coupled with a labeled LDA (L-LDA) model. We conclude by presenting experiments using a 1-year crawl of Twitter data that shows significantly improved topical inference as well as temporal user profile classification when compared to state-of-the-art baselines. Given the popularity of words’ prediction techniques such as Word2vec, we present an additional benchmark to measure the performance of classification.

S. Jaradat · M. Matskin () Royal Institute of Technology (KTH), Stockholm, Sweden e-mail: [email protected]; [email protected] © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_8

209

210

S. Jaradat and M. Matskin

1 Introduction Topic models have gained strong momentum in recent years for their performance and accuracy in analyzing social media data [30]. This has been most visible in Twitter research with applications on topic classification and user recommendation [4]. Topic modeling is mainly concerned with finding and analyzing the latent relationships between collections of documents in order to perform tasks such as data exploration and prediction. To achieve this, the posterior distribution of the model parameters and the latent variables must be approximated because it is intractable to compute them. Markov chain Monte-Carlo (MCMC) sampling techniques and variational inference are the two major approaches used for the approximation of posterior inference [10, 16]. Latent Dirichlet allocation (LDA) [16] is a generative topic model, that has gained attention for its application to mainly supervised topic modeling in social media analytics. There are a number of shortcomings to LDA, specially when used on Twitter data. LDA is an unsupervised algorithm that models each document as a mixture of topics that can be hard to interpret. This makes it difficult to be directly applied to multi-labeled corpora. A remedy to this problem is applying supervision. In order to achieve this, several modifications have been proposed. One of the most attended approaches is using labeled LDA (L-LDA) [28]. L-LDA has shown to be a better option than compared to supervised methods such as supervised LDA [5] and DiscLDA [32], as it does not have the constraint of having only a single label associated with each document [28]. However, L-LDA uses Gibbs sampling as an approximate inference technique, which is one of the MCMC sampling techniques. As such, Gibbs sampling turns out to be less effective than variational inference. In addition to supervision, another shortcoming of LDA is batch execution, which turns into the streaming nature of content being published on Twitter. To deal with this shortcoming, several online algorithms have been proposed for LDA. One of the most efficient ones is the online VB for LDA [10]. It can be used to analyze massive collections, including streaming data, with no need for dealing with ephemeral state of documents after they have been processed. However, existing implementation of this approach is also unsupervised, which can have impact on quality of distilled topics. Variational inference methods can be described as a set of deterministic algorithms that transform the inference problem into an optimization problem. This is done by choosing a simplified distribution that uses a set of free parameters. By solving these parameters and performing optimization steps, the new distribution will be close to the original posterior distribution. It has been shown that this approach is faster than MCMC methods, which calculates the posterior distribution by generating independent samples from the posterior. Existing research have shown that MCMC algorithms can be slow to converge and it is difficult to check their convergence [2, 10, 12, 15]. Word2Vec [25] is one of the recent distributed word representation architectures that showed efficiency in capturing semantic regularities between words. In an

On Dynamic Topic Models for Mining Social Media

211

extension to our work in [13], and given the popularity of Word2vec as well as the relevance to our work, we examine the semantic similarities between words detected per topics in our approach. We also analyze the distributions of words on clusters generated after training the corpus using Word2vec. In our work, we propose for a scalable and online algorithm that can produce high-quality and interpretable topics. Thus, our contributions are: (1) coupling an online VB for LDA [10] with an L-LDA model [28] mainly to improve performance, and (2) using variational inference as an approximate inference technique, which will be enhanced through this work by including supervised learning. The rest of the paper is organized as follows. Section 2 provides a background study on the related research. In Sect. 3, the proposed algorithm is illustrated in detail. There also, relation to the original online VB for LDA and L-LDA algorithms has been shown. Section 4 presents the results of experimental comparisons, accompanied by an analysis of the algorithm’s performance and quality. Section 5 introduces some possible applications of the new algorithm. Finally, the conclusions and the expected future work are presented in Sect. 6.

2 Background In this section, we will briefly present an overview of the previous work related to LDA extensions and topic modeling. This overview is divided into five main subsections: scalable versions of LDA, proposed solutions for adapting LDA to streaming scenarios, supervised extensions of this algorithm, embeddings factorization techniques and their integration with topic modeling, and some relevant applications of LDA extensions in Twitter context.

2.1 Online Learning in Topic Models Some researchers focused on speeding up and offering more scalable versions of LDA using parallelized architectures. The parallelization is achieved either on loosely coupled distributed computers or on tightly coupled multicore CPUs on the same node [19, 21, 29, 34]. Given the importance of dealing with large and streaming data specially in social media analysis, we focus on works that customize topic modeling for streaming scenarios. We hereby focus only on LDA invariants. Domeniconi et al. [9] presents an empirical Bayesian method to incrementally update the model while receiving streaming updates. Suggested by Benzi et al. [3], TM-LDA is an LDA invariant that learns the transition parameters among topics, and uses the learnt transition parameters in the topics distribution prediction task. In the same line, OLDA was proposed by Domeniconi et al. [9], which identifies the emerging topics of text streams and their changes over time. Both [3, 9] are concerned with detecting the patterns in the social behavior. Closely related to our

212

S. Jaradat and M. Matskin

work, there are existing proposals for online versions for LDA [33, 35]. Both works use Gibbs sampling as the approximate inference distribution method. The work introduced in [35] considered running the sampler on new documents only on the new dataset with each update, whereas [33] proposes incremental Gibbs sampling and particle filters for online inference.

2.2 Supervised Learning in Topic Models Existing LDA invariants that incorporate supervision in their process are L-LDA [28], supervised LDA [5], DiscLDA [32]. Nallapati et al. [28] surveys these algorithms and states that supervised LDA and DiscLDA both have limitation of associating a document with a single label (topic) only. We argue that this is not suitable for Twitter where each batch of tweets would need multiple labels to be meaningful. In a closely related project, TweetLDA, [31] proposed L-LDA modeling at the top of Twitter streams. This project highlighted the effectiveness of Labeled-LDA when compared with previous approaches.

2.3 Coupling Supervision and Online Learning in Topic Models Given the fact that we propose for an adaptation of supervision to online learning in social media, we focus on two works that influence this work namely L-LDA [28] and online VB for LDA [10]. L-LDA [28] incorporates supervision in LDA by constraining the topic model to use topics that correspond to the document observed labels. This has shown to be very effective since instead of assigning the document’s words according to their co-occurrence together to latent topics as done by original LDA, it assigns them to labels that have richer meaning, and are based on human classification. The second algorithm [10] is an online variational Bayes algorithm for LDA that can handle massive document collections as well as documents arriving in streams. This approach uses variational inference, as an approximate posterior inference algorithm, which is shown to be faster than MCMC [2, 10, 12, 15]. Such consideration makes it an attractive option for applying Bayesian models to large datasets [10]. In our work, we join the positive aspects of two algorithms by modifying the variational distribution used to approximate the posterior distribution in online VB for LDA algorithm, thus constraining the variational parameters to a specific set of meaningful labels. This in turn will influence the assignment of words to topics according to an interpretable classification mechanism.

On Dynamic Topic Models for Mining Social Media

213

2.4 Word Embeddings and Topic Modeling Word2vec [25] is a neural predictive model for learning word embeddings from raw text. It has shown efficiency in capturing semantic regularities between words [26]. This aids in better text classification due to the semantic vectors provided by Word2vec which preserves the relevant information about the text, yet with low dimensionality. The merge between Word2vec and other traditional topic modeling such as LDA can take advantage of both learning efficient word-level representations and applying it to produce interpretable document topics. LDA2vec [26] is an example of such merge, where the context vector examines not only the word vector but rather the document vector. This helps in producing more meaningful (long- and short-term) predictions that suit the context of the whole document. Other models such as Paragraph2vec and Doc2vec [18] handle variable length of texts (phrases, sentences, or documents) and learn continuous distributed representations of words. Examples of work that applied Word2vec for text clustering is [22]. In [20], Word2vec and TF-IDF models were combined, and their model outperformed TF-IDF, due to the semantic features that can be detected by Word2vec.

2.5 Applications of Latent Dirichlet Allocation in Twitter Among the works that focused on applying LDA in Twitter context are [11] and [23]. They combined LDA with concept mapping, where external sources of knowledge are employed to enhance the topic detection procedure. Authors in [17] incorporated word-pair connection to enhance the topics and produce more interpretable results. Many researchers have analyzed the challenges of hashtags’ clustering and classification in Twitter. In [27], they represent the hashtags by the concatenation of the tweets in which they appear. Then, they apply K-means clustering based on the concept of co-occurrence of terms and hashtags. Spectral clustering was applied in [1] after rounds of filtering to avoid noisy relationships between hashtags. In [14], they have adopted a semantic clustering approach, where they first lookup hashtags from WordNet, if no match found, they split the hashtag into multiple terms, and keep the lookup procedure until a match is found, or otherwise referred to search process from Wikipedia.

3 Latent Dirichlet Allocation Models As stated previously, within this section we give a brief and technical overview of the original LDA model and some extended LDA models: online VB for LDA, L-LDA, and combined online and supervised LDA model, given the context at hand. Table 1 summarizes all the symbols that will be used throughout the text.

214

S. Jaradat and M. Matskin

Table 1 List of symbols Symbol d θ β V zdn wdn D N M K γ φ λ q p Λ T

Meaning Document—sequence of N words Topics’ distribution over documents Words’ distribution over topics Vocabulary used in the model Topics assignments for the word w with index n in d Observed word from document d with index n Corpus—a collection of M documents Number of words in each document Number of documents per corpus List of all topics in the corpus Variational parameter mapped to θ Variational parameter mapped to z Variational parameter mapped to β Approximated variational distribution Posterior distribution of LDA List of labels assigned to d Number of topics assigned to each document

The smoothed LDA model assumes availability of K topics, where each topic is defined as a multinomial distribution over the vocabulary, drawn from a Dirichlet, βk ∼ Dirichlet (η) distribution. For each document, a distribution over topics θd is drawn from Dirichlet (α) where α is a hyper-parameter. Then, for each word i in the document, a topic index zdi ∈ {1, . . . , K} is drawn from the topic weights zdi ∼ θ , and the observed word wdi is drawn from the selected topic wdi ∼ βzdi [16]. Figure 1a shows the graphical representation of the original model of LDA [16]. The posterior distribution of the hidden variables is as follows: p(θ, z|w, α, β) = p(θ,z,w|α,β) p(w|α,β) . This distribution can be approximated using different techniques including variational inference. When variational inference is applied to the approximation of posterior distribution, the LDA model can be transformed into a simpler model as shown in Fig. 1b. Let us assume that this new variational model is named q. In such simplified model, the latent variables θ and β are decoupled by dropping the node w, and removing the edges between z, w, and θ . Then, the free variational parameters γ , φ, and λ are added to the new model to approximate θ , z, and β, respectively. With the new simplified distribution q(z, θ, β), the goal eventually becomes finding the settings of the variational parameters λ, γ , and φ that make q close to true posterior distribution p. Thus, optimization problem can be solved by minimizing the Kullback–Leibler (KL) divergence between the true posterior p and the variational posterior q. The KL divergence cannot be exactly minimized. However, it is possible to maximize an objective function that is equal up to a constant; the ELBO (evidence lower bound) on the log-likelihood, which shall

On Dynamic Topic Models for Mining Social Media

a

η

␤

θ

z

k

g

f

λ

θ

z

␤

215

η

a θ w

N

N

M

(a)

(b)

M

␤

z

k

w

f

N

g

f

θ

z

␤

N

M

(c)

λ

M

(d)

Fig. 1 Graphical representations for (a) latent Dirichlet allocation (LDA) model—source [16], (b) variational distribution used to approximate the posterior distribution in LDA—source [16], (c) labeled LDA (L-LDA) model—source [28], and (d) modified variational distribution used to approximate the posterior distribution in online labeled LDA (OLLDA)

pertain the same effect as minimizing the KL divergence. To achieve this, Jensen inequality is often used to obtain an adjustable lower bound on the log likelihood [10, 16]. This is followed by an iterative method, such as variational EM algorithm to iteratively optimize each variational distribution holding the other parameters fixed. In the expectation step, γ and φ will be iteratively updated while λ fixed. In the maximization step, λ will be updated given φ. Consequently, those updates are guaranteed to converge to a stationary point of the ELBO [10, 16].

3.1 Online VB for Latent Dirichlet Allocation In online VB for LDA [10], the update step is the one in which λ is updated. The expectation step is executed for the newly observed tth vector of words (nt ). In this step, the values of γt and φt are calculated, while holding λ fixed. In this case, the entire corpus is assumed to be consisting of the newly observed document, repeated D times, where D is the number of documents in the batch. The value of λ˜ is calculated under such setting. λ˜ can be used to update λ in the maximization step, which is computed from the weighted average of its previous value, and the newly calculated value λ˜ . Step one in Algorithm 1 is used to give weight to λ˜ , where k ∈ (0.5, 1] controls the rate at which old values of λ˜ are forgotten. The formulas used to calculate Eq [log θtk ] and Eq [log βkw ] are given as follows: K W Eq [log θtk ] = (γdk ) − (i=1 γdi ); Eq [log βkw ] = (λkw ) − (i=1 λki )

where  is the digamma function.

216

S. Jaradat and M. Matskin

Algorithm 1 Online variational Bayes for LDA Define ρ = (τ + t)k Initialize λ randomly for t = 0 To ∞ do E step: Initialize γtk = 1 repeat: Set φtwk ∝ exp Eq [logθtk ] + Eq [logβkw ] Set γtk = α + w φtwk ηtw until K1 k |changeinγtk | < 0.00001 M step: Compute λ˜ kw = η + D ηtw φtwk Set λ = (1 - ρt ) λ + ρt λ˜ end for

3.2 Labeled LDA L-LDA, as visualized in Fig. 1c, incorporates supervision in LDA by constraining the topic model to use only the topics that correspond to a document’s observed label set. As it is seen from the figure, the topic distributions over documents θ do not depend only on the hyper-parameter α but also on Λ which is the set of observed labels. This affects the initial assignment of words to topics and topics to documents [28]. In this model, the number of topics can be dynamically assigned according to the number of labels in the corpus. The vocabulary can be composed from the words of the documents. Whereas in online VB for LDA [10], a fixed vocabulary from English dictionary is often used.

3.3 Online Labeled LDA Using concept of constraining, topic distributions per each document are curbed down to a set of observed labels. To do so, we have modified the variational distribution that is used to approximate the posterior distribution in the online labeled LDA (OLLDA) model as follows. We have the initial values of the variational parameters γ and λ affected by the set of topics observed for the document Λ, as shown in Fig. 1d. This in turn affects the probability of both assigning topics to documents and words to topics. Certain topics will receive higher probability of being assigned to documents than others during the learning process. This is different than the original LDA model that assigns words to topics only based on their ensemble co-occurrence. Also, original LDA model uses a fixed English vocabulary during the assignment process. However, on Twitter most tweets contain hashtags, abbreviations, and names of people which are not part of English vocabulary. This problem is addressed in our project by extracting words from tweets, and assigning them to topics, taking into consideration the prior topics’

On Dynamic Topic Models for Mining Social Media

217

distribution score which we have estimated through the classification service. We believe that this step helps us to achieve more accurate results. As shown later on, when we ran our algorithm with a fixed vocabulary from an English dictionary, the results were not as accurate as the dynamic vocabulary. Algorithm 2 presents the proposed algorithm. Assume that D is the number of documents in the corpus. T is the number of topics considered from the ground truth for each document, which is a subset from Λ. In our research, we considered the top five topics. The algorithm has been experimented to adopt top three and top five topics, and it has been observed that top five topics result in better topical inference output. In our proposed solution, SortedTopT is the set of top T sorted labels associated with each document and provided by the ground truth. At the same time, WordsRatio is the percentage of words allowed to be taken from each document. Each document is partitioned according to the associated topic’s weights, where Pdt represents the size of the partition in document d that will be assigned to topic t, and Pdts indicates the starting index of the partition, while Pdte is an indication of the ending index. To finalize, main steps of the algorithm can be summarized as follows: 1. 2. 3. 4.

Initialization of λ and γ with random values. Topics in corpus are decided from labels associated with documents. Vocabulary is built from the words of documents. For each document, a number of iterations are executed to add value for each top topic in the document, and to assign words to that topic, according to that topic’s weight. 5. Rest of the steps match the original algorithm. In steps 14 to 22, we use the topic’s weight that is provided by the classifier to decide the percentage of the document’s length that will be modified. We calculate the size of each partition, which is used to increase the importance of a certain topic in a document words belonging to a topic according to the following ' & and certain w rule: Pdt = |d|∗t , where Pdt is the size of partition per each topic t, d is the Wd document under study, and tw is the topic’s weight in the document. Wd is the total of all topics’ weights associated to document d. The value of a certain topic in a document, and certain words related to that topic will be incrementally updated during the learning process. This affects their final classification in the context of the whole corpus at the end of the process.

218

S. Jaradat and M. Matskin

Algorithm 2 Online labeled LDA (OLLDA) Define ρ = (τ + t)k Initialize λ Initialize γ for d = 0 to D do for t = 0 to T do if SortedT opT (t) ∈ / K then Add SortedT opT (t) to K end if end for for l = 0 to W ordsRatio do Add d(l) to V end for end for for d = 0 to D do for t = &0 to T do ' opTw Pdt = |d|∗SortedT Wd for i = Pdts to Pdte do λdk ++ γdkwi ++ end for end for end for Normalize data for t = 0 to ∞ do E step: Initialize γtk = 1 repeat: Set φtwk ∝ exp Eq [logθtk ] + Eq [logβkw ] Set γtk = α + w φtwk ηtw until K1 k |change in γtk | < 0.00001 M step: Compute λ˜ kw = η + D ηtw φtwk Set λ = (1 - ρt ) λ + ρt λ˜ end for

4 Experiments 4.1 Dataset Dataset used in this research was gathered by Dokoohaki et al. [8] using Twitter Streaming API1 from February 2014 until mid of October 2014. Some filters were applied while gathering the data such as location which was set to Sweden. The

1 https://dev.twitter.com/streaming/overview.

On Dynamic Topic Models for Mining Social Media

219

total number of tweets is around 7 million, which are mapped to 471,086 users. The chosen sample is 731,676 tweets corresponding to 3061 users. The sample size was reduced due to limitations in the number of allowed requests to consume the classification service, which will be described later. Due to focus on temporal variations of topics, dataset was partitioned. Multiple partitions were constructed for the users, including their tweets from March 2014 until end of September 2014. Each partition represents tweets of all those users during a specific month. The reason for this partitioning is the interest of this work in training the algorithms on the datasets to notice the transitions of topics from month to month. We were also interested in checking the ability of the new algorithm to detect topics in an accurate way, when compared to the ground truth. Within the course of the experiments the proposed algorithm is compared with the online VB for LDA as well as L-LDA. For the purpose of comparison, it was required to apply an update mechanism for L-LDA. According to [24], adding a certain percentage of documents that were used in a previous iteration to the new iteration satisfies the update step in L-LDA. Multiple experiments were conducted, as will be illustrated in the following sections. Experiments have been repeated for different values of the hyper-parameter alpha: 0.1, 0.25, 0.5, 1.0, 1.25, 1.5, 1.75, and 2.0. We observed that in general OLLDA achieves better results for different settings of alpha. Table 2 illustrates the sub-datasets with the number of tweets per each month.

4.2 Experimental Pipeline To adapt the algorithm to the Twitter data at hand, we have proposed for the pipeline visualized in Fig. 2. The goal of this procedure is to annotate tweets of each user with the appropriate labels produced by the classifier. To achieve this, tweets are pooled according to the author, following the “Author-wise pooling” concept described in [7]. Then, stop words are removed from those documents. For our experiment, a translator API was used to translate all the tweets from Swedish to English. For this purpose, Microsoft translator API2 was used, due to its competitive quality of translation. For labels’ assignment to user profiles, Textwise Classifier3 was used to provide the ground truth for the sample dataset. This classifier was chosen due to its accuracy in text classification, when compared to other classifiers, such

Table 2 Number of tweets per each month for the list of users in the sample Mar 44,625

April 83,875

May 159,971

June 109,373

2 https://www.microsoft.com/translator/api.aspx. 3 http://textwise.com/.

July 94,497

Aug 52,479

Sept 186,856

220

S. Jaradat and M. Matskin

Fig. 2 Steps performed before running OLLDA

as Alchemy API.4 For the purpose of this research, we have used an advanced classification service. We could have used some natural language processing tools such as TweetNLP5 for preprocessing steps and enhancing the analysis results.

4.3 Scalability Performance We ran OLLDA, L-LDA, and online VB for LDA on five different user samples containing, respectively: 500, 1000, 1500, 2000, and 3000 users. Each experiment was repeated for five different runs for each sample size. After several runs, we fixed the value of alpha onto 0.25. Figure 3 shows that our proposed algorithm OLLDA scales better than L-LDA. Albeit we noticed that online VB for LDA is faster than OLLDA to small extent. This was expected, as one of the objectives was to achieve more accurate inference, which we aimed to trade off with performance. However as shown, still our proposed approach stands out in the performance as compared to L-LDA.

4 http://www.alchemyapi.com/. 5 http://www.ark.cs.cmu.edu/TweetNLP/.

On Dynamic Topic Models for Mining Social Media OLDA

221

OLLDA

LLDA

2000000

TIME IN MILLISECONDS

1800000 1600000 1400000 1200000 1000000 800000 600000 400000 200000 0 500

1000

1500

2000

3000

NUMBER OF USERS

Fig. 3 Scalability experiment results measured in milliseconds show clearly that OLLDA runs faster than L-LDA. All experiments were executed on a standalone Xeon server with 4 core i7 processors and 16 GB of memory

4.4 Topical Inference Accuracy To compare the quality of topical inference, a number of experiments were customized in several settings for each month. To do so, we divided the data in each execution to a 20% for testing and 80% for validation. This setup was repeated for each consecutive month incrementally. As example, in the first iteration, each algorithm was trained on March partition, and subsequently April partition was used for testing and validation. During the second iteration, March and April were used for training, and the testing was done on May partition, etc. This setup was intentionally done to compare the accuracy of detecting topics of each month with the previous months. Another reason was to figure the transitions in the inferred topics from 1 month to another. After running OLLDA and L-LDA, the nonclassified profiles were compared with the ground truth (which is the data provided by the classifier), and the cosine similarity was estimated for each month. Figure 4 compares the percentage of similarity between the topics provided by the classifier, which reflects the user’s preferences and the detected topics by the algorithms. The results show that OLLDA has higher cosine similarity values when compared to L-LDA. Experiments were executed for different alphas, and in all results OLLDA was better than counterparts. Although L-LDA considers top topics for each profile to augment their chance of representing the document, it does not consider the order and the proportion of each topic for respective profile. In our

222

S. Jaradat and M. Matskin OLLDA

LLDA

0.8

Similarity Value

0.6

0.4

0.2

0.0

April

May

June

July

August

September

Month

Fig. 4 Comparison of cosine similarity estimates between OLLDA and LLDA, with the ground truth for all months Table 3 Precision values for different months

Table 4 Recall values for different months

OLLDA LLDA

April 0.995 0.986

May 0.992 0.977

June 1.000 0.982

July 0.992 0.987

Aug 0.992 0.992

Sept 0.997 0.923

OLLDA LLDA

April 0.202 0.199

May 0.201 0.198

June 0.203 0.199

July 0.201 0.200

Aug 0.201 0.201

Sept 0.202 0.187

approach, we consider those factors while incrementing the value of top topics in each profile during the learning process. To measure accuracy of retrieval, Precision, Recall, F1 scores were estimated for the months of studies. Respective results are illustrated in Tables 3 and 4 rounded to three digits, with the highest observed values observed in June (values in bold). Within our work, we defined the precision as the proportion of profiles having correctly inferred topics, over the total count of profiles (test). Using the same reasoning, recall was estimated as the proportion of profiles having correctly inferred topics over the total count of all profiles (test and validation). Results show very strong precision values for assigning correct topics to each profile under study. Figure 5 plots a comparison of the percentage of correctly classified profiles after running both algorithms for the same testing scenarios and for the top inferred topic. The percentage was calculated based on the comparison between the inferred top topic and the top topic provided by the classifier, assuming that the classifier reflects the user’s top favorite topics. The results show better performance for OLLDA when compared to L-LDA. Table 5 presents the top words that are associated with some of the detected topics in our sample. The majority of users, who are included in the sample, have political and business interests.

On Dynamic Topic Models for Mining Social Media OLLDA

223 LLDA

Percentage Value

1

0.75

0.5

0.25

0 April

May

June

July

August

September

Month

Fig. 5 Comparing correctly classified profiles for top (one) topic between OLLDA and LLDA with the ground truth for all months Table 5 Top words per topics Politics migpol svpol debate

Economics heritage budget revenue

Territorial disputes democrat debate party

Political science immigrant sd eupol

Socialism mingle talent budget

Human rights financial member work

Education belnar talk val

Business attefall budget european

4.5 Word Embeddings and Clustering We followed a similar approach to the one in [22], and we trained our corpus using Word2Vec [25], and then we fed the output into spherical K-means clustering algorithm [6]. The purpose of this experiment is to analyze the distributions of clustered words, and the semantic similarities between the top detected words per topic in OLLDA. The advantage of this strategy is to reduce the dimensionality and speeds the classification process. We applied the clustering using K = 50. Figure 6 plots the distribution of words in clusters for the sub-datasets of the months: April, May, June, July, August, and September. Figure 7 shows that there is one big cluster in each monthly sub-dataset, with other average-size clusters. The sample under study was for political accounts with similar topics and words, which can result in creating a big cluster of similar words, and other clusters for words that are un-related to the main topics of the dataset. Different line colors in Fig. 7 reflect the month which data was clustered. We analyzed the most semantically similar words to the top words that were detected by OLLDA in Table 6. Figure 8 plots the most similar words to the word “svpol.” As expected, most similar words by Word2Vec are related in

224

S. Jaradat and M. Matskin

Fig. 6 Clusters for sub-datasets of 6 months (April–September). The star symbol is the marker of centroid points of the clusters

general to the main word under comparison semantically. However, models such as LDA or LDA2Vec [26] consider the context of the document in addition to the word vector, which produces better and more meaningful distributions of words. We then calculated the average semantic similarity6 between top words detected per each topic in OLLDA (Table 7), and the results confirm the high similarity between those

6 Measurement

of similarity was done using the built-in similarity in Word2vec [25].

On Dynamic Topic Models for Mining Social Media

225

1250

1000

Words

750

500

250

0 10

20

30

40

50

Clusters

Fig. 7 Words distribution across clusters, where the peaks are showing the highest number of words in each cluster Table 6 Most semantic similar words to top detected words by OLLDA

migpol food poll raise people

svpol good alliance policy govern

eupol eu europe year input

debate read talk input today

finance tax week elect big

val svpol talk make good

european day elect eu govern

val svpol talk make good

detected words. In Fig. 9, we examined the classification across months of three top words: migpol, svpol, and eupol (which are related to Politics topic). We have noticed that those words most of the time fall in different clusters, but they all fall in the same cluster when the model is trained on the complete dataset (Fig. 9). This is also expected, as the model got more data during training of all months. Figure 10 visualizes a word cloud for a subset of users with political interests. The word cloud shows the most frequent words during Swedish Elections such as: svpol, vote, elect, and debate.

226

S. Jaradat and M. Matskin

Fig. 8 Similar words to the word “svpol”—political context Table 7 Similarity values between top words that were detected by OLLDA Politics 0.998

Economics 0.886

Territorial disputes 0.999

Political science 0.999

Socialism 0.987

Human rights 0.999

Education 0.933

Business 0.999

5 Online Labeled LDA Applications Twitter data has been used extensively in predicting and explaining a variety of real-life events. One of the interesting applications in this area is analyzing political tweets. Dokoohaki et al. [8] applied a link prediction approach on Twitter data that was gathered along the time line of European and general Swedish elections during 2014, in which they highlighted the correlation between the density of politicians’ conversations and their popularity, that is directly reflected in the estimated vote outcomes. A possible direction of OLLDA could be augmenting such analysis by topic modeling to extract and generate debate topics dynamically thus understanding the major themes in politicians’ conversations. OLLDA can be used to calculate the similarity metrics between users in social networks, which has a direct application for this algorithm in recommendation engines. Similarity in topics and interests can be considered as one of the factors that decides the trust level between users. Therefore, OLLDA can be applied in a

On Dynamic Topic Models for Mining Social Media

227

svpol,migpol and eupol 50

40

30

migpol

20

10 eupol

svpol

0 April

May

June

July

August

September

All

Fig. 9 Classification of top political words in clusters across months (April–September)

framework of tweets and friends recommendation as an example. Another possible application is trend detection which might require enhancements of the algorithm, to distinguish the trending topics. Currently, OLLDA is tested in Twitter context, and it will be interesting to test it in another online social network to verify its ability in producing better detection and inference of topics. Due to time limitations, OLLDA is still not applied in an extended application, which will imply its testing in a runtime scenario as well.

6 Conclusions This manuscript set forth the OLLDA, an enhanced online algorithm for topic modeling on Twitter. In our approach, we modified the approximate inference technique used in the selected online version of LDA, by adding a constraint on the variational parameters used in the approximation. This constraint changed the approach to be supervised, making it useful for streaming scenarios while maintaining accuracy. We conducted experiments on a 1-year Twitter data, with results demonstrating enhanced performance and topical inference quality of the new algorithm, when compared to the state-of-the-art baselines. For future work, we plan to explore applications of framework in recommendation systems.

228

S. Jaradat and M. Matskin

Fig. 10 Word cloud generated for a subset of the sample for users with political interests

References 1. Antenucci, D., Handy, G., Modi, A., Tinkerhess, M.: Classification of tweets via clustering of hashtags. In: EECS, vol. 545, pp. 1–11 (2011) 2. Attias, H.: A variational Bayesian framework for graphical models. In: Advances in Neural Information Processing Systems, vol. 12. MIT Press, Cambridge (2000) 3. Benzi, M., Wang, Y., Agichtein, E.: TM-LDA: efficient online Modeling of latent topic transitions in social media. In: Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 123–131 (2012) 4. Bhattacharya, P., Zafar, M.B., Ganguly, N., Ghosh, S., Gummadi, K.P.: Inferring user interests in the twitter social network. In: Proceedings of the 8th ACM Conference on Recommender Systems, RecSys ’14, pp. 357–360. ACM, New York (2014) 5. Blei, D., McAuliffe, J.: Supervised topic models. In: NIPS (2007) 6. Buchta, C., Kober, M., Feinerer, I., Hornik, K.: Spherical k-means clustering. J. Stat. Softw. 50(10), 1–22 (2012) 7. Buntine, W., Xie, L., Mehrotra, R., Sanner, S.: Improving LDA topic models for microblogs via tweet pooling and automatic labelling. In: Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval (2013) 8. Dokoohaki, N., Zikou, F., Gillblad, D., Matskin, M.: Predicting Swedish elections using twitter: a case for stochastic link structure analysis. In: The 5th Workshop on Social Network Analysis in Applications (SNAA2015), collocated with IEEE/ACM ASONAM 2015 9. Domeniconi, C., AlSumait, L., Barbara, D.: On-line LDA: adaptive topic models for mining text streams with applications to topic detection and tracking. In: Proceedings of the 2008 Eighth IEEE International Conference on Data Mining, pp. 3–12 (2008)

On Dynamic Topic Models for Mining Social Media

229

10. Homan, M., Blei, D., Bach, F.: Online learning for latent Dirichlet allocation. Neural Inf. Proces. Syst. 23, 856–864 (2010) 11. Hu, W., Zhu, M., Wu, O.: Topic detection for discussion threads with domain knowledge. In: Proceedings of International Conference on Web Intelligence and Intelligent Agent Technology, pp. 545–548 (2010) 12. Jakkkola, T., Saul, L., Jordan, M., Ghahramani, Z.: An introduction to variational methods for graphical models. Mach. Learn. 37, 183–233 (1999) 13. Jaradat, S., Dokoohaki, N., Matskin, M.: OLLDA: a supervised and dynamic topic mining framework in twitter. In: 2015 IEEE International Conference on Data Mining Workshop (ICDMW), November 2015, pp. 1354–1359. IEEE, Piscataway (2015) 14. Javed, A.: A hybrid approach to semantic hashtag clustering in social media. PhD Thesis, The University of Vermont and State Agricultural College (2016) 15. Jordan, M., Blei, D.: Variational inference for Dirichlet process mixtures. Bayesian Anal. 1, 121–144 (2006) 16. Jordan, M., Blei, D., Ng, A.Y.: Latent Dirichlet allocation. J. Mach. Learn. Res. 3, 993–1022 (2003) 17. Karapetsas, E., Ota, J., Zhu, D., Fukazawa, Y.: Intuitive topic discovery by incorporating word-pair’s connection into LDA. In: IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology, pp. 303–310 (2012) 18. Le, Q., Mikolov, T.: Distributed representations of sentences and documents. In: International Conference on Machine Learning, pp. 1188–1196 (2014) 19. Liebling, D., Ramage, D., Dumais, S.: Characterizing microblogs with topic models. In: Proceedings of AAAI on Weblogs and Social Media (2010) 20. Lilleberg, J., Zhu, Y., Zhang, Y.: Support vector machines and word2vec for text classification with semantic features. In: 2015 IEEE 14th International Conference on Cognitive Informatics Cognitive Computing (ICCI*CC), pp. 136–140 (2015) 21. Liu, Z., Zhang, Y., Chang, E., Sun, M.: PLDA+: parallel Latent Dirichlet Allocation with data placement and pipeline processing. In: ACM Transactions on Intelligent Systems and Technology (TIST), vol. 2, pp. 1–18 (2011) 22. Ma, L., Zhang, Y.: Using Word2Vec to process big text data. In: 2015 IEEE International Conference on Big Data (Big Data), pp. 2895–2897 (2015) 23. Macskassy, S., Michelson, M.: Discovering users’ topics of interest on twitter: a first look. In: Proceedings of the Workshop on Analytics for Noisy, Unstructured Text Data (AND), pp. 73–80 (2010) 24. McCallum, A., Yao, L., Mimno, D.: Efficient methods for topic model inference on streaming document collections. In: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2009) 25. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. In: Advances in Neural Information Processing Systems, pp. 3111–3119 (2013) 26. Moody, C.E.: Mixing Dirichlet topic models and word embeddings to make lda2vec (2016). arXiv:1605.02019 27. Muntean, C.I., Morar, G.A., Moldovan, D.: Exploring the meaning behind twitter hashtags through clustering. In: Business Information Systems Workshops, pp. 231–242(2012) 28. Nallapati, R., Manning, C., Ramage, D., Hall, D.: Labeled LDA: a supervised topic model for credit attribution in multi-labeled corpora. In: Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, Singapore (2009) 29. Narayanamurthy, S., Smola, A.: An architecture for parallel topic models. In: Proceedings of the VLDB Endowment, Singapore (2010) 30. Pennacchiotti, M., Gurumurthy, S.: Investigating topic models for social media user recommendation. In: Proceedings of the 20th International Conference Companion on World Wide Web, WWW ’11, pp. 101–102 (2011) 31. Quercia, D., Askham, H., Crowcroft, J.: TweetLDA: supervised topic classification and link prediction in Twitter. In: Proceedings of the ACM Web Science, pp. 373–376 (2012)

230

S. Jaradat and M. Matskin

32. Sha Lacoste-Julien, F., Jordan, M.: Disc LDA: discriminative learning for dimensionality reduction and classification. In: NIPS (2008) 33. Shi, L., Canini, K., Griths, T.: Online inference of topics with latent Dirichlet allocation. In: Proceedings of the 12th International Conference on Artificial Intelligence and Statistics (2009) 34. Smyth, P., Asuncion, A., Welling, M.: Asynchronous distributed learning of topic models. In: NIPS, pp. 81–88 (2008) 35. Teo, C., Eisenstein, J., Smola, A., Xing, E., Ahmed, A., Ho, Q.: Online inference for the infinite topic-cluster model: storylines from streaming text. In: Artificial Intelligence and Statistics AISTATS, vol. 15, pp. 101–109 (2011)

Part III

Broader Challenges and Impacts

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis Soon Jye Kho, Swati Padhee, Goonmeet Bajaj, Krishnaprasad Thirunarayan, and Amit Sheth

Abstract Social media provides a virtual platform for users to share and discuss their daily life, activities, opinions, health, feelings, etc. Such personal accounts readily generate Big Data marked by velocity, volume, value, variety, and veracity challenges. This type of Big Data analytics already supports useful investigations ranging from research into data mining and developing public policy to actions targeting an individual in a variety of domains such as branding and marketing, crime and law enforcement, crisis monitoring and management, as well as public and personalized health management. However, using social media to solve domainspecific problem is challenging due to complexity of the domain, lack of context, colloquial nature of language, and changing topic relevance in temporally dynamic domain. In this article, we discuss the need to go beyond data-driven machine learning and natural language processing, and incorporate deep domain knowledge as well as knowledge of how experts and decision makers explore and perform contextual interpretation. Four use cases are used to demonstrate the role of domain knowledge in addressing each challenge.

1 Introduction As the saying goes, data is the new oil in the twenty-first century.1 Data is described as an immense and valuable resource and extracting actionable insights for decision making is its value and has been the key to advancements in various domains. Many

1 https://goo.gl/YQq6pi.

S. J. Kho () · S. Padhee · G. Bajaj · K. Thirunarayan · A. Sheth Kno.e.sis Center, Wright State University, Dayton, OH, USA e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; http://knoesis.org © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_9

233

234

S. J. Kho et al.

companies have realized this and have begun to treat their data as a valuable asset. This data is usually private and is not accessible to the public, resulting in a barrier for researchers. However, there is another source of continuously growing data that is publicly available and has been utilized by the research community for a variety of tasks—social media. Social media serves as a virtual platform for users to share their opinions, report their daily life, activities, opinions, health, feelings, etc., and communicate with other users. In the year 2017, it has been reported that 81% of the public in America use some type of social media2 , and this contributes to the rapid generation of social media data. Twitter alone generates 500 million posts daily. Due to its volume and highly personalized nature, social media has been tapped by researchers for extracting useful individualized information. Various Natural Language Processing (NLP) techniques have been deployed for opinion mining [9, 27], sentiment analysis [10, 20], emotion analysis [17, 39], metadata extraction [1], and user group’s characterization [28]. Just like crude oil is a raw material where its value can only be realized after the refinement process, social media data needs to be analyzed in order to gain actionable insights from it. The volume of social media data is a mixed blessing. From the view of data source, it is undoubtedly a bliss. The users share their views regarding a topic or an issue that can be very diverse and comprehensive in coverage, ranging from entertainment products to politics, and from personal health to religion. This explains why social media data has been used to gauge public perception on various issues and applications, such as brands uptake [36], election prediction3 [5], disaster coordination [3, 30], public health issue such as monitoring depression [41], epidemiological and policy research related to epidemic [23], marijuana legalization [7], and drug abuse surveillance [4]. However, from the view of problem-solving, the content diversity and volume inherent in social media create significant practical challenges for extracting relevant information, as it is akin to searching for a needle in a haystack. To bridge the gap between the available raw data and the needed insights for decision making, researchers have exploited existing knowledge specific to the domain to fill in the gaps to relate pieces of data, determine their relevance to the problem at hand, and power an analytical framework. If the data was the crude oil which is to be refined to gain insights to solve a problem, then the domain knowledge is the instructions on how to use the refinery machines.

2 https://goo.gl/i8n4Jm. 3 https://goo.gl/qYL2kv.

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis

235

2 Challenges in Domain-Specific Social Media Analysis Social media analysis provides complementary insights for decision making. However, social media being an instance of Big Data creates several challenges for researchers to overcome before we can reap the promised fruits. Here, we discuss the major challenges in domain-specific social media analysis with respective examples in Table 1: – Complex real-world domain—Real-world problems are complex in that they involve multiple diverse factors. Understanding their mutual influences and interactions is non-trivial as exemplified many domains including healthcare, especially in the context of mental health disorder, its characteristics, causes, and progression.4 Raw social media data often provides information at a basic level and seldom provides in an abstract actionable form that is necessary for solving the problems. – Lack of context—Online users usually reveal their opinions and feelings [24] to their friends or in public assuming a shared understanding of the situation, which is implicit. Besides, social media such as Twitter impose word limitation further restraining the users from expressing more context. The presence of implicit context complicates a variety of NLP tasks such as entity disambiguation, sentiment analysis, and opinion mining.

Table 1 Examples of social media text that represent the challenges of social media analysis Challenges Complex real-world domain

Social media text I am disgusted with myself

Lack of context

Colloquial nature of language

On sub ive actually felt nothing from +500mg doses for hours

Topic relevance

This new space movie is crazy. you must watch it!

Problem The text implies the user is having low self-esteem. Low self-esteem is one of the depressive symptoms, but it is untenable to diagnose the user with depression based on this individual text The gas pump emoji represents the letter “G.” Without understanding the emojis and the hidden context, it is possible to infer the meaning of the text as: “WHOLE LOTTA GANG SHIT GOING ON” The term “sub” is commonly used to refer Subutex, a brand name of buprenorphine. Using only standard drug name in data collection would fail to capture this tweet as relevant information for drug abuse The term “space movie” refers to different movies in different time periods. The domain knowledge and its temporal salient are crucial in accurately disambiguating the entities

4 https://www.nimh.nih.gov/health/topics/index.shtml.

236

S. J. Kho et al.

– Colloquial nature of language—Due to the informal setting of social media, users tend to use language that is used in ordinary or familiar conversation. Instead of standard terms, users use slang terms and nicknames to refer an entity. Other than that, the language used on social media is prone to grammatical error, unconventional contractions, and spelling mistakes. This creates a challenge for entity recognition as social media data is collected based on defined keywords. If only standard terms are used to crawl for posts, it would fail to capture relevant information. – Topic relevance—Due to the dynamic or periodic nature of certain domains, the same context is relevant to different entities at different time periods. The domain knowledge, as well as their temporal aspect, is important to understand and disambiguate entities in such social media data.

3 Domain Knowledge for Social Media Analysis Domain knowledge has been exploited by researchers to address the challenges mentioned above. Domain knowledge codifies an area of human endeavor or a specialized discipline [13]. It covers a broad range including facts, domain relationships, and skills acquired through experience or education. The use of domain knowledge in Web (semantic) applications such as search, browsing, personalization, and advertising was recognized and commercialized at the turn of the century [11, 32, 33] if not before, and has seen much wider usage when Google relied on its Knowledge Graph for its semantic search in 2013. Artificial Intelligence (AI) researchers have noted that “data alone is not enough” and that knowledge is very useful [8]. Use of knowledge to improve understanding and analysis a variety of textual and non-textual content, compared with the baseline machine learning and NLP techniques, was discussed in [34]. As domain knowledge is normally acquired by humans through years of learning and experience, a domain expert is a scarce resource. Fortunately, there is an abundance of knowledge graphs (KG) [35] which are publicly available. The semantic web community has been putting major effort in producing large and cross-domain (e.g., Wikipedia and DBpedia) as well as domain-specific knowledge graphs (e.g., Gene Ontology, MusicBrainz, and UMLS) to codify well-circumscribed domain of discourse. Knowledge graphs [35] accessible on the web have been increasingly used to incorporate semantics in various applications, such as recommender system [15, 26], named entity disambiguation [37], document retrieval [31], and last but not least, social media analysis. Knowledge graphs are large and creating domain knowledge that can support domain-specific social media analytics can be very challenging. To extract the domain-specific subgraphs from large knowledge graphs, Lalithsena [18] presented an approach by considering entity semantics [19] of the knowledge graphs. Existing approaches extract the subgraph by navigating up to a predefined number of hops in the knowledge graphs starting from known domain entities but this can

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis

237

Fig. 1 Wikipedia category hierarchy for entity US presidential election. Simple n-hop navigation would reach entity “Japan Music Week” which is irrelevant to the US presidential election. PSLbased approach proposed by Lalithsena et al. successfully extracted a subgraph with more relevant entities (green color nodes)

lead to irrelevant entities and relationships in the knowledge graph. The approach proposed by Lalithsena [18] uses the entity semantics of the categories in the Wikipedia category hierarchy to extract domain-specific subgraphs (as shown in Fig. 1). The approach combines the semantic type, lexical, and structural categories using probabilistic soft logic (PSL) [2] to incorporate different forms of evidence. The extracted domain-specific subgraph using the proposed approach reduced 74% of the categories from the simple n-hop expansion subgraph. Researchers have been putting many efforts in advancing the extraction of domain-specific knowledge from large knowledge graphs. The extracted knowledge can be complementary to such domain-specific social media data analysis (e.g., political election). In this article, we are going to demonstrate the role of domain knowledge in addressing the four major challenges and illustrate it using four appropriate use cases.

3.1 Addressing Domain Complexity: Use Case of Mental Health Disorder Mental health disorders encompass a wide range of conditions that are characterized by a change of mood, thinking, and behavior.5 These include but are not limited to obsessive-compulsive disorder, bipolar disorder, and clinical depression. Clinical depression is a common but debilitating mental illness that is prevalent worldwide. 5 https://www.nami.org/learn-more/mental-health-conditions.

238

S. J. Kho et al.

It affects more than 300 million people globally6 and costs over $40 billion every year in the USA on depression treatment [6]. Diagnosis of depression in a patient requires a physician to consider a subset of predefined symptoms that last over a period of time. Typically, depression is detected in a primary care setting through patient health questionnaires7 to screen for the presence and severity of depression. Predefined symptoms of depression constitute the state-of-the-art domain knowledge curated by reputable medical doctors and are incorporated into tools like Patient Health Questionnaire (PHQ-9) [16]. PHQ-9 has been used by clinicians for screening, diagnosing, and measuring the severity of depression symptoms as defined by the American Psychiatric Association Diagnostic and Statistical Manual (DSM-IV). The 9-item measurement tool contains a user-friendly response format, short administration time, and easy scoring [21]. However, studies have shown that these questionnaire-based methods are not accurate because of incomplete information and cognitive biases. First and foremost, there is no guarantee that patients can remember all accounts of depressive symptoms experienced over a certain period of time. Additionally, cognitive biases occur because of the way the questionnaire is phrased or administered, and this can prevent participants from giving truthful responses [12]. Contrary to our expectation that the social stigma associated with a clinically depressed patient will make them highly introverted and unlikely to share their condition publicly, there is a surprising amount of unsolicited sharing of depressive symptoms by users on social media. In fact, online social media provides a convenient and unobtrusive forum to voluntarily document their daily struggles with their mental health issues. This might be due to the anonymity that the social media provides, freeing them from the worries of people judging them. This type of passive monitoring can capture a user’s condition with little burden on the user to report their emotions and feelings. Social media text provides a source for researchers to monitor potential depressive symptoms, but it does not provide a good indication of whether a user is depressed or not. As mentioned above, depression is diagnosed based on the presence of 6 out of 9 symptoms over a period of two weeks as advocated by PHQ9 questionnaire-based diagnosis. Therefore, to model this type of domain practice through social media analysis, one should not just analyze the text in social media post disjointly. Instead, the analysis should perform jointly on every post of a user, as this portrays a more holistic picture of users’ mental health condition. Inspired by the scalability benefits and widespread use of PHQ-9 and social media analysis, Yazdavar et al. [41] emulate traditional observational cohort studies conducted through online questionnaires by extracting, categorizing, and unobtrusively monitoring different depressive symptoms. They developed a statistical model by modeling user-generated content in social media as a mixture of underlying topics evolving over time based on PHQ-9 symptoms. 6 http://www.who.int/mediacentre/factsheets/fs369/en/. 7 https://www.cdc.gov/mentalhealthsurveillance/.

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis

239

To determine the textual cues that reflect the symptoms in each PHQ-9 criteria, Yazdavar et al. [41] generated a lexicon to capture each criterion and leveraged the lexicon as simple background knowledge. Furthermore, given the challenges of understanding the colloquial language on social media, Urban Dictionary8 and the Big Huge Thesaurus9 (a crowd-sourced online dictionary of slang words and phrases) were utilized. For expanding the lexicon, they used the synset of each of the nine PHQ-9 depression symptom categories. The consistency of the built lexicon has been vetted by domain experts. The final lexicon contains over 1620 depression-related symptoms categorized into nine different clinical depression symptoms which are likely to appear in the tweets of individuals suffering from clinical depression (as shown in Table 2). Using the lexicon, Yazdavar et al. identified users with self-reported symptoms of depression based on their profile descriptions. They then formulated a hybrid solution and conducted a temporal analysis of user-generated content on social media for capturing mental health issues. This use of domain knowledge proved valuable in creating a tool with an accuracy of 68% and a precision of 72% for capturing depression symptoms per user over a time interval. To summarize, using the domain knowledge in the form of PHQ-9 categories and associated lexicons, they defined textual cues that could reflect depressive symptoms and use these cues in mining depression clues in social media data. They also uncovered common themes and triggers of depression at the community level. Not only emulating the PHQ-9 simplifies the domain complexity for understanding depression through social media, but it also provides a more intuitive interpretation that can be used to complement physicians in detecting depression.

3.2 Addressing Lack of Context: Use Case of EmojiNet Users share their opinions, activities, locations, and feeling but they seldom express the motivation or intention behind their sharing. This is due to the shared understanding between friends, unwillingness of sharing too much detail, or word limitation by some platform (Twitter only allowed a maximum word count of 140 before they increased it to 280). Lack of context comprises one of the major challenges in social media analysis since context is critical in performing entity disambiguation, sentiment, and emotion analysis. Wijeratne et al. [40] exploit the use of emoji as a means to glean context to enhance understanding of a post. Emoji has been commonly used in social media and is extremely popular in electronic communication.10 People use emoji to show whimsiness and describe emotions that are hard to articulate. As such, emoji pro-

8 https://www.urbandictionary.com/. 9 http://www.thesaurus.com/. 10 https://goo.gl/ttxyP1.

240

S. J. Kho et al.

Table 2 Few examples of lexicon terms and sample tweet for PHQ-9 symptoms PHQ-9 Lack of interest

Feeling down

Sleep disorder

Lack of energy

Eating disorder

Self-hate

Concentration problems

Hyper/Lower activity

Suicidal thoughts

Lexicon terms “Couldn’t care less” “don’t want this” “Used to enjoy” “Zero motivation” “All torn up” “can’t stop crying” “i’m beyond broken” “Shit day” “Can’t sleep” “Sleep deprived” “Sleeping pill” “Crying to sleep” “Drained” “Lassitude” “I am weak” “Tired of everything” “Flat tummy” “Hate my thighs” “Love feeling hungry” “Wanna be thinner” “I am disgusting” “I’m a freak” “Waste of space” “Never good enough” “Overthinking” “Short attention span” “Can’t pay attention” “Brain dead” “spazz” “Paranoid” “Unsettled” “I’m slow moving” “Ending my life” “Overdosing” “My razor” “Sleep forever”

Sample tweet I’ve not replied all day due to total lack of interest, depressed probs

I feel like i’m falling apart

Night guys. Hope you sleep better than me

So tired, so drained, so done

Just wanna be skinny and beautiful

I just let everyone down. why am I even here?

I couldn’t concentrate to classes at all can’t stop thinking

So stressed out I cant do anything

I want summer but then i don’t. . . It’ll be harder to hide my cuts

vides an alternative and effective way to express intention, sentiment, and emotion in a post. A study showed that emoji polarity can improve the sentiment score and understand the meaning of emoji can improve the interpretation of a post [25]. However, a particular emoji can be ambiguous [40]. For example, (face with tears of joy) can be used for expressing happiness (using senses such as laugh and joy) as well as sadness (using senses such as cry and tears) [22]. In order to understand the sense conveyed through an emoji, Wijeratne et al. developed the

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis

241

first machine-readable sense inventory for emoji, EmojiNet11 [40]. They integrated four openly accessible knowledge sources for emoji (i.e., Unicode Consortium, Emojipedia, iEmoji, and The Emoji Dictionary) into a single dictionary of emoji. EmojiNet represents a knowledge inventory of emoji where their senses and definitions are housed. This inventory is used for emoji sense disambiguation application which is designed to automatically learn message contexts where a particular emoji sense could appear. Consider the above example of (face with tears of joy), the emoji sense disambiguation techniques could determine the sense of the emoji (happiness or sadness) based on the context in which it has been used. Using the EmojiNet, researchers are able to understand the meaning of emoji used in a message. Understanding the meaning of emoji hereby enriches the context that could potentially enhance applications that study, analyze, and summarize electronic communications.

3.3 Addressing Colloquial Nature of Language: Use Case of Drug Abuse Ontology Opioids are a class of drugs that are chemically related and interact with opioid receptors in nerve cells. The medical opioid is available legally via prescription and the common medical use of an opioid is to relieve pain. Generally, it is safe to use them over a short-term and as prescribed by the doctor. However, due to its euphoric properties, opioid has been misused which can lead to overdose incidents [38]. The non-medical use of pharmaceutical opioids has been identified as one of the fastest growing forms of drug abuse in the USA. The White House Office of National Drug Control Policy (ONDCP),12 in May 2011, launched America’s Prescription Drug Abuse Crisis [14] initiative to curb prescription drug abuse problem, mainly through education and drug monitoring programs. This underscores the importance of determining new and emerging patterns or trends in prescription drug abuse in a timely manner. Existing epidemiological data systems provide critically important information about drug abuse trends, but they are often time-lagged, suffering from large temporal gaps between data collection, analysis, and information dissemination. Social media offers a platform for users to share their experiences online and this information is useful in timely drug abuse surveillance. To collect social media data that are related to drug use experience, it is important to recognize the mentions of drug-related entities in the tweets. However, entity recognition from social media data is difficult, due to grammatical errors, misspellings, usage of slang term, and street names. For example, Buprenorphine

11 http://emojinet.knoesis.org/. 12 https://www.whitehouse.gov/ondcp/.

242

S. J. Kho et al.

might be referred to as bupe; marijuana concentrate products might be referred to as dabs, earwax, or hash oil [7]. This colloquial nature of the language used in social media causes a decrease in the recall for capturing relevant information. To address the challenge, Cameron et al. [4] manually curated Drug Abuse Ontology (DAO, pronounced dao) with the help of domain expert. They modeled the ontology using web forum posts, domain concepts, and relationships. DAO is the first ontology for prescription drug abuse, and it is used for processing web forum posts along with a combination of lexical, pattern-based and semanticsbased techniques. It contains the mapping of popular slang terms to the standard drug names, and this provides a good base reference for drug entity recognition. This knowledge of slang term mapping and enrichment of vocabulary helps to increase the recall for collecting relevant data, while the reduction in precision is prevented/remedied by using entity disambiguation.

3.4 Addressing Topic Relevance: Use Case of Movie Domain Social media has been used to share information about entities such as movies and books. However, a novel aspect of this communication is that the entities are mentioned implicitly through their defining characteristics. Motivated by this, Perera et al. [29] have introduced and addressed this issue of identifying implicit references to an entity in tweets. An implicit entity is defined as an entity mentioned in a text where its name nor its synonym/alias/abbreviation is not present in the text [29]. For example, an entity (movie “Gravity”) can be explicitly mentioned as The movie Gravity was more expensive to make than the Mars Orbiter Mission. It can also be mentioned implicitly as This new space movie is crazy. you must watch it!. They reported that implicit mention of movie entity is found in 21% of the tweets. This indicates that keyword-based data collection would fail to capture almost onequarter of relevant information and increasing the recall would be unattainable without the use of contextual domain knowledge. Contextual domain knowledge is very important for understanding implicit mentions of entities, but it is complicated by the dynamic nature of some domain (e.g., media, movie, and news). These domains generate a lot of new facts that remain significant only for a short period of time because those new relationships came into prominence due to a related event in the news. Considering the movie domain, movies remain popular only for a limited time, being eclipsed by new movies. In fall of 2013, the mention of “space movie” referred to the movie “Gravity,” whereas, in fall of 2015, it referred to the movie “The Martian.” The temporal salience of this dynamically changing domain knowledge that reflects the loose association between the movie and its referenced property—in this context the “space movie”—is crucial for correctly disambiguating and identifying such implicit mentions. Perera et al. proposed an approach which takes into account the contextual knowledge (common phrases in the tweets relevant to the implicit entity) as well

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis

243

as the factual knowledge (common terms, entities, and relationships relevant to the implicit entity). First, they acquired factual knowledge from DBpedia and extract only relevant knowledge based on their joint probability value with the given entity type. Then, they obtained contextual knowledge from contemporary tweets that explicitly mention the entity. These two types of knowledge are used to create an entity model network (EMN) to reflect the topical relationships among domain entities at a certain time. The EMN identifies relevant domain entities that are relevant in that time period and use this knowledge to identify implicit entity mentions. The combination of factual knowledge and contextual knowledge which takes temporal salience into account enables the system to recognize implicit entity from tweets. Perera et al. evaluated their approach for two domains, viz., movies and books, and showed that the use of contextual knowledge contributed to an improved recall of 14 and 19%, respectively, while the temporal salience improved the accuracy of entity disambiguation task by 15 and 18%, respectively.

4 Discussion Social media platforms have provided people a vehicle for free expressions of opinions, feelings, and thoughts. Researchers and computer scientists have exploited this rich content and the social interactions among its users to enable a deeper understanding of real-world issues and the people’s perspectives on them. The progress in natural language processing techniques and its customization to social media communication provided the initial impetus for the analysis. Semantic Web, or Web 3.0, has emerged as a complementary area of research for machine understanding of web data. This is being harnessed to capture and represent aspects related to human intelligence and cognition, which play a crucial role in human learning by enabling a deeper understanding of content and context at a higher abstraction level. We discussed the importance and significance of domain knowledge in designing and exploiting social media analysis framework. Similar to the roles of knowledge in human intelligence, there are two major roles of domain knowledge in social media analysis: language understanding and information interpretation. Social media, as a platform for human expression and communication, inherently contains ambiguities, gaps, and implicit references similar to human natural language communication that requires domain knowledge for understanding and extracting the context. For example, Sect. 3.3 illustrates the use of different slang terms referring to the same drug entity that can be captured and represented through the use of ontology. This approach enables researchers to capture relevant information and fuse different types of information to obtain actionable insights in the medical context. However, having a deep language understanding is not always sufficient for solving real-world problems that require domain knowledge for understanding the situation before generating actions. This is clearly illustrated in Sect. 3.1, where

244

S. J. Kho et al.

Table 3 Application of domain knowledge in addressing major challenges in social media analysis with examples of domain-specific use case Challenges Use case Complex real-world domain Mental health disorder

Lack of context

EmojiNet

Colloquial nature of language

Drug abuse epidemiology

Topic relevance

Implicit entity linking

Application and improvement Using PHQ-9 in analyzing tweets to simulate the diagnosis practice of physician in diagnosing depression Understanding the meaning of emoji and use it to enrich the context Utilize the mapping of slang terms to standard drug references to improve recall and coverage of relevant tweets collection Extract related factual knowledge and contextual knowledge of a movie from DBpedia. The extracted knowledge and their temporal aspect are then taken into accounts for recognizing and disambiguating the movie/book entity in tweets

domain knowledge is needed to interpret the information extracted from the text for diagnosing depression. Domain knowledge provides an intuitive analysis framework by emulating how human experts practice in real life. In this article, we have discussed four major challenges to domain-specific social media analysis and have highlighted the need to go beyond data-driven machine learning and natural language processing. Table 3 summarizes the specific challenges addressed by the application of domain knowledge, along with the lines of how experts and decision makers explore and perform contextual interpretation, to garner actionable information and insights from social media data. Acknowledgements We would like to thank Sarasi Lalithsena, Shweta Yadav, and Sanjaya Wijeratna for their patient and insightful reviews. We would also like to acknowledge partial support from the National Science Foundation (NSF) award: CNS-1513721: “Context-Aware Harassment Detection on Social Media,” National Institute on Drug Abuse (NIDA) Grant No. 5R01DA03945402: “Trending: Social Media Analysis to Monitor Cannabis and Synthetic Cannabinoid Use,” National Institutes of Health (NIH) award: MH105384-01A1: “Modeling Social Behavior for Healthcare Utilization in Depression,” and Grant No. 2014-PS-PSN-00006 awarded by the Bureau of Justice Assistance. The Bureau of Justice Assistance is a component of the US Department of Justice’s Office of Justice Programs, which also includes the Bureau of Justice Statistics, the National Institute of Justice, the Office of Juvenile Justice and Delinquency Prevention, the Office for Victims of Crime, and the SMART Office. Points of view or opinions in this document are those of the authors and do not necessarily represent the official position or policies of the US Department of Justice, NSF, NIH, or NIDA.

Domain-Specific Use Cases for Knowledge-Enabled Social Media Analysis

245

References 1. Abel, F., Hauff, C., Houben, G.J., Stronkman, R., Tao, K.: Twitcident: fighting fire with information from social web streams. In: Proceedings of the 21st International Conference on World Wide Web, pp. 305–308. ACM, New York (2012) 2. Bach, S.H., Broecheler, M., Huang, B., Getoor, L.: Hinge-loss markov random fields and probabilistic soft logic. J. Mach. Learn. Res. 18(109), 1–67 (2017) 3. Bhatt, S.P., Purohit, H., Hampton, A., Shalin, V., Sheth, A., Flach, J.: Assisting coordination during crisis: a domain ontology based approach to infer resource needs from tweets. In: Proceedings of the 2014 ACM Conference on Web science, pp. 297–298. ACM, New York (2014) 4. Cameron, D., Smith, G.A., Daniulaityte, R., Sheth, A.P., Dave, D., Chen, L., Anand, G., Carlson, R., Watkins, K.Z., Falck, R.: Predose: a semantic web platform for drug abuse epidemiology using social media. J. Biomed. Inform. 46(6), 985–997 (2013) 5. Chen, L., Wang, W., Sheth, A.P.: Are twitter users equal in predicting elections? a study of user groups in predicting 2012 us republican presidential primaries. In: International Conference on Social Informatics, pp. 379–392. Springer, Berlin (2012) 6. Craft, L.L., Perna, F.M.: The benefits of exercise for the clinically depressed. Prim. Care Companion J. Clin. Psychiatry 6(3), 104 (2004) 7. Daniulaityte, R., Nahhas, R.W., Wijeratne, S., Carlson, R.G., Lamy, F.R., Martins, S.S., Boyer, E.W., Smith, G.A., Sheth, A.: “time for dabs”: analyzing twitter data on marijuana concentrates across the US. Drug Alcohol Depend. 155, 307–311 (2015) 8. Domingos, P.: A few useful things to know about machine learning. Commun. ACM 55(10), 78–87 (2012) 9. Ebrahimi, M., Yazdavar, A.H., Salim, N., Eltyeb, S.: Recognition of side effects as implicitopinion words in drug reviews. Online Inf. Rev. 40(7), 1018–1032 (2016) 10. Ebrahimi, M., Yazdavar, A.H., Sheth, A.: Challenges of sentiment analysis for dynamic events. IEEE Intell. Syst. 32(5), 70–75 (2017) 11. Gruhl, D., Nagarajan, M., Pieper, J., Robson, C., Sheth, A.: Context and domain knowledge enhanced entity spotting in informal text. In: International Semantic Web Conference, pp. 260– 276. Springer, Berlin (2009) 12. Haselton, M.G., Nettle, D., Murray, D.R.: The evolution of cognitive bias. In: The Handbook of Evolutionary Psychology. Wiley, New York (2005) 13. Hirschfeld, L.A., Gelman, S.A.: Mapping the Mind: Domain Specificity in Cognition and Culture. Cambridge University Press, Cambridge (1994) 14. House, W.: Epidemic: Responding to America’s Prescription Drug Abuse Crisis. White House, Washington (2011) 15. Kapanipathi, P., Jain, P., Venkataramani, C., Sheth, A.: User interests identification on twitter using a hierarchical knowledge base. In: European Semantic Web Conference, pp. 99–113. Springer, Berlin (2014) 16. Kroenke, K., Spitzer, R.L., Williams, J.B.: The phq-9. J. Gen. Intern. Med. 16(9), 606–613 (2001) 17. Kušen, E., Cascavilla, G., Figl, K., Conti, M., Strembeck, M.: Identifying emotions in social media: comparison of word-emotion lexicons. In: 2017 5th International Conference on Future Internet of Things and Cloud Workshops (FiCloudW), pp. 132–137. IEEE, New York (2017) 18. Lalithsena, S.: Domain-specific knowledge extraction from the web of data. Department of Computer Science and Engineering, Wright State University (2018) 19. Lalithsena, S., Perera, S., Kapanipathi, P., Sheth, A.: Domain-specific hierarchical subgraph extraction: a recommendation use case. In: 2017 IEEE International Conference on Big Data (Big Data), pp. 666–675. IEEE, New York (2017) 20. Liu, B.: Sentiment Analysis and Opinion Mining. Synthesis Lectures on Human Language Technologies, vol. 5, no. 1, pp. 1–167. Morgan and Claypool , San Rafael (2012) 21. Milette, K., Hudson, M., Baron, M., Thombs, B.D., Canadian Scleroderma Research Group: Comparison of the phq-9 and ces-d depression scales in systemic sclerosis: internal consistency reliability, convergent validity and clinical correlates. Rheumatology 49(4), 789–796 (2010)

246

S. J. Kho et al.

22. Miller, H., Thebault-Spieker, J., Chang, S., Johnson, I., Terveen, L., Hecht, B.: Blissfully happy” or “ready to fight”: varying interpretations of emoji. In: Proceedings of ICWSM, vol. 2016 (2016) 23. Miller, M., Banerjee, T., Muppalla, R., Romine, W., Sheth, A.: What are people tweeting about Zika? An exploratory study concerning its symptoms, treatment, transmission, and prevention. JMIR Public Health Surveill. 3(2), e38 (2017) 24. Mukherjee, S., Malu, A., Balamurali, A.R., Bhattacharyya, P.: Twisent: a multistage system for analyzing sentiment in twitter. In: Proceedings of the 21st ACM International Conference on Information and Knowledge Management, pp. 2531–2534. ACM, New York (2012) 25. Novak, P.K., Smailovi´c, J., Sluban, B., Mozetiˇc, I.: Sentiment of emojis. PLoS One 10(12), e0144296 (2015) 26. Ostuni, V.C., Di Noia, T., Di Sciascio, E., Mirizzi, R.: Top-n recommendations from implicit feedback leveraging linked open data. In: Proceedings of the 7th ACM Conference on Recommender Systems, pp. 85–92. ACM, New York (2013) 27. Pak, A., Paroubek, P.: Twitter as a corpus for sentiment analysis and opinion mining. In: LREc, vol. 10 (2010) 28. Pennacchiotti, M., Popescu, A.M.: Democrats, republicans and starbucks afficionados: user classification in twitter. In: Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 430–438. ACM, New York (2011) 29. Perera, S., Mendes, P.N., Alex, A., Sheth, A.P., Thirunarayan, K.: Implicit entity linking in tweets. In: International Semantic Web Conference, pp. 118–132. Springer, Berlin (2016) 30. Purohit, H., Castillo, C., Diaz, F., Sheth, A., Meier, P.: Emergency-relief coordination on social media: automatically matching resource requests and offers. First Monday 19(1) (2013). http:// firstmonday.org/ojs/index.php/fm/issue/view/408 31. Schuhmacher, M., Ponzetto, S.P.: Knowledge-based graph document modeling. In: Proceedings of the 7th ACM International Conference on Web Search and Data Mining, pp. 543–552. ACM, New York (2014) 32. Sheth, A., Avant, D., Bertram, C.: System and method for creating a semantic web and its applications in browsing, searching, profiling, personalization and advertising (Oct 30, 2001). US Patent 6,311,194 33. Sheth, A., Bertram, C., Avant, D., Hammond, B., Kochut, K., Warke, Y.: Managing semantic content for the web. IEEE Internet Comput. 6(4), 80–87 (2002) 34. Sheth, A., Perera, S., Wijeratne, S., Thirunarayan, K.: Knowledge will propel machine understanding of content: extrapolating from current examples. In: Proceedings of the International Conference on Web Intelligence. WI ’17, pp. 1–9. ACM, New York (2017). http://doi.acm.org/ 10.1145/3106426.3109448 35. Singhal, A.: Introducing the knowledge graph: things, not strings. Official google blog (2012) 36. Smith, A.N., Fischer, E., Yongjian, C.: How does brand-related user-generated content differ across YouTube, Facebook, and Twitter? J. Int. Mark. 26(2), 102–113 (2012) 37. Usbeck, R., Ngomo, A.C.N., Röder, M., Gerber, D., Coelho, S.A., Auer, S., Both, A.: Agdistisgraph-based disambiguation of named entities using linked data. In: International Semantic Web Conference, pp. 457–471. Springer, Berlin, (2014) 38. Vowles, K.E., McEntee, M.L., Julnes, P.S., Frohe, T., Ney, J.P., van der Goes, D.N.: Rates of opioid misuse, abuse, and addiction in chronic pain: a systematic review and data synthesis. Pain 156(4), 569–576 (2015) 39. Wang, W., Chen, L., Thirunarayan, K., Sheth, A.P.: Harnessing twitter “big data” for automatic emotion identification. In: Privacy, Security, Risk and Trust (PASSAT), 2012 International Conference on and 2012 International Conference on Social Computing (SocialCom), pp. 587–592. IEEE, New York (2012) 40. Wijeratne, S., Balasuriya, L., Sheth, A., Doran, D.: Emojinet: building a machine readable sense inventory for emoji. In: International Conference on Social Informatics, pp. 527–541. Springer, Berlin (2016) 41. Yazdavar, A.H., Al-Olimat, H.S., Ebrahimi, M., Bajaj, G., Banerjee, T., Thirunarayan, K., Pathak, J., Sheth, A.: Semi-supervised approach to monitoring clinical depressive symptoms in social media. In: Proceedings of the 2017 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining 2017, pp. 1191–1198. ACM, New York (2017)

Privacy in Human Computation: User Awareness Study, Implications for Existing Platforms, Recommendations, and Research Directions Mirela Riveni, Christiaan Hillen, and Schahram Dustdar

Abstract Research and industry have made great advancements in human computation and today we can see multiple forms of it reflected in growing numbers and diversification of platforms, from crowdsourcing ones, social computing platforms (in terms of collaborative task execution), and online labor/expert markets to collective adaptive systems (CAS) with humans-in-the-loop. Despite the advancements in various mechanisms to support effective provisioning of human computation, there is still one topic that seems to be close to neglected both in research and the current design and development of human computation systems, namely privacy. In this work, we investigate this problem. Starting from the fact that user awareness is crucial for enforcing privacy-respecting mechanisms, we conducted an online survey study to assess user privacy awareness in human computation systems and in this paper provide the results of it. Lastly, we provide recommendations for developers for designing privacy-preserving human computation platforms as well as research directions.

1 Introduction Human computation is a concept that has already gained its momentum and its application is evolving in rapid pace with the development of different types of platforms and mechanisms for effective utilization of human intelligence online. Human computation as a term is first coined by von Ahn in [37]. It has been defined as the utilization of human intelligence for tasks, activities, and problems that cannot yet be executed and solved by artificial intelligence. Hence, some call

M. Riveni () · S. Dustdar Distributed Systems Group, TU Wien, Vienna, Austria e-mail: [email protected]; [email protected] C. Hillen Independent Researcher, Nijmegen, The Netherlands e-mail: [email protected] © Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9_10

247

248

M. Riveni et al.

human computation artificial artificial-intelligence. In this work, we put several concepts that include online task execution by people under the umbrella of the Human Computation concept (varying a little from the taxonomy presented in [25] by Quinn et al.). Those concepts are: Crowdsourcing, which involves simple task execution by a large number of (anonymous) people (see a survey in [39]); Social Computing, with which we imply online computations where multiple people are involved in complex task execution (see examples in [9, 26]); Online Labor Markets and Expert Networks; Human-based services in mixed systems in which people provide their services/skills within Service-Oriented Architectures [30] as well as human computation in Collective Adaptive Systems (CAS) [36, 40]. The last type, CAS, are distributed large-scale systems that are flexible in terms of number and type of resources, including human resources, and include complex task execution with a high number of interactions between resources. In these systems, privacy is even more relevant than in crowdsourcing, where tasks are simple and interactions are not common. Moreover, the utilization of cloud services that are inherent in the definition of CAS adds a lot of weight to the importance or privacy; using public clouds requires a high amount of trust on the providers as user data, and generated content and artifacts are usually dependent on using cloud services but are also stored on the Cloud. Security mechanisms for clouds, along with regulations on data management, are of paramount importance if privacy is to be preserved. While there is a solid amount of work on building human computation systems and mechanisms that support efficient management, such as task assignment and management (e.g., routing and delegations), worker management, incentive and payment models for workers online, and quality assurance, little research has been conducted on the privacy implications in these systems. Privacy, however, is a human right and as such these users are also entitled to it. Article 12 of the Preamble of the Universal Declaration of Human Rights states, for example, that “No one shall be subjected to arbitrary interference with his privacy, family, home or correspondence, nor to attacks upon his honor and reputation.” The design of human computations systems as well as everything else on the web needs to be guided by privacy-preserving principles. The key contributions of this paper are: (1) a discussion of why and how user data is utilized in particular human computation systems, (2) an analysis of user privacy awareness on human computation platforms through the results of a user study that we conducted with an online survey, and (3) recommendations for human computation stakeholders, along with some research directions. The remainder of this paper is organized as follows: In Sect. 2, we present related work. Section 3 presents a discussion on privacy implications in human computation, strictly speaking about data collection, data utilization, and privacy risks. In Sect. 4, we present our user study and make an analysis of user privacy awareness on human computation platforms. We present a few recommendations for privacy-preserving mechanisms in Sect. 5 and provide some considerations for possible research direction in Sect. 6. We conclude the paper in Sect. 7.

Privacy in Human Computation: User Awareness Study, Implications for. . .

249

2 Related Work Smith et al. have studied individual privacy concerns in organizations and have identified multiple dimensions of these concerns that they have presented in [15]. In that work, the authors list four factors critical to consider when assessing user privacy concerns: concern over data collection, errors in user data, unauthorized secondary use of user data (e.g., when data is used for other purposes than stated), and improper access to user data. Since then, a number of other models, frameworks of user privacy concerns have been presented (some building upon the work of Smith et al.) such as in [21]. On the other hand, theories and models of how to control privacy and enforce privacy-aware mechanisms are also present in the existing literature. Authors in [22] present such a privacy-control theory and discuss ways of its application in online environments. Fischer-Hübner and Martucci in [13] have inspected privacy implications for Social Collective Intelligence Systems (SCIS) by presenting an overview of the European Data Protection Legal Framework and relating the privacy rules provided by this framework with SCIS systems and mechanisms supported by these systems, such as user profiling for reputation scores, incentive models as well as data provenance. Reputation, incentive models, and data provenance are all listed as risks for user privacy as these mechanisms are inherently designed to work with user profiling. However, the authors list and discuss some of the available tools and technologies that can enable SCIS platform providers to respect and preserve user privacy, mainly via pseudonyms and anonymity. One example is by allowing users to use different pseudonyms for different roles, i.e., context-based pseudonyms that can be used only once per role (e.g., skill type) and thus prevent misbehavior by malicious users. In addition, the authors describe anonymous credential protocols that can be used to create new credentials whenever a user wants, with less or different certificate attributes, which cannot be linked to an original certificate by the verifier and the issuer. Moreover, Fischer-Hübner and Martucci also present Privacy Policy Languages (such as PPL) with which platforms can make negotiations and come up to an agreement with platform users on how, by whom (and what) data can be accessed, processed, and logged. Motahari et al. in [23] have listed privacy threats in ubiquitous social computing by underlining the social inference threats in social computing, where a user can be identified, for example, through contextual information (e.g., location) or social links. Authors in [14] present a privacy model together with a framework for task recommendation in mobile crowdsourcing. The model is based on enabling workers to share information (e.g., location) with a recommendation server by choosing how much and what type of information they wish to share. Task recommendations are based on the information shared by workers. However, authors conclude the obvious, namely that achieving a high efficiency in task recommendations means low level of privacy. Another work that presents a privacy-aware framework is presented in [34] by To et al. The authors present a task-assignment algorithm that preserves location privacy for mobile spatial-crowdsourcing tasks, that is, for tasks that require

250

M. Riveni et al.

workers to be at a specific location. Toch in [35] investigates privacy preferences of users in mobile context-aware applications through crowdsourcing and presents a method to calculate user privacy tendencies. He suggests building distributed systems to tackle privacy risks (with the computation of user privacy tendencies being executed on the client side). Privacy preservation in decentralized systems is discussed in [5]. Privacy activists also advocate for decentralization and zeroknowledge systems; Balkan, for example, has written the Ethical Design Manifesto [4], which states: “Technology that respects human rights is decentralized, peer-topeer, zero-knowledge, end-to-end encrypted, free and open source, interoperable, accessible, and sustainable.” Langheinrich in [19] discusses some Privacy by Design principles, focusing on ubiquitous systems. He states that the Principle of Openness, or Notices, is an important principle during data collection. Users have to be informed when they are being monitored. In addition, Langheinrich discusses that consent should be required in a more flexible way than the “you can use our services only if you consent to our terms/policy.” Users have to be able to use services while opting out of unwanted features.

3 Personal Data on Human Computation Systems 3.1 Collected Data We investigated the collected data from some of the existing crowdsourcing/expertlabor market platforms (by actually creating accounts), such as Amazon Mechanical Turk, Microworkers, Freelancer, Upwork, PeoplePerHour, TopCoder, and uTest. The information required to build up a profile and/or verify a profile sees variations from platform to platform. The following list provides some data required to build and verify profiles generalized through platforms; not every platform requires everything on the list, but everything on the list is required in various platforms.1 – Full mailing address—Sometimes even documents are required to prove address, such as utility bills or bank statements. – A government issued ID—Passport, ID card, or driving license. – Photograph – Code verification along with a user’s face on a photograph—In some platforms, workers need to send a photograph of themselves where they hold a piece of paper with a code provided by a platform written on the paper.

1 We have investigated the information

required by platforms in order to gain a better insight and to guide our discussion in this work along with providing improvement suggestions for researchers and industry alike, which could bring more privacy-aware platforms in the future. We do not intend to imply malicious use of the users’ data by the aforementioned platforms as we have not conducted an investigation regarding the manner of usage of the collected data.

Privacy in Human Computation: User Awareness Study, Implications for. . .

251

– Educational experience—In some of the platforms, filling out at least one educational experience is mandatory. – Job title—In some platforms, filling out a job title is mandatory. – Bank account information – Data from mobile sensing—Depending on the application domain, other sensitive data may be collected at runtime while users are working on tasks or just wearing a smart device. For example, in mobile crowdsourcing applications, such as crowd-sensing, location information, health information (that users share through wearables to applications), and other data may be collected which can be used to identify and profile a user. – Device and connection data—Basic system fingerprinting such as IP address, browser type, and operating system. The EU Data Protection Directive (DIRECTIVE 95/46/EC) [11] stipulates that personal data is “any information relating to an identified or identifiable natural person.” Although this directive is repealed by the General Data Protection Regulation [12] (with the enforcement on 25 May 2018), the principal definition for personal data remains the same, as can be found in Article 4 of the Regulation (under (1)). Moreover, identification is the singling out of an individual within a dataset [24], even if his or her name or other attributes that we typically associate with an identity remain unknown. Consequently, most of the aforementioned information can be used to identify a person and that means that this data is personal and thus should be kept private.

3.2 Reasons for Collecting Personal Data The General Data Protection Regulation (REGULATION (EU) 2016/679) [12] defines profiling as “any form of automated processing of personal data consisting of the use of personal data to evaluate certain personal aspects relating to a natural person, in particular to analyze or predict aspects concerning that natural person’s performance at work, economic situation, health, personal preferences, interests, reliability, behavior, location or movements.” In Human Computation, the collection of user data and building user profiles is usually not conducted for big-profit purposes (such as selling user profiles to advertisers) but for designing and developing mechanisms for effective work management on human computation provisioning systems. In the following, we discuss some of these mechanisms. Task Assignment and Formation of Collectives A number of existing works have presented (expert) discovery and ranking algorithms for service-oriented architectures with human-provided services [10, 29], as well as non-service-oriented systems in which task assignment is based on qualifications. Research in team formation in expert networks is also being investigated [2, 8]. These mechanisms are all based on logged and historical data about workers, for discovering appropriate ones for specific tasks in individual-crowdsourced work or specific collective work.

252

M. Riveni et al.

Management Mechanisms Managing individual and team-based worker performance in human computation is also important for complex systems with automated processes even for the human-in-the-loop coordination. Adaptation mechanisms in human computation within SOA, for example, are discussed in [31], and an adaptation mechanism for elastic collectives based on a trust model is presented in [27]. Algorithms that calculate worker’s performance can be used within delegation mechanisms, where a task is delegated to another worker that may or may not belong to the initial collective. Consequently, a new worker can be added to an existing collective at runtime. Research on both task-assignment and adaptation algorithms that include measurement of worker performance is in big part based on trust and reputation models. Some of these trust and reputation models include not only metrics that can be measured automatically (such as task success rate) but also social trust, which mostly is defined and calculated as a trust score that is given to a worker by collaborators or acquaintances and/or by work requesters/clients based on their satisfaction by the results. Social trust is often subjective and in most cases requires that the person rating a worker knows the worker personally, so in some trust and reputation models worker identities are needed to be known. Hence, in some cases reputation mechanisms are in conflict with privacy, and we need to find ways of bringing these two concepts together and provide privacy aware reputation strategies. Quality of Service Keeping quality at a desired level also requires monitoring of workers. (See an example of service-level agreement (SLA)-based QoS monitoring for crowdsourcing in [16].) Misbehavior Prevention Personal data is sometimes used in building mechanisms for preventing worker misbehavior, such as Sybil attacks—cases in which workers can easily create multiple profiles and make more gain by executing the same tasks multiple times. Incentive Mechanisms Developing and applying appropriate incentive models for workers based on monetary and nonmonetary gains is also based on data collection, and sometimes personal data, for example, when reputation is used as an incentive. (See work on incentives and rewarding mechanisms in Social Computing in [28].) Payments In most platforms that have implemented monetary incentives, user bank/credit account is required so that workers can be paid. However, some platforms allow users to delete this data after they withdraw the required amount from it (such as microworkers.com, in the case of requesters/clients). This should be a standard practice for all data types as well. A regulation example for this is the GDPR’s Right to erasure described in Article 17. We are of the opinion that human computation platforms should consider the provisioning of mechanisms with which subjects would have the possibility to directly modify, update, or erase data without making a request to the controller (the owner of the crowdsourcing platforms), just like Microworkers allows the deletion of bank account information without any explicit requests.

Privacy in Human Computation: User Awareness Study, Implications for. . .

253

All the aforementioned mechanisms require monitoring of workers. However, in research frameworks and systems, no collection of personal data is mentioned explicitly, and the privacy aspect is not tackled, except in specific research presented in Sect. 2. On the other hand, almost all existing platforms in industry require users to share their personal data. Thus, we advocate for all the aforementioned mechanisms to be considered together with their privacy-related implications. In the next section, we examine privacy implications by analyzing a few risk factors that we identified as most relevant.

4 Privacy Risks User Privacy Policy Awareness Users usually agree to Privacy Policies without actually reading them, and they do not read them because they are too long or too complex to understand [38]. This contributes to their use of platforms without being aware of what they are entitled to, regardless of whether their privacy rights are violated or not. Lack of Transparency in Privacy Policies Very often, users are not given complete information about what is considered under personal information, about how their personal data will be used, whether it will be shared with third parties and for how long will this information be stored on a Human Computation platform provider’s servers or on the servers of their service providers. Often, the use of personal data is defined in policies in a vague way. Consider, for example, the statement “we may share certain data. . . ,” using words such as “certain data” without concretely defining what type of data the statement is referring to means getting a clear consent by users to use whatever personal data of users that the provider owns. Many privacy policies also contain phrases such as the following: “we may share information with third parties for industry analysis, research and other similar purposes”; the terms “similar purposes” give the providers the freedom to use personal data in any purpose fitting their needs, without the explicit consent of the platform users. In addition, consider this statement: “we may use your personal information from other services and connect to your account information when necessary,” this is clearly a way to de-anonymize users even if the platform is designed to use pseudonyms, as combining various datasets (e.g., by email addresses) de-anonymizes users. Selling user profiles to the ad industry is also a possibility, although many human computation platforms do not have an ad-based business model. However, companies called data brokers continuously collect data about people from multiple online (and even offline) sources and sell that data to clients in various business domains. Data that they collect can be also retrieved from websites with log-ins, browser, and device fingerprints. Because privacy policies are not straightforward, people cannot be sure whether data brokers are not leveraging some data from human computation platforms as well.

254

M. Riveni et al.

Profiling User information is collected through information that they share with platforms as well as by automatically collecting data by tracking. Among other uses such as computing reputation scores, this data may also be utilized to group people in different categories, by various contexts (e.g., country of residence, and gender). This may be used to set up rules for task assignment, and discriminate certain groups during task assignment and rewarding. As we mention in our study results, there are cases where workers from certain countries are rewarded less than others for the same tasks. In addition, consider, for example, a real danger through crowdsourced tasks for online monitoring of certain locations with the purpose of detecting and reporting criminal activities, or even a different setup, such as identifying wanted offenders from a set of pictures that are posted in a crowdsourced task online. If these platforms are profiling workers, criminals might find ways to identify workers, and workers’ lives could be put at risk. Furthermore, if information from political crises–response crowdsourcing sites2 about people reporting incidents in war-struck areas fall into the wrong hands, it might also pose risks for the reporters, who might be non-tech savvy citizens. These may seem as extreme examples, but may serve well to validate privacy concerns. On the other hand, many companies hide behind the term “anonymity,” for example, they do not require real names and allow people to register with pseudonyms while collecting other personal data, in this way wrongly convincing people that they actually work online without being identified. For example, authors of [6] cite a study revealing that the combination of zip code, gender, and birth-date data had been unique for 216 million US citizens, and consequently citizens can be identified without any other additional data. In the same work, authors also cite another study showing that four data points, such as four sets of time and location data, could be used to uniquely identify people. Thus, leaving out some data while collecting other type of data does not mean that anonymity is achieved, and in most cases people are not transparently informed of this fact. Lack of Control In the current systems, users do not control how their private information is used (whether it is shared, sold, or misused), and have no control over who accesses that information. They have to be content with what they read on privacy policies (when they read them). They are not given control to their own data, to update or delete their data when they want. Moreover, stored information is sometimes not secured enough, rules and regulations are not always respected, and data stored on foreign servers belonging to a different jurisdiction than a person’s residence country (over which the user may not be given a choice) can be misused (e.g., due to security breaches, unencrypted data, unethical employees, or security agencies). Lack of Ownership Users do not own their own data that they have shared with the platforms, rather platform providers do. Similar to the risks in not being able to control data, not owning data means intentional or unintentional sharing with third

2 See,

for example: https://syriatracker.crowdmap.com/.

Privacy in Human Computation: User Awareness Study, Implications for. . .

255

parties, access to user data by unintended parties as well as transferring/selling user information to other parties and monetizing people’s data, which sometimes can be done without users’ knowledge and approval. Lack of Security Last but not the least, in addition to the aforementioned factors, security is of paramount importance for protecting privacy. Data control and ownership do not have any effect on privacy protection if user information is unencrypted. Needless to say, security protocols for securing internet connections and data encryption protocols should be standard features that every human computation platform should support.

5 Study 5.1 Method: Survey Design and Distribution We conducted a study to asses user privacy awareness in human computation with an online questionnaire, hosted on a server at Technische Universität Wien/TU Wien. We asked participants a series of questions that we designed specifically to get their opinion on their private data collected and utilized on the platforms, to get their knowledge on privacy implications on these platforms as well as their concerns. We disseminated our survey in two ways: (1) by sharing it with fellow researchers and colleagues by email, and with acquaintances and friends on social networks by asking them to fill it in (if registered as users on these systems) or send the survey to people that they know are using these systems as requesters or workers; and (2) by creating a task/campaign at Microworkers (https://microworkers.com/) and a HIT batch on Amazon Mechanical Turk, asking workers to fill in our survey (at a given link). We did two rounds of the survey, as we came up with some more questions that we saw relevant during our study. Thus, the first round of the survey had 20 questions, 16 of which were designed to assess user privacy awareness, and 4 were statistical questions to get demographic data. We submitted this survey on Microworkers and to researchers and freelancers through private communication, while the second round of the survey had 5 additional questions, and was conducted only Amazon Mechanical Turk. Where we have less participants for the newly added questions, we mention it when discussing the particular questions. Microworkers has a design that allows requesters to select what user base to choose depending on worker country information and makes payment recommendations according to country-dependent ratings. We created four tasks/campaigns and asked users from four different country groups to fill in our survey. For the first three groups of workers, we paid workers $0.42 per task, as by investigating other studies on these platforms (that have used payments between $0.10 and $1) we concluded that this amount was enough for people who are interested on the topic to accept the task and low enough to discourage misbehavior by those who may want to fill in the survey without interest and spam the results. In spite of recommendations

256

M. Riveni et al.

for lower payment for a group of workers coming from countries rated lower, we decided to pay workers of a lower rated group of countries the same amount of $0.42 and not lower. However, our survey task for a fourth group of workers residents of high rated countries was rejected for that amount as the minimum payment was $1.00 per task for surveys such as ours, so we paid that amount to get our survey completed by a fourth group, as we needed different demographics. In this regard, we strongly encourage ethical payment methods by crowdsourcing platforms and ethical payment behavior by work requesters/clients. Some of these mechanisms could be equal pay per task type for all workers (regardless of country ratings), or individual worker payments based on quality of results. Nevertheless, this side experiment allowed us to qualify the answers that were submitted in reply to our survey, for example, filling in optional questions that required more elaboration. In this context, we noticed no difference in the answers of higher rated countries compared to lower rated groups. In fact, some users who were paid less gave elaborated answers while none of the higher paid participants did this. Workers on Amazon MTurk were paid $1. To filter submissions as well as to avoid spammers and malicious users who fill in the survey without reading the questions and answers, we included a few questions that helped us assess (to some level) the honesty of participant answers. One such “testing” question was added to check a related “yes or no question,” the testing question included radio buttons with more elaborating statements to be chosen if the user answered with “Yes” on the related question, and included a radio button with the statement “I answered with “No” on the previous question” to be selected if a participant has answered with “No” on the related question. In addition, we added a “Yes or No question” asking survey participants if they have read our consent form for the survey and excluded submissions of participants who answered with “No.” A few participants had filled our survey multiple times. We counted only one submission from these participants and excluded all duplicates.

5.2 Results and Analysis Demographics We had a total of 204 participants, the answers of three of which we excluded as a consequence of their negative answers to the question with which we requested participant consent for the survey (through reading our consent form). One hundred and five participants were workers from Microworkers, 78 from Amazon MTurk (engaged in the second round of our study), whereas 21 were participants who we contacted by mail/social networks. Most of our participants were residents of the USA, EU countries, and India. Table 1 shows the level of education and IT knowledge of our participants. Participants with a PhD and Postdoc level of education were users of human computation platforms as requesters of work (for research purposes).

Privacy in Human Computation: User Awareness Study, Implications for. . . Table 1 Demographics of participants

257 Percentage

Education Primary school High school Undergraduate studies/BSc/BA MSc/MA/Specialty training Dr/PhD Postdoctoral researcher Other IT knowledge Expert/professional Medium level (good IT skills but not expert/professional) Knowledge to get around online

1.97 20.6 40.6 15.6 6.86 1.4 12.7 20 59 21

We asked participants to fill in the names of up to three platforms that they use and we got the following variety of platforms as answers: Microworkers, Amazon Mechanical Turk, Upwork, InnoCentive, Elance, Guru, 99designs, CrowdFlower, clickworker, RapidWorkers, ShortTask, Testbirds, cashcrate, fiverr, scribie, TranscribeMe, foulefactory, ideaCONNECTION, and OneSpace. Most of our participants worked on two or three platforms. Privacy Awareness We posed three type of questions assessing user privacy awareness and concerns: the first was related to (a) data collection, control, and ownership of data; the second was related to (b) anonymity online; and the third group of questions was related to (c) regulations and policies. In the following, we discuss the results. (a) Data collection, Usage Concerns, and Security Regarding data collection, we asked participants to state their level of concern regarding the fact that they have to share sensitive data to register and verify their accounts. The level of concern question was set up as a 1–5 Likert scale. 19% of participants stated that they are somewhat concerned with the information that they are obliged to share when they register on the platforms, while 25% were not concerned at all. Table 2 gives a more detailed overview of the answers regarding participant concerns over the collection of their personal data. In addition, to get more detailed information, we listed some of the data types (mentioned in Sect. 2) collected by platforms and asked participants to select which of the given data type they would want to hide. They could choose multiple types. Most of the participants answered that they would prefer hiding: a government issued ID (68%), bank account information (54%), and phone numbers (61%). More details regarding the results for this question are given in Table 3. Next, we asked participants if they ever provide false information when registering and creating their profiles and 17.7% reported that they do. Regarding the reasons for providing false information, 19% reported that they do not feel

258

M. Riveni et al.

Table 2 Data concerns Survey statements How much are you concerned with the personal information that you are obliged to share so that you can register on these platforms? How much are you concerned with the personal information that you need to share so as to build your profile and verify your identity on these platforms? Are you concerned that your information will be misused (by the platform that you are registered with)?

Very Somewhat Neutral Not concerned Not at all Likert score 19% 8% 39% 9% 25% 2.78

19% 22%

25%

12%

22%

2.83

6%

23%

28%

32%

2.16

11%

Table 3 Most common collected data Collected data Name and surname E-mail address Phone number Birthdate Photograph Location information/mailing address Utility bills (sometimes used to verify address) A government issued ID Bank account information None of the above

% of users who want to hide the data 24 22 61 26 54 35 53 68 54 1

comfortable revealing some specific information about them, and 3% reported that they provide some false information in order to create secondary accounts. Users’ knowledge on where their data is stored is important for assessing their privacy awareness; hence, we asked participants whether they know and whether they are concerned if their data is stored on platform providers’ own servers, or if platform providers utilize Cloud services (in which case an agreement should exist between platform providers and Cloud providers for protecting user information by not sharing user data with other parties), and if data is stored in a location with a different jurisdiction (in which case different data protection regulations exist). Participants were given three answer choices and they reported as follows: 32.29% said “I admit I have never thought about these things and frankly I am not concerned.”, 33.86% chose “I admit I have never thought about these things but I

Privacy in Human Computation: User Awareness Study, Implications for. . .

259

Table 4 Security-related statements Survey statements Having in mind that not only my personal information but also content that I produce or expect as a result from engaging on a microwork platform is sensitive, I expect the platform I perform micro-work on to prove on a regular basis (every three months?) that it is secure, by having an independent pen test performed and have the results published “I would like to have the option to receive payouts in a privacy-friendly cryptocurrency.” Please select a choice from 1 to 5, 1 indicating that you are not concerned with secure and private payouts, 5 indicating that you strongly agree with the statement

I strongly agree 15.06%

I agree Neutral I disagree 24.68% 36.98% 9.59%

I strongly disagree 13.69%

10.96%

20.55% 32.88% 10.95%

24.66%

Likert score 3.18

2.82

became concerned now.”, and 33.85% answered with “I have thought about these things and am concerned.” Table 4 shows security-related statements that we added for the second round of the survey and the replies from 78 participants recruited from Amazon MTurk. (b) Anonymity Anonymity is of course fundamentally different from privacy. Privacy means that people may be identified online, but it should be their choice regarding how much and in what way their data is shared and utilized. Nevertheless, the two concepts are invariably related. Hence, we examined opinions on anonymity

260

M. Riveni et al.

as well, and asked participants what would be the reasons in the case they prefer to work anonymously. Some workers who are working on more complex tasks (e.g., projects such as those posted on 99designs, and Freelancer) and not on micro-tasks stated that they would not prefer to work anonymously online, using statements such as “working anonymously is not effective.” Consequently, we assume that these workers do care about reputation as reputation mechanisms bring more clients and work. However, some replied that they would want to work anonymously in cases if they are working on some projects on which they would not want to put their name on but they are well paid, and communication and collaboration with clients is satisfactory. In addition, one participant answered that it would be nice if users are provided with the option to work anonymously online whenever they choose to (opt-in/opt-out). On the other hand, most of the workers who work on micro-tasks answered that they would want to work anonymously for several reasons: they do not want their name to be associated to the type of work they do, to protect their banking information, they do not want the companies with which they work full time to know that they are doing a side job. We quote some answers stating other contexts of concern for anonymity: “I would want to work anonymously so there was no bias towards me based on my demographics and/ or social class. I also would prefer to remain anonymous in case scammers entered the platform pretending to collect data, but instead, they were going to participants homes etc.,” “I like minimizing my digital footprint as much as possible,” “When doing microwork online, you do work for various people, potentially over dozens of people a day. I’d rather not have my sensitive information potentially available to all of them, when I’m forced to provide demographic information for much of the work anyway,” “I’d not want to have that information available for marketers or to be available to be sold. I’d not want other organizations to be able to access such information and use it to send me ads or other materials,” and others. In addition, some answered that they would want to work anonymously because they do not want their earnings to be reflected on their taxes. Related to the latter, one participant stated that he uses foreign money transfer services, such as Payoneer, to avoid taxes for online work. Furthermore, some participants stated their concern of their information being leaked to other parties. One particular participant stated that the reason he would want to work anonymously is that he cannot be certain by whom and how his private information will be used, he added the statement “I want to control my “web” identity as I want.” Thus, we can conclude that workers doing complex tasks are more inclined to identification than workers executing micro-tasks that are easy to execute. Lastly, an interesting answer we encountered was: “I would want to protect my privacy,” even though the specific question was related to anonymity online. Consequently, participants associated anonymity with privacy. (c) Regulations and Policies To assess participant engagement in privacy issues, we went a step further and asked them if they read privacy laws, directives, and policies. Figure 1 provides their reports. Interestingly enough, more than 50%

Privacy in Human Computation: User Awareness Study, Implications for. . .

261

Have you read the data protection law or an online privacy protection law of the country in which you work or provide you services for (or any other regulation for protecting online personal data)? Do you usually read the Privacy Policies on the platforms on which you register online?

Have you read the EU Data Protection Directive? Select “No answer” if you do not live in the EU.

0.00%

Yes

10.00%

No

20.00%

30.00%

40.00%

50.00%

60.00%

Not in EU

Fig. 1 User reports on regulations and policies

of participants reported that they do read privacy policies when they register on platforms. However, around 40% of participants reported that they do not read them. We take this result to be truthful as the number of participants is small. If we had a bigger number of participants we assume that this ratio would be significantly in favor of participants who do not read privacy policies because of complexity, as existing research suggests (see [17, 38]). In order to get participants’ opinion on platforms, research, and regulations on privacy, we included a question with five Likert-scale agreement levels to choose from, for a few statements that we compiled. Most participants agreed that the existing platform providers need to be more transparent about how they use personal information and they also agreed that research and industry should increase their efforts in providing mechanisms that will enable people to control and even own their data. For every statement, we also asked participants if they are knowledgeable on the topics that the statements refer to or not. In total, 57% answered that they need more information on the topics, 38% answered that they have knowledge on the topics and the rest did not answer. Detailed results are given in Table 5.

6 Suggestions 6.1 Recommendations For platform providers, an important privacy-respecting guide in storing personal data is to only store that which is essential to the needs of the platform. With this, we mean the concept of data minimization; if n data points are enough to perform the task for which these points were collected, do not collect > n data points. In the

262

M. Riveni et al.

Table 5 Opinions on regulations, and approaches in research and industry Survey statements Human computation companies/platforms should clearly state under which country/state law they operate Companies/platforms than enable and provide human computation should be more transparent about how they use my personal information I am concerned about the privacy regulations/laws of the country in which I reside and work Research and industry should increase their efforts in giving users more control over the use of their data Research and industry should increase their efforts in enabling tools and mechanisms that will enable users to own their own data, in contrast to current standards where companies own users’ data

I strongly agree 28.13%

I agree Neutral 41.67% 25%

38.54%

29.67% 21.36%

15.1%

I somewhat disagree 2.6%

I disagree 2.6%

Likert score 3.9

7.31%

3.120

3.84

30.73% 31.25%

18.23%

4.69%

3.3

30.73%

38.02% 19.79%

6.77%

4.69%

3.83

29.68%

33.85% 28.13%

4.17%

4.17%

3.81

most general terms, according to Colesky et al., in [7], there are two directions of strategies to protect the privacy of clients: data oriented, and policy oriented. These two directions lead to eight high-level strategies that can be applied to the collection of data in ways that respect the privacy of data subjects. One method of identity protection that lies in between tools for users, and something that developers will have to implement is attribute-based credentials (ABC) [18]. In short, ABC provides the client with a set of credentials such as “over 18” or “holds an MSc in computer science.” The classic example of the use of ABC’s is in buying alcohol: A person must be of legal age in order to buy alcohol,

Privacy in Human Computation: User Awareness Study, Implications for. . .

263

but the only classic way to prove this is by showing an identity card [1]. This card contains more information than is needed, such as exact date of birth, social security number, and full name. The only requirement to know is “over 18,” which in ABC’s can be presented as an attribute. The beauty of this system is that the attribute itself is not trackable, the next time the same client needs to prove the same attribute, this instance is unlinkable to the previous instance. Thus, attributes can be used to hold all information that is needed for the service, placing them under the control of the client, rather than storing them on the server in a user profile. Wellknown anonymization methods are the k-anonymity model, presented in [33], and t-closeness described in [20], which prevents attribute disclosure going beyond the limitations of k-anonymity . However, these strategies do not solve all the practical problems. In the end, a client still has to provide at least some personal data in order to receive a reward for his work, such as a payment method. In this context, one could argue that cryptographic currencies such as Bitcoin could be used to pay rewards, but these too have been shown not to be entirely anonymous [3]. In addition, there are other alternatives of online services that offer means to transfer funds, such as CashU, and Perfect Money. These services typically let one transfer funds from a legitimate source (such as a bank account) to them, and then allow transfer between accounts within the service itself. As such, these services can be seen as a “Trusted Third Party” for money exchange. Although tracking is thus made more difficult, it is still quite possible, as there is a single party that still needs to know enough to be able to transfer funds. Ideally, a completely anonymous client would perform work, and be rewarded in an untraceable way. The technologies exist to make this possible, but as far as we know, no human computation platform has implemented multiple privacypreserving technologies yet, still relying on cheaper, faster, and easier management methods that (might) erode the privacy of clients. The following section mentions a few possible research directions.

6.2 Research Directions Transparency with Rules or Service-Level Agreements System designers, developers and business actors need to come up with more transparent and direct ways of getting user consent (other than the current standard of publishing privacy policies). An interesting open challenge that we will tackle in our future work is our idea of enforcing user consent through SLAs. Depending on the type of (human) tasks and whether their execution can be monitored and measured (see some metrics in [27], introducing SLAs may come as appropriate as a mechanism to monitor, manage, and adapt human computation collectives. In relation to the aforementioned SLA application, a possible research direction is investigating the inclusion of privacy clauses (e.g., from privacy policies) in SLAs so that users will be obligated to read them and give consent when negotiating SLAs.

264

M. Riveni et al.

This could be a two-way negotiation, employers could regulate personal data and content/artifact privacy in relation to the workers as well as the system, and workers could regulate their personal data in relation to employers, other workers, and the system. Privacy-Preserving Workflows Human-based computing in general and crowdsourcing in particular, in addition to issues with personal information, have issues with sensitive artifacts and data submitted for tasks. People who submit tasks may want to reveal only a part of the data. Thus, the design of workflows that provide enough knowledge for workers to be able to execute tasks but do not disclose the full context of requesters’ work/interest is an open (domain-dependent) research question (example work presented in [32]). Payment Methods When people work in socio-technical systems individually and do not belong to an organization, even with the most efficient anonymization methods, e.g., on the assumption that all worker data is private, the payment methods are still an open question as they can be still used to identify a person. Location Let us assume that a person consents to his/her location data being collected, e.g., in a crowdsourced traffic management of a city. In this case, developers need to pay attention, for example, to set some location checkpoints, which would not be used to infer sensitive information, such as, for example, religion (checkpoints near religious buildings), hospitals, houses, and other institutions. Evaluation Methods An interesting research challenge are also evaluation methods for software, evaluating the included privacy-preserving mechanisms. Raising People Awareness About Privacy Methods and techniques for raising awareness on privacy should not be a question tackled by experts working on social and legal areas only; it is crucial that computer science researchers approach these challenges as they develop software and disseminate their research. This section provides only a discussion of possible mechanisms for preserving privacy, and possible research challenges, and it is not an attempt to provide an extensive list of tools, strategies, and problems; the goal is to provoke and motivate researchers of human computation systems to tackle privacy challenges.

7 Conclusions The goal of this research was to get an insight into user privacy awareness for human computation platforms. As the reported results show, we may conclude that users are moderately concerned for their privacy on these platforms. This is partly because they are willing to show their reputation publicly, and partly because they are not informed enough about how their personal data is collected and processed. Most of our participants stated that they became concerned after they read the statements in our survey concerning privacy implications in these systems.

Privacy in Human Computation: User Awareness Study, Implications for. . .

265

The lack of privacy awareness is a key factor why corporations today can leverage the power of personal data collection and analysis, which in the majority of cases is done without peoples’ knowledge or consent. Thus, academia, industry, and the civil society needs to focus more on improving awareness. In addition, privacy protection laws and regulations need to be enforced for privacy policies to make sense. Lastly, we recommend system developers and businesses to be guided by principles which respect users’ privacy and include privacy-preserving settings by default. Exploring and building mechanisms that will encourage users to read privacy policies and to prove that they have read them is an important research direction, because awareness and consent are the most important elements in privacy protection and preservation. Generalizing the importance of the topic, we advocate that research from every area in computer science should progress with having the context of privacy in mind, as the way we build applications and systems affects the direction in which our societies will further develop. Technology and society are in an interminable process of mutual effect on their change and transformation, in which process the construction of reality takes place; it is up to us to choose and build the type of reality we want to live in, and what better than when progress is accompanied by transparency, trust, and consequently, autonomy. Acknowledgements We would like to thank all participants of our survey, for providing us with their opinions and enabling us to gain an overview of their privacy concerns.

References 1. Alpár, G., Jacobs, B.: Credential design in attribute-based identity management. In: 3rd TILTing Perspectives Conference on Bridging Distances in Technology and Regulation, pp. 189–204 (2013) 2. Anagnostopoulos, A., Becchetti, L., Castillo, C., Gionis, A., Leonardi, S.: Power in unity: forming teams in large-scale community systems. In: CIKM, pp. 599–608 (2010) 3. Androulaki, E., Karame, G.O., Roeschlin, M., Scherer, T., Capkun, S.: Evaluating user privacy in bitcoin. In: International Conference on Financial Cryptography and Data Security, pp. 34– 51. Springer, Berlin (2013) 4. Balkan, A.: Ethical design manifesto Ind.ie. Online available: https://ind.ie/ethical-design/. Accessed 22 July 2017 5. Buchegger, S., Crowcroft, J., Krishnamurthy, B., Strufe, T.: Decentralized systems for privacy preservation (Dagstuhl Seminar 13062). Dagstuhl Rep. 3(2), 22–44 (2013) 6. Christl, W., Spiekermann, S.: Networks of Control, 1st edn. Facultas, Vienna (2016) 7. Colesky, M., Hoepman, J.h., Hillen, C.: A critical analysis of privacy design strategies. In: IWPE. IEEE, San Jose (2016) 8. Dorn, C., Dustdar, S.: Composing near-optimal expert teams: a trade-off between skills and connectivity. On the Move to Meaningful Internet Systems: OTM 2010, pp. 472–489 (2010) 9. Dustdar, S., Bhattacharya, K.: The social compute unit. IEEE Internet Comput. 15, 64–69 (2011) 10. Dustdar, S., Truong, H.L.: Virtualizing software and humans for elastic processes in multiple clouds – a service management perspective. Int. J. Next-Generation Comput. 3(2) (2012)

266

M. Riveni et al.

11. European Directive: 95/46/ec of the European parliament and of the council of 24 October 1995 on the protection of individuals with regard to the processing of personal data and on the free movement of such data. Official J. EC 23(6) (1995) 12. European Regulation: Regulation (eu) 2016/679 of the European parliament and of the council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46/ec (general data protection regulation). Official J. EC 119 (2016) 13. Fischer-Hübner, S., Martucci, L.A.: Privacy in social collective intelligence systems. In: Social Collective Intelligence: Combining the Powers of Humans and Machines to Build a Smarter Society, pp. 105–124. Springer, Cham (2014) 14. Gong, Y., Guo, Y., Fang, Y.: A privacy-preserving task recommendation framework for mobile crowdsourcing. In: IEEE Global Communications Conference, GLOBECOM, Austin, pp. 588– 593, 8–12 December 2014 15. Jeff Smith, H., Milberg, S.J., Burke, S.J.: Information privacy: measuring individuals’ concerns about organizational practices. MIS Q. 20(2), 167–196 (1996) 16. Khazankin, R., Psaier, H., Schall, D., Dustdar, S.: Qos-based task scheduling in crowdsourcing environments. In: Proceedings of the 9th International Conference on Service-Oriented Computing, ICSOC’11, pp. 297–311. Springer, Berlin (2011) 17. Kluver, L.: ICT and Privacy in Europe. Experiences from technology assessment of ICT and Privacy in seven different European countries. Final report October 16, 2006, European Parliamentary Technology Assessment network (EPTA), Wien (2006) Online available: http:// epub.oeaw.ac.at/?arp=0x0013038d. Accessed 26 June 2016 18. Koning, M., Korenhof, P., Alpár, G.: The ABC of ABC – an analysis of attribute-based credentials in the light of data protection, privacy and identity. Petsymposium (2014). https:// www.petsymposium.org/2014/papers/Koning.pdf 19. Langheinrich, M.: In: Privacy by Design – Principles of Privacy-Aware Ubiquitous Systems, pp. 273–291. Springer, Berlin (2001) 20. Li, N., Li, T., Venkatasubramanian, S.: t-closeness: privacy beyond k-anonymity and l-diversity. In: Chirkova, R., Dogac, A., Özsu, M.T., Sellis, T.K. (eds.) ICDE, pp. 106–115. IEEE Computer Society, Los Alamitos (2007) 21. Malhotra, N.K., Kim, S.S., Agarwal, J.: Internet users’ information privacy concerns (IUIPC): the construct, the scale, and a causal model. Inf. Syst. Res. 15(4), 336–355 (2004) 22. Moloney, M., Bannister, F.: A privacy control theory for online environments. In: 42st Hawaii International Conference on Systems Science (HICSS-42 2009), Proceedings (CD-ROM and online), 5–8 January 2009, Waikoloa, Big Island, pp. 1–10 (2009) 23. Motahari, S., Manikopoulos, C., Hiltz, R., Jones, Q.: Seven privacy worries in ubiquitous social computing. In: Proceedings of the 3rd Symposium on Usable Privacy and Security, SOUPS ’07, pp. 171–172. ACM, New York (2007) 24. Party, D.P.W.: Opinion 4/2007 on the concept of personal data. European Commission, Brussels (2007) 25. Quinn, A.J., Bederson, B.B.: Human computation: a survey and taxonomy of a growing field. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI ’11, pp. 1403–1412, ACM, New York (2011) 26. Riveni, M., Truong, H.L., Dustdar, S.: On the elasticity of social compute units. In: Jarke, M., Mylopoulos, J., Quix, C., Rolland, C., Manolopoulos, Y., Mouratidis, H., Horkoff, J. (eds.) Advanced Information Systems Engineering. Lecture Notes in Computer Science, vol. 8484, pp. 364–378. Springer, Cham (2014) 27. Riveni, M., Truong, H.L., Dustdar, S.: Trust-aware elastic social compute units. In: Trustcom/BigDataSE/ISPA, vol. 1, pp. 135–142. IEEE, Piscataway (2015) 28. Scekic, O., Truong, H.L., Dustdar, S.: Incentives and rewarding in social computing. Commun. ACM 56(6), 72–82 (2013) 29. Schall, D., Dustdar, S.: Dynamic Context-Sensitive PageRank for Expertise Mining. In: Social Informatics: Second International Conference, SocInfo 2010, Laxenburg, 27–29 October 2010. Proceedings, pp. 160–175. Springer, Berlin (2010)

Privacy in Human Computation: User Awareness Study, Implications for. . .

267

30. Schall, D., Truong, H.L., Dustdar, S.: Unifying human and software services in web-scale collaborations. IEEE Internet Comput. 12(3), 62–68 (2008) 31. Skopik, F., Schall, D., Psaier, H., Dustdar, S.: Adaptive provisioning of human expertise in service-oriented systems. In: Proceedings of the 2011 ACM Symposium on Applied Computing, SAC ’11, pp. 1568–1575. ACM, New York (2011) 32. Swaminathan, S., Fok, R., Chen, F., (Kenneth) Huang, T.-H., Lin, I., Jadvani, R., Lasecki, W.S., Bigham, J.P.: WearMail: on-the-go access to information in your email with a privacy-preserving human computation workflow. In: Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, UIST 2017, Quebec City, QC, Canada, October 22–25, 2017, pp. 807–815 (2017). https://doi.org/10.1145/3126594.3126603 33. Sweeney, L.: k-anonymity: a model for protecting privacy. Int. J. Uncertainty Fuzziness Knowledge Based Syst. 10(5), 557–570 (2002) 34. To, H., Ghinita, G., Shahabi, C.: A framework for protecting worker location privacy in spatial crowdsourcing. Proc. VLDB Endow. 7(10), 919–930 (2014) 35. Toch, E.: Crowdsourcing privacy preferences in context-aware applications. Pers. Ubiquit. Comput. 18(1), 129–141 (2014) 36. Truong, H.L., Dustdar, S.: Context-aware programming for hybrid and diversity-aware collective adaptive systems. In: Business Process Management Workshops: BPM 2014 International Workshops, Eindhoven, The Netherlands, September 7–8, 2014, Revised Papers, pp. 145–157. Springer, Cham (2015) 37. Von Ahn, L.: Human Computation. PhD thesis, Pittsburgh (2005) AAI3205378 38. Williams, T.L., Agarwal, N., Wigand, R.T.: Protecting Private Information: Current Attitudes Concerning Privacy Policies. Academy of Science and Engineering, Los Angeles (2015). ISBN: 978-1-62561-003-4 39. Yuen, M.C., King, I., Leung, K.S.: A survey of crowdsourcing systems. In: Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third International Conference on Social Computing (SocialCom), pp. 766–773. IEEE, Piscataway (2011) 40. Zeppezauer, P., Scekic, O., Truong, H.L., Dustdar, S.: Virtualizing communication for hybrid and diversity-aware collective adaptive systems. In: Service-Oriented Computing – ICSOC 2014 Workshops: WESOA; SeMaPS, RMSOC, KASA, ISC, FOR-MOVES, CCSA and Satellite Events, Paris, France, 3–6 November 2014, Revised Selected Papers, pp. 56–67. Springer, Cham (2015)

Index

A Access and visibility control, 150 AFOCS, 178, 195–199 Aggression supervised classifier, 87 Air-Traffic network, 121, 124, 125 Amazon Mechanical Turk, 250, 255–257 American Psychiatric Association Diagnostic and Statistical Manual (DSM-IV), 238 Anomalous @mention behavior, 27–29 Anonymity, 254, 258–260 Apache Ant class collaboration network, 163 Apache Commons Graph, 165 Approximate inference distribution method, 212 Arbitrary subgraphs, 111 Artificial Intelligence (AI), 4, 5, 236 Attribute-based credentials (ABC), 262–263 “Author-wise pooling” concept, 219 Automatic Information classification, 148–149 Automatic privacy assessment system, 143 Average number of followers (ANF), 169–170 B Bag-of-tokens model, 11 Bank account information, 251, 257 Barabási-Albert Preferential Attachment (PA) model, 108, 122, 125, 126 Bayesian framework, 12 Bayesian Information Criterion, 162 Bayesian method, 211, 212 Behavioral mechanisms, 133 Bi-clique communities, 160 Big Huge Thesaurus, 239

Biological networks, 121, 124 Bipartite graph, 58, 61, 162 Bitcoin, 263 Browser, 253 Bullying, 6, 8, 12, 86 C Cascading classifier (CC), 87 Cashcrate, 257 CashU, 263 Centrality metrics, 72–73 CEP, see Community evolution prediction CFinder, 165 Chemical molecule graphs, 121 Chemical network, 121, 124, 125 Citation network, 121 Clarifai, 74 Clarifiers, 45, 46 Clickworker, 257 Clique percolation methods, 160, 163, 165 CLiZZ, 156, 176, 187–190 Closed-privacy model, 144, 149 Closeness centrality, 73, 196 Cloud services, 248, 257 Coarse-grained predictive analysis, 68, 75 real-world event, outcomes and trends of, 76–77 signals, 76, 77 US gun reform debate 2018, 78–79 US 2016 presidential election, 77 Collective adaptive systems (CAS), 248 Co-mention graph, 20, 23, 27–29, 31, 34, 35, 38 CommTracker, 163

© Springer International Publishing AG, part of Springer Nature 2019 N. Agarwal et al. (eds.), Emerging Research Challenges and Opportunities in Computational Social Network Analysis and Mining, Lecture Notes in Social Networks, https://doi.org/10.1007/978-3-319-94105-9

269

270 Community-aware ranking algorithms, 156, 157, 162, 171–172, 190, 194 Community detection algorithms, 107, 158, 165, 182 Community evolution prediction (CEP), 167, 172–174, 185 accuracies, 198–199 community properties, 194–196 significance of features, 196–197 Community influencers, 23, 29–30 Community mapping, 54, 162–163, 165 Community update and tracking (CUT) algorithm, 162 Conditional Random Fields (CRF), 80 Configuration model (Conf), 122, 125, 126 Confirmation bias, 46 Content-based intent mining, 10–11 Context-driven automatic privacy assessment, 148–150 Convergence analysis, 122–123 Convolutional and gated recurrent neural network (CGRNN), 80 Credibility network, 47, 48 network initialization, 55 network optimization, 55–56 Credit reporting, 146–147 CrowdFlower, 257 Crowdsourcing, 248, 250 Cyberbullying, 12, 86 Cyborg accounts, 25, 47

D Data brokers, 253 Data fusion techniques, 161 DBdata, 49 DBLP, 156, 161, 189, 196–199 DBpedia, 236, 243 age estimation, 90 for coverage and generality, 70 engagement analysis, 81 harassment, 86 location estimation, 88 Deceiving, 8 Deep learning, 74, 80, 83, 91, 95 Deep neural networks, 44, 59 Dense subgraph detection method, 30–31, 35 Detailed balance condition, 118 Device fingerprints, 253 Dictionary of Affect in Language (DAL), 87 Diffusion network, 47, 48 Digamma function, 215

Index Disassortative degree mixing (DMID), 155, 168–171, 187–190, 196–199 Disassortative vector (DV), 169 Document retrieval, 236 Doc2vec model, 213 Domain-specific use cases, social media analysis colloquial nature of language, drug abuse ontology, 236, 241–242, 244 complex real-world domain, mental health disorders, 235, 237–239, 244 lack of context, EmojiNet, 235, 239–241, 244 topic relevance, movie domain, 236, 242–244 Drug Abuse Ontology (DAO), 69, 81, 236, 241–242, 244 Drug-related tweets, 84 E Echo chambers, 22, 23, 45 E-coli network, 121, 124, 125 Edge set induced subgraphs, 111 eDrugTrends, 84 Eigenvector centrality, 73, 173, 196 Elance, 257 ElasticNet, 90 Electronic communication, 239 Email, 156 EM algorithm, 162 Emoji, 71 EmojiNet, 71, 235, 239–241, 244 Emoticons, 71 Emotional analysis, 79–81 Emotional supporting, 7 Emotion-related hashtags, 70 Engagement analysis, 81 Enron network, 122, 124, 196–199 Entity disambiguation, 236 Entity model network (EMN), 243 Equifax, 146 Erdös-Renyi (ER) model, 108, 122, 125 EU Data Protection Directive (DIRECTIVE 95/46/EC), 251 Evidence lower bound (ELBO), 214–215 Experian, 146 Expert identification, 156, 161–162 Expertise sharing, 8 F Facebook, 132, 134, 136, 137, 154, 156, 161, 196–199

Index Fake news challenges, 44 detection, 44 friendship network embedding, 53–56 interaction network embedding, 50–52 knowledge network matching, 57–59 temporal diffusion representation, 52–53 heterogeneous networks, 48–49 homogeneous networks, 47–48 mitigation aim of, 44–45 network intervention, 62–63 network size estimation, 61–62 user identification, 59–61 negative societal impacts, 44 network properties echo chamber, 45 filter bubble, 46 individual users, 45–46 malicious accounts, 46–47 FANMOD, 110, 116 Fantasy premier league (FPL), 89 Feed-forward (FF) network, 95 Filter bubble, 46 Fine-grained predictive analysis, 68, 75, 76 Firibinome social influence bot network, 20, 21, 34–36 Fiverr, 257 Follower–followee network, 94 Foulefactory, 257 Freelancer, 250 Frequency heuristic, 45 Frequent Subgraph Mining (FSM), 110 Friend of a Friend (FOAF), 134 Friendship network embedding, 53–56 Friendship networks, 47–48

G GANXiS, 164 Gender estimation, Twitter, 90–91 Gene interaction network, 121 Gene Ontology, 236 General Data Protection Regulation (REGULATION (EU) 2016/679), 251 General trust, 46 Geometric random graphs, 125 GeoNames, 88 GeoText, 83 Gibbs sampling, 210, 212 Google, 49

271 Google+, 154 Government issued ID, 250, 257 Gradient Boosted rank, 94 Graphlets, 110, 125 Graph patterns, 109 Graph stream, 165 Group Evolution Discovery (GED) technique, 156, 163 Gullibility, 46 Gullible users, 45, 46 Guru, 257

H Harassment, 86–87 Hashtags, 70, 85, 213 HeatMap, 196 Help offering, 7 Heterogeneous networks, 48–49 Hill-climbing algorithm, 61 HITS algorithm, 156, 157, 161, 162, 171–172, 178, 194, 195 HOCTracker, 163 Homogeneous networks, 47–48 Homophily theory, 47–48 Host network, 109 Human computation, privacy in artificial artificial-intelligence, 248 awareness, 264 CAS, 248 cloud services, 248 crowdsourcing, 248, 250 decentralized systems, 250 definition, 247 demographics, 256–257 evaluation methods, 264 location, 264 mobile context-aware applications, 250 mobile spatial-crowdsourcing tasks, 249 online labor markets and expert networks, 248 payment methods, 264 personal data on collected data, 250–251 incentive mechanisms, 252 management mechanisms, 252 misbehavior prevention, 252 payments, 252 quality of service, 252 task assignment and formation of collectives, 251 PPL, 249 Principle of Openness/Notices, 250

272 Human computation, privacy in (cont.) privacy awareness anonymity, 258–260 data collection, usage concerns and security, 257–258 regulations and policies, 260–261 privacy-preserving workflows, 264 privacy risks control, lack of, 254 ownership, lack of, 254–255 profiling, 254 security, lack of, 255 transparency in privacy policies, lack of, 253 user privacy policy awareness, 253 recommendations ABC, 262 item Bitcoin, 263 item CashU, 263 subitem data minimization, 261 k-anonymity model, 263 Perfect Money, 263 t-closeness, 263 SCIS systems and mechanisms, 249 social computing, 248, 249 survey design and distribution, 255–256 transparency with rules/service-level agreements, 263–264 zero-knowledge systems, 250 Hybrid-interaction and history-based privacy system, 144 Hypertext Transfer Protocol (HTTP), 179 I IdeaCONNECTION, 257 Implicit entity, in tweets, 236, 242–244 Independent cascade model (ICM), 60, 62 Individual bots, 20, 23, 25 Induced subgraphs, 111, 116 InfoMap, 165 Information diffusion strategy, 170 Information sensitive sentimental analysis model, 150 InnoCentive, 257 Instagram, 132 Integrated Development Environment (IDE), 164 Intelligent data integration module, 150 Intent, 3–4 definition, 5 intentional stance, 5

Index mining methods content-based intent mining, 10–11 network-based intent mining, 12–13 query intent, 9 user profile-based intent mining, 11–12 types of, 4 intent for mixed social bad and ugly uses, 8 intent for mixed social good and bad uses, 9 intent for social bad use, 8 intent for social good use, 7–8 intent for social ugly use, 8 OSN messages, examples of, 5, 6 Intentional stance, 5 Interaction network, 49

J Jaccard similarity, 163 Jensen inequality, 215 JGraphT, 165 JGraphX, 165 Jmod, 165 JUNG, 164

K Kavosh, 110 K-means clustering, 213, 223 K-Nearest Neighbor (KNN), 93 K-Nearest Neighbor with Dynamic Time Warping (KNN-DTW), 93 Knowledge graph (KG) DBPedia age estimation, 90 for coverage and generality, 70 engagement analysis, 81 harassment, 86 location estimation, 88 domain-specific knowledge graphs, 236–237 Knowledge network, 48–49 linked open data, 49 matching flow optimization, 57 path finding, 56–57 stance network aggregation, 57–59 shared knowledge network, 22 K-shells, 28 Kullback–Leibler divergence (KLD), 80, 214

Index L Labeled LDA (L-LDA) model, 210, 212, 215, 216 Label Propagation Algorithm (LPA), 175 Latent Dirichlet allocation (LDA), 162, 210 applications, in Twitter, 213 approximate inference technique, 210 batch execution, 210 DiscLDA, 210, 212 ELBO, 214–215 free variational parameters, 214 Gibbs sampling, 210, 212 KL divergence, 214 labeled LDA, 210, 212, 216 multinomial distribution, 214 OLLDA model, 216–218, 226–227 online VB, 210, 212, 215–216 supervised LDA, 210, 212 transition parameters, 211 variational EM algorithm, 215 variational inference, 214 Latent/Hierarchical Dirichlet Allocation (LDA/HDA), 74 Latent Semantic Analysis (LSA), 74 Latent Semantic Indexing, 74 LDA, see Latent Dirichlet allocation Leadership-based approach, 61 Least Absolute Shrinkage and Selection Operator (LASSO), 90 Least-effort algorithms, 94 Least square optimization algorithm, 90 LFR method, 165, 185, 189 LIBLINEAR, 80 Linear regression, 90, 94 age estimation, 90 anomaly and popularity prediction, 91, 94 Link communities, 158, 160–161, 177, 187 LinkedGeoData, 85 LinkedIn, 146, 154, 161 Link farming, 13, 24 Link prediction task, 11 Location Name Extraction (LNEx), 88, 89 Long Short Term Memory (LSTM), 95 Louvain method, 165 M Machine learning, 11, 12, 25 deep learning, 74 LIBLINEAR, 80 Multinomial Naive Bayes, 80 statistical features, 73 Malicious accounts, 46–47 Malicious intent (spamming), 12

273 Manipulating, 8 Map Equation, 165 Marketing, 9 Markov Chain Edge Set Sampling (MCESS), 116–119, 122–123, 127–128 Markov chain Monte-Carlo (MCMC) sampling techniques, 210 Marvel network, 121, 124, 125 Mean absolute error (MAE), 80 Mean absolute percentage error (MAPE), 88 Mean average precision (MAP), 156 Mean reciprocal rank (MRR), 156, 186 Mental health disorders, 235, 237–239, 244 Merging of overlapping communities (MONC), 177–178, 187–190 Metropolis–Hastings methodology, 116 MFinder, 110, 113 Microworkers, 250, 252, 255–257 MinDFS code, 119 Mitigation campaign, 62–63 Modularity modeling, 54 Modularized nonnegative matrix factorization (MNMF), 54 Motif Detection algorithm, 110, 116 Multi-entry neural network architecture (MENET), 83 Multinomial Naive Bayes, 80 MusicBrainz, 236 MySpace, 143

N Naïve Bayes (NB) gang communities and members, 87 gender estimation, 90 harassment, 86 location estimation, 88 sales and stock price prediction, 95 sentiment analysis, 80 Naïve realism, 46 Nash equilibrium, 159 Natural language processing (NLP) techniques, 68, 70, 220, 234, 243, 244 NETMODE, 110 Network-based intent mining, 12–13 Network initialization, 55 Network optimization, 55–56 News dissemination ecosystem, 44 News embedding, 50 NING social network, 164 Nonnegative matrix factorization (NMF), 50 Normalized mutual information (NMI), 180, 184–185, 187

274 O OCD, see Overlapping community detection OneSpace, 257 Online labeled LDA (OLLDA) model, 216–218, 226–227 Online social networks (OSNs) adoption of, 3 authentication system, 142 automatic privacy assessment system, 143 defensive behaviors, 133 diverse usage of, 3–4 experiments and analysis Facebook datasets, 136, 137 Twitter dataset, 141–142 weighted cliques, 139–141 weighted edges, 137–138 weighted reputations, 139 future challenges hybrid information filtering, 14 malicious intent, fixing erroneous spreading of, 14 profiling anonymous identities, 14 transforming social bots, 14 information privacy, 133 intent (see Intent) MySpace, 143 network anonymization, 142 organization’s visibility permissions, 132 privacy, 133 recommendation systems, 142 reputation model, 132, 134–136 rule-based access control, 143 scoring algorithms, 132 security concerns, 143 sharing enforcement model, 143 trust (see Trust) user’s personal attributes, 132 users’ postings, 132 user’s relational attributes, 132 websites, 132 Open-privacy model, 144, 149 OpenStreetMap, 88 OSNs, see Online social networks OSoMe, 68 Overlapping community detection (OCD) AFOCS, 178 algorithms, 155–157 clique percolation methods, 160 CLiZZ, 176 community evolution prediction accuracies, 198–199 community properties, 194–196 significance of features, 196–197

Index content-based and attribute-based approaches, 158 expert finding mean average precision and mean reciprocal rank values, 190 spearman correlation values, 194 expert ranking baselines, 178–179 hierarchical clustering, 159 leader-based techniques, 159 line graphs and link communities, 160–161 merging of, 177–178 random-walks and label updates method, 159–160 REST, 200 results on real-world networks, 188–192 synthetic networks, 187–188 SLPA, 175 SSK, 175–176 static community detection algorithms, 158 symbols used for, 168 tools and web services, 164–165 universal metrics, 158 Web service, 179–181 Overlapping community structures analysis, 154 community-aware ranking algorithms, 157, 171–172 community evolution and prediction analysis, 163–164 community mapping over time, 162–163 datasets, 183 community analysis, 181 expert ranking question–answer forums, 181–182 detection of, 154 diffusion of opinions, 155, 157 disassortative degree mixing, 155, 157 DMID, 155–157 dynamism, 154 evaluation protocol and metrics classification context, precision in, 186 information retrieval context, precision in, 185–186 LFR synthetic networks, 185 mean reciprocal rank, 186 modularity, 182, 184 NMI, 184–185 evolution analysis and prediction community evolution prediction, 172–173 GED technique, 173–175

Index expert identification, 156, 161–162, 171–172 HITS algorithm, 156 methodology DMID, 168–171 problem formulation, 166–167 OCD AFOCS, 178 algorithms, 155–157 clique percolation methods, 160 CLiZZ, 176 content-based and attribute-based approaches, 158 expert ranking baselines, 178–179 hierarchical clustering, 159 leader-based techniques, 159 line graphs and link communities, 160–161 merging of, 177–178 random-walks and label updates method, 159–160 real-world networks, results on, 188–190 SLPA, 175 SSK, 175–176 static community detection algorithms, 158 symbols used for, 168 synthetic networks, results on, 187–188 tools and web services, 164–165 universal metrics, 158 Web service, 179–181 PageRank algorithm, 156 RESTful Web services, 157 social network analysis, 166 synthetic network generators, 156

P PageRank algorithm, 28, 29, 31, 32, 156, 157, 161, 162, 172, 179, 194, 195 Paragraph2vec model, 213 Partition density (PD), 177 Part-of-speech (PoS) tagging, 70, 87 Path finding, 56–57 Patient Health Questionnaire (PHQ-9), 238–240 Pattern-aided supervised classification approaches, 11 PeoplePerHour, 250 Perfect Money, 263 Persuaders, 45, 46 Phone numbers, 257 Poisson process, 94

275 Polynomial uniform sampling algorithm, 113 Prediction accuracy (PA), 186, 198, 199 Predictive analysis, Twitter data applications, 68 anomaly and popularity prediction, 91, 93–94 community on social media, 89 demographics, 89–91 and evaluation, comparative analysis, 92–93 healthcare, 82–83 location estimation, 88–89 political issues, 84–85 public health, 83–84 sales and stock price prediction, 95 social issues, 85–87 transportation, 88 Twitris platform, 82 coarse-grained analysis, 68, 75 real-world event, outcomes and trends of, 76–77 signals, 76, 77 US gun reform debate 2018, 78–79 US 2016 presidential election, 77 fine-grained analysis, 68, 75, 76 human expert guidance, 75 signals, extraction of emotional analysis, 79–81 engagement analysis, 81 sentiment analysis, 79, 80 topical analysis, 79, 81 systematic framework, 68, 69 tweets, language understanding of centrality metrics, 72–73 closeness centrality, 73 emoji, 71 emoticons, 71 hashtags, 70 machine learning algorithms, 73–74 natural language processing, 70 PoS tagging, 70 profile and header image, 74 statistical features, 73 stop words, removal of, 70 tweet metadata, 71, 72 URL, 70–71 user metadata, 71, 72 Word2Vec, 74 PRescription Drug abuse Online Surveillance and Epidemiology (PREDOSE), 84 Privacy assessment system, 148 Privacy, in human computation, see Human computation, privacy in Privacy Policy Languages (PPL), 249

276 Privacy recommendation system, 150 Probabilistic soft logic (PSL), 237 Probabilistic topic model, 83 Probability theory, 112 Professional referrals, 145–146 Provenance paths, 59–60 Proximity mapping, 54 Publisher–news interactions, 51–52 Pure echo-chambers, 22 Pylouvain-igraph, 165

Q Query intent, 9

R Radial-volume centralities, 28 Random forest classifier anomaly and popularity prediction, 91, 93, 94 gang communities and members, 87 sales and stock price prediction, 95 Random graph models configuration model, 122, 126 degree distribution of network, 108, 126 Erdös-Rényi model, 108, 122, 125 Preferential Attachment random graph models, 108, 122, 125, 126 total variation distances, 125–126 RapidWorkers, 257 Received gains (RG), 170 Reciprocal @mention network, 25 Recurrent neural networks (RNN), 52–53, 91, 95 Reddit, 68, 83 Relation Extraction Corpus (GREC), 49 REpresentational State Transfer (REST), 179 Reputation model, 132, 135–136 system applications context-driven automatic privacy assessment, 148–150 credit reporting, 146–147 friends’ recommendation system, 144–145 hybrid-interaction and history-based privacy system, 144 invitations from “strangers,” 145 networks evolution, assessments of, 148 positive/negative interactions, 148 professional referrals, 145–146 spam detection, 147 Resistance threshold (RT), 170

Index Restarting edge set sampling (RESS), 113–116, 122–123 Retweeting, 72 RMSprop, 95 Root means square error (RMSE), 88 Rule-based classification approach, 10 Rumoring, 8

S Scribie, 257 SECOM, 160 Semantic clustering approach, 213 Semantic filtering mechanism, 81 Semi-supervised approach, 83 Sensationalizing, 9 Sentimental analysis methods, 148 Sentiment analysis, 79, 80 Service-level agreements (SLA), 252, 263–264 Shared hashtag network, 22 ShortTask, 257 SIBN, see Social influence bot network Singular Vector Decomposition technique, 12 SLPA, see Speaker Listener Label Propagation Social botnets, see Social influence bot network Social bot network impact of, 26 link farming, 24 SIBN (see Social influence bot network) tolerance-based defense schemes, 25 Social bots, 20, 46 Social Collective Intelligence Systems (SCIS), 249 Social credibility, 45 Social identity theory, 45 Social influence bot network (SIBN), 39–40 alt-right community (ALT16) dataset, 27–29, 31–33 co-mention network, 20, 23 community-level, 21 content-level, 20–21 core bots, 22 dense subgraph detection, 30–31 echo-chambers, 22, 23 Euromaidan community (EUR17) dataset, 27, 32, 33 Euromaidan Image Sharing SIBN, 36–38 Firibinome SIBN, 20, 21, 34–36 graph-level perspective, 24 @mention network, 20 non-core bots, 22 promoted accounts vs. community influencers, 23, 29–34

Index pure echo-chambers, 22 shared hashtag network, 22 social graph features, 25 social network, 22 Syrian revolution (SYR15) dataset, 27, 34 topic groups, 22, 23 Twitter follower network, 20 user-level, 21, 23 Social influence theory, 48 Social media analysis, domain knowledge for AI researchers, 236 domain-specific use cases, challenges in colloquial nature of language, drug abuse ontology, 236, 241–242, 244 complex real-world domain, mental health disorders, 235, 237–239, 244 lack of context, EmojiNet, 235, 239–241, 244 topic relevance, movie domain, 236, 242–244 entity semantics, 236 knowledge graphs, 236, 237 language understanding and information interpretation, 243 natural language processing techniques, 243 PSL, 237 recommender system, 236 social media communication, 243 Twitter, 234 Web (semantic) applications, 236 Wikipedia category hierarchy, 237 Social network analysis, 13, 110 Social trust, 252 Spam detection, 147 Speaker Listener Label Propagation (SLPA), 159, 164, 175, 187, 188, 190, 196–199 Spectral clustering, 213 Stance network, 49 Stance network aggregation, 57–59 Stanoev, Smilkov and Kocarev (SSK), 175–176, 187 Subgraph distributions arbitrary subgraphs, 111 challenges, 109 FSM, 110 graph patterns, 109 local topological features, 108–109 Motif Detection, 110 network motifs, 110, 111 pattern distribution advantage of, 109 black box model, 120

277 host graph, 109, 111–112 MCESS, 116–119, 122–123, 127–128 polynomial uniform sampling algorithm, 113 random graph models (see Random graph models) real-world network datasets, 121–125 RESS, 113–116, 122–123 two distributions, total variation of, 119, 120 Supervised learning classifier, 11, 12 Supervised machine learning methods, 70 Support Vector Machine (SVM) age estimation, 90 aggression/loss tweets, 87 anomaly and popularity prediction, 91, 94 gender estimation, 91 harassment, 86–87 healthcare, 83 sales and stock price prediction, 95 sentiment analysis, 80 Support vector regression (SVR), 88, 90 Sybil attacks, 252 Sybil-detection algorithms, 25 Symbolic Aggregate approXimation with Vector Space Model (SAX-VSM), 93

T Technische Universität Wien (TU Wien), 255 Temporal user engagements, 52–53 Testbirds, 257 TF-IDF model, 74, 91, 213 Time-dependent Hawkes process (TiDeH), 94 Time series analysis, 72, 73 Tolerance-based approach, 25 TopCoder, 250 Topical analysis, 79, 81 Topical hashtag-level sentiment analysis, 70 Topical inference accuracy, 221–223 Topic models, social media analytics coupling supervision and online learning, 212 experiments dataset, 218–219 experimental pipeline, 219–220 scalability performance, 220–221 topical inference accuracy, 221–223 word embeddings and clustering, 223–226 LDA, 210 applications, in Twitter, 213 approximate inference technique, 210

278 Topic models, social media analytics (cont.) batch execution, 210 DiscLDA, 210, 212 ELBO, 214–215 free variational parameters, 214 Gibbs sampling, 210, 212 KL divergence, 214 labeled LDA, 210, 212, 216 multinomial distribution, 214 OLLDA model, 216–218, 226–227 online VB, 210, 212, 215–216 supervised LDA, 210, 212 transition parameters, 211 variational EM algorithm, 215 variational inference, 214 MCMC sampling, 210 online learning, 211–212 supervised learning, 212 variational inference methods, 210 word cloud, 225, 228 word embeddings, 213 Word2Vec, 210 Toponym, 88 TranscribeMe, 257 Transfer learning paradigm, 11 Transition matrix, 159 TransUnion, 146 Trolls, 46 Trust asymmetric, 133 context specific, 133 definition, 133 dynamic, 133 FOAF, 134 interaction and communication models, 133 non-transitive, 133 and reputation models, 252 reputation ranking formula, 134 subjective, 133 Tweets. Twitter centrality metrics, 72–73 closeness centrality, 73 emoji, 71 emoticons, 71 hashtags, 70 machine learning algorithms, 73–74 natural language processing, 70 PoS tagging, 70 profile and header image, 74 statistical features, 73 stop words, removal of, 70 tweet metadata, 71, 72

Index URL, 70–71 user metadata, 71, 72 Word2Vec, 74 Twitris, 68, 81, 82, 84 Twitter, 122, 132, 141–142, 154, 161 classification and user recommendation (see Topic models, social media analytics) daily active users, number of, 67 data, predictive analysis (see Predictive analysis, Twitter data) link farming, 13 monthly active users, number of, 67 shared hashtag network, 22 threatening intent, 12 topic groups, 22 user-generated data, 67, 68 TwitterWorld, 83

U UMLS, 236 Uniform Resource Identifier (URI), 179 Unsupervised learning approach, 12 Upwork, 250, 257 Urban Dictionary, 239 User embedding, 50–51 User metadata, 71, 72 User–news interactions, 51 User profile-based intent mining, 11–12 uTest, 250 UTGeo11, 83

W WebOCD framework, 181, 200 Web service, 179–181 Web-Stanford network, 121, 125 Weighted degree centrality, 28 Weighted trust model, 136 Wikipedia, 236 Wikipedia editors, 13 Wisdom of Crowd (WoC), 89 Word cloud, 225, 228 Word embeddings, 74, 213, 223–226 WordNet, 213 Word2Vec, 74, 89, 210, 223

Y Yeast network, 121, 124 YouTube, 86, 87, 95