The Role of Social Capital in Information Di usion over Twitter: a Study Case over Brazilian posts

Abstract

Social Capital is the resulting advantage of the individual's localization in a social structure. It can be measured by traditional complex networks metrics, or speci c ones, such as information capital, brokerage and bridging. Our goal is to verify which users have high information capital, bridging and brokerage for providing and spreading information. To do so, we rst categorize Twitter users into seven types: typical users, primary media, secondary media, independent experts, fan accounts, fake accounts and potential bots. Then, we analyze their proles on trending topics. Our results show potential bots and fan accounts as the main information spreaders in Brazil, a very concerning result given the upcoming presidential election in October 2018.

Keywords

• Social Networks Social Capital Information Di usion

-

Online Social Networks (OSNs), such as Facebook, Twitter, LinkedIn and Instagram, have achieved unprecedent growth in recent years. Current statistical data show Facebook has over 2.2 billion monthly active users, while Instagram, Twitter and LinkedIn have over 813 million, 330 million and 260 million monthly active users respectively [ 38 ]. Such huge volume of users and relationships is a motivation for several researches in the areas of Complex Networks, Big Social Data, and Urban Computing [ 16,22,23,36 ].

Among many usages, OSNs are powerful media for information di usion [ 1,18,30,27 ]. Information di usion occurs when there is a ow of information from one individual to another. The information may be retained by an individual or spread out on the network [ 39 ]. Individuals who are in a privileged position in the OSN have more social capital for information di usion.

Social capital is a comprehensive concept, without one single de nition or metric to capture all its facets [ 31 ]. In Social Network Analysis (SNA), it is the resulting advantage of the individual's right localization in a social structure [ 9 ]. For instance, the individual who holds information has the power to change what happens in one environment and to understand its surroundings [ 42 ].

In this scenario, new players arise and become relevant: the \human sensors", or citizens who share information about their environment via OSNs, supplementing, complementing or even replacing information as measured by physical sensors [ 13 ]. Human involvement is particularly useful in detecting multiple processes in complex personal, social and urban spaces where traditional embedded sensor networks have limitations [ 37 ]. The society engagement in the OSNs, whether individually or in small groups, can be facilitated through its social capital [ 34 ]. Overall, humans are relevant data sources, acquiring and spreading information on their own [ 37 ].

Such human sensing for information di usion can be useful for emergency events management, crime detection, urban administration, intelligent transportation, smart tra c control, public healthcare, political engagement, among others [ 26,37 ]. In addition, the information can change the view of the recipient, motivating him/her to join the social network of the information provider [ 37 ]. Also, trust among communicating individuals strongly a ects the reach and impact of information [ 19 ], and trust is an important social capital aspect [ 3,34 ].

However, OSNs have become signi cant spreaders of false facts, urban legends, fake news, or, more generally, misinformation. Misinformation drives the emergence of a post-truth society, where the debate becomes damaged by the repetition of discussions refuting the primary media or independent experts (ofcial sources of truthful information) [ 27 ].

Even so, people still rely on the news published in social media. In 2017, according to Reuters [ 32 ], despite the predominance of TV stations in the media environment, social media played an essential role in the consumption of news, being the primary sources of news within the Brazilian urban context. For instance, in 2016, the impeachment of President Dilma Rousse drew attention for its repercussion in Brazilian social media. However, only 30% of people believe that the social media is free from undue political in uence.

Overall, the accurate information di usion contributes to building smart urban spaces [ 29 ], promoting improvement in the quality of life of their citizens. Consequently, to ensure reliability to the citizen, the primary media and independent experts must take the lead in OSNs. Henceforth, we address the social capital forms as related to individuals' abilities to acquire and spread information. We model a network of Twitter users relating them through retweets in messages that contribute to a subject to become Brazilian Twitter Trend. We analyze and compare social capital metrics to verify the importance of the primary media and independent experts compared to typical users, fan accounts, and potential bots, in the Brazilian information di usion context. Although considering only messages written in Brazilian Portuguese, our methodology is broad enough to be applied to any language without loss of generality. 2

Related Work

This section is divided in two parts: general work on social capital and online social networks (Section 2.1), and speci c metrics for social capital (Section 2.2).

Social Capital and Online Social Networks

In SNA, social capital is the resulting advantage of the individual's right localization in a social structure [ 9 ]. There are many de nitions of social capital. For instance, Bourdieu [ 6 ] emphasizes social capital as the collective resources used by individual members to obtain services and bene ts either in the absence or conjunction with their economic capital. For Coleman [ 11 ], social capital is a neutral resource that facilitates any action, where the individuals are responsible for achieving their objectives. In turn, Putnam [ 34 ] de nes social capital as characteristics of the social organization, such as social networks, social norms, and social trust, which facilitate coordination and cooperation for mutual benet and civil engagement. Moreover, Bertolini and Bravo [ 3 ] de ne social capital as the resources individuals have access in a given network.

Authors also have di erent ways of thinking about how to classify social capital types. Putnam [ 33 ] de nes two forms of social capital: bonding and bridging. In [ 21 ], Jackson introduces seven types: information, brokerage, bridging, coordination, favor, reputation, and community capital. Moreover, Burt [ 9 ] de nes two activities related to social capital: brokerage (an individual self placing in a privileged position in the network), and closure (the coordination of a closed group of individuals in the network).

Regarding social capital over OSNs, there are metrics to visibility, reputation, popularity, and authority. Bertolini and Bravo [ 3 ] describe that reputation and authority comprise the sum of knowledge and information disseminated within a given group. Measuring the relationships' strength is also relevant (identifying weak ties), as well identifying hubs and in uencers. Then, there are traditional metrics such as: ego-network measure, the structural hole measure [ 10 ], homophilia measure [ 5 ], and the standard centrality measures [ 14 ]. Also, Kang and Shen [ 25 ] de ne quantitative metrics for social capital, not validated in OSNs, and Michalak et al. [ 31 ] present measures of social capital based on techniques of cooperative game theory.

Social capital is also used for explaining information di usion processes in social networks. Authors in [ 2,17 ] show that the presence of weak ties and hubs accelerates the information di usion on social networks. In turn, Kleinberg [ 14 ] models the networks' cascading behavior like an in uence process among individuals. Di erently, many authors deal with the essential nodes detection (in uencers) in OSNs as being the individuals who maximize spreading [ 20,24,35 ].

The aforementioned works do not rely on social capital for understanding the process of information di usion over OSNs. Here, we assume that information capital, bridging and brokerage are skills from individuals that are well placed on the network structure and take advantage of their position to acquire and control the information ow. Furthermore, we apply the hub and authorities concepts proposed by Kleinberg in [ 28 ] as tool for identifying the use of social capital on information di usion process.

Metrics for Social Capital

The previous section presented general work on social capital and networks. We now discuss existing metrics for social capital. We begin by providing notation that helps to de ne the metrics. Let G denote a directed graph, where V (G) is its set of vertices (or nodes), and E(G) its set of edges. Also, n = jV (G)j and m = jE(G)j. G is represented by its adjacency matrix A 2 [0; 1]n n, where Au;v=1 indicates the existence of an edge between u and v, and Au;v=0 otherwise. Given No(u) as the set of u's outgoing neighbors and Ni(u) the set of u's incoming neighbors, then jNo(u)j is u's out-degree and jNi(u)j its in-degree. Finally, N (u) is the set of u's neighbors, where N (u) = No(u) [ Ni(u) and jN (u)j is u's degree.

A path in G between two nodes u and v is a succession of distinct nodes u ) u0; u1; u2; :::; up ) v such that Ak;k+1 = 1 (8k j 0 k < p). A geodesic (shortest path) between nodes u and v is a path such that no other path between them involves a smaller number of edges. Let guv the number of geodesics connecting u to v, and guv(i) the number of geodesics that node i is on. Lastly, let d the graph diameter, that is, the largest geodesic distance between any pair of nodes. Information Capital. Decay Centrality (DC) has been applied to measure individual's information capital [ 21 ]. In summary, DC measures the number of individuals reached by a speci c individual, regardless the path length to arrive to them. DC counts paths of di erent lengths, i.e., how many people one can reach at di erent distances. The decay of information with distance is captured via a parameter p, with 0 < p < 1. Furthermore, DC favors individuals that reach the largest number of neighbors in up to T hops, known as information's endurance. The DC metric of a node i, is given by Equation 1, where N l(i) is the set of individuals at maximum distance l from i in G.

DC(i) =

T X pl:jN l(i)j: l=1

From Equation 1, a node has high information capital if its broadcast information reaches a large number of nodes. Otherwise, a node has low information capital. Although it has a simple calculation, DC does not consider all the possible paths that information might take [ 21 ]. The Eigenvector Centrality (EC) [ 4 ], or simply EigenCentrality, is an alternative to solve such a limitation [ 21 ]. For EC metric, the node importance depends on its neighborhood importance. Moreover, EC is a node in uence measure in the network and given by Equation 2, where is the largest eigenvalue of R, and R is an eigenvector of A (the graph's adjacency matrix).

EC(i) = 1 X EC(t):

t2N(i) Bridging Capital. Individuals who are bridges in the network topology are special nodes that can control the information ow [ 21 ], and/or accelerate its (1) (2) di usion [ 17 ]. Here, we follow Granovetter's theory [ 17 ] that de nes a bridge as a weak tie in the network. For labeling a node as a bridge, we apply the Neighborhood overlap metric (NO) as de ned in Equation 3. Then, following Brandao and Moro [ 7 ], we classify a tie (edge, link) between nodes u and v as weak if: 0 NO(u; v) 0:2.

N O(u; v) =

jNo(u) \ No(v)j jNo(u) [ No(v) fu; vgj Brokerage Capital. From [ 10,21 ], nodes with high Betweenees Centrality (BC) play the role of brokers in the network. A normalized BC metric (for directed graphs) is given by [ 15 ] as Equation 4.

BC(i) =

P (n

gjk(i) j;k2E(G) gjk : 1):(n 2) Hubs and Authorities. In order to rank the most important nodes based on their outgoing and incoming links, we apply the HITS algorithm proposed in [ 28 ]. As we discuss in Section 2.1, this approach enables pro ling nodes with special roles on the information di usion process. In summary, HITS algorithm computes two types of ranking: (i ) the authority ranking estimates the node importance based on the incoming links; and (ii ) the hub ranking estimates the node importance based on the outgoing links. We refer the reader to [ 28 ] for more details on such an algorithm. 2.3

Main Contributions

Our contributions over the related work are: { We apply network topological metrics for analyzing information capital (eigenvector centrality), bridging (neighborhood overlap), and brokerage (betweenness centrality); { We apply decay centrality for measuring information capital; { We apply the HITS algorithm for pro ling nodes in Twitter; and { We provide a comparative analysis of the top 10 Twitter users, regarding network topological metrics, decay centrality, and the HITS algorithm. 3

Methodology

Our work evaluates social capital over a social network as extracted from posts in an online microblogging platform. To do so, we rst build a dataset through collecting real posts in Twitter over a week. In order to be able to easily qualify the posts and our analyses results, our collecting process is limited to tweets in Brazilian Portuguese, our native language. Nonetheless, our methodology is (3) (4) broad enough to be applied to any language without loss of generality. Then, in order to better qualify our analysis, we categorize the users in two di erent forms. Next, we implement the previously discussed metrics and apply them over a network modeling. Finally, we are able to evaluate such metrics and the potential relevance of users according to their behavior. The next sections discuss respectively: the dataset building process, our categorization for user types, metrics implementations and parameters setting, and the network modeling considered. 3.1

Dataset

Our work evaluates social capital over a social network as extracted from tweets. Tweet is a Twitter3 post (limited to 280 characters, in Brazil), and retweet (RT) is a re-posting of a tweet, which helps quickly sharing a given tweet. Moreover, Twitter Trends (TT) are topics that have become immediately popular (as opposed to topics have been popular for a while or on a daily basis)4.

We initially collect Brazil's top 50 TTs. For each TT, we collect 100 most popular tweets. For each tweet, we collect 100 most recent RTs. The number of collected posts is 100 due to search limitations in the Twitter API5. We collect continuously for seven days (2018/04/20 to 2018/04/26), to acquire a more signi cant amount of data.

After cleaning and data processing, the dataset6 contains 165,936 and 371,612 distinct users and messages, respectively. In this fase, 1,648 distinct subjects appeared in TTs. In the top 10 TTs collected, there are subjects related to TV, entertainment, musicians, sports, and a single related to commemorative dates (\Tiradentes"). There is no political or economical subject in the overall top 10 (surprinsingly given the current Brazilian crisis), as presented in Table 1.

Now, considering only political/economic TT, the top 10 topics regard the current Brazilian political crises and the upcoming presidential election in October 2018, as shown in the bottom half of Table 1. In summary, the TT cover: Lula Livre and Pris~ao de Lula are popular clamor over the arrest of former Brazilian president Lu s Inacio Lula da Silva; Ciro Nogueira, Palloci, Rocha Loures, Mantega, and Temer are Brazilian politicians who are currently targets for corruption investigations; Odebrecht is a large Brazilian construction company, also investigation target; Marcos Valerio is a publicist involved (and delator) in corruption schemes; PSDB is a political party; and STF and STJ are the federal supreme and superior courts in Brazil. 3.2

Twitter User Types

We categorize Twitter users in two ways. The rst type is called posting categories and regards user behavior over posting tweets and retweets. There are 3 https://www.twitter.com 4 https://help.twitter.com/en/using-twitter 5 https://developer.twitter.com/en/docs 6 Dataset available at http://homepages.dcc.ufmg.br/~mirella/projs/apoena/ datasets.html users who tweet more than retweet (providers ), users who retweet more than tweet (spreaders ), and users who balance both actions (neutral ). Therefore, we de ne p ratio as being the ratio between the tweets (out-degree) of a user, and the total number of tweets and retweets (degree), as shown in Equation 5. Thus, users can be spreaders (0 p ratio 0:25), neutral (0:25 < p ratio < 0:75), or providers (0:75 p ratio 1). In our speci c dataset (Section 3.1), most users are spreaders (164,884 users), followed by providers (1,043 users), and then neutral (9 users).

p ratio =

out-degree out-degree + in-degree : (5)

The second way, called social categories, distributes users into seven types regarding the social characteristics observed in the user's timeline on Twitter: i Potential bots for users who have Botometer Score7 bigger then 2.5; ii Independent experts for users who are experts in a given theme, which usually include journalists acting independently; iii Fake accounts for users who assume a false identity, passing through another personality; 7 Botometer checks the activity of a Twitter account and gives it a score based on how likely the account is to be a bot. Botometer Score ranges from 0 to 5. Higher scores are more bot-like (http://botometer.iuni.iu.edu) [ 12 ]. iv Fan accounts for users who identify themselves as celebrity fan accounts, or use the social network just to demonstrate their fanaticism (about a football team or a political party for example); v Primary media for users who represent the major channels of communication in the Brazilian context of news publishing; vi Secondary media for users who act as lower relevance media, comparing to primary media; and vii Typical users for other users who do not fall into the previous types. 3.3

Implementations and Parameters Setting

Our experiments use implementations8 and parameters settings as follows. { DC: we implement DC in Python9, and we run it setting p=0:5 (moderated decay) and T =d=8 (maximum endurance). The DC calculated values are normalized (0 DC(i) 1); { EC: we use Python NetworkX10 library, which is an iterative algorithm with two parameters: (i) max iter = 100, which de nes the maximum number of iterations in power method; and (ii) " =1e-06, which is the error tolerance used to check convergence in power method iteration. Calculated values are also normalized (0 EC(i) 1); { BC: we use Brandes Algorithm[ 8 ] with normalization (0 BC(i) 1); { NO: we implement NO in Python; and { HITS: we use Python NetworkX library, an iterative algorithm with three parameters: (i) max iter = 100, which de nes the maximum number of iterations in power method; (ii) " =1e-06, which is the error tolerance used to check convergence in power method iteration; and (iii) normalized = T rue, which normalizes results by the sum of all of the values. 3.4

Network Modeling

We model the online social network as a directed graph G, where nodes u and v (u; v 2 V (G)) represent Twitter users and an edge (v ! u) 2 E(G) means that user u retweeted a message from user v. Such a modeling provides social capital for users who publish (out-degree) or retweet (in-degree) messages. The graph G built from our dataset has jV (G)j=165,936 nodes and jE(G)j =280,898 edges.

Fig. 1 shows a toy example of our network modeling. It illustrates providers (red color nodes), spreaders (blue color nodes), and neutral (yellow color nodes). Another example is given by Fig. 2 with a small piece of our dataset that shows a potential bot (@Felipe 100) acting on the network. 8 Code available at

datasets.html 9 https://python.org 10 https://networkx.github.io/documentation Here, we analyze and compare metrics for social capital facets regarding the information di usion process to explore the importance of primary media and independent experts compared to typical users, fan accounts and potential bots in the Brazilian context of information di usion. First, we analyze our types of users according to the information capital measures (Section 4.1), an important contribution of our work in the context of misinformation in urban centers. Then, we expand such evaluation by considering the other forms of calculating social capital measure regarding: bridging (Section 4.2), brokerage (Section 4.3), and hubs and authority index (Section 4.4). We start our analyses with Decay Centrality (DC), which favors users that reach the largest number of neighbors in up to T hops. Thus, the providers that transmit information for other providers have higher DC. Table 2 presents the results, in which sports and entertainment news providers stand out (seven out of 10). Therefore, users have high EC if they are related to other users who are also well connected to the network. In our context, users who retweet the highly retweeted information are privileged. Hence, typical users, fan accounts and potential bots stand out, as the results in Table 3. Unfortunately, these are the user types that contribute to turning a subject into a trend, depreciating the quality of the main information served on Twitter, in the midst of the Brazilian upcoming elections.

As seen earlier, DC and EC are centrality metrics that reveal information capital of individuals in the network. However, both metrics present di erent results. DC reveals the main providers because they are nodes that start long sequences of broadcasts and relays. Meanwhile, EC reveals the main spreaders because they retweet many messages of important nodes (providers). Therefore, DC and EC do not measure the same event, and there is no correlation between them ( = 0; 095, p-value = 0)11. Our modeling process induces the formation of disconnected components because we create a component for each message, where this message points to its RTs. However, users tend to connect as a new TT arises. Then, bridges link weakly connected components, allowing broadcast through such weak tie. Overall, in our dataset, there are seven weakly connected components and 165,901 strongly connected components.

Fig 3 shows a subgraph that contains the top weak ties (red color edges) and their neighborhood. Each weak tie is an edge (v ! u), where v is a provider (represented as red nodes), and u is a spreader (represented as blue nodes). Nodes 11 is Spearman Rank Correlation Coe cient [ 41 ] u and v are bridges because they allow the connection between u and v's neighborhoods. Since bridging capital refers to the ties' strength, it does not make sense to correlate it with the aforementioned node centrality metrics. Speci cally in our dataset, bridging capital highlights ve political party fan accounts as spreaders 12, two independent experts as spreaders 13, three primary media as providers 14, and three independent experts as providers 15. Moreover, the potential bot \@paiva tv" stands out acting as secondary media in the network.

Bridging and brokerage are similar in the sense that both focus on the singular position of an individual in the network. Nevertheless, they are still di erent and measured through distinct forms. 4.3

Brokerage Capital Measure

Nodes with high BC are important brokers in communication and information di usion [ 21,40 ]. Table 4 shows that BC presents the primary media and independent experts as main brokers in the Brazilian information di usion context, where six out of 10 users belong to the \Grupo Globo"16, and all top 10 are providers. Thus, the information tends to circulate in the network passing through the providers, mainly belonging to a single group. We also analyze the correlation between BC and DC (not shown due to space constraints). There is a weak positive correlation between them ( = 0; 322, p-value = 0), as both measures identify information providers over network. The HITS algorithm [ 28 ] calculates Hub Index (HI) and Authority Index (AI), where HI measures social capital of users who produce much information (like providers) and AI measures social capital of users who mostly retweet, that is, users who spend a lot of time browsing the network (like potential bots). Table 5 presents the results in two parts. First, Table 5a shows the hubs as the main providers, where the sports information providers stand out (nine out of 10). Meanwhile, Table 5b shows the potential bots and football fan accounts stand out (seven out of 10) as authorities. Interestingly, all football fan accounts are fans of the \Clube de Regatas Flamengo (CRF)" (the largest football fan club in Brazil). The potential bots suspended17 by Twitter were also CRF fans. Probably, the suspended accounts were linked to tweetdecking18 practice to in ate popularity (social capital facet).

Furthermore, there is a weak positive correlation between AI and EC ( = 0; 261, p-value = 0) because both metrics use eigenvector concept and capture similar events (not equivalent). As seen in Section 4.1, providers who broadcast information for another providers have high DC values (like a hub). Hence, there is a very strong positive correlation between HI and DC ( = 0; 995, p-value = 0), and there is a weak positive correlation between HI and BC ( = 0; 324, pvalue = 0) { not illustrated due to space constraints. 5

Conclusion

Categorizing users in OSNs is an important task for understanding how information spreads over society. Humans as data sources can streamline the sharing of relevant events, whether for emergency events management, crime detection, urban administration, intelligent transportation, smart tra c control, 17 https://help.twitter.com/pt/managing-your-account/

suspended-twitter-accounts 18 http://www.newsweek.com/tweetdecking-why-twitter-suspended-multiple-accounts-840031 public healthcare, or even political engagement. It is a matter of major concern when bots or malicious users assume such activities.

Here, we analyzed the Twitter Brazilian users behavior in publishing and sharing Twitter trend topics to understand how important information ows over the network. We do so by addressing the social capital forms as related to individuals' abilities to acquire and spread information. We analyzed and compared social capital metrics to verify the importance of the primary media and independent experts compared to typical users, fan accounts, and potential bots, in the Brazilian information di usion context.

In general, potential bots and fan accounts are users who spread information through retweets, and they are the main authorities in the social network. Potential bots have automated behavior. Then, they can be programmed for malicious purpose. Furthermore, fan accounts exacerbate a fanaticism sentiment. For instance, the retweets may in ate the \hate speech" or spread fake news. This is very concerning given the upcoming Brazilian presidential election in October 2018. However, despite the Media Groups monopoly, we also found the primary media and independent experts as the main information providers, which may represent there is still hope in controlling misinformation over the network.

Acknowledgements.

Work funded by CAPES, CNPq and FAPEMIG, Brazil.

References

1. Ahsan , M. , Kumari , M. , Singh , T. , Pal , T.L. : Sentiment based information di usion in online social networks . IJKDB 8 ( 1 ), 60 { 74 ( 2018 )

2. Barabasi , A. ; Frangos , J. : Linked: the new science of networks science of networks . Perseus Books Group ( 2014 )

3. Bertolini , S. , Bravo , G.: Social capital, a multidimensional concept . In: Euresco Conference ( 2001 )

4. Bonacich , P. : Power and centrality: A family of measures . American journal of sociology 92(5) , 1170 { 1182 ( 1987 )

5. Borgatti , S. , Jones , C. , Everett , M. : Network measures of social capital . Connections 21 ( 1 ), 27 { 36 ( 1998 )

6. Bourdieu , P.: The forms of capital . In: Richardson, J . (ed.) Handbook of Theory and Research for the Sociology of Education , pp. 241 { 258 . Greenwood , New York ( 1986 )

7. Branda~o, M.A. , Moro , M. M.: Analyzing the strength of co-authorship ties with neighborhood overlap . In: International Conference on Database and Expert Systems Applications . pp. 527 { 542 . Springer ( 2015 )

8. Brandes , U. : A faster algorithm for betweenness centrality . Journal of Mathematical Sociology 25 ( 2 ), 163 { 177 ( 2001 )

9. Burt , R.S.: Brokerage and closure: An introduction to social capital . Oxford University Press ( 2005 )

10. Burt , R.S.: Structural holes: The social structure of competition . Harvard University Press ( 2009 )

11. Coleman , J.S.: Social capital in the creation of human capital . American Journal of Sociology 94 , S95 { S120 ( 1988 )

12. Davis , C.A. , Varol , O. , Ferrara , E. , Flammini , A. , Menczer , F. : Botornot: A system to evaluate social bots . In: Int'l Conf on World Wide Web, Companion Volume. pp. 273 { 274 . Montreal, Canada ( 2016 )

13. Doran , D. , Severin , K. , Gokhale , S. , Dagnino , A. : Social media enabled human sensing for smart cities . AI Communications 29 ( 1 ), 57 { 75 ( 2016 )

14. Easley , D. , Kleinberg , J.: Networks, crowds, and markets: Reasoning about a highly connected world . Cambridge University Press ( 2010 )

15. Freeman , L.C. : A Set of Measures of Centrality Based on Betweenness . Sociometry 40 ( 1 ), 35 { 41 ( 1977 )

16. Gao , S. , Janowicz , K. , Couclelis , H.: Extracting urban functional regions from points of interest and human activities on location-based social networks . Trans. GIS 21 ( 3 ), 446 { 467 ( 2017 )

17. Granovetter , M.S.: The Strength of Weak Ties . vol. 78 , p. 1360 { 1380 . JSTOR ( 1973 )

18. Hu , Y. , Song , R.J. , Chen , M. : Modeling for information di usion in online social networks via hydrodynamics . IEEE Access 5 , 128 { 135 ( 2017 )

19. Hui , C. , Goldberg , M. , Magdon-Ismail , M. , Wallace , W.A. : Simulating the di usion of information: An agent-based modeling approach . IJATS 2 ( 3 ), 31 { 46 ( 2010 )

20. Iannelli , F. , Mariani , M.S. , Sokolov , I.M.: Network centrality based on reactiondi usion dynamics reveals in uential spreaders . CoRR abs/ 1803 .01212 ( 2018 )

21. Jackson , M.O.: A typology of social capital and associated network measures . CoRR abs/1711 .09504 ( 2018 )

22. Jero^nimo, C.L.M. , Campelo , C.E.C. , de Souza Baptista , C. : Using open data to analyze urban mobility from social networks . JIDM 8 ( 1 ), 83 ( 2017 )

23. Kadar , C. , Brungger, R.R. , Pletikosa , I. : Measuring ambient population from location-based social networks to describe urban crime . In: Social Informatics . pp. 521 { 535 ( 2017 )

24. Kang , C. , Kraus , S. , Molinaro , C. , Spezzano , F. , Subrahmanian , V. : Di usion centrality: A paradigm to maximize spread in social networks . Arti cial Intelligence 239 , 70 { 96 ( 2016 )

25. Kang , S. , Shen , L. : A quantitative measure for meal-mate social capital networks . In: Int'l Conf. on Intelligent Environments . pp. 124 { 131 ( 2016 )

26. Kim , J. , Bae , J. , Hastak , M. : Emergency information di usion on online social media during storm cindy in U.S . Int J. Information Management 40 , 153 { 165 ( 2018 )

27. Kim , J. , Tabibian , B. , Oh , A. , Scholkopf, B. , Gomez-Rodriguez , M. : Leveraging the crowd to detect and reduce the spread of fake news and misinformation . CoRR abs/1711 .09918 ( 2017 )

28. Kleinberg , J.M.: Authoritative sources in a hyperlinked environment . J. ACM 46 ( 5 ), 604 {632 (Sep 1999 )

29. Kukka , H. , Kostakos , V. , Ojala , T. , Ylipulli , J. , Suopajarvi, T., Jurmu , M. , Hosio , S. : This is not classi ed: everyday information seeking and encountering in smart urban spaces . Personal and Ubiquitous Computing 17 ( 1 ), 15 { 27 ( 2013 )

30. Li , M. , Wang , X. , Gao , K. , Zhang , S.: A survey on information di usion in online social networks: Models and methods . Information 8 ( 4 ), 118 ( 2017 )

31. Michalak , T.P. , Rahwan , T. , Moretti , S. , Narayanam , R. , Skibski , O. , Szczepanski , P.L. , Wooldridge , M.: A new approach to measure social capital using gametheoretic techniques . SIGecom Exchanges 14 ( 1 ), 95 { 100 ( 2015 )

32. Newman , N. , Fletcher , R. , Kalogeropoulos , A. , Levy , D.A. , Nielsen , R.K. : Reuters institute digital news report 2017 . Reuters ( 2017 )

33. Putnam , R.D.: Tuning in, tuning out: The strange disappearance of social capital in america . PS: Political science & politics 28(4) , 664 { 683 ( 1995 )

34. Putnam , R.D.: Bowling alone: the collapse and revival of american community . In: ACM Conference on Computer Supported Cooperative Work . p. 357 ( 2000 )

35. Saito , K. , Kimura , M. , Ohara , K. , Motoda , H.: Super mediator - a new centrality measure of node importance for information di usion over social network . Inf. Sci. 329 (C), 985 { 1000 ( 2016 )

36. Smarzaro , R., de Lima, T.F.M. , Jr. , C.A.D. : Could data from location-based social networks be used to support urban planning? In: Proceedings of the 26th International Conference on World Wide Web Companion, Perth, Australia, April 3-7 , 2017 . pp. 1463 { 1468 ( 2017 )

37. Srivastava , M. , Abdelzaher , T. , Szymanski , B. : Human-centric sensing . Phil. Trans. R. Soc. A 370 ( 1958 ), 176 { 197 ( 2012 )

38. Statista: Most famous social network sites worldwide as of april 2018, ranked by number of active users , https://www.statista.com/statistics/272014/ global-social -networks-ranked-by-number-of-users , accessed: 2018-04-24

39. Stimmel , C.L. : Building smart cities: analytics, ICT, and design thinking . CRC Press ( 2015 )

40. Tang , L. , Liu , H.: Community detection and mining in social media . Synthesis lectures on data mining and knowledge discovery . Morgan and Claypool ( 2010 )

41. Wayne , W.D., et al.: Applied nonparametric statistics. Boston, MA: PWS-Kent ( 1990 )

42. Zheng , Y. , Capra , L. , Wolfson , O. , Yang , H.: Urban computing: Concepts, methodologies, and applications . ACM Trans. Intell. Syst. Technol . 5 ( 3 ), 38 :1{ 38 : 55 ( 2014 )