A Framework for Quantifying Controversy of Social Network Debates Using Attributed Networks: Biased Random Walk (BRW)
Hanif Emamgholizadeh, Milad Nourizade, Mir Saman Tajbakhsh, Mahdieh Hashminezhad, Farzaneh Nasr Esfahani
AA Framework for Quantifying Controversy of Social NetworkDebates Using Attributed Networks: Biased Random Walk (BRW)
Hanif Emamgholizadeh a , Milad Nourizade b , Mir Saman Tajbakhsh c , Mahdieh Hashminezhad d and Farzaneh Nasr Esfahani e a [email protected] b [email protected] c [email protected] d [email protected] e [email protected] A R T I C L E I N F O
Keywords :Social Media, Polarization, Contro-versy, Social Networks, Random Walk
A B S T R A C T
All societies have been much more bipolar over the past few years, particularly after the emer-gence of online social networks and media. In fact, the gap between two ends of social spectrumis going to be even deeper after the spread of new media. In this circumstance, social polariza-tion has been a growing concern among socialists and computer science experts because of thedetrimental impact which online social networks can have on societies by adding fuel to the fireof extremism.Several researches were conducted for proposing measures to calculate controversy level in so-cial networks, afterward, to reduce controversy among contradicting viewpoints, for example,by exposing opinions of one side to other side’s members. Most of the attempts for quantifyingsocial networks’ controversy have considered the networks in their most primary forms, withoutany attributes. Although these kinds of researches provide platform-free algorithms to be used indifferent social networks, they are not able to take into account a great deal of useful informationprovided by users (node attributes).To surmount this shortcoming, we propose a framework tobe utilized in different networks with different attributes. We propeled some Biased RandomWalks (BRW) to find their path from start point to an initially unknown end point with respect toinitial energy of start node and energy loss of nodes on the path. We extracted structural attributeof networks, using node2vec, and compared it with state-of-the-art algorithms, and showed itsaccuracy. Then, we extracted some content attributes of user and analyze their effects on theresults of our algorithm. BRW is compared with another state-of-the-art controversy measuringalgorithm. Then, its changes in different level of controversy in Persian Twitter is considered toshow how it works in different circumstance.
1. Introduction
The surge of online social network and media in the first years of new century seemed to be so promising not onlyfor personal relations but also for social and political freedom. Then, gradually the darker side emerged. Although inan individual level social network provides wider communication, users try to make relations with like-minded people[9, 49]. In social level bots, cyborgs and trolls attempt to manipulate public opinion and even determine elections’results [5]. One of the growing concerns in this regard is the users’ behaviors in controversial topics. Usually, con-troversial topics separate each community into, at least, two well-separated sub-communities with disparate opinionsabout the current issues, and online social networks escalate the controversy level by filter bubble and also users se-lective exposure , in which different sides of a debate are exposed only to the contents produced by other users withsimilar ideas [43]. This can make society even further radical and bipolar.Quantifying controversy was the objective of social researches for the past decades [34, 20, 23, 8]. To do so, sociol-ogists considered qualitative data and defined polarization as people’s distance from ideological perspective [8]. Theyconducted survey-based research for evaluating participants’ opinions about controversial topics. Afterward, conclu-sion were drawn by analyzing the results. The advent of social media in new century and huge free-to-use informationprovided by networks and users pushed sociology and computer science toward finding new methods for evaluating
ORCID (s):
Hanif Emamgholizadeh et al.
Page 1 of 23 a r X i v : . [ c s . S I] O c t iased Random Walk (BRW) polarization in societies. Adamic and Glance’s research [1] was the first study which made a bridge between offlinepolarization research and online world, their results were supported by Hargittai et al.s’ qualitative and quantitativeanalyses [31]. These two researches concentrated on weblog’s links for assessing users’ behavior regarding controver-sial issues. Conover et al. [18] extended previous research to social media. Recent researches also have confirmed thepolarization, especially in political discussions [6].After proving the existence of polarization, the next step is evaluating the amount of topic controversy which di-vides the society into two bipolar communities. A variety of measures from different perspectives have been proposed.On one hand, sociologists have been trying to find the distance of rival opinions [13]. On the other hand, computerscientists have been working with quantitative data in graphs or provided content in social networks [26, 44].Assessing and quantifying controversy is practical for detecting and predicting controversial topics [48], reducingpolarization [25], and also finding fake news [53]. Additionally, for better prediction of information diffusion [32]controversial topics should be considered deeply. It seems that polarization and controversy detection is going to bethe focus of more research in near future.The past researches about detecting controversial topic contained some shortcomings. For example, some of themconsidered only a small part of a network, which are boundary nodes [29]; others only rely on network structure [26].Finally, some of the algorithms are useful for special networks, such as signed networks [11]. However, all of thesemeasures have a common drawback; all of them ignore invaluable content and profile information provided by usersin social networks.To overcome this main drawback, we propose a framework for using in different platforms and networks with dif-ferent kinds of attributes. Our framework is based on a biased random walk (BRW) in which there is an initial energyto start a journey by random walker, and in each step a special amount of energy, called as energy loss, is reduced bythe node through which random walker passes. Based on the network and expected usage, there are several methods forassigning initial and energy loss; however, we have some suggestions for initial and energy loss, which our assessmentsdemonstrate to be useful for measuring controversy level of a topic.The rest of this paper is organized as follows. In the section 2, we provide an extensive review on the proposedmeasures for quantifying controversy. In section 3, we introduce our framework and suggest some variations forcalculating controversy in social networks. In section 4, we evaluate our variation of framework and compare it withone of the state-of-the-art methods. Finally, section 5, concludes this paper.
2. Related works
Before being considered from computer science perspective, polarization was a big concern in sociology. Dif-ferent researches have been conducted to evaluate polarization changes in societies [20, 23, 8]. Sociologists evaluatepublic opinion by quantitative data which is gathered from different surveys. Based on their definition which con-siders polarization as distance between opponent groups’ opinions, sociologists draw their results about the amountof polarity and its change. There are different measures to be used for evaluating quantitative data for calculating po-larization. Bramson et al. provided a set of measures to detect opinion distance between two contradicting groups [13].By the emergence of social networks, computer scientists started to consider this topic from their own perspectives,using new methods of text mining [35]. They tried to overcome the limitations of tradition methods using huge amountof invaluable information provided by users in social networks.Adamic and Glance [1] considered blogosphere in US 2004 presidential election. They made their basic networkusing the established links among the active blogs before the election day and found two well-separated groups ofsub-networks related to the Democrats or Conservative. Hargittai et al. [31] took both quantitatively and qualitativelydata into account and found similar behavior in users, like what reported by Adamic and Glance [1]. This line wasfollowed by other researchers too; one of the last researches has investigated Brazilian Political Crisis in 2016 [6]. Inthe lack of long term research on social polarization in social networks, Garimella et al. [27] focused on the long term
Hanif Emamgholizadeh et al. Page 2 of 23iased Random Walk (BRW)
Table 1
Researches for detecting controversy in offline and online communities.Type FeaturesName Network ContentOffline Dimaggio et al. [20] - × Fiorina et al. [23] - × Baldassarri et al. [8] - × Bramson [13] - × Online Adamic et al. [1] × -Hargittai et al. [31] × × Alves et al. [6] × -Garimella et al. [27] × - and persistent topics and provided a long term analysis for polarization in social media. All of the aforementionedstudies proved that there exist some kinds of segregations between contradicting political and social groups in socialnetworks, which makes exposure to other side’s points of view rare to the users, and this makes the situation evenworse, since for alleviating the amount of extremism, contradicting public opinion should be exposed to each other.Table 1 summarizes the research done to detect controversy on social media or real life.Studies on measuring controversy can be divided into three main groups: Content based, notwork based, andcombined models. If an approach uses network science tools, like community detection, random walk, and so on,it is called network based method. If it uses textual or qualitative information analyzing tools, for instance, NLP orstatistical metrics, the approach is called content based method. Finally, if an approach uses algorithms for analyzingtextual information; at the same time, it uses network science approaches for analyzing network data, the approach isknown as a combined method.• Content based: This group of methods go over the content produced by the network’s users and measure thecontroversy using hand labeling, crowd source, and NLP methods (Sociological methods also lay in this group)[20, 23, 8, 41, 15, 54, 13, 4].• Network based: This approach utilizes relations made by users and draws the results using only network topo-logical information [26, 44, 29, 17].• Combined methods: These methods have combined two aforementioned approaches for getting better results[24, 19, 41, 15].In the rest of this section, we will review proposed algorithms for measuring polarization and controversy in socialmedia. Content based approach includes two sets of methods, which consider contents produced by users in social net-works, opinion influence, and survey results. The first class of methods contains sociological perspectives in whichafter gathering qualitative information from surveys, some statistical methods are used for extracting controversy andbipolarity level in the society. This approach is beyond the scope of our work, so we are not going to investigate itdeeply (the interested readers are referred to [20, 23, 8, 13]). Among these methods, the most similar measure, whichhas a viewpoint near computer science perspective was proposed by Wojatzk et al. [54]. They put forward a method formaking crowdsourcing claims and ordering them. The claims are contents that collected from the social networks andranked using the scores assigned by participants. The results are evaluated by NLP methods. The most rudimentarywork in this regard is extracting and assigning scores for controversial words and phrases [40, 42]In the second class, contents are analyzed using the methods provided by NLP methods; these methods similarto the algorithms of the first class are network-free methods, i.e., they do not need network’s structural information.Although Garimella et al. [26] showed that bag-of-word and sentiment analysis are not distinctive measures for findingcontroversial issues, Al-Ayyoub et al. [4] defined some sentiment based methods and proved their usefulness in finding
Hanif Emamgholizadeh et al. Page 3 of 23iased Random Walk (BRW) and quantifying controversial topics.Controversy detection is a flourishing topic in the web sphere as well. The main purpose of this branch of contro-versy detection is making users aware of controversial issues in their web exploration. The articles of wikipedia, thewell-known online encyclopedia, have been considered for extracting controversy level of them [39, 12]. The amountof the controversy in the web, also, has been considered deeply [21, 36]. Dori-Hacohen et al. have used the mostfrequent words of a web page as inquiry for retrieving the wikipedia pages. Next, the controversy scores are assignedto retrieved wikipedia pages, and controversy level of the web page is elicited. Standing on this cornerstone research,Jong et al. [37] proposed a probabilistic method not only to find the controversial topics, but also to rank them. Firstly,they provided a probabilistic model [39], and in the next step, a language modeling approach is presented for detectingcontroversy. Jang et al. [36] have remarked two main drawbacks [21], namely, the query problem and score shortcut,and suggested a new method to address these shortcomings.Using sentiment analysis and NLP methods have made the bases of Tsytsarau et al. works [52] in which a frame-work for finding contradiction in news articles has been proposed, including three phases: Detecting topics, detectingsentiments, and analyzing the sentiments using mean and variance of the sentiment’s distribution.A set of content based researches relies on sentimental signs in documents. Choi et al. [16] considered publishedarticles from June 2001 to May 2002 for extracting controversial topics and their sub-topics. The first step was ex-tracting keywords to be used in the queries. The next step was assigning sentiment score for the negative and positivesignals in the documents. Finally, the last step was labeling the extracted topics as controversial or noncontroversial.They also devised a regression model for extracting the subtopics. In similar line, Ignatow et al. [33] used LSA [22]for finding topics of news articles and analyzing them based on sentiment analysis models.Beelen et al. [10] used articles’ comments for finding their controversy. They extracted four sets of features, namely,structural, linguistic, emotional, and WikiPedia similarity. Then, Random Forest and Support Vector Machine wereused for predicting controversy news.With regard to the lack of researches considering news articles in addition of social network content, Sriteja et al.[50] tried to combine the available data in news media’s articles on a topic with social media users reaction toward thetopic. Features, such as user opinion, controversial words, and topic intensity, which contain sentiment analysis, lexi-cal analysis, and users interaction, like the number of comment, are taken into account for calculating controversy score.
In the network based methods, the only instrument to be used is network. These methods purposefully neglectcontent data as it is deeply context-dependent. Different cultures need various questions in a survey and different NLPmethods, and trainers are needed for different languages. In addition, quantifying internal opinions is really a drudgery,if possible. Instead, these methods rely on structural features of networks, which are context-free and easy to use.Guerra et al. [29] were the first who used topological attributes for extracting controversy level. After finding twodisjointed community (for further study of community detection algorithms refer to [46]), they divided the nodes intotwo disjointed groups, the "boundary nodes" and the "internal nodes", the former set includes the nodes, which have atleast one edge to other community members (cross-group edges), and the latter set contains the nodes, which do havejust edge to the other nodes within their own community (internal edges). The edges that are neither cross-group norinternal are neutral, and then discarded. By defining 𝑑 𝑖 as the internal degree of a node or number of internal edgesand 𝑑 𝑏 as boundary degree or number of cross-group edges the polarity measures is defined as: 𝑃 = 1 | 𝐵 | ∑ 𝑣 ∈ 𝐵 [ 𝑑 𝑖 ( 𝑣 ) 𝑑 𝑏 ( 𝑣 ) + 𝑑 𝑖 ( 𝑣 ) − 0 . ] , where 𝐵 is the set of boundary nodes. Finally, the authors showed that the polarized topic has low density of high de-gree nodes in the boundary set. In other words, density of influential users (users with high degree in overall network) Hanif Emamgholizadeh et al. Page 4 of 23iased Random Walk (BRW) is low in the boundary set. Although this method overcomes the drawbacks of modularity measure [45] used in [18]as polarization measure, it has very limited perspective to the network and considers only a small set of users.Morales et al. [44], inspired by the electronic dipole moment, proposed a method similar to the well-known com-munity detection method, label propagation. They found two communities of the network and assigned the mostpossible amount of controversy (-1 and 1) to the most influential nodes of these communities. Then the process ofdistributing the assigned amounts of controversy for the selected users continued until convergence. Ultimately, thedistance between positive and negative distribution is considered as the controversy measure. The main shortcomingof this method is considering the influencers as the producer of content; however, based on controversy definition fromcomputer science viewpoint, the amount of users’ exposure to rival group content should be considered. Consequently,all users should play their own role in the distribution process.The most similar measure to our work is Garimella et al.’s [26] method. They also used a community detectionmethod for separating two polar groups (METIS [38]). Then a random walk is used for computing controversy measure.Lets 𝑃 𝐴𝐵 = 𝑃 [ 𝑠𝑡𝑎𝑟𝑡𝑖𝑛𝑝𝑎𝑟𝑡𝑖𝑡𝑖𝑜𝑛𝐴 | 𝑒𝑛𝑑𝑖𝑛𝑝𝑎𝑟𝑡𝑖𝑡𝑖𝑜𝑛𝐵 ] be conditional probability for a random walk; then, RWC (RandomWalk) is defined as: 𝑅𝑊 = 𝑃 𝑋𝑋 𝑃 𝑌 𝑌 − 𝑃 𝑌 𝑋 𝑃 𝑌 𝑋 , (1)where 𝑋 and 𝑌 are two communities of network. The random walk starts from each node and continues its journey toreach one of the influential nodes of network from both communities. The authors also presented a measure for eachuser’s controversy. This state-of-the-art method makes clear distinction between controversial and non-controversialissues; however, it suffers from a big shortcoming. The graphs used in this work, similar to the Morales et al. [44],are simple graph without any extra information on the edges or nodes. Undoubtedly, the presented information in thesocial networks, regardless of our ability to capture and measure them, can improve the final outcomes. For instance,Castillo et al [14] showed that combining social network’s information with structural information can improve ourability in detecting piece of fake information in social networks. To the best of our knowledge, the only proposedcontroversy measuring method for a network with other type of information except the relation is [11] which takes intoaccount sign of edges, as well.The entire aforementioned network based methods used a community detection method to segregate two opponentgroups. To overcome this problem, Coletto et al. [17] put forward a motif based and content independent methodwhich, used structural features and motifs for finding controversial issues. This method also built its result only onstructural information in the networks.Another usage of network for finding polarization is appeared in Al Amin et al.’s [3] work. The authors used net-work to reveal polarization too, but their made network is totally from a different nature. A bipartite graph in whichone of the sets includes sources or users who have produced a post or retweeted it, and the other set contains the poststhat are circulating in the network sphere. Then, using this network a matrix is made in which each entry is the combi-nation of a source activity probability in a camp and a post probability for circulation in it. The next step in this methodis factorizing this matrix into two separated matrices, which show the user activity and post circulation probability,separately. This factorization is achieved by optimizing an objective function using gradient decent method. Beforethis work Akoglu [2] also exploited bipartite graph; however, the used method for extracting controversy is different.Trying to find leaning the users in addition to a measure for ranking their partisan level, a bipartite graph was formed.The members were users and subjects. Markov Random Field was utilized for providing an objective function, and anapproximating method for solving the objective function, that is Loopy Belief Propagation, is used.Another opinion based method, which also utilizes the structure of network, is proposed by Amelkin et al. [7].Although the method puts forward a measure for calculating individual’s polarity level, the main aim of the paper wasproposing a distance method for opinion changes in a network. They have practiced transportation problem solutionfor finding the difference between two sides distribution on the consequent time instance. Their methods are useful forfinding anomalies in a network, probably injected by a number of malicious users. Hanif Emamgholizadeh et al. Page 5 of 23iased Random Walk (BRW)
The third class has something in common with both network and content based methods. On one hand, they workwith content and concepts; on the other hand, they use network for measuring controversy.Two of the opinion formation methods [24, 19] have made the foundation of opinion based models. Friedkin [24]proposed a method for opinion formation with two variable, 𝑠 𝑖 and 𝑧 𝑖 . The 𝑠 𝑖 is the internal opinion of user 𝑖 and 𝑧 𝑖 isthe expressed opinion of 𝑖 . Unlike Friedkin’s model [24], in DeGroot’s model [19] user’s opinion changes gradually ( 𝑠 𝑖 and 𝑧 𝑖 play the same role in this method too). Friedkin’s model was used further in the controversy detection methods.Matakos et al. [41] made a network of influence, and then executed a random walk to calculate 𝑧 𝑖 for each users as"polarization index". Polarization index is used further for evaluating controversy of a network. Friedkin’s idea wasalso used by Chen et al. [15]. They produced a controversy measure using z internal production ( z is vector of all 𝑧 𝑖 s in the network). Our algorithm is different from these opinion based algorithm in the information it utilizes. Ouralgorithm combines structural information provided by social network (and not opinion network) with user-providedinformation(and not their opinion).In one of the most recent studies aiming at quantifying controversy in social networks, Bonchi et al. [11] made aeigenvector based methods for evaluating controversy in signed graphs. This algorithm is different from the proposedalgorithm in this paper in the type of information that it uses. This algorithm only considers sign of edge; however,our algorithm takes all the provided information by the nodes into account.To the best of our knowledge, there is not any algorithm presented to calculate controversy which directly computescontroversy level in attributed graphs. Indeed, our approach tries to combine content and structural information, at thesame time, provide a flexible framework to be used in different situations and conditions.
3. Methodology
In this section, we intend to introduce a method for combining two sources of invaluable information, namely,content or textual data and structural data. We combine these sources of information for broadening horizons of con-troversy measuring algorithms, and also for surmounting present shortcomings in the proposed algorithms.If 𝑁 is a network; 𝑁 and 𝑁 , where 𝑁 ∩ 𝑁 = ∅ and 𝑁 ∪ 𝑁 = 𝑁 are two communities of 𝑁 . We want toevaluate the average level of nodes in 𝑁 to which the idea of 𝑖 as a member of 𝑁 will be exposed.Before introducing our methodology, we provide some definitions, which are going to be used in the rest of thisstudy. Controversial topics:
Controversial topics are the topics, which divide societies into at least two contradicting groupsabout the issue.
Attributed graph:
Attributed graph is a graph, which has some extra information for the nodes and/or edges. Forinstance, number of friends of the user, number of posts, etc. In attributed graph, each node or edge has a vector whichcontains the corresponding node’s information.
People’s political stance, most of the time, arise from their personal characteristics and psychology. Thus, personalinformation, which is extractable from users profile, is a dependable source of information. For instance, it is notsurprising to see a socialist political activist to support a candidate from the left side of a political spectrum. In additionto the previous information, as user attribute, the used terminology in the produced text and even used hashtags bytwo sides of a controversial topics shed light on the level of controversy in social network. For instance, Fig. 1illustrates the frequency of used hashtags by two sides of a controversial topic (Iran November 2019 protests) in theword clouds. As it is clear, there are some shared hashtags used frequently by two sides, and in the same breath,there are other representative hashtags, which make two sides of a controversial topic distinguishable. For instance,
Hanif Emamgholizadeh et al. Page 6 of 23iased Random Walk (BRW)
Table 2
Classification of researches for presenting a controversy measure.Type FeaturesName Network ContentSociological Bramson et al. [13] - × Dimaggio et al. [20] - × Fiorina et al. [23] - × Baldassarri et al. [8] - × Wojatzki et al. [54] - × Klenner et al. [40] - × Mejova et al. [42] - × NLP Al-Ayyoub et al. [4] - × Kittur et al. [39] - × Borra et al. [12] - × Dori et al. [21] - × Jang et al. [36] - × Jang et al. [16] - × Tsytsarau et al. [52] - × Ignatow et al. [33] - × Beelen et al. [10] - × Sriteja et al. [50] - × Network Guerra et al. [29] × -Morales et al. [44] × -Garimella et al. [26] × -Coletto et al. [17] × -Al Amin et al. [3] × -Amelkin et al. [7] × -Akoglu [2] × -Combined Friedkin [24] × × Chen et al. [15] × ×
Bonchi et al. [11] × ×
BRW × ×
Figure 1:
Word clouds of used hashtags by two contradicting sides of political debates in Persian tweets, November 2019.
Persian hashtags, are frequently utilized hashtags by pro-government Twitter users. In contrast, protesters frequentlyused
Hanif Emamgholizadeh et al. Page 7 of 23iased Random Walk (BRW) opinion, introduced by a user, in the social network; the more central node, the more ability to be heard. For in-stance, the probability of being a highly retweeted post or idea, which has been shared by a celebrity is much morethan an ordinary user. Therefore, it is reasonable to expect that the amount of being exposed by other users for aspecial idea, likely depends on the position of producer of the idea. The position of the producer of a node can becalculated by the structural or content features (or both of them) of nodes compressed in a vector. Additionally, a postor idea do not endlessly circulate in a network and have a lifetime. To take into account these characteristics of thepost or idea, we simulate the lifetime of an idea by reducing its initial energy in each step based on the energy lossof the node through which the random walker passes. This is called energy loss of node and is proportionate to theposition of a node. Due to being interested in the random walk which passes through the boundary edges, we estimatethe position of nodes for calculating energy loss with respect to the average position of the nodes in other communities.Moreover, in Garimella et al.’s algorithm [26], they considered only the random walks, which start from influentialusers and reach to other influential users on the other sides of a debate. Meanwhile, we believe that the amount ofcontroversy measures should be calculated based on their ability for penetrating the other communities. Networks canbe structured in different layer, and depth of each layer can be calculated with respect to its distance from the boundarynodes. Thus, the ability of a random walk in penetrating the community, other than the one which it belongs to, isequal to the maximum layer of nodes (belong to other communities) which it passes through.Boundary nodes are defined as the nodes in a community, which have at least one edge to a node in another com-munity. Starting from these boundary nodes and putting their level as zero, the set of level 1 contains the nodes, whichonly have one or more edges to the nodes in level 0 and nodes in set of level 2. Generally, the level of a node is 𝑛 if ithas at least one link to a node in level 𝑛 −1 and not fewer level, at the same time, at least one link to a node in level 𝑛 +1 ,if exist (though this definition similar to Breadth First Search (BFS), but it is different in having more than one startingpoint). Now, we can define a random walker’s ability to drilling down the other side’s community as the maximumlevel of nodes, in the other community, through which the random walker passes. Figure 2 indicate the flowchart ofthis framework.To sum up, historically, the controversy level of a topic from a social network perspective has been defined as theaverage amount of being exposed to contradicting opinion, regardless of corresponding users’ reactions (acceptance orrejection) [26, 29, 44]. Users establish closer and stronger echo-chambers for controversial topics. For instance, figure3, indicates a community with a boundary and two levels. In this figure, nodes in level 1 are exposed to informationproduced by boundary nodes. Thus, the users in level 1 communicate with more like-minded people in comparisonto nodes in the boundary. Having propagated a contradicting point of view, boundary nodes have a much moderateviewpoint compared to the nodes in level 1, which only have exposed to the boundary nodes idea. Additionally, thenodes in level 1 have weaker echo-chamber, and therefore, a more moderate point of view compared to the node inlevel 2, which only sees ideas of the nodes in level 1. Thus, the further from the boundary nodes, the stronger andcloser echo-chamber, and the more radical opinion will be.Having this in mind, if an idea is able to go to the deeper levels in the contradicting communities, its controversylevel is fewer. Consider 𝑖 distributes Hanif Emamgholizadeh et al. Page 8 of 23iased Random Walk (BRW)
Figure 2:
Framework flowchart. tweet or idea 𝑡 . Then, we have 𝑟 𝑖 = ( 𝑖, 𝑡 ) , where 𝑟 𝑖 is the corresponding random walker. From its starting point, 𝑟 𝑖 ’sresponsibility is simulating the diffusion pattern of ( 𝑖, 𝑡 ) . In each step, 𝑖 idea’s energy is reduced because the probabilityof being retweeted when a content goes further from its producer decreases. It seems that this energy should be refilledif the content is retweeted by an influencer. However, if 𝑡 is retweeted by influencer ( 𝑗 ), then random walk which startsfrom 𝑗 and distributes idea 𝑡 , i. e. 𝑟 𝑗 = ( 𝑗, 𝑡 ) , refills this energy and simulates its distribution pattern. At the same time, 𝑟 𝑖 = ( 𝑖, 𝑡 ) as the simulator of 𝑖 ’s idea continues its work to death. In fact, the random walk 𝑟 𝑖 energy should have beenretweeted if there were only a set of selected nodes for initiating the random walks, whereas, in our framework, eachnode has its own random walker. In fact, our algorithm algorithm takes into account influencer and non-influencer’sopinion distribution, separately.There are several ways to assign initial energy and energy loss in attributed graphs. Here, we introduce two basicmethods, which can be freely used in any kind of attributed network; notwithstanding, the framework is flexible enoughto be used in different ways.There are several approaches for assigning initial energy for each node, such as degree centrality, betweennesscentrality, or eginvector centrality. However, for our version of this framework, we adapt a data mining and machinelearning point of view.Influencers have higher centrality and more audiences. Therefore, the distance from the central points shows theability to attract more audiences; these distances determine the initial energy of a node, that is, the energy by which arandom walker starts its journey. Furthermore, influencers of the other communities attract more like-mind audiences.Thus, the more distance from other community’s influencers, the fewer audiences with the contradicting point of viewwill be, eventually, the more energy loss for that node we will have. It is worth mentioning that we want to find the Hanif Emamgholizadeh et al. Page 9 of 23iased Random Walk (BRW)
Figure 3:
Structure of a community as one side of a debate in a social network. ability of a random walker to reach other community; hence, we take the contradicting community’s central point intoaccount for calculating energy loss.We have attributes of the nodes; therefore, we can map each of these nodes to a point in the space. We know thatthese nodes have been assigned to two disjointed sets using a community detection algorithm; hence, we are easilyable to find the central nodes of these two node sets using these attributes, which are representative vector of nodes.Having mapped nodes on the space and calculated the central nodes of each group, we define the initial energy andenergy loss as follows:
𝐼𝑛𝑖𝑡𝑖𝑎𝑙𝐸𝑛𝑒𝑟𝑔𝑦 = 1
𝐷𝑖𝑠𝑡 ( 𝑛𝑜𝑑𝑒, 𝐶𝐶 𝑖𝑛 ) + 1 𝐷𝑖𝑠𝑡 ( 𝑛𝑜𝑑𝑒, 𝐶𝐶 𝑜𝑢𝑡 ) (2) 𝑙𝑜𝑠𝑠𝐸𝑛𝑒𝑟𝑔𝑦 = 𝐷𝑖𝑠𝑡 ( 𝑛𝑜𝑑𝑒, 𝐶𝐶 𝑜𝑢𝑡 ) (3)Where 𝐶𝐶 𝑖𝑛 , 𝐶𝐶 𝑜𝑢𝑡 are central node of communities, which corresponding node belongs to and does not belongto, respectively, and 𝐷𝑖𝑠𝑡 ( ., . ) is euclidean distance (or any other distance measure). Intuition for this measure is asfollows: the ability to be heard (initial energy) depends on the distance from current community’s average point of viewto start circulation of a piece of information, and distance from the average point of view (central point of community)of other community (because we are interested in the ability of a random walk in penetration of other community). Forenergy loss, the intuition is as follows: the farther from middle point of contradicting community, the lower chance toreach to another community, and therefore, to be heard. To avoid trapping in the local neighborhood, the initial energycan be multiplied to a sufficiently big number, but it should be noted that to compare a different network, the sameconfiguration has to be used. Fig. 4 shows the position and distance of node from central points. Hanif Emamgholizadeh et al. Page 10 of 23iased Random Walk (BRW)
Figure 4:
Central and initial points and a node position. Red and blue ellipses are two communities. In this figure we have
𝐼𝑛𝑖𝑡𝑖𝑎𝑙𝐸𝑛𝑒𝑟𝑔𝑦 = 𝑎 + 𝑏 and 𝑙𝑜𝑠𝑠𝐸𝑛𝑒𝑟𝑔𝑦 = 𝑏 . In our implementation, especially for being able to compare to other methods, we change network structural in-formation to attributes. In fact, there are several methods for the representation of network [30]. New representationmethods are introduced to map networks to a lower dimension space in which the geometric position of nodes reflecttheir structural attributes. Therefore, in this representation, we expect to see similar nodes near each other. This sim-ilarity in pure networks, without any attributes, is based on the edges; thus, we believe that community nodes shouldbe much closer to each other in newly mapped versions.There are several representation algorithms to map a node on space [28, 51, 47]. We selected node2vec [28] to mapour network to space. The first objective of node2vec is preserving network’s neighborhood structure, and it performsthis using a biased random walk. Having this in mind, we try to assign a vector to each node, and this vector is theoutput of node2vec algorithm. Now, we have structural information of a node as its attribute.Additionally, we have other attributes as vector for nodes. By concatenating on these features, such as used hashtagfrequency or TF-IDF of used words, we can have broader perspective toward the node.However, we have more things to do for achieving a better loss function. As mentioned before, we have twosets of nodes: Boundary nodes and non-boundary nodes (which we call them internal nodes). Our loss function, forsimulating the ability of a node in transferring a piece of information from belonging community to other community,has two kinds of perspectives, local perspective and global perspective. In local perspective, we just consider positionof a node in the corresponding community with its neighbors. Two different measures for two different kinds of nodes,boundary and internal, are available. First, conductance [55] for a node is defined as the ratio of of edges, which stayinside the community, to the edges which go to the other community:
𝐶𝑜𝑛𝑑 𝑖 = | 𝑒𝑑𝑔𝑒 𝑜𝑢𝑡 || 𝑒𝑑𝑔𝑒 𝑖𝑛 | , (4)where | 𝑒𝑑𝑔𝑒 𝑜𝑢𝑡 | , | 𝑒𝑑𝑔𝑒𝑖𝑛 | are the number of edges, which go out of the community that node 𝑖 belongs to and thenumber of edges that stay inside the community, which node 𝑖 belongs to, respectively. Concentrating on the definitionof internal nodes which do not have any links to a node belonging to another community, it is evident that conductanceis only meaningful for the boundary nodes and is equal to zero for internal nodes. Indeed, conductance is the abilityof a node for passing a piece of information to another community. For internal nodes, this ability lays on the level towhich the node belongs. For example, ability of a node which is in the level 𝑛 for passing a piece of information toother community is more than a node which lay in level 𝑛 + 1 since it is a level far from a node in another community.Thus, we need another measure for internal nodes, and it is exactly the level to which a node belongs. We call thismeasure closeness to demonstrate how much a node is close to another node with different community: 𝑐𝑜𝑙𝑠𝑒𝑛𝑒𝑠𝑠 𝑖 = 𝑙𝑒𝑣𝑒𝑙 𝑖 , (5)where 𝑙𝑒𝑣𝑒𝑙 𝑖 is the level which node 𝑖 belongs to. Hanif Emamgholizadeh et al. Page 11 of 23iased Random Walk (BRW)
For global perspective, we regard the distance of a node from the central node of another community using attributesof the nodes represented by vectors. The more distance to contradicting community’s central point, the more loss inenergy, and the less ability to pass a piece of information to other side: 𝑑𝑖𝑠𝑡 𝑐𝑒𝑛𝑡𝑒𝑟 𝑖 = 𝐷𝑖𝑠𝑡 ( 𝑖, 𝐶𝐶 𝑜𝑢𝑡 ) . (6)We know that distance from the other community’s central point, and closeness has direct relation with energy loss,and conductance have inverse relation with it. Finally, we have this relation for the loss function: 𝑙𝑜𝑠𝑠 𝑖 = 𝛼 ∗ 𝑐𝑙𝑜𝑠𝑒𝑛𝑒𝑠𝑠 𝑖 𝑚𝑎𝑥 _ 𝑐𝑙𝑜𝑠𝑒 + 𝛽 ∗ 𝑑𝑖𝑠𝑡 𝑐𝑒𝑛𝑡𝑒𝑟 𝑖 + ( 𝛾 ∗ 𝑚𝑎𝑥 _ 𝑐𝑜𝑛𝑑 − 𝐶𝑜𝑛𝑑 𝑖 𝑚𝑎𝑥 _ 𝑐𝑜𝑛𝑑 ) (7)where 𝑚𝑎𝑥 𝑐 𝑜𝑛𝑑 is maximum conductance in the graph and 𝑚𝑎𝑥 _ 𝑐𝑜𝑛𝑑 − 𝐶𝑜𝑛𝑑 𝑖 𝑚𝑎𝑥 _ 𝑐𝑜𝑛𝑑 is used for normalization. 𝑚𝑎𝑥 _ 𝑐𝑙𝑜𝑠𝑒 is themaximum closeness for the nodes whose corresponding node is in the same community with them. 𝛼, 𝛽, 𝛾 can be usedfor tuning the effect of each elements (we put all of them equal to 1 in our version).Algorithm 1 shows how the Biased Random Walk works. In each step, if the current energy is more than 0 oneof the neighbors is chosen as target node and the current energy will be reduced proportionate to energy loss of thetarget node. It is worth to mention that due to being direct graph, it is not surprising to see some nodes in each level asdangle nodes, which do not have any outgoing edge. In these cases for the next step, the BRW choose one of the nodein the same level as the current node, then, the current energy is reduced proportionate to target node energy loss, andBRW continue its work to run out its energy ( 𝑐𝑢𝑟𝑟𝑒𝑛𝑡 𝑒 𝑛𝑒𝑟𝑔𝑦 < ). Data:
An attributed graph
Result:
Maximum level of a node, belonging to other side of debate, which random walk pass through.Pre-processing: Assign Initial energy and energy loss for each node;MaxList = []; while
There is an unprocessed node do Maximum level = 0; Current Node = Starting Node of a random walk; Current energy = Initial Energy ofCurrent Node; while
Current energy> 0 do current node = randomly selected neighbor of current node;Current energy = Current energy - current node’s energy loss; if Maximum level < Current node’slevel then
Maximum level = Current node’s level / maximum level of community; endend
Add Maximum level to MaxList end return MaxList;
Algorithm 1:
Framework for calculating controversy level of an attributed graph.The next step is extracting the ability of each random walker in penetration the other community. After starting thejourney from a node, the maximum level of nodes, laying in other community, divided into the maximum level of thatcommunity can be considered as the random walker’s exposure level. There are two versions of this exposure level.• First, the average on all journeys, regardless of their ability in crossing boundaries or staying on current commu-nity, (in this version if random walk stays in its own community the exposure level is equal to zero. If it crossesthe boundary nodes the exposure level is one unit more than the maximum level of node through which the ran-dom walker passes, for making this controversy level distinguishable from the previous condition by assigning1 to the boundary nodes of other community). We call this measure which is bigger for non-controversial topicsand smaller for controversial topics as R andom W alk P enetration R ate (RWPR).• Second, the average only on the random walks which crosses the boundary of communities (in this version,as well, the exposure level is one unit more than node’s level), known as B oundary C rossed R andom w alk P enetration R ate (BCRPR). For considering both of these measures, we calculate RWPR and BCRPR average ,too. Hanif Emamgholizadeh et al. Page 12 of 23iased Random Walk (BRW)
It is clear that our implementation of this framework is not the only one and each user based on her or his needscan adapt the framework and have its own version of this framework. In the evaluation section, we also concatenateon textual attributes to the structural attributes vector to evaluate the other attribute role in finding controversy.
Com-plexity.
For each node a predefined number of walks, 𝑤 , start. Each of the walks starts with a specific start energy,which depends on the start node’s features and other nodes’ position in the space; we call average start energy of allnodes as ̄𝑠 . Then, the walk goes forward until it run out of energy, and this depends on loss energy of nodes. We callaverage loss energy of network as ̄𝑙 . Now, average complexity of algorithm is: 𝑂 ( 𝐵𝑅𝑊 ) = 𝑛𝑤 ̄𝑠̄𝑙 , (8)where 𝑛 is number of nodes in the network.
4. Evaluation
In this section, we are going to evaluate our implementation of framework (BRW). Two kinds of evaluations areused to show the capabilities of BRW. First, to show why we need new information source to get more reliable results,we compare BRW with RW [26], which has been shown to outperform all other state-of-the-art algorithms becausewe perform our evaluation on the data used in [26] and in that work, the authors showed that their measure outperformother methods, we only compare our methods with RW. Although our methods, unlike other modern methods, arenot directly presented for structural features, the results are not worse than new random walk based model [26]. Inaddition, we deeply analyze the behavior of the BRW using the collected data from Persian tweets to show how changesin structural and content information are reflected in the outcomes of BRW.
As mentioned in section 2, one of the last methods for evaluating controversy level of a topic is RW based method,which Garimella et al. introduced in [26]. This method has been presented to directly deal with structural features of anetwork. In this section, we compare our methods with the RW. Even though we overlook some precious informationby compressing the structural information in a much smaller vector using Node2vec, our results in detecting controver-sial topic is not worse than RW. Indeed, the RW, like our method, is not able to be more efficient than random choicein detecting controversial topics, based on the data collected by Garimella et al. [26], this is one of our motivations inconsidering another source of information.
For assessing our method with Random Walk, we use the dataset presented in [26]. The datasets contain follower-followee network in addition to retweet network; however, there is a consensus in the literature, which the best networkfor detecting controversial topics is retweet network [18]. Therefore, we only work with retweet network which isshown in the Table 3 and Table 4 as controversial and non-controversial networks, respectively .In these datasets, the hashtags are the keywords for collecting data from Twitter. Table 3 contains topics whichare controversial, and Table 4 includes the information of non-controversial topics (In fact, we tried to extract textualinformation of tweets because the tweet IDs were available; however, we could only collect about 30 percent of tweetsdue to being deleted or private account. Therefore, the information was not enough to be used in the content featurebased evaluation). The final results of controversy level for RW and RWPR for BRW is shown in Table 5 and Table 6 for controversialand non-controversial topics, respectively. We used published code of RW, by the authors , for this assessment. Thedefault configuration was used for RW in which the random walk repeated for 100 times. There are much more thingsto do in our algorithm. However, for simplicity, we tried to keep everything as simple as possible. We used node2vec We have used these network as it has been presented; however, there is only one exception. Mothers’ day network was prohibitively big, todeal with this problem we took half of the nodes randomly and made their induced graph. https://github.com/gvrkiran/controversy-detection/archive/master.zip Hanif Emamgholizadeh et al. Page 13 of 23iased Random Walk (BRW)
Table 3
Information of controversial topics networks [26].Hashtag
Table 4
Information of non-controversial topics’ network [26] networksHashtag for getting embedded vector of each node. The random walk repeated 50 times for each node, and in this special case,we do not multiply the initial energy to a big number.In Fig. 5, we show scatter plots of controversy measure for RW and BRW. There are 10 controversial (red) and 10non-controversial (blue) topics in each plot. After calculating the measures, we separated 10 first points (10 secondpoints) in RW (BRW) plot as non-controversial topics, and the rest as controversial topics. As it is evident bothof the methods are not able to detect controversial topics better than chance, precision of both of the methods are0.5. However, it should be noted that our method is not an only structural based method (unlike RW), and during thecompression process (using node2vec), we have to neglect some information. This poor performance pushes us towardtaking into account other available sources of information in the networks. Iran has been one of the most bipolar societies in the world due to the evergrowing gap between government and aconsiderable portion of the people, and this makes this country a good source of data for controversial topics. Over thepast few years, there have been a number of protests and riots for different reasons. One of the last protests embarkedbecause of a great increase in the petrol price in November 15th of 2019 initiated a chain of protests and riots in differentcities of Iran. After a while, this gap has been filled as an aftermath of Iran popular general’s death in January fifthof 2020 by the United State government. We collected Persian tweets from 1 November 2019 to 1 February 2020 forcovering all the important events in Iran political environment. We extracted different tweets for different aims, whichare elaborated upon in the rest of this section. The penetration level for BRW, unlike controversy level for RW, is big when topic is not controversial.
Hanif Emamgholizadeh et al. Page 14 of 23iased Random Walk (BRW)t
Table 5
RW and BRW for controversial topics.Hashtag Random Walk BRW
Table 6
RW and BRW for non-controversial topics.Hashtag Random Walk BRW (a) Controversy measure for RW.(b) Controversy measure for BRW.
Figure 5:
Controversy measure for RW and BRW.
In the midnight of November 15th the government suddenly announced that the price of petrol would increase fromthe next day. Iranians who had been living in a hard life under the United State sanctions started a chain of protestson the day after this announcement. The protests lasted for several days and a number of people were killed in theriots. We collected Persian tweets 15 days before and 15 days after this event. As a preprocessing phase, we extractedall the related hashtags to this event (political hashtag with more than one occurrence), and using these hashtags alltweets containing these hashtags were elicited. After establishing the corresponding network, the largest connectedcomponent containing 1216 nodes and 2676 edges was used as the network for assessment.
Hanif Emamgholizadeh et al. Page 15 of 23iased Random Walk (BRW)
For extracting the attributes of nodes, we only used existing hashtags in the tweets of each user (other informationlike profile information and word TF-IDF also can be used, but we restricted our data to hashtags). For each user, therewas a table with two rows, the first one containing the list of used hashtags by all users, and the second row, includingthe number of times the corresponding hashtags had been used by the user.In this section of evaluation, we are going to show how RWPR changes when we impose some noise on the node’sattributes and structure. The imposed noises are:• Structural noise: With a probability for each node, the noise adds an edge from the corresponding node to arandom node in the graph.• Attribute noise: With a probability in this kind of noise, the hashtag table of corresponding node is replacedwith the hashtag table of another randomly selected node.For clearly assessing the behavior of BRW, we made three versions of network for this data set. In the first version,we only took into account the structural features of network, extracted by
Node2vec . Indeed, Node2vec was used toproduce a vector with 20 dimensions as the features of nodes. In the second version, hashtags were the only featureof nodes. For having equal dimension with Node2vec version, we utilized PCA to reduce the dimension of hashtagtable for each user to be 20, known as hashtag network. Finally, we produced a vector with 10 elements for each user,provided by Node2vec, and put it beside the reduced vector (with 10 elements) of hashtags to make a vector with 20dimensions, we call this version both .As mentioned before, we have three measures: RWPR, BCRPR, and RWPR and BCRPR average ( for shorteningaverage). We illustrate these measures’ changes under different circumstances.First of all, we consider Node2vec version 6. Clearly, the amount of RWPR, BCRPR, and average increase bythe increase of noise. This demonstrates that the ability of penetration increases as the number of randomly assignededges rise. We know that one of the main problems of controversial topics is the lack of boundary nodes to help thediffusion of information from one side to another side. The more randomly added edges, the more boundary nodes,and the bigger RWPR will be. As we see below, the only clear increasing behavior for BCRPR is in Node2vec versionbecause randomly added edges by connecting nodes from different level compacts, and the depth of the graph makethe nodes in contradicting communities near to each other; therefore, the crossed random walk can easily achieve thefarther level of contradicting community.Fig. 7 indicates the average initial energy and energy loss in Node2vec network. Although initial energy’s behavioris in line with what we expect, the energy loss increases. It seems that this behavior stems from the dissimilarityimposed by new edges among the nodes. However, this increase in the amount of energy loss can be compensated by theincrease in the number of boundary nodes. We know that boundary nodes are the nodes which play as bridges betweentwo communities. New randomly added edges increase the number of crossing edges between two communities;consequently, increase the number of boundary nodes. The more boundary nodes we have, the more random walkerswhich go through one side to other side of a network will be. Fig. 8 shows the increase in number of boundary nodeswhich compensates the increase in energy loss.The next considered network is hashtag network in which the only features for each node are the reduced vector,by PCA, of hashtag usage. Fig. 9 shows the behavior of BRW for hashtag network. Fig. 9a indicates the RWPR. Thegraph increases slowly in the first steps and suddenly rises significantly at bigger noises. Because it is not evident, wealso demonstrate the exact number of Fig 9a graph in Table 7. Fig. 9b illustrates that the BCRPR depends on networkstructure and not other kinds of features, as described before. Fig. 10 shows that as we expect the initial energy ofnodes increases when their hashtags become similar, at the same time, the energy loss of nodes decreases, which isindicative of nodes’ gentility in passing information from one side of network to another side.Now we consider the behavior of the network when both structural (Node2vec) and content information (hashtags)are taken into account (we call this network as both_hashtag since we do not have noise on hashtag features to distin-guish it with the next subsection’s networks in which we have noise on both structure and content information). In otherwords, in this special case, we track behavior of network when both network and content behavior exist, but we have
Hanif Emamgholizadeh et al. Page 16 of 23iased Random Walk (BRW) (a) RWPR of Node2vec network. (b) BCRPR of Node2vec network. (c) Average of RWPR and BCRPR.
Figure 6:
Graph of BRW behavior with different level of noise for Node2vec network. These graphs indicate reduction inthe controversy level of network by injecting noise to the network. (a) Average initial energy of Node2vec network fordifferent level of noise. (b) Average energy loss of Node2vec network for dif-ferent levels of noise.
Figure 7:
Average initial energy and energy loss in Node2vec network. The graphs demonstrate that central points ofcommunities by injecting noise approach to each other. Hence, we have increase in the average initial energy. noise only on structural features. Similar to Node2vec network, in this case, also, RWPR, BCRPR, and consequentlyaverage metric increase when we randomly add edges to push well-separated communities toward each other 11. Theinitial energy and energy loss simulate the behavior of Node2vec network, as well, Fig. 12.Finally, we evaluate the changes in the behavior of BRW when we impose noise on both structural and contentinformation. For this case, we add five levels of structural noise to the networks(0.2, 0.4, 0.6, 0.8, 1.0), then, foreach case, we calculate their controversy level by adding noise to content features. Fig 13 indicates the behaviorof five network produced by adding noise to their structures. Additionally, x axis shows the imposed noise on thecontent feature of networks. By adding the noise; therefore, making the nodes similar to each other increase RWPR,but BCRPR remain almost the same. Although initial energy increase in bigger structural noise, there is no sign ofinfluence imposed by content noise in this case, Fig 14a. Maybe this originates from sparseness of content date. Energyloss for the network decreases, as expected 14b.
For comparing the changes of political and non-political topics controversy level in Persian tweets we made 5,versions of networks. Then, the results for Node2vec, Hashtag, Both, in addition to RW is presented for comparingthe results of our algorithm with RW as the baseline method, which outperforms other controversy measuring methods[29].
Hanif Emamgholizadeh et al. Page 17 of 23iased Random Walk (BRW)
Figure 8:
Number of boundary nodes in Node2vec graph for different level of noise. When we inject noise to the networksome of the nodes from different communities connect to each other and join to boundary node’s set. Therefore, numberof boundary nodes increase. t
Table 7
Table of RWPR Noise PNC0.0 0.0021870.1 0.0021810.2 0.0021910.3 0.0022470.4 0.0022720.5 0.0024600.6 0.00254030.7 0.0034830.8 0.0052050.9 0.0147291.0 0.018345
First of all, we extracted all the tweets which contained at least one of the hashtags which happened more than150times among all the tweets, collected from November first to January 29th. The established network on this data set isthe network which we call it
All in the table. All_politic is the network which we made this network by extracting allpolitical hashtags with more than 150 occurrences among the all tweets (pay attention that in the section 4.1, we con-sidered all the tweets with political hashtags regardless of the hashtags frequency). Finally, we separated
All networkwith respect to the tweet date to Nov., Dec., and Jan. (Table 8).In the past few years, Iran has encountered several protests regarding the political and political economic issues.
Hanif Emamgholizadeh et al. Page 18 of 23iased Random Walk (BRW) (a) RWPR of hashtag network. (b) BCRPR of hashtag network. (c) Average of hashtag network.
Figure 9:
Graph of BRW behavior with different levels of noise for hashtag graph. These graphs indicate that RWPR,unlike BCRPR and Average of hashtag network, reflect the inject the nature of inject noise correctly. (a) Average initial energy of hashtag network for dif-ferent levels of noise. (b) Average energy loss of hashtag network for dif-ferent level of noise.
Figure 10:
Average initial energy and energy loss in hashtag network. As it is clear, when we inject noise to hashtagnetwork the average initial energy increase significantly, and the average energy loss decrease.
Therefore, we have to consider the political issues much more controversial, as they end up with street protestations asthe maximum level of controversy compared to daily life events.Due to bipolarity level in Iran and the small number of boundary nodes, most of the random walks could not endup in another community, if any. For having more realistic perspective in this section, we multiplied the initial energyto 200, this is the case for all the runs; the other variables of configuration remained the same as before.
Table 9 shows the RWPR and RW of the networks. As expected, the network containing all topics is less contro-versial, and this is evident in bigger RWPR for
All network compared to All_politic network. Nov. has bigger RWPRin comparison to other months, and this is because of the protests which occurred in this month as reaction to increasein petrol price. Jan. has the smallest RWPR due to the death of Iran’s popular General which followed by one of thebiggest and glorious funeral ceremony in Iran history. The results match with the real atmosphere in Iran, and this istrue for hashtag, and both but not Node2vec. The result of Node2vec is precisely similar to RW as they both take intoaccount only structural information.Table 9 indicates that RW correctly demonstrates
All and All_politic networks relations. However, it is unable to
Hanif Emamgholizadeh et al. Page 19 of 23iased Random Walk (BRW) (a) RWPR of both_hashtag network. (b) BCRPR of both_hashtag network. (c) average of both_hashtag network.
Figure 11:
Graph of BRW behavior with different levels of noise for both_hashtag network. The graphs show that whenwe inject noise to the networks, as we expect, the controversy level of networks decrease. This is true for all RWPR,BCRPR, and average of both_hashtag networks. (a) Average initial energy of both_hashtag networksfor different levels of noise. (b) Average energy loss of both_hashtag network fordifferent level of noise.
Figure 12:
Average initial energy and energy loss in both_hashtag networks. This graphs indicate significant increase inthe average initial energy compared to smooth increase in the average energy loss. show Nov. and Dec. relations, correctly. We know that due to nationwide protestations Twitter atmosphere besidethe social atmosphere was controversial in Nov. Whereas, in Dec. Iran had only occasional protestations. Thus, theTwitter atmosphere beside the social environment was less controversial. But RW, similar to Node2vec is not able toreflect this relation in their results.In fact, the difference between Nov. and Dec. indicates the changes in controversy level of the frequently discussedtopics in Persian tweets. Therefore, inability of RW and Node2vec reflects importance of other sources of informationfor improving precision of the algorithms. The correct order of controversy level, detected by our algorithm, not onlydemonstrates that our algorithm out performs RW, but also indicates that more information should be taken into ac-count for better results.These results prove that, in some cases, using only structural data, we are not capable of evaluating the controversylevel of the social network environment. In turn, by combining textual and profile, we can overcome this shortcoming.
Hanif Emamgholizadeh et al. Page 20 of 23iased Random Walk (BRW) (a) RWPR of both network. (b) BCRPR of both network. (c) average of both network.
Figure 13:
Graph of BRW behavior with different levels of noise for both network. This graphs show that RWPR reflectinjected noise effect to the network correctly. (a) Average initial energy of both network for differ-ent level of noise. (b) Average energy loss of both network for differentlevel of noise.
Figure 14:
Average initial energy and energy loss in both network. These graphs show the effect of noise on the averageinitial energy and average energy loss when we inject the noise to both the hashtags and network.
Table 8
Information of Persian tweets networks.Name
5. Conclusion
This paper propose a flexible framework for taking into account the non-structural features of network providedin social media context. Previous research focused on structural information, while they neglected precious informa-tion source provided by users in social networks. Our implementation of the proposed frame work (BRW) is a biasedrandom walker, which starts with an initial energy proportionate to the position of starting node and loss its energy ineach step proportionate to the energy loss of each node on the path. The level to which a random walk reaches in thecontradicting community (the community which does not belong to) is its ability in penetration, in other words, it isbeing heard. We call this controversy of a topic.
Hanif Emamgholizadeh et al. Page 21 of 23iased Random Walk (BRW)
Table 9
These table indicate controversy level of networks, measured by RWPR and RW. These networks are the networks createdusing all Persian tweets, all political Persian tweets, Persian tweets in Nov. Dec., and Jan. Iran’s society experienceddifferent level of bipolarity in this period of time.Name Node2vec Hashtag Both RW DescriptionAll 0.02165 0.0044 0.0034 0.582 All Persian eventsAll_politic 0.0299 0.0013 0.0013 0.74 Political eventsNov 0.03083 0.0019 0.0014 0.68 Nationwide protestationsDec 0.0233 0.0023 0.0021 0.76 Occasional protestationsJan 0.0530 0.0025 0.0024 0.65 Iran’s general Funeral
It seems that community detection based models for assessing controversial topics have the inherent defects de-pending on the basic community detection algorithm. For future research, we are going to release controversy detectionmeasures from community detection algorithms to have more realistic perspective to controversial topic issue.
References [1] Adamic, L.A., Glance, N., 2005. The political blogosphere and the 2004 us election: divided they blog, in: Proceedings of the 3rd internationalworkshop on Link discovery, ACM. pp. 36–43.[2] Akoglu, L., 2014. Quantifying political polarity based on bipartite opinion networks, in: Eighth International AAAI Conference on Weblogsand Social Media.[3] Al Amin, M.T., Aggarwal, C., Yao, S., Abdelzaher, T., Kaplan, L., 2017. Unveiling polarization in social networks: A matrix factorizationapproach, in: IEEE INFOCOM 2017-IEEE Conference on Computer Communications, IEEE. pp. 1–9.[4] Al-Ayyoub, M., Rabab’ah, A., Jararweh, Y., Al-Kabi, M.N., Gupta, B.B., 2018. Studying the controversy in online crowds’ interactions.Applied Soft Computing 66, 557–563.[5] Allcott, H., Gentzkow, M., 2017. Social media and fake news in the 2016 election. Journal of economic perspectives 31, 211–36.[6] Alves, P.F., Mutsvairo, B., 2019. Together and separate? an exploratory study of political polarization on social media during the 2016brazilian political crisis, in: Reporting Human Rights, Conflicts, and Peacebuilding. Springer, pp. 243–263.[7] Amelkin, V., Bogdanov, P., Singh, A.K., 2017. A distance measure for the analysis of polar opinion dynamics in social networks, in: 2017IEEE 33rd International Conference on Data Engineering (ICDE), IEEE. pp. 159–162.[8] Baldassarri, D., Gelman, A., 2008. Partisans without constraint: Political polarization and trends in american public opinion. AmericanJournal of Sociology 114, 408–446.[9] Barberá, P., Jost, J.T., Nagler, J., Tucker, J.A., Bonneau, R., 2015. Tweeting from left to right: Is online political communication more thanan echo chamber? Psychological science 26, 1531–1542.[10] Beelen, K., Kanoulas, E., van de Velde, B., 2017. Detecting controversies in online news media, in: Proceedings of the 40th InternationalACM SIGIR Conference on Research and Development in Information Retrieval, ACM. pp. 1069–1072.[11] Bonchi, F., Galimberti, E., Gionis, A., Ordozgoiti, B., Ruffo, G., 2019. Discovering polarized communities in signed networks, in: Proceedingsof the 28th ACM International Conference on Information and Knowledge Management, ACM. pp. 961–970.[12] Borra, E., Weltevrede, E., Ciuccarelli, P., Kaltenbrunner, A., Laniado, D., Magni, G., Mauri, M., Rogers, R., Venturini, T., 2015. Societalcontroversies in wikipedia articles, in: Proceedings of the 33rd annual ACM conference on human factors in computing systems, ACM. pp.193–196.[13] Bramson, A., Grim, P., Singer, D.J., Fisher, S., Berger, W., Sack, G., Flocken, C., 2016. Disambiguation of social polarization concepts andmeasures. The Journal of Mathematical Sociology 40, 80–111.[14] Castillo, C., Mendoza, M., Poblete, B., 2011. Information credibility on twitter, in: Proceedings of the 20th international conference on Worldwide web, ACM. pp. 675–684.[15] Chen, X., Lijffijt, J., De Bie, T., 2018. Quantifying and minimizing risk of conflict in social networks, in: Proceedings of the 24th ACMSIGKDD International Conference on Knowledge Discovery & Data Mining, ACM. pp. 1197–1205.[16] Choi, Y., Jung, Y., Myaeng, S.H., 2010. Identifying controversial issues and their sub-topics in news articles, in: Pacific-Asia Workshop onIntelligence and Security Informatics, Springer. pp. 140–153.[17] Coletto, M., Garimella, K., Gionis, A., Lucchese, C., 2017. Automatic controversy detection in social media: A content-independent motif-based approach. Online Social Networks and Media 3, 22–31.[18] Conover, M.D., Ratkiewicz, J., Francisco, M., Gonçalves, B., Menczer, F., Flammini, A., 2011. Political polarization on twitter, in: Fifthinternational AAAI conference on weblogs and social media.[19] DeGroot, M.H., 1974. Reaching a consensus. Journal of the American Statistical Association 69, 118–121.[20] DiMaggio, P., Evans, J., Bryson, B., 1996. Have american’s social attitudes become more polarized? American journal of Sociology 102,690–755.[21] Dori-Hacohen, S., Allan, J., 2015. Automated controversy detection on the web, in: European Conference on Information Retrieval, Springer.pp. 423–434.
Hanif Emamgholizadeh et al. Page 22 of 23iased Random Walk (BRW) [22] Dumais, S.T., 2004. Latent semantic analysis. Annual review of information science and technology 38, 188–230.[23] Fiorina, M.P., Abrams, S.J., 2008. Political polarization in the american public. Annu. Rev. Polit. Sci. 11, 563–588.[24] Friedkin, N.E., Johnsen, E.C., 1990. Social influence and opinions. Journal of Mathematical Sociology 15, 193–206.[25] Garimella, K., De Francisci Morales, G., Gionis, A., Mathioudakis, M., 2017. Reducing controversy by connecting opposing views, in:Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, ACM. pp. 81–90.[26] Garimella, K., Morales, G.D.F., Gionis, A., Mathioudakis, M., 2018. Quantifying controversy on social media. ACM Transactions on SocialComputing 1, 3.[27] Garimella, V.R.K., Weber, I., 2017. A long-term analysis of polarization on twitter, in: Eleventh International AAAI Conference on Web andSocial Media.[28] Grover, A., Leskovec, J., 2016. node2vec: Scalable feature learning for networks, in: Proceedings of the 22nd ACM SIGKDD internationalconference on Knowledge discovery and data mining, ACM. pp. 855–864.[29] Guerra, P.C., Meira Jr, W., Cardie, C., Kleinberg, R., 2013. A measure of polarization on social media networks based on communityboundaries, in: Seventh International AAAI Conference on Weblogs and Social Media.[30] Hamilton, W.L., Ying, R., Leskovec, J., 2017. Representation learning on graphs: Methods and applications. arXiv preprint arXiv:1709.05584.[31] Hargittai, E., Gallo, J., Kane, M., 2008. Cross-ideological discussions among conservative and liberal bloggers. Public Choice 134, 67–86.[32] Hoang, T.B.N., Mothe, J., 2018. Predicting information diffusion on twitter–analysis of predictive features. Journal of computational science28, 257–264.[33] Ignatow, G., Evangelopoulos, N., Zougris, K., 2016. Sentiment analysis of polarizing topics in social media: News site readers’ comments onthe trayvon martin controversy, in: Communication and Information Technologies Annual: [New] Media Cultures. Emerald Group PublishingLimited, pp. 259–284.[34] Isenberg, D.J., 1986. Group polarization: A critical review and meta-analysis. Journal of personality and social psychology 50, 1141.[35] Jain, V.K., Kumar, S., Fernandes, S.L., 2017. Extraction of emotions from multilingual text using intelligent text processing and computationallinguistics. Journal of computational science 21, 316–326.[36] Jang, M., Allan, J., 2016. Improving automated controversy detection on the web, in: Proceedings of the 39th International ACM SIGIRconference on Research and Development in Information Retrieval, ACM. pp. 865–868.[37] Jang, M., Foley, J., Dori-Hacohen, S., Allan, J., 2016. Probabilistic approaches to controversy detection, in: Proceedings of the 25th ACMinternational on conference on information and knowledge management, ACM. pp. 2069–2072.[38] Karypis, G., 1997. Metis: Unstructured graph partitioning and sparse matrix ordering system. Technical report .[39] Kittur, A., Suh, B., Pendleton, B.A., Chi, E.H., 2007. He says, she says: conflict and coordination in wikipedia, in: Proceedings of the SIGCHIconference on Human factors in computing systems, ACM. pp. 453–462.[40] Klenner, M., Amsler, M., Hollenstein, N., 2014. Verb polarity frames: a new resource and its application in target-specific polarity classifica-tion .[41] Matakos, A., Terzi, E., Tsaparas, P., 2017. Measuring and moderating opinion polarization in social networks. Knowledge Discovery 31,1480–1505.[42] Mejova, Y., Zhang, A.X., Diakopoulos, N., Castillo, C., 2014. Controversy and sentiment in online news. arXiv preprint arXiv:1409.8152 .[43] Messing, S., Westwood, S.J., 2014. Selective exposure in the age of social media: Endorsements trump partisan source affiliation whenselecting news online. Communication research 41, 1042–1063.[44] Morales, A., Borondo, J., Losada, J.C., Benito, R.M., 2015. Measuring political polarization: Twitter shows the two sides of venezuela.Chaos: An Interdisciplinary Journal of Nonlinear Science 25, 033114.[45] Newman, M.E., 2006. Modularity and community structure in networks. Proceedings of the national academy of sciences 103, 8577–8582.[46] Papadopoulos, S., Kompatsiaris, Y., Vakali, A., Spyridonos, P., 2012. Community detection in social media. Data Mining and KnowledgeDiscovery 24, 515–554.[47] Perozzi, B., Al-Rfou, R., Skiena, S., 2014. Deepwalk: Online learning of social representations, in: Proceedings of the 20th ACM SIGKDDinternational conference on Knowledge discovery and data mining, ACM. pp. 701–710.[48] Popescu, A.M., Pennacchiotti, M., 2010. Detecting controversial events from twitter, in: Proceedings of the 19th ACM international conferenceon Information and knowledge management, ACM. pp. 1873–1876.[49] Quattrociocchi, W., Scala, A., Sunstein, C.R., 2016. Echo chambers on facebook. Available at SSRN 2795110 .[50] Sriteja, A., Pandey, P., Pudi, V., 2017. Controversy detection using reactions on social media, in: 2017 IEEE International Conference onData Mining Workshops (ICDMW), IEEE. pp. 884–889.[51] Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J., Mei, Q., 2015. Line: Large-scale information network embedding, in: Proceedings of the24th international conference on world wide web, International World Wide Web Conferences Steering Committee. pp. 1067–1077.[52] Tsytsarau, M., Palpanas, T., Denecke, K., 2011. Scalable detection of sentiment-based contradictions. DiversiWeb, WWW 1, 9–16.[53] Vicario, M.D., Quattrociocchi, W., Scala, A., Zollo, F., 2019. Polarization and fake news: Early warning of potential misinformation targets.ACM Transactions on the Web (TWEB) 13, 10.[54] Wojatzki, M., Mohammad, S., Zesch, T., Kiritchenko, S., 2018. Quantifying qualitative data for understanding controversial issues, in:Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018).[55] Yang, J., Leskovec, J., 2015. Defining and evaluating network communities based on ground-truth. Knowledge and Information Systems 42,181–213.[22] Dumais, S.T., 2004. Latent semantic analysis. Annual review of information science and technology 38, 188–230.[23] Fiorina, M.P., Abrams, S.J., 2008. Political polarization in the american public. Annu. Rev. Polit. Sci. 11, 563–588.[24] Friedkin, N.E., Johnsen, E.C., 1990. Social influence and opinions. Journal of Mathematical Sociology 15, 193–206.[25] Garimella, K., De Francisci Morales, G., Gionis, A., Mathioudakis, M., 2017. Reducing controversy by connecting opposing views, in:Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, ACM. pp. 81–90.[26] Garimella, K., Morales, G.D.F., Gionis, A., Mathioudakis, M., 2018. Quantifying controversy on social media. ACM Transactions on SocialComputing 1, 3.[27] Garimella, V.R.K., Weber, I., 2017. A long-term analysis of polarization on twitter, in: Eleventh International AAAI Conference on Web andSocial Media.[28] Grover, A., Leskovec, J., 2016. node2vec: Scalable feature learning for networks, in: Proceedings of the 22nd ACM SIGKDD internationalconference on Knowledge discovery and data mining, ACM. pp. 855–864.[29] Guerra, P.C., Meira Jr, W., Cardie, C., Kleinberg, R., 2013. A measure of polarization on social media networks based on communityboundaries, in: Seventh International AAAI Conference on Weblogs and Social Media.[30] Hamilton, W.L., Ying, R., Leskovec, J., 2017. Representation learning on graphs: Methods and applications. arXiv preprint arXiv:1709.05584.[31] Hargittai, E., Gallo, J., Kane, M., 2008. Cross-ideological discussions among conservative and liberal bloggers. Public Choice 134, 67–86.[32] Hoang, T.B.N., Mothe, J., 2018. Predicting information diffusion on twitter–analysis of predictive features. Journal of computational science28, 257–264.[33] Ignatow, G., Evangelopoulos, N., Zougris, K., 2016. Sentiment analysis of polarizing topics in social media: News site readers’ comments onthe trayvon martin controversy, in: Communication and Information Technologies Annual: [New] Media Cultures. Emerald Group PublishingLimited, pp. 259–284.[34] Isenberg, D.J., 1986. Group polarization: A critical review and meta-analysis. Journal of personality and social psychology 50, 1141.[35] Jain, V.K., Kumar, S., Fernandes, S.L., 2017. Extraction of emotions from multilingual text using intelligent text processing and computationallinguistics. Journal of computational science 21, 316–326.[36] Jang, M., Allan, J., 2016. Improving automated controversy detection on the web, in: Proceedings of the 39th International ACM SIGIRconference on Research and Development in Information Retrieval, ACM. pp. 865–868.[37] Jang, M., Foley, J., Dori-Hacohen, S., Allan, J., 2016. Probabilistic approaches to controversy detection, in: Proceedings of the 25th ACMinternational on conference on information and knowledge management, ACM. pp. 2069–2072.[38] Karypis, G., 1997. Metis: Unstructured graph partitioning and sparse matrix ordering system. Technical report .[39] Kittur, A., Suh, B., Pendleton, B.A., Chi, E.H., 2007. He says, she says: conflict and coordination in wikipedia, in: Proceedings of the SIGCHIconference on Human factors in computing systems, ACM. pp. 453–462.[40] Klenner, M., Amsler, M., Hollenstein, N., 2014. Verb polarity frames: a new resource and its application in target-specific polarity classifica-tion .[41] Matakos, A., Terzi, E., Tsaparas, P., 2017. Measuring and moderating opinion polarization in social networks. Knowledge Discovery 31,1480–1505.[42] Mejova, Y., Zhang, A.X., Diakopoulos, N., Castillo, C., 2014. Controversy and sentiment in online news. arXiv preprint arXiv:1409.8152 .[43] Messing, S., Westwood, S.J., 2014. Selective exposure in the age of social media: Endorsements trump partisan source affiliation whenselecting news online. Communication research 41, 1042–1063.[44] Morales, A., Borondo, J., Losada, J.C., Benito, R.M., 2015. Measuring political polarization: Twitter shows the two sides of venezuela.Chaos: An Interdisciplinary Journal of Nonlinear Science 25, 033114.[45] Newman, M.E., 2006. Modularity and community structure in networks. Proceedings of the national academy of sciences 103, 8577–8582.[46] Papadopoulos, S., Kompatsiaris, Y., Vakali, A., Spyridonos, P., 2012. Community detection in social media. Data Mining and KnowledgeDiscovery 24, 515–554.[47] Perozzi, B., Al-Rfou, R., Skiena, S., 2014. Deepwalk: Online learning of social representations, in: Proceedings of the 20th ACM SIGKDDinternational conference on Knowledge discovery and data mining, ACM. pp. 701–710.[48] Popescu, A.M., Pennacchiotti, M., 2010. Detecting controversial events from twitter, in: Proceedings of the 19th ACM international conferenceon Information and knowledge management, ACM. pp. 1873–1876.[49] Quattrociocchi, W., Scala, A., Sunstein, C.R., 2016. Echo chambers on facebook. Available at SSRN 2795110 .[50] Sriteja, A., Pandey, P., Pudi, V., 2017. Controversy detection using reactions on social media, in: 2017 IEEE International Conference onData Mining Workshops (ICDMW), IEEE. pp. 884–889.[51] Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J., Mei, Q., 2015. Line: Large-scale information network embedding, in: Proceedings of the24th international conference on world wide web, International World Wide Web Conferences Steering Committee. pp. 1067–1077.[52] Tsytsarau, M., Palpanas, T., Denecke, K., 2011. Scalable detection of sentiment-based contradictions. DiversiWeb, WWW 1, 9–16.[53] Vicario, M.D., Quattrociocchi, W., Scala, A., Zollo, F., 2019. Polarization and fake news: Early warning of potential misinformation targets.ACM Transactions on the Web (TWEB) 13, 10.[54] Wojatzki, M., Mohammad, S., Zesch, T., Kiritchenko, S., 2018. Quantifying qualitative data for understanding controversial issues, in:Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018).[55] Yang, J., Leskovec, J., 2015. Defining and evaluating network communities based on ground-truth. Knowledge and Information Systems 42,181–213.