Does News Make Birds of a Feather Flock Together? The Role of Ideological
News Web Sites on Twitter Echo Chambers in the 2018’s Brazilian Election
submitted in partial fulfillment for the degree of master of science Pieter Attema Zalis
11824891
master information studies data science
faculty of science university of amsterdam
2018-06-28
Internal Supervisor Second Examiner Title, Name Dr Maarten Marx Dr Judith Moeller Affiliation UvA, FNWI, IvI UvA, ASCoR, IVIR Email maartenmarx@uva.nl J.E.Moller1@uva.nl
ABSTRACT
This paper has two main goals: to investigate the exis-tence of echo chambers and to measure the role of on-line ideological news inside these echo chambers during the second turn of the 2018’s Brazilian presidential elec-tion. From a collection of tweets streamed between Oc-tober 8, 2018, and OcOc-tober 27, 2018, we built a network and detected, by analyzing patterns of retweets with political relevant hashtags, two main clusters highly segregated on political preferences. One cluster favors Fernando Haddad, the center-left candidate, and the second favors Jair Bolsonaro, the right-wing candidate who was elected. The analysis of the role of ideologi-cal news has two steps. First, we demonstrated that, in the pro-Bolsonaro echo chamber, individuals retweet more ideological news than traditional news. In the pro-Haddad echo chamber, individuals retweet more traditional news than ideological. Second, retweets that contain ideological news were removed to see if it is possible to establish causality between ideological news and the formation of echo chambers. The effects of the experiment were almost nonexistent as the strength of the two clusters remained the same. This study does not provide further evidence to the current literature, which indicates that the different patterns of news diet based on political preferences might cause political po-larization.
KEYWORDS
echo chambers; ideological news; political communica-tion; Twitter; social networks, Brazil
1
INTRODUCTION
By 2017, the world woke up to a problem which journal-ists had seen coming for some time: news - in the way we previously understood - was partially broken [38]. The problem today has its taxonomy of terminologies in literature. Some name it fake news [27], while others prefer using concepts such as false stories [42], disin-formation [9], indisin-formation disorder [43] or ideological online news [16]. Next to it, the second problem is that, with the abundance of the information sources avail-able on the internet today, individuals can create their Daily Me [34], a personalized news diet in which they
expose themselves only to information that satisfies their pre-existent beliefs [32].
If separately the two problems raise high concerns, together they might make democracies ungovernable [46]. A reality in which citizens discuss politics based on lies, on alternative facts and avoid divergent views can lead societies to fragmentation and polarization [40]. As a consequence, the capacity of achieving political consensus, one of the cornerstones of liberal democracy, becomes harder or nearly impossible. As influential philosophers - Hannah Arendt [2], Jürgen Habermas [17] and John Stuart Mill [31] - agree, a proper debate can only flourish when individuals are exposed to cross-ideological opinions.
This research investigates the existence of echo cham-bers on social media and what can be classified as ideo-logical oriented news sites [16]. Echo chamber refers to the idea that, due to the abundance of choices and news recommendation algorithms, people only read what pleases them [40]. By doing so, they cluster themselves into communities of like-minded people. Ideological oriented news sites are online news domains that offer coverage consistently favoring one political perspec-tive over others [16]. The contextual setup is Twitter data collected in the second turn of the 2018’s Brazilian presidential election.
If we consider that Donald Trump’s election in 2016, together with the Brexit referendum, is the tipping point in the discussion about information disorder [43], Brazil naturally becomes a relevant country to have a deeper understanding. The similarities between the elected president, Jair Bolsonaro, and Trump are evident. Both are right-wing anti-establishment populist candidates with high social media presence and a particular relish to criticize traditional media. International media out-lets got to the point of nicknaming him as the "Tropical Trump"1. Additionally, roughly four out of ten Brazil-ians also recognize that they live in their bubble on the internet and that they only look for biased opinions they already agree with [19].
To sum up, we split the analysis into three research questions.
1
• (RQ 1) To what extent does ideology segregate Brazil-ian Twitter users into different echo chambers of political communication?
• (RQ 2) To what extent does the role of ideological news in those echo chambers is bigger or smaller than the role of traditional news?
• (RQ3) Do echo chambers disappear without the retweets of ideological news?
This paper goes as follows. In the next section, we start with an overview of the related work on the core concepts of the study: ideological oriented news site, echo chambers, and how the two are connected. In sec-tion three, we give details of the data and explain the methodology used to answer the research questions. Section four reports the findings. We finish by summa-rizing the main conclusions, stating the limitations and providing directions to future studies in section five.
2
RELATED WORK
2.1
Echo Chambers on Twitter
While the concerns on how the internet could lead to a “cyberbalkanization” emerged already in the middle of the 90’s [41], the first empirical result would only confirm this hypothesis almost a decade later. In a pio-neering study, Adamic and Glance [1] showed a clear division in the American blogosphere between liberal and conservatives blogs during the 2004’s American elections. Years later, Wojcieszak and Mutz [45] found a similar pattern in online chat rooms dedicated exclu-sively to politics. However, those patterns of homophily - the principle that contact between similar people oc-curs at a higher rate than among dissimilar people [30] - tended to disappear when political discussions
hap-pened in non-political chat rooms. This result, suggest the authors, indicated that cross-cutting political dis-course could happen when politics became incidentally the center of discussion.
From 2010 on, as blogs and chat rooms lost ground for social media, studies followed this change and fo-cused their data collection on those emerging online environments. Most of the studies have been conducted on Twitter as it is the only major social media plat-form that makes a large amount of data available for research. The evidence of echo chambers in Twitter appears to be mixed and indicates that the results are
highly dependent on the metrics used in each study [21].
Twitter users tend, for instance, to follow other in-dividuals of the same political conviction. Under this assumption, Barberá [5] was able to develop an algo-rithm that derives the policy positions of American individuals from the structure of following links be-tween individual Twitter users and relevant political actors. Echo chambers were also found when analyzing patterns of retweets between liberals and conserva-tives in the United States [13]. These findings, however, were not replicated by the authors when connections between users are built through mentions, i.e., when a user only mentions another user on a Twitter message.
The content in the topic of discussion also matters. In a similar direction of what Wojcieszak and Mutz sug-gested in chat rooms, Nagler et al. [6] found that events unrelated to politics can stimulate the discussion among individuals who differ in their political opinions. On the other hand, in strictly political debates as elections, individuals tend to close themselves in like-minded communities, a finding also corroborated by Barberá [5].
2.2
News in Echo Chamber
For many decades, the risks that media could lead in-dividuals to cluster themselves into like-minded com-munities did not raise much concern. Even if, at least since the 1940s, researches have noticed that partisans individuals encountered congenial messages more of-ten than uncongenial ones [25], the scarcity of media options in the era of mass media made people fed their media diet with very similar sources. To put it shortly, the idea of an almost-monopolistic media market char-acterized part of the 20th century [8]. Add two charac-teristics. Traditional outlets followed the idea of "con-sonance", i.e., they tended to portray controversial is-sues in a homogeneous manner [36]. Values such as detachment, nonpartisanship, facticity, and balance also became the norm in newsrooms [46].
The era of online news consumption arrived, how-ever, at the beginning of the 21st century and a new, different avenue of possibilities opened on how citizens consume information and on how journalists can pro-duce content. As abundance replaces scarcity, people now had the chance to rely on heuristic clues, such as their political preferences, to get news [26]. Thus,
some argue, the age of niche media replaced the age of mass media [20]. Moreover, social media, the platform which drives much of traffic to news websites, were engineered such that every time a user posts content, their brain releases a tiny hit of dopamine [44]. As Face-book stated in 2016, "if you look through thousands of stories every day and choose ten that were the most important to you, which would they be? The answer should be your News Feed. It is subjective, personal, and unique - and defines the spirit of what we hope to achieve."2
The replacement of the economics of objectivity by the "economics of emotions" [3] and the capacity offered by the internet in reaching niche audiences explain the emergence of ideological oriented news websites. They can be problematic because, first, previous studies show that they have the capacity of encouraging inaccurate beliefs [16]. Second, they segregate news consumers into alternative sources perceived as more congenial to their preferences [20]. In other words, partisan media may lead to the formation of echo chambers.
Some claim that higher levels of political polarization could be a consequence of the emergence of those ide-ological news domains [39]. Even if the debate about the causality is far from pacified, it seems that, to a cer-tain extent, the connection between echo chambers and ideological news exists. Faris et al. [7] talk, for exam-ple, about the formation of an insular right-wing media ecosystem highly segregated from all other media out-lets during the 2016’s American election. Bhatt et al. [10] extend the finding by showing a similar pattern in the left-wing media, even if the levels of segregation in the right-wing outlets are still higher. In Brazil, Ribeiro and Ortellado [33] also conclude, by analyzing 12 mil-lion Brazilians following 500 different political pages on Facebook, that ideological news became a valuable combat tool in a battle of information between two highly polarized clusters on Facebook.
3
METHODOLOGY
3.1
Network Construction
We start answering the research questions through the construction of a network. The nodes in this network are Twitter users. The edges are the retweets. If a user
2
https://newsroom.fb.com/news/2016/06/building-a-better-news-feed-for-you/
retweets a message of another user, we connect these two by an edge. The edges do not have weights or direction: the network is undirected.
The data was collected through Twitter’s Streaming API, which, based on a query of specific keywords and hashtags, gives in real time a sample of all messages posted on Twitter. Between October 8, 2018, and Octo-ber 27, 2018, we collected a total amount of 19,649,008 tweets based on the query containing the names of the two candidates, Haddad and Bolsonaro, and the abbre-viations of their respective parties, PT and PSL. This data set contains 880,655 unique users and 78.57% of the tweets are retweets. The median number of retweets per user is 1 and the median number of retweets per message is 2. These two low medians indicate that a small set of users and messages concentrate a signifi-cant amount of retweets. In the case of users, 14% of the individuals are responsible for 80% of all traffic of retweets. In the case of the messages, the top 3% most retweeted messages caused 80% of all traffic of retweets.
Our network was created following the methodol-ogy developed by Conover et al. [13]. It starts with the selection of the two most popular hashtags: one from the left (#elenao) and another from the right (#bol-sonaro). Then, another hashtag is considered relevant if its Jaccard similarity coefficient with either #elenao or #bolsonaro is higher than 0.005. For X a hashtag, let TX
denote the set of all tweets with hashtag X . The Jaccard similarity between hashtags A and B equals |TA∩ TB|
divided by |TA∪ TB|.
This process resulted in a total of 40 politically rel-evant hashtags occurring in 459,107 tweets. This data forms the network. The graph has 178,833 nodes (20% of all unique users). There are 4,217 components in the graph, but one, main component contains 94% of all nodes. The clustering coefficient is 0.0015, the density is 0.00005 and the average shortest path distance is 4.85 nodes.
3.2
Cluster Analysis
The next step is to find the clusters, also called com-munity structures, inside the network. When a set of nodes in the network has a higher probability of being connected together than connected to other nodes, it is said that they form a community [12, 14]. Following Conover et al. [13], the communities in our network are determined by clustering nodes through Rhaghavan’s
label propagation [37] seeded with labels determined by Newman’s leading eigenvector modularity maximiza-tion [12]. The later aggregates nodes by calculating the leading non-negative eigenvector of the modularity ma-trix of the graph, and the former updates the labels by majority voting in the neighborhood of the node [23]. Previous research has used modularity as a property to confirm the robustness of the communities [18]. It measures when the division is meaningful, in the sense that there are many edges within communities and only a few between them [14, 35]. The literature suggests that modularity values higher than 0.4 [18] or even 0.3 [23] indicate the formation of relevant communities.
If modularity helps us finding significant communi-ties, it does not allow us to make any inference about the formation of like-minded communities, echo-chambers. Thus, we finish the cluster analysis by performing a con-tent analysis of the messages retweeted in the network. Inside each community, we extracted a random sample of tweets and manually labeled them as pro-Bolsonaro (right), pro-Haddad (left) or undefined. The goal, here, is to find possible evidence of echo-chambers, in the sense that pro-Haddad messages or pro-Bolsonaro messages are grouped in the same clusters. Two coders perform the task of manually labeling the messages. Cohen’s Kappa coefficient is 0.80, which indicates that there is a good agreement between the two coders.
3.3
Ideological News Inside Echo
Chambers
We define ideological news as online media domains which show a coverage consistently favoring one polit-ical perspective over the others [16]. Practpolit-ically speak-ing, news that favors Bolsonaro or attacks Haddad is classified as right-wing, while news that favors Haddad or attacks Bolsonaro is classified as left-wing. The ideol-ogy classification is done on the news domain level and not for each article individually. The key element of the content analysis is that ideological news has differ-ent processes and intdiffer-entions than traditional news [27]. While public interest guides the intentions of traditional news [24], to favor political groups is what motivates ideological news. In the process of traditional news, facts and opinions stay separately [24]. The process of ideological news mixes facts and opinions.
The content analysis worked as follows. Two coders performed the analysis in a list of 10 random articles
per domain found in the data set of retweets. They de-fined the domains as ideological news or not. In the end, 44 domains were classified as ideological news do-mains, with a Cohen’s Kappa coefficient of 0.93. Then, the coders needed to label ideological domains as right-wing or left-right-wing. In the end, 25 were labeled as left and 19 as right ideological news domains. Cohen’s Kappa co-efficient is 0.91. In the collection of 15,439,985 retweets, individuals shared 475,815 URLs of news domains clas-sified as right-wing news and 306,384 URLs of left-wing domains.
The traffic of ideological news will be compared with the traffic of traditional media. News domains were classified as traditional if they fall into one of the follow-ing criteria: national newspaper, regional newspaper, TV channel, news magazine, radio station, established news portal or international news agency with content in Portuguese. These criteria lead us to a final list of 29 traditional news domains, which were shared 798,843 times inside the collection of retweets. To answer the second research question, we perform a Fisher Exact Test to check if the differences between traditional and ideological news in the formed communities are statis-tically significant.
The last, third research question looks for a possible causality relationship between ideological news and the formation of echo-chambers. All research designs sug-gested so far prevent any inference of causality. A better suggestion is an experiment in which we manipulate a condition and compare different outcomes between the control group and the experimental group. The control group is the communities found for research question one. The condition of the experiment is the removal of all retweets that contains articles of ideological news do-mains. After doing so, the analysis is performed again following the same procedures of the control group. The assumption is that, if the like-minded communities of RQ1 disappear without the retweets of ideological news, we can claim a causality relationship between ideological news and echo-chambers.
4
FINDINGS
4.1
Echo Chambers
Relying upon Conover et al. [13] clustering method in the major component of the network leads to the formation of two communities. Cluster 1 has 112,948 members and Cluster 2 has 55,604. A modularity value
Figure 1:A visual representation of the two communi-ties. Red indicates Cluster 1, pro-Haddad (N=112,948), while blue Cluster 2, pro-Bolsonaro (N=55,604)
of 0.49 indicates that the strength of the division in the network is strong and that the divisions between the clusters are meaningful. The assortativity coefficient of 0.99 confirms the robustness of the results, as values closer to 1 indicate that similar nodes, i.e., from the same cluster, tend to connect through edges and val-ues close to -1 indicate that dissimilar nodes, i.e., from different clusters, tend to connect more often. Addition-ally, from all edges in the network, only 0.5% of the edges connect to nodes from the other cluster. To sum up, we have enough evidence to suggest the formation of two significant communities.
Table 1: Connectivity Measures Measures Network Cluster 1 Cluster 2 Av. Shortest Path 4.85 4.88 3.76 Cluster Coefficient 0.00156 0.00157 0.00156 Density 2.86x10−5 3.33x10−5 0.000124
Table 1 provides further measurements related to the connectivity of the overall network and the two com-munities formed in the main component. There are no main differences in the clustering coefficient. On other measures, however, the higher value of density and the lower value of the average shortest path on Cluster 2 indicate that this community is more robust and con-nected than Cluster 1. The robustness and connectivity of the later do not differ much from the entire network. To sum up, Cluster 2 seems to be a smaller, more cohe-sive community, while Cluster 1 is a bigger, more diffuse community with less connectivity between its members. Interestingly, the higher average shortest path of Clus-ter 1 compared to the overall network indicates that some of the shortest paths between two nodes inside that cluster go through Cluster 2 nodes.
The next question, then, is if the two communities are split by political preferences. The manual content analysis of 1,000 random tweet messages - which 514 comes from Cluster 1 and 486 comes from Cluster 2 - indicates that there is a high degree of political
ho-mophily splitting Clusters 1 and 2 into two groups of po-litically like-minded users. From the messages tweeted by users in Cluster 1, 89.31% had content classified as pro-Haddad, 6.03% pro-Bolsonaro and 4.66% undefined. Cluster 2 was almost the opposite, with 93.83% mes-sages favoring Bolsonaro, 1.64% favoring Haddad and 4.53% undefined. To sum up, the answer to the first re-search question is that there is enough evidence to claim that ideology can segregate Brazilian Twitter users into echo-chambers of political communication. While Clus-ter 1 is pro-Haddad, ClusClus-ter 2 is pro-Bolsonaro.
Table 2: Labels of Messages Pro-Bolsonaro Undefined Pro-Haddad N Cluster 1 31 (6.03%) 24 (4.66%) 459 (89.31%) 514 (100%) Cluster 2 456 (93.83%) 22 (4.53%) 8 (1.64%) 486 (100%)
4.2
News inside the Echo Chambers
The investigation on the news diet habits also shows different behaviors between the two communities. The analysis of retweets of ideological and traditional news
shows that Cluster 1, pro-Haddad, had 38,33% of ideo-logical news and 61,77% of traditional news. In Cluster 2, pro-Bolsonaro, there is almost the opposite pattern: 59,99% of the shares with labels correspondent to ideo-logical news, while 40,01% corresponds to traditional news. Fisher’s Exact Test confirms that the difference is statistically significant (p < 0.01). The straightforward answer to the second research question, then, is that while, in the pro-Haddad Cluster, the role of traditional news is bigger than the role of ideological news, in the pro-Bolsonaro Cluster, the role of ideological news is bigger than the role of traditional news.
The analysis of news inside the clusters also helps to provide further evidence for our first research question. In echo chambers, we would expect, for instance, to have users in the pro-Haddad Cluster to share a tiny fraction of right-wing ideological news compared to left-wing ideological news. The opposite pattern would be equally expected in the pro-Bolsonaro Cluster. In both Clusters, this pattern is confirmed. Only 0.64% of the labeled data inside Cluster 2 is left-wing news and 59.36% is right-wing news. In Cluster 1, only 0.32% of the labeled data is right-wing ideological news, while 37.91% is left-wing ideological news.
Table 3: News Types per Cluster Traditional News Right-Wing News Left-Wing News N Cluster 1 5,571 (61.77%) 30 (0.32%) 3,416 (37.91%) 9,017 (100%) Cluster 2 3,129 (40.01%) 4,642 (59.35%) 50 (0.64%) 7,821 (100%)
4.3
Echo Chambers without
Ideological News
Caution is necessary to interpret the previous results as we should not overestimate the importance of news inside the two communities. As Figure 2 demonstrate, both traditional news and ideological news corresponds to a tiny percentage of what the two clusters retweet. To be more specific, the domains labeled as ideological and traditional news represent together only 4.34% of the retweets in Cluster 1, while they represent only 4.93% of the retweets in Cluster 2. This percentage already gives clues for the third research question. With such a
small set of ideological news, probably users would still aggregate themselves through like-minded communi-ties independently of the existence of ideological news domains.
Figure 2:Purple indicates retweets of messages that weren’t classified. Green indicates traditional news and orange indicates ideological news
The validity of this assumption is tested by removing retweets of ideological news domains and rerunning the network analysis with the same setup as before. The findings confirm that the effects of removing the ideological news are small, barely nonexistent. Again Cluster 1 and Cluster 2, which form the main compo-nent of the network, represent the same 94% of the nodes in the network. Cluster 1 has now 112,062 users and Cluster 2 has 53,810. Only 0.4% of the nodes are labeled differently from the previous setup. The modu-larity value of the communities, again, is 0.49. In other words, to answer the final research question, which asked if echo chambers would disappear without the retweets of ideological news, we have strong evidence to support the fact that ideological news has no effects on the formation of echo chambers.
5
CONCLUSION
This paper started with a question in its title: does ide-ological news make birds of a feather flock together? The existence of two echo chambers was confirmed (RQ1): one favoring the left-wing candidate, Fernando Haddad, another favoring the right-wing candidate,
Jair Bolsonaro. Inside each chamber, different patterns of media consumption emerged (RQ2). While the pro-Bolsonaro Cluster had more ideological news than tra-ditional news, on the pro-Haddad cluster, individuals shared more traditional news than ideological. Finally, it was shown that the echo-chambers would exist inde-pendently of users retweeting ideological news (RQ3). To sum up, the answer to the title’s question is that news does not drive people into echo-chambers. They might reinforce or catalyze polarization, but do not cause it [32].
In this sense, one of the main findings of this paper is that it does not confirm the conclusion of previous stud-ies that advocated for a causal relationship between dif-ferent news diets and political polarization [20, 39]. So, how should we interpret the role of ideological news? The question goes beyond the causality analyzed in this paper, but it is possible to hypothesize some explana-tions. Since 2013, there has been a political turmoil in Brazil. Major public protests happen regularly and the Car Wash Operation revealed a vast corruption scandal that led to one president impeached, two former presi-dents sent to jail and to the resurgence of a conservative political movement that culminated with Bolsonaro’s election. If no polls show empirical evidence that these events are leading people to a more polarized society, it is unimaginable to suggest that these events do not bring consequences to the public debate. In this context, what might explain the role of ideological news is that they are a combat tool [33], in which, by sharing on Twitter, users make clear for what they politically stand for and for what political tribe they belong.
The second main conclusion is that right-wing indi-viduals rely more on ideological news than left-wing individuals. Previous studies had similar results in other countries, such as the US [7] and Germany [28]. Un-derstanding why this pattern emerges goes beyond the scope of this study, but it is a point that future work could investigate. A suggestion is to explore the con-nection between ideological news and Affective Intelli-gence, a theory that connects emotions to politics. Mar-cus, Neuman and MacKuen suggest that anger strength-ens the influence of ideological identification on individ-uals [29]. Thus, a possible explanation is that, by being angrier than left-wing individuals, Bolsonaro support-ers rely more on ideological news than traditional news to strengthen their political identity. This hypothesis
could be tested, for instance, by running on tweets dic-tionaries, such as Linguistic Inquiry and Word Count (LIWC), which measures the number of anger words in a piece of text.
Another critical discussion related to future work and limitations of this research regards the size of the network. By following the method of Conover et al. and their idea of political relevant hashtags, only 40 hashtags were selected. This amount led to a network with only 20% of the unique users in the entire collected data. To surpass this limitation, other methods, which would increase the exploitation of hashtags in the data and consequently increase the number of nodes, could be tested.
A good start to future work on this matter is to think of hashtags as indicators of political orientation. Re-search in the past has demonstrated that left-wing and right-wing individuals tend to cluster separately around different hashtags [11]. Thus, a possible solution to in-crease the retweet network is to start the analysis of echo chambers with the construction of a hashtag co-occurrence network in which edges connect two hash-tags if they appear in the same tweet. After the network is constructed, we could cluster nodes in the main com-ponent of the network. Hashtags are now considered political relevant if they belong to the same cluster of #elenao and #bolsonaro, the two most common hash-tags. In other words, if here we used the Jaccard simi-larity coefficient to define politically relevant hashtags, in future work, it could be possible to try the usage of network communities to define political relevancy.
Finally, the last comment should be made on echo chambers. The confirmation in this study of their exis-tence, based on retweets, does not mean that individuals live in completely closed narrow-minded communities. Previous studies demonstrated that preferring opinion-reinforcing information does not mean that individuals avoid completely their exposure to opinion-challenging information [4, 15]. Individuals might indeed avoid ac-tively sharing opinion-challenging information. This does not necessarily mean, however, that they do not see messages of users from a different political perspec-tive; that they do not seek dissimilar information out-side social media platforms; or that they do not engage in a conversation with dissimilar minds when the dis-cussion has a non-political topic. There is even evidence that digital media can drive people to more diverse sources of information [22, 32]. To sum up, what we
showed here is the higher propensity to share opinion-reinforcing information, but we can not confirm, to any extent, if pro-Haddad and pro-Bolsonaro supporters did not cross by opinion-challenging information outside their bubbles of retweets.
ACKNOWLEDGEMENTS
I would like to express my very great appreciation to Dr. Maarten Marx, who gave all his support as a supervisor, and Dr. Judith Moeller, who accepted to be the second reader and will certainly enrich this work with her acute comments. To my computer scientist friends, Guillaume Corda and Rafael Pierre, a sincere thanks for being patient with my mathematical illiteracy.
REFERENCES
[1] Adamic, L., and Glance, N. The political blogosphere and the 2004 u.s. election: divided they blog. In Proceedings of the 3rd international workshop on link discovery (Aug 2005), ACM, pp. 36–43.
[2] Arendt, H. Between past and future : eight exercises in political thought. Penguin Books, New York, NY, USA, 2006.
[3] Bakir, V., and Mcstay, A. Fake news and the economy of emotions: Problems, causes, solutions. Digital Journalism 6, 2 (2018), 154–175.
[4] Bakshy, E., Messing, S., and Adamic, L. A. Exposure to ideologically diverse news and opinion on facebook. Science 348, 6239 (Jun 2015), 1130–1132.
[5] Barberá, P. Birds of the same feather tweet together: Bayesian ideal point estimation using twitter data. Political Analysis 23, 01 (jan 2015), 76–91.
[6] Barberá, P., Jost, J. T., Nagler, J., Tucker, J. A., and Bon-neau, R. Tweeting from left to right: Is online political com-munication more than an echo chamber? Psychological Science 26, 10 (2015), 1531–1542.
[7] Benkler, Y., Roberts, H., Faris, R. M., Etling, B., Zucker-man, E., and Bourassa, N. Partisanship, propaganda, and disinformation: Online media and the 2016 u.s. presidential election. Tech. rep., Berkman Klein Center for Internet Society, 2017.
[8] Bennett, W. L., and Iyengar, S. A new era of minimal effects? the changing foundations of political communication. Journal of Communication 58, 4 (2008), 707–731.
[9] Bennett, W. L., and Livingston, S. The disinformation or-der: Disruptive communication and the decline of democratic institutions. European Journal of Communication 33, 2 (2018), 122–139.
[10] Bhatt, S., Joglekar, S., Bano, S., and Sastry, N. Illuminating an ecosystem of partisan websites. In Companion Proceed-ings of the The Web Conference 2018 (Republic and Canton of Geneva, Switzerland, 2018), WWW ’18, International World Wide Web Conferences Steering Committee, pp. 545–554.
[11] Bode, L., Hanna, A., Yang, J., and Shah, D. V. Candidate networks, citizen clusters, and political expression: Strate-gic hashtag use in the 2010 midterms. The ANNALS of the American Academy of Political and Social Science 659, 1 (2015), 149–165.
[12] Clauset, A., Newman, M. E. J., and Moore, C. Finding com-munity structure in very large networks. Phys. Rev. E 70 (Dec 2004), 066111.
[13] Conover, M., Ratkiewicz, J., Francisco, M., Goncalves, B., Menczer, F., and Flammini, A. Political polarization on twitter. In Proceedings of the 5th International AAAI Conference on Weblogs and Social Media Political (Jul 2011).
[14] Fortunato, S., and Hric, D. Community detection in net-works: A user guide. Physics Reports 659 (2016), 1–44. Com-munity detection in networks: A user guide.
[15] Garrett, R. K. Echo chambers online?: Politically motivated selective exposure among internet news users. Journal of Computer Mediated Communication 14, 2 (2009), 265–285. [16] Garrett, R. K., Weeks, B. E., and Neo, R. L. Driving a wedge
between evidence and beliefs: How online ideological news ex-posure promotes political misperceptions. Journal of Computer Mediated Communication 21, 5 (2016), 331–348.
[17] Habermas, J. The structural transformation of the public sphere : an inquiry into a category of bourgeois society. MIT Press, Cambridge, MA, USA, 1989.
[18] Himelboim, I., Smith, M., and Shneiderman, B. Tweeting apart: Applying network analysis to detect selective exposure clusters in twitter. Communication Methods and Measures 7, 3-4 (2013), 195–223.
[19] Ipsos. Global advisor: Fake news. Tech. rep., Ipsos, Paris, FR, Set 2018.
[20] Iyengar, S., and Hahn, K. S. Red media, blue media: Evidence of ideological selectivity in media use. Journal of Communica-tion 59, 1 (3 2009), 19–39.
[21] Jungherr, A. Twitter use in election campaigns: A systematic literature review. Journal of Information Technology & Politics 13, 1 (2016), 72–91.
[22] Kleis Nielsen, R., and Selva, M. More important, but less robust? five things everybody needs to know about the future of journalism. Tech. rep., Reuters Institute for the Study of Journalism, Jan 2019.
[23] Kolaczyk, E. D., and Csárdi, G. Statistical Analysis of Network Data with R, 2014 ed., vol. 65 of Use R. Springer New York, New York, NY, 2014.
[24] Kovach, B., and Rosenstiel, T. The elements of journalism : what newspeople should know and the public should expect. Three Rivers Press, New York, NY, USA, 2007.
[25] Lazarsfeld, P. F. The people’s choice : how the voter makes up his mind in a presidential campaign, 2nd ed ed. Columbia University Press, New York, NY, USA, 1948.
[26] Lazer, D., Baum, M., Grinberg, N., Friedland, L., Joseph, K., Hobbs, W., and Mattsson, C. Combating fake news: An agenda for research and action. Tech. rep., Shorenstein Center On Media, Politics and Public Policy, May 2017.
[27] Lazer, D. M. J., Baum, M. A., Benkler, Y., Berinsky, A. J., Greenhill, K. M., Menczer, F., Metzger, M. J., Nyhan, B.,
Pennycook, G., Rothschild, D., Schudson, M., Sloman, S. A., Sunstein, C. R., Thorson, E. A., Watts, D. J., and Zittrain, J. L. The science of fake news. Science 359, 6380 (mar 2018), 36–44.
[28] Majó-Vázqez, S., Nurse, J. R. C., Simon, F. M., and Nielsen, R. K. Digital-born and legacy news media on twitter during the german federal election. Tech. rep., Reuters Institute for the Study of Journalism, Oxford, UK, Oct 2017.
[29] Marcus, G. E., Neuman, W. R., and MacKuen, M. Affective intelligence and political judgment. University of Chicago Press, Chicago, IL, USA, 2000.
[30] McPherson, M., Smith-Loving, L., and M Cook, J. Birds of a feather: Homophily in social networks. Annual Review of Sociology 27 (2001), 415–444.
[31] Mill, J. S. On liberty, people’s edition ed. Longman, London, UK, 1865.
[32] Moeller, J., and Helberger, N. Beyond the filter bubble: concepts, myths, evidence and issues for future debates. Tech. rep., University of Amsterdam, Den Haag, NL, May 2018. [33] Moretto Ribeiro, M., and Ortellado, P. From fake news
to hyper-partisan media. SUR 27 15, 27 (Jul 2018), 69–81. [34] Negroponte, N. Being digital. Hodder Stoughton, London,
UK, 1995.
[35] Newman, M. E. J. Modularity and community structure in networks. Proceedings of the National Academy of Sciences 103, 23 (2006).
[36] Noelle-Neumann, E., and Mathes, R. The ‘event as event’ and the ‘event as news’: The significance of ‘consonance’ for media effects research. European Journal of Communication 2, 4 (1987), 391–414.
[37] Raghavan, U. N., Albert, R., and Kumara, S. Near linear time algorithm to detect community structures in large-scale networks. Phys. Rev. E 76 (Sep 2007), 036106.
[38] Rusbridger, A. Who broke the news? Guardian (Aug 2018). [39] Stroud, N. J. Polarization and partisan selective exposure.
Journal of Communication 60, 3 (2010), 556–576.
[40] Sunstein, C. R. Republic.com 2.0. Princeton University Press, Princeton, New Jersey, USA, 2007.
[41] Van Alstyne, M., and Brynjolfsson, E. Could the internet balkanize science? Science 274, 5292 (1996), 1479–1480. [42] Vosoughi, S., Roy, D., and Aral, S. The spread of true and
false news online. Science 359, 6380 (2018).
[43] Wardle, C. The need for smarter definitions and practical, timely empirical research on information disorder. Digital Journalism 6, 8 (2018), 951–963.
[44] Wardle, C., and Derakhshan, H. Information disorder: To-ward an interdisciplinary framework for research and policy making. Tech. rep., Council of Europe, Sep 2017.
[45] Wojcieszak, M. E., and Mutz, D. C. Online groups and polit-ical discourse: Do online discussion spaces facilitate exposure to political disagreement? Journal of Communication 59, 1 (2009), 40–56.
[46] Yochai, B., Robert, F., and Hal, R. Network Propaganda : Manipulation, Disinformation, and Radicalization in American Politics. Oxford University Press, Cambridge, MA, 2018.
