Controversy on Social Media: Collective Attention, Echo Chambers, and Price of Bipartisanship

Texte intégral

(1)Controversy on Social Media: Collective Attention, Echo Chambers, and Price of Bipartisanship Gianmarco De Francisci Morales ISI Foundation with. Kiran Garimella Aristides Gionis Michael Mathioudakis.

(2) Controversy: from Latin contra (against) vertere (turn) “turned against, disputed”.

(3)

(4)

(5)

(6)

(7) The gap grows.

(8) The gap grows.

(9) Goal Understand how controversies unfold in social media.

(10) Outline.

(11) Outline. Quantify.

(12) Outline. Quantify. Graph.

(13) Outline Polarization Measure. Quantify. Graph.

(14) Outline Polarization Measure. Quantify. Graph. Evolve.

(15) Outline Polarization Measure. Quantify. Evolve. Graph. Time.

(16) Outline Polarization Measure. Collective Attention. Quantify. Evolve. Graph. Time.

(17) Outline. Cause. Polarization Measure. Collective Attention. Quantify. Evolve. Graph. Time.

(18) Outline Polarization Measure. Collective Attention. Cause. Quantify. Evolve. Opinion. Graph. Time.

(19) Outline Echo Chambers. Polarization Measure. Collective Attention. Cause. Quantify. Evolve. Opinion. Graph. Time.

(20) Quantifying Controversy in Social Media WSDM 2016, TSC 2018.

(21) Black/Blue or White/Gold?.

(22) Desiderata In the wild Not necessarily political No domain knowledge Language independent Allows comparison.

(23) Problem Formulation Graph-based unsupervised formulation Conversation graph for a topic (endorsements) Find partition of graph (represents 2 sides) Measure distance between partitions (random walks).

(24) Endorsement Graph. #марш. #sxsw.

(25) Pipeline. • • • •. Retweets Follow Mentions Content. • • •. METIS Spectral Label propagation. • • • •. Random walk Edge betweenness 2d embedding Sentiment variance.

(26) Example.

(27) Example. #beefban. #марш. #sxsw. #germanwings.

(28) Example. #beefban. #марш. Controversial. #sxsw. #germanwings. Non controversial.

(29) RWC Rationale Random Walk Controversy score Zaller's RAS model (Receive, Accept, Sample) "The Nature and Origins of Mass Opinion" Response Axiom: "Individuals form opinions by averaging across the considerations that are immediately salient or accessible to them" Authoritative (influential) users with high degree set opinions Measure likelihood of user to be exposed to opinions from influential users on either side.

(30) Random Walk.

(31) Random Walk. X. Y.

(32) Random Walk. X. Y.

(33) we select one partition at random (each with (a) (b) walk that starts from a random vertex in tha Partitions obtained for (a) #beefban, (b) #russia march by using the hybrid graph RWC Definitionvertex visits any high-degree (from ch. The partitions are more noisy than those in Figures 3(a,b). either sid dom Walk Controversy (RWC ) measure as fo Subsequently, we select one partition at random (each with probability 0 ending in partition X and one ending in par Consider two random walks, one ending in partition X and one ending in er a random walk that starts from a random vertex in that partition. Th partition Yvisits , RWCany is the difference ofvertex the probabilities of twoside). events: (i) nates when it high-degree (from either obabilities of two events: (i) both random wa walks started from the partition ended inas andfollows. (ii) both“Consi defineboth therandom Random Walk Controversy (RWC they ) measure inwalks, and (ii) both random walks started pa random started a partitionXother than one they inin Ya, RWC m onewalks ending in in partition and one the ending in ended partition nce of the probabilities of two events: (i) both random walks started fr The measure is quantified as on they ended in and (ii) both random walks started in a partition other t. ey ended in.” The measure is quantified as. RWC RWC = P=XX P P PY Y XX. YY. PPY X, PXY ,. PY X. XY. PAB Y , A,}Bis 2 {X, Y } conditional is the conditional probability X, the probability. PAB = Pr [start in partition A | end in partition B].. = Pr [start in partition A |desirable end in partition AB orementioned probabilities have the following properties: (i) they. d by the size of each partition, as the random walk starts with equal pro ach partition, and (ii) they are not skewed by the total degree of vertices. probabilities have the following desirable prop.

(34) RWC Properties Probabilities are conditional on ending in either partition Random walks end on either side with equal probability Not skewed by size of each partition Not skewed by total degree of vertices in each partition Close to 1 when probability of crossing sides low (high controversy) Close to 0 when probability of crossing comparable to that of staying (low controversy).

(35) Controversy Detection.

(36) Controversy Score.

(37) A:19. (a). (b). Fig. 10: RWC scores for synthetic Erdös-Rényi graphs planted with two communities. p1 is the intra-community edge probability, while p2 is the inter-community edge probability.. generate random Erdös-Rényi graphs with varying community structure, and compute the RWC score on them. Specifically, to mimic community structure, we plant two separate communities with intra-community edge probability p1 . That is, p1 defines how dense these communities are within themselves. We then add random edges between these two communities with probability p2 . Therefore, p2 defines how connected the two communities are. A higher value of p1 and a lower value of p2 create a clearer two-community structure. Figure 10 shows the RWC score for random graphs of 2000 vertices for two different settings: plotting the score as a function of p1 while fixing p2 (Figure 10a), and vice-versa. Planted Synthetic Graphs.

(38) Summary RWC: a measure for how controversial a discussion on a topic is on social media Graph-based measure: no domain knowledge, language agnostic Intuitive semantics founded on opinion formation models Captures controversy better than state-of-the-art User-level polarization measure easy to derive.

(39) The Effect of Collective Attention on Controversial Debates on Social media WebSci 2017 (Best Paper Award).

(40) The Effect of Collective Attention on Controversial Debates on Social media.




(44) "Trump taxes" on Google.

(45) "Trump taxes" on Google Rachel Maddow show on 2005 tax return.

(46) Obamacare on Twitter.

(47) Gun Control on Twitter.

(48) Literature so far. Controversial debates examined in isolation As static snapshots.

(49) Contribution Controversial debates are dynamic They change with collective attention. Analyze controversial debates over time Particularly when collective attention increases When external ‘event’ happens.

(50) Data Twitter 4 longitudinal polarized topics Obamacare, Abortion, Gun control, Fracking 5 years (2011 -- 2016) Hundreds of thousands of users Millions of tweets.

(51) Definitions. Retweet Graph Reply Graph Core Users.

(52) Retweet Graph.

(53) Reply graph.

(54) Core.

(55) Core. Core Users.

(56) Experiments.

(57) Experiments Compare these two points.

(58) Retweet Graph.

(59) Retweet Graph 1) New users enter the discussion.

(60) Retweet Graph 1) New users enter the discussion. 2) Most retweets to existing core users.

(61) Retweet Graph 1) New users enter the discussion. 3) Cross-side retweets decrease. 2) Most retweets to existing core users.

(62) Retweet Graph 1) New users enter the discussion. 3) Cross-side retweets decrease. 2) Most retweets to existing core users. 4) Within-side retweets increase.

(63) Controversy Measure. Figure 2: RWC score as a function of the activity in the retweet network. An increase in interest in the controversial topic corresponds to an increase in the controversy score of the retweet network.. 5.1 Network. F t r s w.

(64) Core-Periphery Openness. Figure 12: Core–periphery openness as a function of activity in the retweet network. As the interest increases, the number of core-periphery edges, normalized by the expected number of edges in a random network, increases. This suggests a propensity of periphery nodes to connect with the core nodes when interest increases..

(65) Reply Graph. Attention increases. Cross-side edges increase: more discussion.

(66) Attention Increase. Normal Condition. Content Pro Choice Pro Life.

(67) Content Pro Life. Normal Condition. Pro Choice. Attention Increase. Content becomes uniform across the sides.

(68) Long-Term Polarization.

(69) Summary Controversial debates during external events Polarization increases Retweet graph becomes hierarchical (core-periphery) More replies across sides Content becomes more uniform Many more results in the paper!.

(70) Political Discourse on Social Media Echo Chambers, Gatekeepers, and the Price of Bipartisanship WWW 2018.

(71) Political Discourse on Social Media Characterized by heavy polarization Emergence of echo chambers ("Hear your own voice") Might hamper deliberative process in democracy Lack of shared world view Concern expressed by former US Presidents, Facebook, Twitter, and more.

(72) Polarization Cause Selective exposure? People see only content that agrees with their preexisting opinion Biased assimilation? People pay more attention to content that agrees with their pre-existing opinion.

(73) Echo Chamber Definition Echo = opinion Chamber = network Joint content + network definition Echo chamber = political leaning of content that users receive from network agrees with that of content they share to the network.

(74) Production/Consumption Consumption What you receive in your feed What your followees tweet Production What you tweet.

(75) Political Leaning Scores Based on source of the content (500 domains) Score derived by self-declared affiliation of sharers on FB FoxNews.com is aligned with conservatives (CP = 0.9), HuffingtonPost.com is aligned with liberals (CP = 0.17).

(76) Production/Consumption Scores Polarity scores based on “content” leaning (from source) Production score Average political leaning of the content the user tweets Consumption score Average political leaning of the content the user receives on their feed Results of selection by the user.

(77) δ-partisanship. f s nr e m., s n. e e k. Figure 1: Example showing the de�nition of -partisan users. The dotted red lines are drawn at and 1- . Users on the left of the leftmost dashed red line or right of the rightmost one are -partisan..

(78) δ-{partisan,consumer,gatekeeper} δ-partisan: produces content with polarity beyond δ δ-bipartisan: produces content with polarity within δ δ-consumer: consumes content with polarity beyond δ δ-gatekeeper: δ-partisan but not δ-consumer consumes from both sides but produces content aligned with only one side blocks information flow towards its community.

(79) Network Measures Network-based latent-space user polarity Based on following politicians with aligned ideology Network centrality (PageRank) Local clustering coefficient Retweet/favorite rates and volumes.

(80) Correlation. (a). (b). (c). (d). (e) (e). (f). (g). (h). (i). (j) (j). Figure 3: Distribution of production and consumption polarity, for P�� (�rst row) and and N��P�� N��P�� (second (second row) row) datasets. The scatter plots display the production (x-axis) and consumption ( -axis) polarities of of each each user user in in aa dataset. dataset. Colors Colors indicate user polarity sign, following [6] (grey = democrat, yellow = republican). The one-dimensional one-dimensional plots plots along along the the axes axes show the distributions of the production and consumption polarities for democrats and republicans. republicans..

(81) Correlation: Gun Control.

(82) Variance (f). (g). (h). (i). (j). Figure 3: Distribution of production and consumption polarity, for P�� (�rst row) and N��P�� (second row) datasets. The scatter plots display the production (x-axis) and consumption ( -axis) polarities of each user in a dataset. Colors indicate user polarity sign, following [6] (grey = democrat, yellow = republican). The one-dimensional plots along the axes show the distributions of the production and consumption polarities for democrats and republicans.. (a). (b). (c). (d). (e). (f). (g). (h). (i). (j). Figure 4: Top: Production polarity variance vs. production polarity (mean). Bottom: Consumption polarity variance vs. consumption polarity (mean). However, di�erently from the rest of the side they align with, they show a lower clustering coe�cient, an indication that they are not completely embedded in a single community. Given that they receive content also from the opposing side, this result is to be. Finally, given that both partisans and gatekeepers sport higher centrality, we compare their PageRank values directly and �nd that there is a signi�cant di�erence: partisans have a higher PageRank compared to gatekeepers (�gure not shown). This e�ect is more.

(83) Variance. (b). (c).

(84) Combined. Guncontrol. Obamacare. Abortion. 2.5. Large. 2.0. 2.0. 1.0. 1.0. 0.2. (a). 0.2. (b). 0.2. partisan bipartisan 0.3 0.4 Threshold δ. (c). 0.0. partisan bipartisan 0.3 0.4 Threshold δ. 0.0. partisan bipartisan 0.3 0.4 Threshold δ. −0.5. 0.2. partisan bipartisan 0.3 0.4 Threshold δ. 0.0. 0.0. 0.5. 1.0. 1.0. 1.5. 2.0. 2.0. Price of Bipartisanship. 0.2. (d). partisan bipartisan 0.3 0.4 Threshold δ. (e). Figure 5: Absolute value of the user polarity scores for -partisan and -bipartisan users.. (a). 0.3 0.4 Threshold δ. 0.2. 0.3 0.4 Threshold δ. (b). partisan bipartisan. 1e−04. 1e−03. 1e−03. partisan bipartisan. 1e−04. 2e−04 0.2. Abortion. 0.2. 0.3 0.4 Threshold δ. (c). 1e−05. 0.3 0.4 Threshold δ. 2e−05. 2e−05. 5e−07 0.2. partisan bipartisan. Obamacare. 1e−05. partisan bipartisan. 2e−04. 2e−06. partisan bipartisan. Guncontrol 2e−03. Combined 2e−03. 1e−05. Large. 0.2. (d). 0.3 0.4 Threshold δ. (e). Figure 6: Pagerank for -partisan and -bipartisan users.. ble 3: Comparison between -gatekeeper users and a ranm sample of normal users. A 3 indicates that the correonding property is signi�cantly higher for gatekeepers < 0.001) for at least 4 of the 6 thresholds used. A mis next to the checkmark (-) indicates that the property is ni�cantly lower.. Table 4: Accuracy for prediction of users who are pa sans (p) or gatekeepers ( ). (net) indicates network and p �le features only, (n-gram) indicates just n-gram featur The last two columns show results for all features combin p (net). (net). p (n-gram). (n-gram). p.

(85) hreshold δ. (b). Threshold δ. Thresh. (c). Price of Bipartisanship: PR. (d). value of the user polarity scores for -partisan and. (b). 1e−04 0.2. 0.3 0.4 Threshold δ. (c). 1e−05. 2e−05. 0.3 0.4 hreshold δ. Obama 1e−03. partisan bipartisan. 2e−04. partisan bipartisan. Guncontrol 2e−03. ombined. 0.2. 0.3 Thresh. (d).

(86) ions d of g of. prource prothe and ddifor ties. (for and. mpalso prorces. ach The. Table 2: Comparison of various features for partisans & bipartisans and gatekeepers & non-gatekeepers. A 3 indicates that the corresponding feature is signi�cantly higher for the group of the column (p < 0.001) for at least 4 of the 6 thresholds used, for most datasets. A minus next to the checkmark (-) indicates that the feature is signi�cantly lower.. Partisans vs Bipartisans Gatekeepers vs Non-gatekeepers Features PageRank clustering coe�cient user polarity degree retweet rate retweet volume favorite rate favorite volume # followers # friends # tweets age on Twitter. Partisans. Gatekeepers. 3 3 3 3 3 3 3 3 7 7 7 7. 3 (-) 3 (-) (-) 3 (-) 3 7 7 7 7 7 7 7 7. datasets).9 A “3 (-)” means that the property is signi�cantly lower.

(87) Summary Find echo chambers in political discussion on Twitter Definition of echo chambers with two elements: Content (echo) + Network (chamber) Data supports the selective exposure theory Bi-partisan users pay a price in terms of network centrality and content appreciation.

(88) Conclusions How do controversies unfold on social media? Measuring is the first step (RWC) Controversies are dynamic (time is an important factor) Collective attention increases polarization Echo chambers associated with controversies Evidence of selective exposure and price of bi-partisanship.

(89) What's next? Joint opinion formation + network generation model Adding data to opinion dynamics models Temporal dynamics of the process Application to other contexts (Reddit, Facebook) Interventions: can we do something about it?.

(90) Thanks! Questions please! @gdfm7 [email protected]. 64.

(91)