Zurich Open Repository and Archive University of Zurich Main Library Strickhofstrasse 39 CH-8057 Zurich www.zora.uzh.ch

Year: 2012

Political Polarization and Popularity in Online Participatory Media: an Integrated Approach

Garcia, David ; Mendez, Fernando ; Serdült, Uwe ; Schweitzer, Frank

DOI: https://doi.org/10.1145/2389661.2389665

Posted at the Zurich Open Repository and Archive, University of Zurich ZORA URL: https://doi.org/10.5167/uzh-98855 Book Section

Originally published at: Garcia, David; Mendez, Fernando; Serdült, Uwe; Schweitzer, Frank (2012). Political Polarization and Popularity in Online Participatory Media: an Integrated Approach. In: Weber, Ingmar; et al. PLEAD 12: Proceedings of the First Edition Workshop on Politics, Elections and Data. New York: Association for Computing Machinery ACM, 3-10. DOI: https://doi.org/10.1145/2389661.2389665 Political Polarization and Popularity in Online Participatory Media: An Integrated Approach

David Garcia Fernando Mendez, Uwe Serdült Frank Schweitzer Chair of Systems Design ZDA, Universität Zürich Chair of Systems Design ETH Zurich {Fernando.Mendez, ETH Zurich [email protected] Uwe.Serdult}@zda.uzh.ch [email protected]

ABSTRACT short time. But the main difference is that in the latter, We present our approach to online popularity and its appli- users can also reach large numbers of other users by posting cations to political science, aiming at the creation of agent- publicly accessible comments. Commonly, we see collective based models that reproduce patterns of popularity in par- patterns that are emergent phenomena resulting from the ticipatory media. We illustrate our approach analyzing a interaction of many users. One of these is the popularity of dataset from Youtube, composed of the view statistics and online content, which is difficult to explain by analyzing the comments for the videos of the U.S. presidential campaigns “representative user”, due to the inherent heterogeneity of of 2008 and 2012. Using sentiment analysis, we quantify the the users of participatory media. Despite these individual collective emotions expressed by the viewers, finding that differences, the dynamics of votes in Digg [1], and of video democrat campaigns elicited more positive collective emo- views in Youtube [6, 25], show the existence of statistical tions than republican campaigns. Techniques from compu- regularities of content popularity. tational social science allow us to measure virality of the A common approach is to measure online popularity through videos of each campaign, to find that democrat videos are the amount of users that viewed some content. Because user shared faster but republican ones are remembered longer in- interaction plays a fundamental role in online communities, side the community. Last we present our work in progress the sheer amount of views is not able to grasp other rele- in voting advice applications, and our results analyzing the vant features of popularity. Some examples are how fast the data from choose4greece.com. We show how we assess the content is shared among users, or the collective emotions policy differences between parties and their voters, and how expressed in open discussions. In particular, user emotional voting advice applications can be extended to test our agent- expression and positive/negative votes, such as likes and based models. dislikes, are essential for the political sciences. In addition to the number of views or votes, this information contains Categories and Subject Descriptors new levels of complexity. For example, the patterns of pos- J.4 [Computer Applications]: Social and behavioral sci- itive and negative votes for the content posted in Reddit, a ences—Psychology, Sociology; H.1.2 [Information Systems]: social bookmarking website, reveal that the way users vote Models and principles —User/machine Systems depends on the votes given by other users [31]. This kind of emotional influence goes beyond objective information shar- Keywords ing, and require psychological explanations. Popularity, politics, agent-based modeling, emotion In our approach, we aim at a unified view of online popu- larity that links collective and individual behavior, applying 1. INTRODUCTION it to political parties and candidates. Our statistical anal- Online participatory media, such as social networking sites ysis of user behavior in online communities reveals robust or discussion forums, play a key role in current political cam- patterns, such as the diffusion of popular content and the paigns, and offer the chance to gather large datasets of voter collective emotions expressed towards it. One step further, expression and behavior. For example, the messages posted we also aim at reproducing these patterns by agent-based by thousands of Twitter users allowed the analysis of the models based on emotional influence [23]. Agent-based mod- differences in social interaction between parties in the U.S. els are starting to be used as a tool in social psychology and elections [4, 5]. A common feature of traditional mass me- emotion research [24]. We have applied these kind of mod- dia and online participatory media is that both offer politi- els to reproduce certain patterns of emotional interaction cians a way to reach a very large numbers of users in very in chatrooms [9], and product reviews [10]. Agent-based models provide a tractable link between the collective hu- man behavior and the interactions of individuals, which is of particular value for political sciences [18]. Agent-based Permission to make digital or hard copies of all or part of this work for models are based on assumptions that describe the indi- personal or classroom use is granted without fee provided that copies are vidual behavior and the interaction between agents. These not made or distributed for profit or commercial advantage and that copies assumptions need to be empirically testable in order to be bear this notice and the full citation on the first page. To copy otherwise, to integrated in a larger scientific perspective, and to be ap- republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. plied in real-world applications. Having access to the online PLEAD'12, November 2, 2012, Maui, Hawaii, USA. traces of users in social networks is usually not enough to Copyright 2012 ACM 978-1-4503-1713-9/12/11 ...$15.00. design the model at a fine-grained level. In our approach we In general, it achieves human accuracy, i.e. it can estimate go beyond the publicly visible layer of the user, integrating emotional content from text with a similar error rate as a other data sources of individual behavior. With respect to human reader. It has been recently used for the analysis of political applications, we plan to test our models with data emotional expression in Yahoo answers [14], Twitter trends on individual activity in voting advice applications (VAAs). [26], and chatroom communication [9]. SentiStrength uses VAAs match the policy preferences of real voters (who visit a lexicon of emotional-bearing terms combined with the de- the website) with that of candidates and/or political parties tection of negations, amplifiers and diminishers, providing (that are already encoded in the online system), through the two values, a positive score p ∈ [+1, +5], and a negative voter’s answers to a set of predefined questions. While de- score n ∈ [−1, −5]. As the text of Youtube comments is signed to provide advice to the users, VAAs also provide an very short (max. 500 characters), we aggregate both values experimental platform to gather data on voter preferences, to a measure of polarity. A comment c is classified as posi- which cannot be easily retrieved through online traces or tive (ec = +1) if p + n > 0, negative (ec = −1) if p + n < 0, election surveys. or neutral (ec = 0) if p = n and both have an absolute value In this article, we present our approach to popularity of lower than 4. This way, comments with high and equal posi- political campaigns, and how we can apply our findings in tive and negative scores are difficult to classify, as we assume political sciences. We illustrate our work in progress by that the text is too short to express the coexistence of emo- quantifying the success of the campaigns in Youtube for the tional states. In fact, only 612 comments were detected as U.S. presidential elections of 2008 and 2012. Our results the pathological cases of [+4, −4] or [+5, −5], which we dis- show how we quantify the content diffusion of a campaign card from our analysis as they form less than 0.3% of the and the collective emotions towards candidates. We follow total. with a description of our integrated approach, composed of data analysis from online participatory media, agent-based 2.2 Collective emotions in political discussions modelling of social interaction in online communities, model validation, and applications to VAAs. The article closes pre- In addition to the emotional expression in user comments, senting how we can test the patterns of user interaction in we also study collective emotions through the likes and dis- political contexts through VAAs, and how such applications likes for the videos in our Youtube dataset. Each video has can be improved by our models and statistical analysis. a fixed amount of likes and dislikes at the moment of the data retrieval. The distribution of these values per cam- paign provide information about how positive or negative 2. YOUTUBE CAMPAIGNS is the collective response of the Youtube community. There is no limit in the amount of likes and dislikes a video can 2.1 Data collection and emotion classification receive, so their values are not bounded from the right. Fur- We want to illustrate our work in progress by showing thermore, these distributions show a large variance, ranging a relevant example of the application of our approach to from 1 to 10 million. For this reason we choose to analyze the online popularity of political campaigns, in particular these distributions in a logarithmic scale, as shown in Fig. to the Youtube channels of the candidates for the US presi- 1. These plots already allow us to discard that they follow dential elections in 2008 and 2012. For this application, we power-law or log-normal distributions, requiring more data have retrieved information for all videos in the official chan- to infer some possible underlying parametric distribution. nels for Barack Obama (barackobamadotcom), Mitt Romney Fig. 1 helps us to detect differences in the way users re- (mittromney), and John McCain (johnmccaindotcom). In sponded to the set of videos of a political campaign. The total, 2865 videos where available at the date of the data Obama2008 campaign had a clearly larger average amount of retrieval, June 4th 2012. For each video, the following data likes than amount of dislikes, while these two values are com- was available: i) time series of views since its creation (more parable for the Obama2012 campaign. As the 2008 videos than 190 million views), ii) amount of likes and dislikes have been available for about four years, the larger amount (1024402/249986), and iii) comments with text (244010). of likes might have its source in later likes coming after We divide our dataset in four subsets referring to each of the Obama’s victory. Our analysis will continue after the presi- four studied campaigns. The Obama2008 and McCain2008 dential elections of 2012 to assess if this pattern of positivity subsets contain all the videos in the barackobamadotcom and emerges for the winning candidate, or if it might be taken johnmccaindotcom channels up to the 2008 presidential elec- as support for the detection of different collective emotions. tions day. The Obama2012 and Romney2012 subsets con- We aggregate the emotions expressed in the set of com- tain all the videos in the barackobamadotcom and mittrom- ments of a video Cv by calculating the ratios of positive e =1 e = 1 Pc∈Cv c Pc∈Cv c − ney channels created up to a year before the data retrieval v v P = |Cv | , negative N = |Cv | , and neutral date. e =0 Pc∈Cv c v The programming interface of Youtube limits the access U = |Cv | comments. These three values allow us to to up to 1000 comments per video, but provides the total map each video in a space that can equally represent videos number of comments for each video. From this we infer that that did not elicit collective emotional responses (high Uv), 70 of the videos of our dataset (2.5% of the total), have created positive or negative collective emotions (high Pv or more than 1000 comments. The text of each comment has high Nv), or created polarized emotional reactions (both been processed using SentiStrength [28], a sentiment anal- high Pv and Nv). In Fig. 2 we plot each video as a point ysis tool for the extraction of emotional content from short inside a triangle with a distance to the vertices inversely pro- texts. SentiStrength is the state-of-the-art tool for lexicon- portional to Uv for the upper vertex, Nv for the lower left based analysis of short, informal text. This is the case of vertex, and Pv for the lower right vertex. Given the over- Youtube comments, for which SentiStrength’s accuracy is all ratios of emotions in the retrieved comments ,P,¯ N,¯ U¯, above 88% when classifying positive from negative text [27]. we can perform nonparametric statistical tests to determine tv mtos ntecs fMt onyscanl we channel, Romney’s Mitt of case neg the collective one elicited In videos of the two to emotions. time pattern similar ative this overall but be 2008, the to in starts that seen two 2012 see the in expression can Comparing emotional we emotions. some campaigns having collective Obama channel, discus- negative his in negative of a uploaded Mc cases videos more John of few much the hand, case in triggering other sions no were the videos and On Cain’s responses emotion. in- clearly collective collective numerous were negative positive having videos of expression, Obama’s stances positive Barack towards of skewed comments dataset. the this in for bipolar find a not elicited do video we passes the video which that a emotion, conclude if collective Finally, we (yellow). tests, is three we it all the response, type third, which emotional the not of collective nor not a second but but created the video third, not (red) the but emotion the that test know collective first and the negative passes first a it the only emotion If created collective passes it positive it second, a If only created (green). video the third, inlsaei h omnt.Ti akcniti three emo- in collective consist task a This reflect community. video the in a state of tional comments the the the for whether of dislikes(red) campaigns. logarithm studied and four the the (green) in of likes videos Distributions of amount 1: Figure et tte9%cndneinterval: confidence 95% the at tests ouigo h 08cmags h mtosexpressed emotions the campaigns, 2008 the on Focusing the not but test, second the and first the passes video a If .Testing 3. .Testing 2. .Testing 1. 2012 2008 httevdopoue eaiecletv emotion. collective negative a produced video the that httevdopoue oiiecletv emotion. collective positive a produced video the that mtoa epne n epoedwt h ettwo next the with proceed tests. we and response, emotional l necag flnst te ieso websites. or videos other to conclude links we of If exam- exchange for an (gray), ple response underemotional an elicited i that sis epne(lc) fw conclude we If (black). response density density

0 20 40 60 80 0 100 200 300 400 500 600 700 .51218222832384.2 3.8 3.2 2.8 2.2 1.8 1.2 0.25 .51218222832384.2 3.8 3.2 2.8 2.2 1.8 1.2 0.75 Obama Obama N U P U v v v v versus ≃ versus versus U ¯ U h ie i o lcta emotional an elicit not did video the , v U P N < ¯ ¯ ¯ fw antrjc h hypothe- the reject cannot we if : fw find we if : fw find we if : U ¯ h ie rae collective a created video the ,

density density 0 5 10 15 20 25 30 0 20 40 60 80 100 N P .512182228323.8 3.2 2.8 2.2 1.8 1.2 0.25 . . . . . 3.3 2.9 2.5 2.1 1.7 1.3 Romney v Mc Cain Mc v U > > v > N P ¯ ¯ econclude we , econclude we , U ¯ h video the , χ 2 - . ie ttime at video iesaepsil,ie asi hscs.Ti etu 1310 the us for left videos This 138 and case. campaign. campaign fastest this Romney2012 Obama2012 the in the in days for users videos i.e. of possible, activity scale we the and time investigate points, to 100 of like resolution would maximum a with series time . esrn h vrlt’o oiia cam- political of ’virality’ the Measuring 2.3 candidate. indeed need the can not about campaigns discussions does online attract negative and Popularity to encourage popularity, politician positive videos. a be these for to to desirable the always attention of is large response it emotional whether of of type Mc the Youtube kind John of This to negativity. quantification of distri- similar of values the high seems but reaching without expression emotions, Cain’s collective emotional of of case bution any find not did four the in videos the campaigns. for studied comments collective the the in of emotions representation Triangular 2: Figure h erea ae h esnfrti sta h histori- the that is this for the reason in before available The days data cal 100 date. 2012, the retrieval 25th on the the February to after focusing analysis uploaded campaigns, our videos Romney2012 restrict We and and of Obama2012 Obama2012 campaigns. the the series their of of of time traces videos dynamical success the for the explore looking for campaigns, we Romney2012 views be- section of exchange of this amount information success In the this online on The users. lies tween [17]. campaign community political other online by a adoption the the of in and diffusion users campaign the measuring the defined marketing, about be viral information can to campaign analogy For political an community. a as the of of ’virality’ the activity example the but views, of only of properties amount not time the dynamical is like into as metrics popularity stationary such look before, of composed mentioned to expression, the we such us As by allowed of aggregates. only elicited properties data stationary emotions available the collective the the but video, about insights us edefine We h mtosepesdi h omnso ie gave video a of comments the in expressed emotions The 2012 2008 N paigns N ieso oiia apinoestequestion the opens campaign political a of videos 0.8 0.8 Obama Obama N

t 0.6 0.6 o ahday, each For . v

(

0.2 t 0.2

stetettlaon fvesfra for views of amount total the the as ) 0.4 0.2 0.4 0.6 0.8

U 0.4 0.2 0.4 0.6 0.8

U 0.4

0.2 0.4

0.2

0.6

Youtube 0.6

0.8 0.8 P P n v N aeo h iesincludes videos the of page N ( t steaon fviews of amount the is )

0.8 0.8 Romney Mc Cain Mc

0.6 0.6 0.2

0.4 0.2

0.2 0.4 0.6 0.8 0.4 0.2 0.4 0.6 0.8 U

U

0.4 0.4

0.2 0.2

0.6

0.6

0.8 0.8 P P within the more than 100 videos uploaded in the channel. This means that every time a video was uploaded and fea- tured in the website, the sharing rate of the users involved kept the viewing activity alive in the following days. These density videos keep alive in the community for a long time, which is in line with the stronger social interaction of right leaning 0.0 0.2 0.4 0.6 0.8 0.6 0.4 0.2 0.0 1 2 3 users in Twitter [4]. The case of the Obama2012 campaign class is different, as there is a much larger ratio of videos of class 1, meaning that some videos of this campaign did not trigger

log10(r(t)) the sharing response of the users, who forgot about them quickly. Nevertheless, the unscaled counts show that the Obama Obama2012 contained 710 videos of class 2, while the Rom- Romney

−3 −2 −1 0 1 ney2012 campaign only 112. This teaches us that, in order to get a large amount of critical responses, or ’viral’ videos, a campaign needs to produce a large amount of videos, of which some might not reach this point of criticality and have 0.5 1.0 1.5 2.0 a lower impact in the community as a whole. The previous analysis allowed us to detect when the videos log10(t) [days] of a campaign are above or below the threshold of ’viral- Figure 3: Average growth rates of the videos of ity’, but we can still ask the question of whether we can the Obama2012 and Romney2012 campaigns. Error measure the overall virality of the campaign. To do this, bars show standard error over the available videos we calculate the growth rate of the amount of views per for each campaign. Inset: Distribution of the classes day for each one of the videos of the campaign defined as of view response functions. rv(t) = nv(t)/Nv(t−1). The growth rate rv(t) measures the amount of new views of a video per previous view, serving that happened only during the day t after the upload of t ′ as an estimator of the infection rate of a video in the com- a video, so Nv(t) = P ′ nv(t ). We apply a statistical t =1 munity, or in other words, the average amount of new views model for the analysis of the complex dynamics of Youtube triggered by each time a user viewed the video before. Pre- users [6], classifying the different responses of the commu- vious analysis of the time evolution of these growth rates nity through the time series nv(t). This technique explores of video views [25] showed patterns in the way the over- two different facets of the collective response to the videos. all Youtube community behaves. In our case, we focus on The first one is the nature of the initiation of the response, the videos of the two studied political campaigns, to draw which can be exogenous if there was some central mecha- particular pictures for the users involved in the viewing of nism that promoted the video from its creation, for example such videos. To analyze each campaign, we average over the being featured by Youtube. An endogenous response would whole set of available videos at time t, taking r(t) = hrv(t)i. be produced by the interaction of a large amount of users, The time evolution of hr(t)i contains key information on the bringing up videos previously unknown to the larger com- degree of sharing induced by each campaign. munity. The second feature we explore is the criticality of Fig. 3 shows the time evolution of the growth rates of the response, which we use as a way to study the ’virality’ of the views for the Obama2012 and Romney2012 campaigns. the video. A video elicits a critical response if Youtube users The growth rate of the views of the videos in the Obama2012 share and recommend the video at a higher rate than they campaign is significantly larger than the ones in the Rom- forget about it, leading to a long response in the views. On ney2012 campaign for the first two weeks after the video is the other hand, this response is subcritical when the shar- posted. This difference is especially high in the first week, ing rate is lower than the decay of attention to the video, when the growth rates of Obama2012 videos are usually leading to a very rapid decreasing amount of views. The sta- above 1, i.e. on average, videos receive more new views tistical model introduced in [6] compares the total amount every day than in all the previous days. In the case of Rom- of views of a video, Nv, with the maximum amount of views max ney2012, only the second day brings an amount of views in a single day, nv . This analysis classifies the type of col- larger than the first, and then all the growth rates are below lective response to each video into three classes: exogenous 1. On the other hand, the slope of the growth rates shows a subcritical (1), exogenous critical (2), and endogenous (3). stronger decay for Obama2012 than for Romney2012, as af- The inset of Fig. 3 shows the rescaled distributions of the ter the second week, Romney’s videos attracted more views different response classes for the videos of the Obama2012 in comparison with the previous amount. This difference is and Romney 2012 campaigns. In both campaigns we no- in line with the fact that we did not find videos of class 1 in tice that more than 80% of the videos are in the exogenous Romney2012, as the videos of this campaign seem to cause classes 1 and 2. This is not surprising, as the channels of more moderated responses that last longer in time than the the presidential candidates should receive a special treat- ones in the Obama2012 campaign. Obama’s videos grow ex- ment in Youtube due to their public interest, and the videos plosively in the first two weeks, but they do not keep this uploaded in them are very likely to be featured or appear in growth rate as probably most of the community willing to the main pages of Youtube. This poses a very different sce- view them has already done so. Our data retrieval technique nario as in the case of user uploaded videos, where the vast did not impose any demographic or geographic bias, but the majority cannot receive an external influence that triggers communities addressed by this videos might have different an exogenous response. locations, age and gender distributions. The demographics Focusing on the Romney2012 campaign, we notice that we did not find any video of class 1 (exogenous subcritical) of Youtube might play a role in the origin of these results, 3.2 Modeling user interaction and their influence is open for future research. Most likely, the stylized facts found in our analysis of OPM will not be possible to be explained as a superposition of the 3. AN INTEGRATED APPROACH TO ON- behavior of individual users. Agent-based models provide a link between these macroscopic effects, and the microscopic LINE POPULARITY interaction among individuals. In this article we have pre- In this section we present our approach to popularity in sented preliminary results in the analysis of the popularity online participatory media, and its applications to political of political campaigns in Youtube. The agent-based models sciences. Our work is divided in three areas: data anal- that reproduce the stylized facts of campaign virality and ysis from online participatory media (OPM), agent-based collective emotions of Section 2 are current work in progress, modelling of the interaction between users (ABM), and ap- following our previous approaches for chat communication plications to improve and analyze voting advice applications and product reviews [9, 10]. The design of these ABM will (VAA). VAAs are online applications deployed during elec- allow the execution of simulations that span the possible toral campaigns which match the policy preferences of vis- scenarios of the community. In addition, some models allow iting users with those of candidates and/or political parties the usage of tools from statistical physics that provide ana- that are already stored in the system. The crucial point lytical results on the equations that rule the behavior of the about VAAs for our study is that they are becoming in- agents, and these in turn can be formulated as predictions creasingly popular -millions of visitors in many cases- and about the behavior of the social system as a whole. generate large datasets of mass public opinion (see more de- Each ABM is based on a set of assumptions that need to tailed discussion in Section 4). Our results in each of these be empirically testable, so the design of our models will be areas (OPM, ABM, and VAAs) is aimed at generating an driven in a way such that VAAs can provide data for model impact in the other two areas. This requires a special mul- validation at the individual level. The assumptions of our tidisciplinary effort, because each of these areas address dif- models will be phrased as hypotheses that should be tested ferent scientific domains. In particular our analysis of OPM against the individual user data, and VAAs can be mod- is relevant for statistics, our ABMs for computer science, ified and extended to provide this support. Agent-based and our VAAs for political sciences. models can also provide different metrics that summarize properties of the social interaction between users of VAAs, 3.1 Emergence of popular content including their patterns or communication and their posi- The central aim of our research is to understand the emer- tioning towards political issues and candidates. VAAs can gence of popularity and polarization in online participatory be improved with recommendation modules that make use media and study their influence in politics. We have pre- of network metrics inspired in the social interaction present sented our preliminary results on Youtube in Section 2, but in our ABM. Different models of social dynamics allow the we do not plan to restrict our analysis to the videos of U.S. formulation of network metrics that can, for example, mea- political campaigns. Websites like Reddit and Digg pro- sure the influence among users from different points of view. duced public datasets on the way users voted for articles, For example, different assumptions of social behavior can including those of political content across countries. In ad- be used for spammer detection in Twitter [11], by apply- dition, specialized websites like politnetz.ch can provide ing their related network metrics. Furthermore, analytical useful information about the support achieved by politicians results of ABMs can be applied to mechanism design, pro- participating in online social networks. All these communi- viding possible improvements in the decision space of the ties include commenting functionalities that allow the users designers of online communities. Simulations of these mod- to discuss about the content posted on them. To extend our els can deliver insights of the impact of different decisions understanding of user behavior in such participatory media, and policies followed by community managers, for example we plan to perform text analysis on the user comments, as we in order to maximize the impact of political campaigns, or did with SentiStrength on the comments for Youtube videos. to trigger discussions that allow voters to deepen in their Different tools can be used to extract particular knowledge own opinions. from different kinds of text, and the choice of such tools heavily depends on the length, formality and language of the texts. 3.3 Measuring individual dynamics Our target in this area is to find statistical regularities of For many users, political leaning is a private concern that popularity that can be compared across communities, and is not shared publicly. This leads to numerous problems, help to predict future behavior. These regularities, com- like self-selection bias, which need to be accounted for when monly known as stylized facts are macroscopic patterns that studying collective expression of political opinions [12]. This can be observed at the collective level, but are difficult to silent majority can compose a large part of the voting com- predict from the behavior of individual users. In particu- munity, and data of this kind is rather scarce. In our work, lar we will focus in the different patterns of user emotional we want to overcome the limitations of analyzing a vocal reaction to individual political topics, parties, and candi- minority that might not be representative of the voter com- dates, as well as how the behavior of users changes when munity as a whole. Our aim is to explore deeper than the they can perceive content popularity. If large enough, the public behavior of general users, studying the dynamics of data contained in OPM can also be used to detect topics of individual users and voters. The high degree of privacy of special relevance, which is of particular importance for the VAAs, in comparison with comments on participatory me- efficiency of voting advice applications (VAA). In Section 4, dia, allow users to freely explore their political convictions we explain how this output of topic selection can be used to without the fear of being exposed. When using a VAA, users improve VAAs. allow the usage of their anonymized data for research, but no user account, email or any other kind of personal data is citizens to better define their own subjective, political pref- required to use the platform. erences and to match these with the stated (or expert coded) We count with previous implementations of VAAs that preferences of candidates or political parties. The end result attracted significant amounts of users for local and national should be a more informed vote choice among the range of , Cyprus, London, Scotland, Peru, and parties/candidates competing in the political space. Brazil. We plan to develop some extensions of their function- ality to enhance the data they produce in order to address 4.2 Insights on voter behavior the hypotheses generated by our ABMs, as model validation Our VAA implementations provide substantial datasets is an essential block of the modeling cycle. To explore the on the individual voter and party policy questionnaires. Our individual dynamics of emotions and opinions, we plan to analysis of these datasets focuses on one question that has implement three new mechanisms in our VAA platform: i) occupied a sub-field of political science for many years: how emotional feedback, ii) user discussions, and iii) user to user best to estimate the positions of political parties in the ideo- invitations. The emotional feedback module will consist on logical space. Typically, two dominant strategies have been a set of inquiries to the users that request them to provide a employed to estimate the ideological positions of parties: self-report of emotions in a Likert scale, in particular related expert surveys (usually sent to a handful of political scien- to political topics, candidates, and the output of the VAA it- tists) or the content analysis of party manifestos. It is much self. The user discussion module will consist on a forum that rarer to find the use of mass surveys to estimate political contains discussions for the different political topics of the party positions. VAA-generated datasets are ideally suited VAA, on which the users can make completely anonymous to this task for three reasons: i) they typically contain a comments. This data will allow us to couple the internal much more extensive set of policy statements than the sur- emotions of the users with their political leaning, their vot- veys conducted by national election teams, ii) the datasets ing intention, and their public online expression. In the next they produce are very large, containing millions of users, section, we explain our previous and current work on VAAs, and iii) most VAAs also contain supplementary questions, leaving our work related to ABMs and data visualization for such as the socio-demographic profile of users and their vote future publications. intention or party affiliation. With such data it is possible to map the positions of political parties in the ideological 4. APPLICATIONS TO POLITICAL SCIENCES space, based on the mass opinions of their supporters. Fig. 4 shows the ideological space of the recent elections 4.1 Voting Advice Applications in Greece on May 6th 2012. The scatter plot shows the VAAs are online applications that gather the policy pref- data provided by the users visiting the Greek elections VAA erences of the visiting user, matching them to the policy choose4greece.com, and the expert coding of the political statements of candidates and/or political parties. Developed parties. This political map consists of a left-right x axis, in The Netherlands in the 1990s, and initially paper-based, and a y axis in which the VAA designers incorporated a VAAs have now proliferated across many European coun- new dimension that structured the Greek elections of 2012 tries during election campaigns. Furthermore, participatory -pro versus anti positions to the Greek bailout conditions media has further increased the dissemination potential of imposed by the European Union and IMF. Squares in Fig. such tools across the wider electorate. To date it appears 4 show the mean position of partisan supporters based on that VAAs have become more ’institutionalized’ in a number their answers to the policy statements that loaded on to each of specific national electoral settings: Belgium, Germany, of the dimensions. Circles represent the expert coding of the The Netherlands, and Switzerland. What is striking about corresponding parties, connected to the squares aggregating the deployment of VAAs in some of these pioneer cases is the users willing to vote for the corresponding party. Party the sheer magnitude of voting ’recommendations’ issued, be- supporters occupy a less polarized position than the aca- tween 20-40 per cent of the electorate in some elections and demically coded positions of the political parties. Whilst it around 5 million recommendations in the case of German is to be expected that the average partisan supporter is less elections in 2005. These countries share important electoral consistent across an extensive set of policy statements, it is featureas, namely the fact that they are multi-party sys- important to note that all the mean positions of the parti- tems. Indeed, it has been argued that VAAs are especially san supporters are in the correct quadrants. Such dynam- well suited to multi-party electoral settings [22]. The same ics could be relevant to broader debates about processes of appears to be true for systems where elections are candidate- party dealignment or realignment among political scientists. centred rather than party based [29, 16]. This result is in line with the increasing trend towards party Another reason for the increasing online popularity of dealignment and growing numbers of floating voters [32, 16], VAAs is their relative simplicity. The mechanism is rather which in turn is an additional argument to use VAAs as a straightforward. Prior to an election, candidates/parties tool for electoral studies. provide their policy positions by filling in a questionnaire The lower polarization of the average party supporters containing an extensive set of policy statements. In some versus the expert-coded policy position of the parties sug- cases, the policy positions of candidates/parties are coded gests that voters prefer parties that take more polarized po- by political science experts. When the VAA is launched dur- sitions, and not ones that are necessarily closest to them ing the election campaign, citizens can then fill in the same in the political space. This leads us to a concern about policy questionnaire. An algorithm matches the responses VAAs of relevance for computational science, which is how of candidates/parties with the ones provided by the citizens to produce a best ’match’ between voters using the appli- using the tool. The core output is an ordered ranking of cation and the parties/candidates. This is especially impor- candidates/parties according to the degree of similarity to tant since VAAs can be used in some cases by a significant users’ preferences. In short, the aim of the tool is to allow share of the electorate. In short, the choice of algorithm stream policy agenda. We plan to improve VAAs through a mechanism of topic selection based on information retrieval

SYRIZA from OPM. While the wording of a statement is a delicate PASOK issue that requires political science expertise, the set of top- KKE ics addressed by the statements can be improved automati- Greens cally. We plan to implement a feedback system that moni- LAOS tors participatory media, focusing on a large set of possible Dem. Alliance relevant topics. The system will measure the polarization Social Pact DIMAR of user generated content on the topic, and provide a rank of the most significant topics for the VAA. This tool would be a new channel for incorporating policy options that are typically neglected by mainstream media, reshaping users’ perceptions of the political landscape. Such an approach would incorporate elements of the ’contestatory’ model of democracy, which emphasizes the need for alternative view- Figure 4: Representation of the two-dimensional points to mainstream opinion as a crucial element in a well projection of the questionnaire responses for par- functioning democracy [21]. ties (circles) and their voters (squares) in the VAA With the advent of social network platforms, in partic- for the first Greek elections 2012. The horizontal ular Facebook and Twitter, VAAs can quickly become a axis refers to left vs. right oriented policies, and the viral affair. Some of this can be gleaned from the thou- vertical axis to the positioning regarding the bailout. sands of ’likes’ and ’tweets’ displayed by VAA sites. Indeed, used by VAA designers could have electoral consequences. a VAA can also be considered as a socio-technical system, Presently, all VAAs are based on a social choice theory of capable of generating meaningful mass network data. Such democracy [7], a rational model of voter behavior that goes data could lead researchers to develop a better understand- back to Downs’ influential economic theory of democracy [7]. ing of opinion dynamics during online political campaigns. The voter is assumed to compare her stance with that of the In a first step, we will develop a network component for our candidates/parties on all of the issues and chooses the can- VAA that would allow the user (ego) to share obtained re- didate/party which offers, on balance, the largest number sults with friends (alteri) and to invite them to participate of preferred policies. Over the years these ideas have been as well. This approach aims at overcoming the simplified formalized into a spatial theory of electoral choice where view of a VAA user, which often is limited to its answers to voters choose the party/candidate whose policy position is the questionnaire and surveys. The addition of social net- most proximate to theirs [33, 30]. These ’proximity’ models work and interaction data allows an extended view of the form the theoretical core of a VAA, typically using a City user beyond its relation to party policies. Block measure or, less commonly, a Euclidean metric. But there are other theories of voting which are predicated 5. CONCLUSIONS on different assumptions. Directional theories conceptualize We have presented our preliminary results on the statis- the policy dimension differently [15, 33, 30], proposing that tical analysis of popularity in participatory media, through what matters most is for the voter and candidate/party to the collective emotions and ’virality’ of political videos in be on the correct ’side’ of the argument. According to di- Youtube. These results shed light on the different mecha- rectional logic, voters do not tend to distinguish fine policy nisms of emotional interaction and political information dif- gradations and are more concerned that their candidate is fusion in online communities. We introduced our broader on the correct side of the argument. Indeed, the theory sug- perspective on the topic of popularity in participatory me- gests that, within a boundary of acceptability, voters prefer dia and its applications to the political sciences, integrat- more extreme candidates. The metric used in a directional ing collective and individual dynamics through agent-based theory is the scalar product. VAA generated data is ideally models. Our current platforms of VAAs provide novel data suited for testing the predictive power of these competing that complements what we can extract from the online traces hypotheses of voting behavior. A preliminary analysis by of Internet users. This data can be used to test different hy- Mendez [20] finds that directional inspired algorithms per- potheses of voter behavior, and to support the assumptions form better in predicting the voter intention of partisan VAA of our ABMs. Furthermore, future VAAs can include a se- users than proximity metrics in four electoral settings. ries of socially-aware tools, like topic selection based on user discussions. 4.3 Improving VAAs through social media Future analysis of various online communities as well as A crucial element to VAA design is the selection of policy new instances of VAAs can provide large amounts of high statements. Policy statements should be chosen that best quality data on voter behavior. We expect processes of so- polarize the parties/candidates as well as the voters using cial comparison, cognitive dissonance and homophily to be the tools. Here the problem is that the choice of policy at work not only in the real but also in the virtual world statements (and their wording) is not necessarily a neutral [8, 19]. VAA network data can then be analyzed with the affair. Simulation results [32] show that the selection of par- help of concepts and indicators well known in social network ticular statements has a considerable impact on the voting analysis [3]. Reciprocity within social networks of VAA users recommendations that are produced. The configuration of would for example be expected to go hand in hand with sim- the policy questionnaire could benefit some parties dispro- ilar political beliefs and values [2, 13]. portionately. Through their selection of items, it could be argued that VAA designers are helping to reinforce the main- 6. ACKNOWLEDGEMENTS [18] J. Lorenz. Zur Methode der agenten-basierten Authors acknowledge A. Abisheva for her assistance on Simulation in der Politikwissenschaft am Beispiel von data retrieval. Meinungsdynamik und Parteienwettstreit. In Jahrbuch fur¨ Handlungs- und Entscheidungstheorie. Band 7: 7. REFERENCES Experiment und Simulation. VS Verlag fur¨ [1] S. Asur, B. A. Huberman, G. Szabo, and C. Wang. Sozialwissenschaften, 2012. Trends in Social Media: Persistence and Decay. In [19] M. McPherson, L. Smith-Lovin, and J. M. Cook. ICWSM’11, Palo Alto, 2011. Birds of a feather: Homophily in social networks. [2] P. F. Berelson, Bernard R. Lazarsfeld and W. N. Annual Review of Sociology, 27:pp. 415–444, 2001. McPhee. Voting: A Study of Opinion Formation in a [20] F. Mendez. Matching voters with political parties and Presidential Campaign. University of Chicago Press, candidates: An empirical test of four algorithms. 1954. International Journal of Electronic Governance, [3] S. P. Borgatti, A. Mehra, D. J. Brass, and (Special issue on Voting Advice Applications and the G. Labianca. Network analysis in the social sciences. State of the Art), forthcoming. Science, 323(5916):892–895, 2009. [21] P. Pettit. Democracy, electoral and contestatory. In [4] M. D. Conover, B. Gonc, A. Flammini, and I. Shapiro and S. Macedo, editors, Designing F. Menczer. Partisan Asymmetries in Online Political Democratic Institutions. Nomos, NYU Press, New Activity. EPJ DataScience, 1(6):1–17, 2012. York, 2000. [5] M. D. Conover, B. Goncalves, J. Ratkiewicz, [22] A. Ramonait´e. Voting advice applications in lithuania: A. Flammini, and F. Menczer. Predicting the Political Promoting programmatic competition or breeding Alignment of Twitter Users. In SocialCom’11, pages populism? Policy and Internet, 2(1):117–141, 2010. 192–199, 2011. [23] F. Schweitzer and D. Garcia. An agent-based model of [6] R. Crane and D. Sornette. Robust dynamic classes collective emotions in online communities. The revealed by measuring the response function of a European Physical Journal B, 77(4):533–545, 2010. social system. Proceedings of the National Academy of [24] E. R. Smith and F. R. Conrey. Agent-based modeling: Sciences of the United States of America, a new approach for theory building in social 105(41):15649–15653, Oct. 2008. psychology. Personality and social psychology review : [7] A. Downs. An economic theory of political action in a an official journal of the Society for Personality and democracy. The Journal of Political Economy, Social Psychology, Inc, 11(1):87–104, Feb. 2007. 65(2):135–150, 1957. [25] G. Szabo and B. A. Huberman. Predicting the [8] L. Festinger. A theory of social comparison processes. popularity of online content. Communications of the Human Relations, 7:117–140, 1954. ACM, 53(8):80, Aug. 2010. [9] A. Garas, D. Garcia, M. Skowron, and F. Schweitzer. [26] M. Thelwall, K. Buckley, and G. Paltoglou. Sentiment Emotional persistence in online chatting communities. in Twitter events. Journal of the American Society for Scientific Reports, 2:402, May 2012. Information Science and Technology, 62(2):406–418, [10] D. Garcia and F. Schweitzer. Emotions in Product 2011. Reviews–Empirics and Models. In SocialCom’11, [27] M. Thelwall, K. Buckley, and G. Paltoglou. Sentiment pages 483–488, 2011. strength detection for the social web. Journal of the [11] D. Gayo-Avello. Nepotistic Relationships in Twitter American Society for Information Science and and their Impact on Rank Prestige Algorithms. In Technology, 63(1):163–173, Jan. 2012. Congreso Espa˜nol de Recuperaci´on de Informaci´on, [28] M. Thelwall, K. Buckley, G. Paltoglou, D. Cai, and page 31, Madrid, Spain, 2010. A. Kappas. Sentiment strength detection in short [12] D. Gayo-avello. I Wanted to Predict Elections with informal text. Journal of the American Society for Twitter and all I got was this Lousy Paper - A Information Science and Technology, Balanced Survey on Election Prediction using Twitter 61(12):2544–2558, Dec. 2010. Data. arXiv:1204.6441v1, 2012. [29] J. Thurman and U. Gasser. Three case studies from [13] R. Huckfeldt. Interdependence, density dependence, switzerland: Smartvote. Berkman Center Research and networks in politics. American Politics Research, Publication, No. 2009-03.3, 2009. 37(5):921–950, 2009. [30] M. Tomz and R. P. Van Houweling. Candidate [14] O. Kucuktunc, B. B. Cambazoglu, I. Weber, and positioning and voter choice. American Political H. Ferhatosmanoglu. A large-scale sentiment analysis Science Review, 102(03):303–318, 2008. for Yahoo! answers. In WSDM’12, page 633, 2012. [31] P. Van Mieghem. Human Psychology of Common [15] D. Lacy and P. Paolino. Testing proximity versus Appraisal: The Reddit Score. IEEE Transactions on directional voting using experiments. Electoral Studies, Multimedia, 13(6):1404–1406, Dec. 2011. 29:460–471, 2010. [32] S. Walgrave, M. Nuytemans, and K. Pepermans. [16] A. Ladner and J. Pianzola. Do voting advice Voting aid applications and the effect of statement applications have an effect on electoral participation selection. West European Politics, 32(6):1161–1180, and voter turnout? Electronic Participation, pages 2009. 211–224, 2010. [33] A. Westholm. Distance versus direction: The illusory [17] J. Leskovec, L. A. Adamic, and B. A. Huberman. The defeat of the proximity theory of electoral choice. The dynamics of viral marketing. ACM Transactions on American Political Science Review, 91(4):865–883, the Web, 1(1):39, May 2007. 1997.