Embed Size (px)
Citation preview
Rising tides or rising stars?: Dynamics of shared attention on Twitterduring media events
Yu-Ru Lin1,*, Brian Keegan1, Drew Margolin2, and David Lazer1,2
1College of Social Sciences and Humanities, Northeastern University, Boston, MA 02115, USA2College of Computer and Information Science, Northeastern University, Boston, MA 02115, USA
*To whom correspondence should be addressed. Email: [email protected]
July 8, 2013
Abstract
“Media events” such as political debates generate conditions of shared attention as many users simultaneouslytune in with the dual screens of broadcast and social media to view and participate. Are collective patterns of userbehavior under conditions of shared attention distinct from other “bursts” of activity like breaking news events?Using data from a population of approximately 200,000 politically-active Twitter users, we compare features of theirbehavior during eight major events during the 2012 U.S. presidential election to examine (1) the impact of “mediaevents” have on patterns of social media use compared to “typical” time and (2) whether changes during mediaevents are attributable to changes in behavior across the entire population or an artifact of changes in elite users’behavior. Our findings suggest that while this population became more active during media events, this additionalactivity reflects concentrated attention to a handful of users, hashtags, and tweets. Our work is the first study ondistinguishing patterns of large-scale social behavior under condition of uncertainty and shared attention, suggestingnew ways of mining information from social media to support collective sensemaking following major events.
Introduction
Twitter plays an increasingly important role dissemi-nating information for both breaking news events aswell as ceremonial media events. Examples of thesemedia events include political campaigns, sportingevents, and television shows where the dates are sched-uled and the stakes are known. Increasingly, these me-dia events invite viewers to “dual screen” by usingtheir mobile devices to share their reactions on socialmedia during live television broadcasts. These back-channel forms of communication facilitate processes ofinformation sharing as audiences respond to narrativesdriven by the content on both screens.
Social media have the potential to allow individu-als to openly communicate with large audiences overtopics of shared interest. However, socio-technical sys-tems like Twitter are also characterized by large dis-parities where the vast majority of activity is concen-trated around a few “stars.” This would suggest mediaevents should reproduce the concentrated attention to-wards elite users rather than bringing in new voices.
However, the aphorism “a rising tide lifts all boats” re-flects how changing systems can benefit all the mem-bers of that system.
This paper addresses two related questions. The firstexamines whether shared attention changes user behav-ior: do Twitter users behave differently during a mediaevent as compared to breaking news events or “normal”time? On one hand, media events could simply repro-duce existing behavioral patterns at larger volumes ofactivity but on the other hand, they could also causesignificant changes in users’ communication patternsof using hashtags, mentions, and retweets. If these be-havioral differences exist, this leads to a second ques-tion: are behavioral differences the result of behaviorshifting across the entire population or do some userspreferentially benefit from behavioral changes duringmedia events? Shared attention to an event could letnew voices emerge and create a “rising tide” where be-havior across all users changes in similar ways. Alter-natively, shared attention to an event could reinforcethe popularity of elite users and create “rising stars”that drown out alternative voices.
1
arX
iv:1
307.
2785
v1 [
cs.S
I] 1
0 Ju
l 201
3
Drawing on the concepts of “media events” and “so-cial foci” from media studies and sociology, we pro-pose a theory of media event-driven behavioral changeto explain how conditions of shared attention alter col-lective behavior in social media. Our research designuses eight events related to U.S. politics that occurredbetween late August and mid-October 2012 as naturalexperiments to analyze changes in the behavior of ap-proximately 200,000 politically-engaged Twitter users.Media events like the national political conventions andpresidential debates are compared to the news eventssurrounding the Benghazi attacks and “47%” contro-versy as well as four “’typical” days preceding eachof the debates. Examining the changes in the rate andconcentration of communication patterns, connectiv-ity, and user responsiveness, the findings provide sup-port for our theory of media event-driven behavioralchange by demonstrating that (1) fundamental patternsof social media activity are significantly altered dur-ing media events compared to non-media events and(2) these differences are largely attributable to an in-creasing concentration of attention to elite users ratherthan an increasing distribution of activity and attentionacross all users.
Background“Media events” are defined by Dayan and Katz as cer-emonies or performances that interrupt routines and in-vite audiences to join in a collective experience [5].These events monopolize media coverage and displaceother regularly scheduled forms of media, are transmit-ted in real time as they occur, are organized by actorsoutside of the control of the media, are pre-planned andadvertised to viewers in advance of their occurrence,are treated by the media with reverence and ceremony,and enthrall large audiences through a norm of sharedviewing. Examples of media events include sportingevents and political speeches that temporarily create anoccasion for large portions of society to ignore divi-sions of geography, culture, and status and focus to-gether on a single event that displaces all other events.
Other events share some but not all of these features.News events may be large enough to displace cover-age, occur outside of media control, and are coveredlive, but they are not pre-planned and thus lack theceremonial aspects or large, enthralled audiences. Ex-amples of news events include disasters and scandalswhere information is scarce and the implications are
uncertain. Nevertheless, significant changes in socialmedia use are observed during major events such asprofessional sports [17], television shows [2], and nat-ural disasters [21]. These back-channel sources of com-munication during other events highlight the potentialof these technologies to support rapid information shar-ing, broad participation, and collective sensemaking asaudiences develop and respond to narratives which aredriven by the content they are consuming on both theirtelevision and mobile device screens [20].
The extensive adoption of “dual screening” prac-tices also influences how citizens are consuming po-litical information: 11% of live debate watchers “dualscreened” the debate by using social media on a com-puter or mobile device simultaneously with watchingtelevision coverage [18]. Users’ tweets during politi-cal debates encode important shifts in sentiment [6]and political candidates’ behavior on Twitter predictstheir likelihood of winning elections [16]. The spreadof political hashtags is also a type of complex conta-gion [19].
Prior work has examined how temporal patternsof collective attention and information propagationon social media are affected by users’ finite atten-tion influence when attending to competing informa-tion sources [22], the novelty of the information it-self [23], the reactions of early users [14], and the bal-ance between rates of spreading versus forgetting [4].Retweets play an important role both labeling contentand signaling membership in a community [24]. Usersare more likely to retweet content that requires lesscognitive effort to act upon by focusing their attentionon the most recent and visible content [3, 10].
With some exceptions [13], analyses of collective at-tention and information spreading have generally as-sumed the system is operating at a steady state and donot account for exogenous shocks that change manyusers’ communication and information sharing behav-iors. Does dual screening during media events lead tosignificant changes in behavioral patterns across a pop-ulation of politically-active Twitter users? If it does,how do these differences manifest themselves acrossdifferent levels of this population?
Media event-driven behavioral changeMedia events could potentially create larger volumes ofactivity without altering the type or distribution of un-derlying behavioral patterns. For example, while there
2
may be more people generating more tweets, the under-lying rates of mention, hashtag, and retweet use couldremain unaltered. However, we argue the social andcultural contexts of information consumption duringmedia events interact with repertoires of technologyuse such as dual screening and lead to what we callmedia event-driven behavioral change.
The conditions of shared attention created by mediaevents displace normal content and become foci aroundwhich individuals collectively orient their behavior andorganize new social relations [7]. When there are nomedia events or conditions for shared attention, indi-viduals’ contexts and activities are fragmented among avariety of different places, groups, and interests. Theiraggregate patterns of using language, hashtags, men-tions, and retweets exhibit some behavioral regulari-ties [11, 9], but are largely uncorrelated between dif-ferent foci and the communities surrounding them.
Media events are not simply live events with largeaudiences, they are rare and enthralling and ceremonialoccasions that demand shared attention. They disruptprevailing fragmented behavioral patterns by creatinga single focal activity to which people pay attention.This shared attention in turn displaces normal contextsand the behaviors associated with them. In effect, themedia event is an external force that causes the popu-lation of users to go through a phase transition from arelatively disordered state to a highly ordered state [1].This “highly ordered” state is characterized by usersattending to a single shared focal activity rather thanmany focal activities and behaving in unusually ac-tive ways such as dual screening. Furthermore, usersin this highly ordered state produce and consume in-formation from an audience that is reflexively aware oftheir shared exposure to the same information. In otherwords, everyone knows that everyone is watching thesame thing.
Like television in its early years [12], social mediacan magnify the significance viewers ascribe to me-dia events and create a liminal space inaccessible tothe audience members physically present at the event,viewers watching only television coverage of the event,or users solely monitoring social media. These limi-nal spaces generate novel understandings and emerg-ing consensus on the meaning of the media event. Forexample, concepts like “big bird”, “binders”, and “bay-onets” are meaningless unless they are interpreted inthe context of their enthusiastic sharing and improvisa-
tion by social media users during media events like the2012 U.S. Presidential debates. This single shared fo-cus and intense activity will qualitatively alter prevail-ing behavioral patterns and communication structures.
We expect that the enthralling nature of media eventswill not only increase tweet volumes, but events char-acterized by high levels of shared attention will changebehavioral patterns such as using retweets, mentions,and hashtags. Users participating in the liminal, dualscreening space between traditional and social mediawill be more receptive to content from other users alsoengaged in dual screening.
Behavioral changes resists concentration
If behavioral differences exist during the shared at-tention of media events, are these differences spreadacross the population or do some segments like elitespreferentially benefit from this increased attention?Users’ increased attention to a few dominant userscould lead to rapid convergence around topics as elitesguided collective attention to specific interpretationsand frames. For example, users retweeting elite userswould lead to increased concentration of novel retweetsas individuals adopted these leaders’ messages. Al-ternatively, users’ may become more receptive to alltypes of messages in which case replies, mentions, andretweets will increase but will not change the concen-tration of mentions, replies, hashtags, or retweets.
We argue that media event-driven behavioral changewill disrupt patterns of behavior across all classes ofusers and will prime users to share and communicatemore. Conditions of shared attention will create oppor-tunities for new and more diverse voices to be heardas fragmented social foci begin to merge and overlap:average users making average tweets who are neverretweeted more than two or three times might receivedozens of retweets during a media event. Conversely,elite users who already tweet regularly will see theirshare of attention decline as more users have largernet changes in activity than the elites, which in turnincreasingly displaces elite content. Thus, despite in-creasing levels of activity and changes in aggregate be-havioral patterns, the merging of fragmented social fociduring media events should cause the structure of men-tions, retweets, and hashtags to become increasinglydecentralized.
3
Research designWe identified eight events related to the 2012 U.S.Presidential campaign that occurred over the approx-imately six-week period of time between late Augustand mid-October. Six media events were identifiedduring this time: the Republican National Convention(RNC) from August 27 through August 30 (“CONV1”), the Democratic National Convention (DNC) fromSeptember 4 through 6 (“CONV 2”), three debates onOctober 3 (“DEB 1”), 16 (“DEB 3”), and 22 (“DEB 4”)involving the presidential candidates, and single vicepresidential debate on October 11 (“DEB 2”). We con-trast these media events with two news events that oc-curred in the same span of time: the terrorist attack onthe American consulate in Benghazi that killed Ambas-sador J. Christopher Stevens on September 11 (“NEWS1”) and the release on September 18 of a video in whichMitt Romney argues “47 percent” of Americans are“dependent upon government” (“NEWS 2”). Both ofthese news events were major stories that dominatedmedia attention for several days.
Together, these eight events make up a continuum ofvarying shared attention: (1) news events that shouldexhibit low levels of media event-driven behavioralchanges since these have diffuse audiences and lowmutual awareness of audience members, (2) the na-tional political conventions that should exhibit mediumlevels of media event-driven changes since partisansselectively expose themselves to the conventions re-flecting their political beliefs, and finally (3) the de-bates that should exhibit the highest levels of mediaevent-driven change as its live and ceremonial naturedrives intense shared interest.
Each of these types of events provide conditions ofa natural experiment to understand how large and ex-ogenous shocks cause changes in a well-defined popu-lation. To provide a baseline for these events, we com-pared the activity of this population during the peaksof these events to the peak levels of activity four daysbefore each of the debates when there were no majormedia or news events. The six media events and twonews events that make up the treatment group and four“typical” events that make up the control group give usa total 12 distinct observations of how the behavior ofthe same population of Twitter users changed duringthese events. Although tweet volumes vary regularlythroughout the week, these pre-events fell on different
days of the week during each of their 96-hour windowsreducing the systematic bias of these events. In general,users’ behavior during the “typical” time preceding thedebate events might have been impacted by the excite-ments of expected debates and other campaign events,leading to a conservative comparison of changing be-havior. This conservative comparison is more appropri-ate because it ensures that the change we measure is nota result of long-term behavioral drift.
Because our design requires tracking behavioralchange across multiple treatments, random samplingfrom the “garden hose” is inappropriate. Instead,we identified a specific sub-population of politically-engaged Twitter users and created a large “computa-tional focus group” to track their collective behaviorover time as a panel [15]. While this limits the gen-eralizability of our findings, it also provides importantinsights to users within the political domain. If a usertweeted using a hashtag like “#debate” or mentionedone of the candidates’ Twitter accounts during any ofthe four presidential debates and their tweet appearedin the Twitter “garden hose” streaming API1, the userwas selected into our user pool. Next, we collected thecomplete tweeting history for these users going backto mid-August using Twitter’s REST API2. Becausethese queries are expensive owing to rate limits, weprioritized users who tweeted during more of the de-bates. Thus users who tweeted during all four debatesare more likely to be represented in the sample thanusers who tweeted during only one of the debates. Theresulting corpus has 290,119,348 tweets from 193,532unique users including elites such as politicians, jour-nalists, and pundits as well as non-elite partisans andaspiring comedians.
For each of the eight events, we examined tweetsmade during a 48– to 96–hour window covering theevent itself and its aftermath. Within these windows,we examined tweet volumes and identified the hourcontaining the peak level of cumulative activity. Allsubsequent analyses in the paper will look at aggre-gate behavior in this one hour window of peak activity.Descriptive statistics for the number for the time of thewindow, unique users, tweets, retweets, mentions, andhashtags observed in each of the 12 events are sum-marized in Table 1. A “political relevance ratio” was
1https://dev.twitter.com/docs/streaming-apis2https://dev.twitter.com/docs/api/1.1
4
PRE NEWS CONV DEB
description Pre-debate baselineBenghazi attack,47% controversy
Republican Nat’l Conv.Democratic Nat’l Conv.
Presidential debates
time4 days before each debate(20:00-20:00 EDT)
2-day news cycle(14:00-14:00 EDT)
3 days(08:00-14:00 EDT)
4 hours(20:00-02:00 EDT)
duration 96 hours×4 48 hours×2 66 hours×2 6 hours×4peak tweet volume 441,168 131,636 296,138 1,591,513peak unique users 58,823 30,684 38,864 114,663political relevance ratio 0.08 0.16 0.50 0.63shared attention none low medium high
Table 1: Summary of datasets.
also calculated to validate the differences between po-litical events. This ratio is the fraction of tweets dur-ing each of the events that containing the names (e.g.,“Obama” or “Romney”), candidates’ twitter handles(e.g., “barackobama” or “mittromney”), or any of thethe events (e.g., “DNC”, “RNC”, “debate”, “benghazi”,“47 percent”, etc.) at the peak time.
Figure 1 plots the distributions of tweet volumes forthe hours preceding and following the one-hour win-dow we analyzed. We show one instance for each cat-egory. The peaks for the debates correspond to thefirst hour of each debate, the peak for the RNC cor-responds to Clint Eastwood’s famous keynote directedat an empty chair at 21:00 EDT on August 30, the peakfor the Benghazi attacks corresponds to 20:00 EDT onSeptember 11 when the first American fatality was re-ported.
Features of communicationTweets encode a variety of relational features such ashashtags, mentions, replies, and retweets. Using thesefeatures, we construct networks to measure the distri-butions of users’, hashtags’, and tweets’ centralities.
• Hashtags (#) – This user-to-hashtag networkmodels when user Alice mentions hashtag “#foo”in a tweet. Alice’s hashtag out-degree reflects thenumber of hashtags she has referenced and thehashtag in-degree of “#foo” reflects the numberof users referencing it.
• Mentions (@) – This user-to-user network mod-els when user Alice addresses user Bob anywherein her tweet. This notifies Bob and lets anyonefollowing her see the tweet. Alice’s mention out-
degree reflects the number of unique users she hasmentioned in her tweets and Bob’s mention in-degree reflects the number of unique users whohave mentioned him in their tweets.
• Replies – This user-to-user network models whenuser Alice addresses user Bob as the start of hertweet. This notifies Bob and only lets the usersfollowing both of them see the tweet. Alice’s replyout-degree reflects the number of unique users shehas replied to in her tweets and Bob’s reply in-degree reflects the number of unique users whohave replied to him in their tweets.
• Retweets (RT) – This user-to-tweet networkmodels when user Alice repeats a tweet fromuser Bob. Alice’s retweet out-degree reflects thenumber of tweets she has re-tweeted and Bob’stweet in-degree reflects the number of users whohave retweeted that tweet. The in-degrees of all ofBob’s tweets can be summed to make a Retweetuser in-degree (RT user) reflecting the numberof times all of Bob’s tweets have been retweetedby all other users.
ResultsWe review the results of our design by comparingthe changes in communication patterns across all fourtypes of events. We unpack these differences in activ-ity levels across event types by analyzing the changesin the rate of adopting novel content for each typeof event. To understand whether these differences arethe result distributed activity across all users or theconcentration of activity around a few users, we ex-amine whether the distributions of activity for media
5
(a) DEB 1 (b) PRE 1 (c) CONV 1 (RNC) (d) NEWS 1 (Benghazi)
0
10000
20000
30000
10/0400:00
10/0401:00
10/0402:00
10/0403:00
10/0404:00
10/0405:00
10/0406:00
twee
t vol
ume
per
min
ute
3000
4000
5000
10/0222:00
10/0223:00
10/0300:00
10/0301:00
10/0302:00
10/0303:00
10/0304:00
twee
t vol
ume
per
min
ute
1000
2000
3000
4000
5000
08/3023:00
08/3100:00
08/3101:00
08/3102:00
08/3103:00
08/3104:00
08/3105:00
twee
t vol
ume
per
min
ute
1000
1500
2000
2500
09/1723:00
09/1800:00
09/1801:00
09/1802:00
09/1803:00
09/1804:00
09/1805:00
twee
t vol
ume
per
min
ute
Figure 1: Tweet volume per minute. Number of tweets per minute in four of the 12 datasets. (a) The six hours during the firstdebate event (“DEB”). For other categories, we plot the six hour volume centering around the peak within the data range: (b)Normal period prior to the first debate evening (“PRE”). (c) Republican National Convention (RNC) event (“CONV”). (d)Breaking political news event regarding Benghazi attack (“NEWS”).
events differ significantly from news events and thepre-event baselines. Finally, we analyze the relation-ships between users’ audience size and their positionsin these activity networks.
Changes in communication
Figure 2 plots the changes in communication volumesfor the twelve observations grouped by event type(typical, news event, national convention, and debate).Results show an increase in topical communicationand a decrease in inter-personal communication dur-ing shared attention. Tweet volumes for the debates arethree to four times greater than during the other typesof events (2a) and the rate of hashtag use nearly dou-bles during media events over the non-media event rate(2b). Because hashtags are an ad hoc way to create asub-community by affiliating a tweet with a label oth-ers are using, the rise of this behavior during mediaevents suggests users are broadcasting diffuse interestsin topics rather than engaging in interpersonal discus-sions.
The ratio of tweets that include any mentions ofusers in the tweet (2c) and the fraction of tweets thatwere replies to one or more users (2d) declines sub-stantially during media events like the debates. This40% decline in directed communication suggests me-dia events may not only dominate attention but alsochange social media behavior to become less inter-personal and more declarative. Finally, the ratio ofretweets among all tweets (2e) increases 20% over thetypical and news events during the more media event-like convention and debates even though these occurredat different times and with different levels volumes ofactivity.
Changes in concentrationThe previous sections demonstrated significantchanges in the aggregate behavior of the users’ popu-lation . However it is unclear whether these differencesare driven by most users becoming more receptiveto elite messaging (“rising stars”) or if all users arebecoming more engaged (“rising tides”).
We use Lorenz curves to plot the cumulative distri-bution of activity in the system for each of the fourtypes of events. A Lorenz curve shows for the bot-tom x% of users or tweets, the percentage y% of theactivity they generated. Thus more equally-distributedactivity will be more linear while more highly con-centrated activity will more parabolic. Relative to thetypical pre-debate events, a top-down pattern of “risingstars” should be indicated by a concentration of activityin a few users while a bottom-up pattern “rising tides”should be indicated by a flattened distributions.
We measure the level of degree concentration inthese Lorenz curves using the Gini coefficient. It is de-fined as the ratio of the area that lies between the line ofequality (the line at 45 degrees) and the Lorenz curveover the total area under the line of equality. The Ginicoefficient for a set of users or tweets P with degreesyi (i = 1, ..., n) and probability function f(yi) is givenby:
G = 1−∑n
i=1 f(yi)(Si−1 + Si)
Sn,
where Si =∑i
j=1 f(yj)yj and S0 = 0. The Gini coef-ficient is a measure for identifying preferential patternsin general, as opposed to measures such as power-lawexponent which can only apply to networks followingpower-law distribution.
Figure 4 plots the in- and out-degree Lorenz curves
6
(a) tweet volume (b) hashtag ratio (c) mention ratio (d) reply ratio (e) RT ratio
●
●
●●
●●●
●
●
●
●
●
400000
800000
1200000
1600000P
RE
NE
WS
CO
NV
DE
B
twee
t vol
ume
●
●
●●
●●
●
●●
●
●●
0.15
0.20
0.25
0.30
0.35
0.40
PR
E
NE
WS
CO
NV
DE
B
# ra
tio
●
●
●●
●
●
●
●
●
●
●
●0.48
0.52
0.56
0.60
0.64
PR
E
NE
WS
CO
NV
DE
B
@ r
atio
●
●
●
● ●●
●
●
●●●
●
0.16
0.20
0.24
0.28
PR
E
NE
WS
CO
NV
DE
B
Rep
ly r
atio
●
●●
●
●
●
●
●
●
●
●
●
0.22
0.26
0.30
0.34
PR
E
NE
WS
CO
NV
DE
B
RT
rat
io
Figure 2: Changes in communication volume. Twitter communication volume change during the peak time in different events.Diamond shapes indicate the mean value of each category. (a) The tweet volumes at the peak hour in the 12 events. (b) Theratio of tweets with at least one hashtag to the total tweets at the peak hour. (c) The ratio of tweets mentioning a user to the totaltweets at the peak hour. (d) The ratio of tweets replying to users to the total tweets at the peak hour. (e) The ratio of retweets tothe total tweets at the peak hour.
(a) Adopt hashtags (b) Adopt mentions (c) Reply users (d) Adopt RTs (e) Adopt users’ RT
10
15
20
25
30
PR
E
NE
WS
CO
NV
DE
B
min
utes
10
15
20
25
30
PR
E
NE
WS
CO
NV
DE
B
min
utes
10
15
20
25
30P
RE
NE
WS
CO
NV
DE
B
min
utes
10
15
20
25
30
PR
E
NE
WS
CO
NV
DE
B
min
utes
10
15
20
25
30
PR
E
NE
WS
CO
NV
DE
B
min
utes
Figure 3: Adoption lags for different event types. Average lag time per user (in minutes) to adopt novel items (hashtags,mentions, users, retweets).
for the four networks of hashtags, mentions, replies,and retweets. The out-degrees for all four types ofactivity networks show significant similarities acrossthe four event types and comparatively high levels ofconcentrated activity: the top users are responsible formost hashtag references (4e), mentioning other users(4f), replying to other users (4g), and retweeting users’content (4h). These findings together suggest the con-centration of users’ attention to content in the networksis very similar regardless of media events.
The in-degrees show more substantial differencesbetween event types as well as truncated distributions.The top 20% of hashtags make up more than 50% ofall observed hashtags (4a), with this concentration be-ing more exaggerated around the national conventionsand more relaxed around the news events. This reflectsthe dominance of a official hashtags like “#nbcpoli-tics” during the conventions and the presence of manyunrelated but popular hashtags like “#moviesyoucant-
dislike” during the unscheduled news events. The con-vention and debate media events also drove increasedconcentration of mention (4b) and reply activity (4c)around top users as compared to pre-events and newsevents. This is suggestive of many users directing theirtweets towards elite users rather than each other. Fi-nally, retweet activity (4d) on the six media eventsshows the largest differences from the pre-debate base-line (see Table 2). The top 25% of users’ tweets ac-counted for approximately 75% of all retweet activity,clear evidence that users’ behavior under conditionsof shared attention become increasingly concentratedaround elites rather than increasingly distributed acrossmany users.
In addition to the differences within the distribu-tions of for a given type of event, the Lorenz curvesin Figure 4 are also qualitatively different across thein-degree versus out-degree distributions: out-degreesshow a pattern of consistent inequality throughout the
7
(a) Hashtag in-degree (b) Mention in-degree (c) Reply in-degree (d) RT in-degree
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
(e) Hashtag out-degree (f) Mention out-degree (g) Reply out-degree (h) RT out-degree
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
0
25
50
75
100
0 25 50 75 100% of population
% o
f vol
ume
dbname
PRENEWSCONVDEB
Figure 4: Lorentz curves for cumulative degree distributions of activity. Increasing equality converges toward diagonal linefrom the origin to the upper-right and increasing inequality converges toward a hyperbola rising to 100% of volume at the 100thpercentile.
population while in-degrees show a pattern of even dis-tribution until rising sharply at the 90th percentile. Fig-ure 5 also captures these differences in in-degree versusout-degree concentration: in-degree concentrations (y-axis) vary considerably across event types while out-degree concentrations are consistent across event types.We discuss this observation in more detail in the fol-lowing section.
To test whether the concentration of activity differedsignificantly from the pre-debate event baselines, wemeasured the deviation of each event type’s Lorenzcurve from the pre-debate baseline events using a two-sample Kolmogorov-Smirnov (K-S) test [8]. The K-S statistics in Table 2 confirm the Lorenz curves fornetworks of mentions, replies and retweets are signif-icantly different during the news events, national con-ventions, and debates. The statistics for the differencesbetween the out-degrees are generally much larger thanthe statistics for the in-degrees, reflecting the larger dif-ferences between these curves discussed above. Takentogether, users reproduce similar behaviors of focus-
ing their activity on certain tweets, users, and hash-tags across all types of events even though the tweets,users, and hashtags that are the focus of this attentionbecomes more concentrated during media events.
Hashtag Mention Reply RT user
in-d
eg.
NEWS 0.133 0.074 0.040 0.108CONV 0.113 0.280 0.145 0.339DEB 0.019a 0.158 0.162 0.274
out-
deg.
NEWS 0.062 0.060 0.067 0.095CONV 0.066 0.052 0.011b 0.091DEB 0.037 0.083 0.091 0.125
Table 2: Kolmogorov-Smirnov test (K-S test) for comparingthe PRE curves with the remaining three curves in Figure 4.All the statistics D listed here have p-values p < 10−6 un-less reported otherwise: ap = 0.0464, bp = 0.592 (n.s.).
8
in-d
egre
e
(a) Hashtag (b) Mention (c) Reply (d) RT (e) RT user
●●●
●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●●●●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●●●●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●
●●
●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●
●●
●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
out-
degr
ee
(f) Hashtag (g) Mention (h) Reply (i) RT (j) RT user
●●●●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●● ●●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●●●●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●●●
0.2
0.4
0.6
2 4 6 8 10average degree
Gin
i
dbname
●
PRENEWSCONVDEB
●●●
0.2
0.4
0.6
2 4 6 8average degree
Gin
i
dbname
●
PRENEWSCONVDEB
Figure 5: Connectivity-concentration state spaces. For each of the twelve observed events, the Gini coefficient for the network’sdegree distribution is plotted on the y-axis and the average degree of the network is plotted on the x-axis.
Patterns of connectivity and concentration
To unpack the question of whether the behavioralpatterns observed above are driven more by changesacross the system of users (“rising tides”) or changesarising from the concentration of activity (“risingstars”), we examine the relationship between these net-works’ connectivity and concentration. Figure 5 showsthe Gini coefficient and the average degree of users inthe system for each of the activity networks (hashtags,mentions, replies, retweets). System-level changes dur-ing media events should be reflected in the increasedaverage degree as more hashtags, users, and tweetsbecome more well-connected (x-axis). Alternatively,individual-level changes during media events should bereflected in the increased Gini coefficient as users con-centrate their activity around fewer hashtags, users, andtweets (y-axis). Networks in which the users are evenlybut poorly connected would cluster around the lower-left, networks with poor connectivity but high levelsof centralization would cluster in the upper-left, net-works with an even distribution of highly connectednodes would cluster in the lower-right, and networkswith highly connected but nevertheless highly concen-trated activity would cluster in the upper-right.
The absolute position of each network observation in
this space as well as their position relative to networksfor other types of events both warrant discussion. In theuser-to-hashtag network (5a,5f), all twelve events showmoderate amounts of centralization and low levels ofconnectivity. Users do not use many unique hashtagson average (5a) nor are there more hashtags in circula-tion on average (5f) during the media events comparedto other types of events. There is more centralizationamong the hashtags than users, reflecting the domi-nance of official hashtags about events, but the differ-ences between types of events are not major.
In the user-to-user mention network, media eventshave a greater tendency for a few users to receive manymentions from other users than other types of events(5b) but most users still receive fewer than two men-tions during the peak hour for all types of events. Asimilar pattern is found in the user-to-user reply net-work (5c), both of which can be interpreted as changesat the level of the system concentrating activity with-out corresponding changes at the level of the users in-creasing connectivity. Returning to the mention net-work, the mention out-degree spaces show that usersmention between three and five other users on averageover the course of the peak hours for all events (5g)but much of this mentioning behavior in concentratedin a few, very “chatty” users communicating with very
9
many other users. The reply network’s out-degree (5h)has half the average connectivity of the mention net-work’s out-degree, indicating users reply to only twoother users during peaks across and this effect holdsacross all types of events.
Retweet behavior shows major differences betweenevent types that were not observed in the other types ofactivity networks. The in-degrees of retweets for mediaevents in the user-to-retweet network (5d) and user-to-user retweet network (5e) both show limited increasesin average connectivity but extremely large increasesin centralization. In other words, few users and theirtweets “go viral” (the average retweeted tweet is onlyretweeted between one and three times) but there arealso users and tweets that are retweeted so widely, theyaccount for a disproportionate amount of observed ac-tivity. This pattern on media events differs significantlyfrom the pre-debate baseline and news events whichhave low connectivity as well as low centralization anddemonstrates a stronger shift in the behavior of the sys-tem in addition to moderate changes in the behavior ofusers.
The out-degree behavior in the user-to-tweet (5i)and user-to-user (5j) retweet networks also show ma-jor changes for the debates in particular compared tothe baseline and news events. The average user morethan doubled their retweeting activity from approxi-mately 4 retweets and 3 users during a peak hour tobetween 8 and 9 tweets (5i) and 6 and 7 users (5j)during the debates. This retweet activity is consistentlyand highly centralized across all event types suggestingsome users’ focus predominantly on actively retweet-ing other content.
Across these activity types, the in-degrees saw con-sistent patterns of increasing centralization rather thandegree in response to media events while the out-degrees saw patterns of increasing degree rather thanconcentration in response to media events. Taken to-gether, this suggests that while users across the sys-tem become more active during media events, this ad-ditional activity predominately benefits a handful ofusers, hashtags, and tweets.
Changing user responsivenessThe prior sections examined the behavioral changesby aggregating all users together despite the signif-icant differences in their activity. While activity be-came more centralized, it is unclear who benefitted
from this centralization: were users with larger audi-ences the focus of more activity? This section segmentsusers into three classes based upon their audience size:“elites” are in the 90th percentile for number of fol-lowers (≥ 1805), “rookies” are in the 10th percentilefor number of followers (≤ 88), and “typicals” are themiddle 80%.
Based on this segmentation, Figure 6 plots the dis-tributions for several of the activity types related to theconcepts analyzed above. The “tweet adoption index”is the difference in average time to adopt retweet duringthe pre-debate baseline and the average time to retweetduring the debate. High values for tweet adoption in-dex indicate users adopted novel content in the debatemore quickly than they did for the pre-debate baselinewhile low values indicate they adopted novel contentat the same rate. For retweeting behavior, users withfew followers retweeted much more rapidly during thedebates than they did during the baseline and this be-havior decreased as users’ audience size grew (6a).
Analogously, the “user retweet adoption index” isthe difference in time to retweet a user during the pre-debate baseline and the time to retweet the same userin the debate. High values again indicate egos adoptedalters’ content more quickly during the debate than dur-ing normal time while low values indicate there is nodifference in the time to adopt alters’ content. Differ-ences are less pronounced in the distribution of this in-dex and audience size (6d). This suggests media eventslike the debate drive all users, regardless of audiencesize, to seek out and quickly retweet content from usersthey have previously retweeted.
Unsurprisingly, users with more followers are sub-stantially more likely to have their content retweeted(6b) and to be mentioned by other users (6e) during thedebates. Users with hundreds of thousands of followersretweeted less than 10 tweets during the debates whileusers with fewer than 10 followers retweeted more than40 tweets during the debates (6c). While popular usersretweeted less frequently than other types of users, theydid engage in more interpersonal communication thanusers with fewer followers (6f).
Audience size has strong effects across the commu-nication and information sharing activities we exam-ined. These elite users retweet less frequently, adopttweets more slowly, and reply much more often thanother types of users during the peak hours of mediaevents. This suggests elites’ cultivation of large audi-
10
(a) (b) (c)
●● ●
●
●
●
●
●
●
●
●
●
4
8
12
16
10 100 1000 10000 1e+05follower
retw
eet l
ag
group ● elite typicals rookies
●●
●●
●●
●●
●
●
●
1
10
100
100 1000 10000 1e+05 1e+06follower
retw
eet u
ser
in−
degr
ee
group ● elite typicals rookies
● ● ●●
●● ● ● ●
●
●
● ●●
●
10
20
30
40
50
10 100 1000 10000 1e+05follower
user
−R
T o
deg
group ● elite typicals rookies
(d) (e) (f)
●
● ●
●
●
●
●
●
●
10
15
20
10 100 1000 10000follower
lag
retw
eet u
ser
group ● elite typicals rookies
●●
●
●●
●●
●●
●
1
10
100
100 1000 10000 1e+05 1e+06follower
repl
y in
−de
gree
group ● elite typicals rookies
●●
●●
●
●●
●
●
●
●●
●
●
●
4
6
8
10
10 100 1000 10000 1e+05follower
repl
y ou
t−de
gree
group ● elite typicals rookies
Figure 6: Responsiveness during debates. The x-axis plots the number of followers on a log-scale.
ences is highly strategic: they only share some infor-mation with their audiences despite being the focus ofsubstantial amounts of attention. At the other extreme,rookies are profligate and early adopters of novel con-tent despite having little substantive impact on eithertheir audiences.
DiscussionPrevious work examining the dynamics of Twitterusers’ behavior generally assumed the underlyingsocio-technical system was stable and bursts of viralactivity were functions of endogenous popularity dy-namics related to the novelty or recency of the con-tent. However, not all bursts of activity on social me-dia are the same. Major media events like presiden-tial debates or national political conventions are exoge-nous shocks that not only dominate the collective at-tention of Twitter users, but also significantly changethe behavior of users and the system as a whole. Whilewe found that information sharing behaviors like us-ing hashtags or retweeting increased during these me-dia events, interpersonal communication behaviors likementioning and replies decreased in the aggregate. Thislends support to our theory of media-event driven be-
havioral change that predicted users would enter more“highly ordered” states as conditions of shared atten-tion and intense activity displaced the communicationpatterns of their normal social foci.
We examined whether these media events created“rising tides” that changed the behavior similarlyacross the system or if these events created “risingstars” that reinforced the attention and audience foralready-elite users. While these media event-drivenchanges in activity are partially attributable to changesin behavior among the average user, they primarilyreflect an increasing concentration of activity. Refer-ences to hashtags, mentions of users, and retweetsbecame significantly more centralized during mediaevents without correspondingly large changes in the av-erage behavior of users. Crucially, the beneficiaries ofthis newfound attention were not distributed through-out users with different numbers of followers, but con-centrated among users with the largest audiences. De-spite the potential for social media to increase our col-lective communicative capacity by having larger “pub-lic squares” with more diverse voices speaking, occa-sions for large-scale shared attention such as mediaevents appear to replace existing social dynamics with
11
increased collective attention to “rising stars”.The analyses we employed have several limitations
that are opportunities for future work. Our data in-cluded only eight events across a relatively brief six-week period of time on topics related to politics, lim-iting the generalizability of these findings to other do-mains. Future work might explore whether similar pat-terns are found in other types of media events suchas sports (e.g., Super Bowl) and awards ceremonies(e.g., Academy Awards) or across longer spans of timesuch as an entire political campaign. Despite the sizeof cohort of users whose behavior we analyzed and ourintent to capture the behavior of politically-engagedusers, the sampling strategy we employed potentiallyoversampled on users active during the debates. Al-ternative sampling strategies might uncover weaker ordifferent social dynamics. A variety of more advancedmetrics and features such as waiting times and assor-tative degree mixing could be used to analyze socialdynamics in more detail. These analyses also involvedquantitative analyses of aggregated behavioral tracesthat omitted analysis of the content of these tweets andthe motivations of users that could be revealed by par-ticipant observation, interviews, and other qualitativemethods.
ConclusionsThis paper articulated a theory of media-event drivenbehavioral change to explain for why some bursts ofactivity generate different patterns of behavior thanother bursts. By considering not only changes in theoverall level of activity, but changes in the structure ofthe networks of users, tweets, and hashtags, we iden-tified the influence of several processes operating atmicro- and macro- levels. Our findings demonstratethat changes in the aggregate levels of activity duringmedia events are driven more by “rising stars” as eliteusers become the focus of collective attention ratherthan being driven by “rising tides” as users distributetheir attention more broadly to new and diverse voices.These findings have implications for collective actionunder conditions of uncertainty and shared attentionwhile suggesting new ways of mining information fromsocial media to support collective sensemaking follow-ing major events. Social media like Twitter are notonly sites for political communication among politi-cians and their supporters, they are increasingly be-coming spaces for otherwise segmented audiences to
come together in a third space to participate in conse-quential events.
AcknowledgmentsOmitted for review.
References[1] Claudio Castellano, Santo Fortunato, and Vittorio Loreto.
Statistical physics of social dynamics. Rev. of ModernPhysics, 81(2):591, 2009.
[2] F. Ciulla, D. Mocanu, A. Baronchelli, B. Goncalves, N. Perra,and A. Vespignani. Beating the news using social media: thecase study of American Idol. EPJ Data Science, 1(8), 2012.
[3] S. Counts and K. Fisher. Taking it all in? visual attention inmicroblog consumption. In Proc. ICWSM’11, pages 99–104.AAAI, 2011.
[4] R. Crane and D. Sornette. Robust dynamic classes revealedby measuring the response function of a social system. Proc.of the National Academy of Sciences, 105(41):15649–15653,2008.
[5] D. Dayan and E. Katz. Media events: The live broadcastingof history. Harvard University Press, 1992.
[6] N.A. Diakopoulos and D.A. Shamma. Characterizing de-bate performance via aggregated Twitter sentiment. In Proc.CHI’10, pages 1195–1198. ACM, 2010.
[7] Scott L Feld. The focused organization of social ties. Ameri-can Journal of Sociology, 86(5):1015–1035, 1981.
[8] MH Gail and JL Gastwirth. A scale-free goodness-of-fit testfor the exponential distribution based on the gini statistic. J.Royal Statistical Society, pages 350–357, 1978.
[9] Scott A Golder and Michael W Macy. Diurnal and seasonalmood vary with work, sleep, and daylength across diversecultures. Science, 333(6051):1878–1881, 2011.
[10] N.O. Hodas and K. Lerman. How visibility and di-vided attention constrain social contagion. arXiv preprintarXiv:1205.2736, 2012.
[11] Haewoon Kwak, Changhyun Lee, Hosung Park, and SueMoon. What is Twitter, a social network or a news media?In Proc. of WWW’ 10, pages 591–600. ACM, 2010.
[12] K. Lang and G.E. Lang. The unique perspective of televisionand its effect: A pilot study. American Sociological Review,pages 3–12, 1953.
[13] J. Lehmann, B. Goncalves, J.J. Ramasco, and C. Cattuto. Dy-namical classes of collective attention in Twitter. In Proc.WWW’12, pages 251–260. ACM, 2012.
[14] K. Lerman and T. Hogg. Using a model of social dynamicsto predict popularity of news. In Proc. WWW’10, pages 621–630. ACM, 2010.
[15] Yu-Ru Lin, Drew Margolin, Brian Keegan, and David Lazer.Voices of victory: A computational focus group frameworkfor tracking opinion shift in real time. In Proc. WWW’13.WWW, 2013.
12
[16] A. Livne, M.P. Simmons, E. Adar, and L.A. Adamic. Theparty is over here: Structure and content in the 2010 election.In Proc. ICWSM’11, pages 201–208. AAAI, 2011.
[17] J. Nichols, J. Mahmud, and C. Drews. Summarizing sportingevents using Twitter. In Proc. IUI’12, pages 189–198. ACM,2012.
[18] Pew Research Center for the People and the Press. One-in-ten ”dual-screened” the presidential debate. Technical report,2012.
[19] D.M. Romero, B. Meeder, and J. Kleinberg. Differencesin the mechanics of information diffusion across topics: id-ioms, political hashtags, and complex contagion on Twitter.In Proc. WWW’11, pages 695–704. ACM, 2011.
[20] D.A. Shamma, L. Kennedy, and E.F. Churchill. Tweet thedebates: understanding community annotation of uncollectedsources. In Proc. SIGMM’09, pages 3–10. ACM, 2009.
[21] S. Vieweg, A.L. Hughes, K. Starbird, and L. Palen. Mi-croblogging during two natural hazards events: what Twittermay contribute to situational awareness. In Proc. CHI’10,pages 1079–1088. ACM, 2010.
[22] L. Weng, A. Flammini, A. Vespignani, and F. Menczer. Com-petition among memes in a world with limited attention. Sci-entific Reports, 2, 2012.
[23] F. Wu and B.A. Huberman. Novelty and collective attention.Proc. National Academy of Sciences, 104(45):17599–17601,2007.
[24] L. Yang, T. Sun, M. Zhang, and Q. Mei. We know whatyou# tag: does the dual role affect hashtag adoption? In Proc.WWW’12, pages 261–270. ACM, 2012.
13
