Anda di halaman 1dari 8

2016 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM)

An Analysis of Sentiments on Facebook during the


2016 U.S. Presidential Election
Saud Alashri Srinivasa Srivatsav Kandala Vikash Bajaj
CIDSE Decision Theater Network CIDSE
Arizona State University Arizona State University Arizona State University
Tempe, Arizona Tempe, Arizona Tempe, Arizona
salashri@asu.edu sskandal@asu.edu vbajaj2@asu.edu

Roopek Ravi Kendra L. Smith Kevin C. Desouza


CIDSE College of Public Service School of Public Affairs
Arizona State University & Community Solutions Arizona State University
Tempe, Arizona Arizona State University Phoenix, Arizona
rravi5@asu.edu Phoenix, Arizona kev.desouza@gmail.com
klsmit40@asu.edu

Abstract—Social networking sites (SNS), such as Facebook and public thanks to the Internet and social networking sites (SNS).
Twitter, are important spaces for political engagement. SNS have SNS such as Facebook and Twitter provide an opportunity for
become common elements in political participation, campaigns, individuals to learn information, share information, communi-
and elections. However, little is known about the dynamics
between candidate posts and commentator sentiment in response cate with networks, and develop new networks, all of which
to those posts on SNS. This study enriches computational create a new, personalized form of political engagement. The
political science by studying the 2016 U.S. elections and how presidential election of 2008 was the first time in U.S. history
candidates and commentators engage on Facebook. This paper that a candidate (President Obama) utilized social media as
also examines how online activity might be connected to a campaign strategy. In addition to his own campaign web-
offline activity and vice versa. We extracted 9,700 Facebook
posts by five presidential candidates (Hillary Clinton, Donald site (barackobama.com), President Obama used an extensive
Trump, Bernie Sanders, Ted Cruz, and John Kasich) from SNS that included profiles on such platforms as Facebook,
their official Facebook pages and 12,050,595 comments on those Twitter, MySpace, YouTube, Flickr, LinkedIn, Digg, Eons,
posts. We employed topic modeling, sentiment analysis, and BlackPlanet, AsianAve, Faithbase, and several others. He used
trends detection using wavelet transforms to discover topics, those online communities to complement what was happening
trends, and reactions. Our findings suggest that Republican
candidates are more likely to share information on controversial offline, and vice versa [1].
events that have taken place during the election cycle, while President Obama’s success shed light on the possibilities
Democratic candidates focus on social policy issues. As of social media in U.S. elections and set a trend for fu-
expected, commentators on Republican candidate pages express ture candidates to emulate; that trend goes beyond the U.S.
negative sentiments toward current public policies as they Narenda Modi, Prime Minister of India, was the second most
seldom support decisions made by the Obama administration,
while commentators on democratic candidate pages are more “liked” politician on Facebook, only behind President Obama.
likely to express support for continuation or advancement SNS are important for campaigning because of their diverse
of existing policies. However, the significance (strong/weak) user demographics. According to the Pew Research Center
and nature (positive/negative) of sentiments varied between (2014), 74% of Internet users use SNS with little variation
candidates within political parties based on perceived credibility across education, income, gender, and ages 18-49 [2]. SNS
of the candidate’s degree of credibility on a given issue.
Additionally, we explored correlation between online trends of also provide a unique opportunity for politicians to connect
comments/sentiment and offline events. When analyzing the with youth voters; a group that historically has low political
trend patterns, we found that changes in online trends are participation. According to the Pew Research Center (2016),
driven by three factors: 1) popular post, 2) offline debates, and youth voters (or Millennials) were most likely to get their
3) candidates dropping out of the race. political news from SNS [3].
Keywords—Social Networking Sites, Facebook, US Elections, The use of SNS for elections has changed drastically from
Time Series Analysis, Sentiment Analysis, Topic Inference. 2008 to 2016. Unlike the 2008 U.S. elections where President
Obama leveraged SNS as a competitive weapon, during the
I. I NTRODUCTION 2016 elections all candidates have exploited SNS as a critical
The world of politics has progressed to a new level of element of their campaign. The focus of our paper is to
personalized engagement between candidates and the voting investigate the following: 1) variations in activity between
IEEE/ACM ASONAM 2016, August 18-21 2016, San Francisco, CA, USA
978-1-5090-2846-7/16/$31.00 ©2016 IEEE

795
the candidates on Facebook, 2) the dynamics of activity by 230,000 the week before President Mubarak’s resignation with
candidates and their commentators on Facebook, and 3) how the majority of tweets (225,000) being sent from outside of
online activity might be correlated to offline activity and the country [7].
events. Offline user features such as demographic characteristics
We are particularly interested in understanding how candi- like income, gender, and education complement online par-
dates get their message out, how commentators react to them, ticipation. For instance, individuals with higher education and
and what factors explain variations in activities by candidates income attainment are more likely to engage in civic activities
and their commentators on Facebook. due to more knowledge and fewer barriers to participation.
Moreover, we want to examine if commentators’ reactions, Gender accounts for slight differences whereas men are more
in terms of volume and sentiment, track in the same direction. active in politics while women are equally or more active in
To do so, we conducted an analysis of interactivity patterns community activities and grassroots movements [8]. Online
between commentators and the official Facebook pages of the features of participation include offline characteristics such
top five presidential candidates: Hillary R. Clinton (Clinton), as network, size, strength, and diversity that all combine to
Donald Trump (Trump), Bernie Sanders (Sanders), Ted Cruz impact participation. For instance, scholars have found that
(Cruz), and John R. Kasich (Kasich). Internet usage can strengthen existing interpersonal ties as well
Our approach in this paper can be summarized as: topic as create new networks that continue offline [9], [10]. Williams
inference for candidates’ posts, sentiment analysis of com- and Gulati [11] concluded that online activity is an indicator
mentators’ comments, trends detection of oblique cumulative (albeit a very important one) along with other indicators such
curves of comments and sentiment using wavelet transforms. as media coverage, expenditures, and organizing activities for
The rest of the paper is organized as follows. Section II is success.
a literature review on SNS and elections. The methodology Further, the type of SNS used can be an indicator of
of this study is described in Section III. Section IV presents participation. Williams and Gulati [12] note that Facebook is
the experimental results and the qualitative analysis. The final the leading SNS for political campaigning because it offers
section concludes the paper and discusses future work. candidates an additional method of mobilizing voters. Its
ubiquity and high volume of younger users, is particularly
II. L ITERATURE R EVIEW useful and valuable. Further, Facebook is unique because of its
The pervasiveness of SNS in everyday life changes how features such as the “newsfeed” and “wall” which allow users’
the public sphere is conceptualized and utilized [4]. Enli and thoughts and opinions to be displayed for their networks,
Skogerbø [5] contend that SNS enables the personalization of which is likely to spur more interaction [13]. Williams and
politics by which candidates and citizens are able to connect in Gulati [11] assert that the number of Facebook supporters may
new and different ways, which create different forms of politi- be a valid predictor of electoral success.
cal participation. Bennett [6] contends that the personalization Studies that have utilized various methods of analysis on the
of political participation “. . . is importantly channeled through relationship between candidates, election outcomes, and online
often dense social networks over which people can share their activity to understand political support have only arrived at
own stories and concerns through the pervasive use of social tentative findings. Véronis [14] found that counting a candi-
technology which enables individuals to become important date’s mentions can be a better predictor of election success
catalysts of collective action processes as they activate their over election polls. Robertson et al. [15] found in a study
own social networks”(p. 22). of Facebook wall-commentators during the 2008 election that
SNS signal a degree of relevance and importance to the most individuals that posted a comment on a candidate’s page
complementary relationships between online and offline activ- usually only did so once and that President Obama received
ities. The uprising of young Arabs in (mainly) Tunisia and the most attention in political dialogues during the campaign.
Egypt known as the Arab Spring in 2011 is a recent example For individuals that posted on multiple candidates’ Facebook
of SNS working as a conduit for collective action. The uprising wall, they tended to be more verbose, post more often, provide
was the result of many causes steeped in opposition over more outside links, direct comments to others rather than about
political oppression; SNS played a central role in shaping themselves, and refer to the candidates directly and more often.
protests and political debates surrounding the issue through Standberg [16] studied SNS (Twitter and Facebook) in Fin-
the pushing of information to the masses, which eventually land to understand the use of social media in the 2011 election
led to the resignation of President Hosni Mubarak. Howard cycle. The study concluded that differences in social media had
et al. [7] found that the political conversations being had much to do with independent variables such as age, household
about the uprising were happening mostly online and amongst income, gender, political interest, and Internet use. Further,
young, urban, and well-educated individuals who were mostly candidates used social media as an “add-on” to increase
women. Additionally, spikes in online conversations were mobilization. Hong and Nadler [17] found a strong relationship
often preceded by major offline events and that SNS helped to between the numbers of times a politician is mentioned by
spread information across borders. Perhaps one of the biggest traditional media with the number of mentions they get on
signs of SNS saturation is that tweets grew from 2,300 to Twitter. Although they found no statistical significance, they

796
did highlight the potential for a candidates’ Twitter activity to refer to as commentators). The dataset spans a period of
increase the traditional media mentions. time between January 1st , 2015 (when candidates announced
O’Connor et al. [18] used sentiment analysis to measure they will run for presidential race) to February 28th , 2016, a
consumer confidence (consumers’ feeling about health of fourteen month period. We employed python code to collect
economy and their finances) and political opinion over a these posts and comments, removing posts or comments that
year. They found correlations to sentiment word frequencies are not written in English.
in Twitter messages which revealed the potential of utilizing
text streams from social media as a substitute to traditional B. Topic Inference for Candidates Posts
polling. However, there should be some caution exercised To discover topics discussed on candidates posts, we applied
when considering social media as an indicator. Social media Latent Dirichlet Allocation (LDA) [22] on all candidates’
manipulation does happen by candidates, propagandist, or posts. We utilized LDA because it is one of the most popular
interest groups and can be a significant challenge to prediction probabilistic topic inference models. This model assumes each
and understanding [19]. document (post in our case) as a mixture of K latent topics, and
Despite the possibilities for manipulation, SNS have a great each topic is a probability distribution over words. Therefore,
opportunity to help us learn more about public deliberation. a topic is the clustering of co-occurred words together. For
Tumasjan et al. [20] studied the 2009 German federal election instance, the topic Planned Parenthood is described by co-
to investigate Twitter usage and whether online messages occurred words (e.g., Planned, Parenthood, Women, defund,
match offline public sentiment. They found that the number God) We utilized Gensim LDA implementation [23] to infer
of mentions on Twitter reflect election results. Gibson and latent topics from posts and then use the resultant model to
McAllister [21] found in the 2004 Australian national elections assign each post to its most probable topic.
that the online component of a campaign had a strong positive
C. Sentiment Analysis of Commentators
impact on candidate’s share of the votes and was integral for
victory. They concluded that the web is a major vote attractor We conducted sentiment analysis on comments to determine
for candidates but questioned what aspects of a candidate’s whether reactions to posts were positive, negative, or neutral.
web presence actually converts interest into votes. Effing et We pre-processed comments to remove those that are not
al. [1] studied social media use in the Netherlands and found written in English. Also, any comment that had a repetitious
that during the 2010 national elections, candidates with higher pattern was considered as spam and was removed, as an exam-
social media engagement got the most votes but this was not ple (VOTE FOR SANDERS, VOTE FOR SANDERS). Next,
the same at the local level in 2010/2011. we utilized Stanford CoreNLP [24] as an open-source tool
The unique contribution of this paper can be found in 1) for calculating sentiment score. This is a well-known tool for
the study of the 2016 U.S. presidential elections where social analyzing sentiment, and it generates sentiment scores from 0
media use is at all time high and all candidates are active to 4, with 0 being the most negative, 4 being the most positive,
on SNS and, 2) the examination of alignment between offline and 2 being neutral. Next, we studied comment sentiment as
events and online activities. Moreover, we hope to contribute oblique cumulative curves over time (next section) as well as
to knowledge on computational political science both with sentiment of inferred topics per candidate.
our findings as well as our methodological approach (e.g. D. Oblique Cumulative Curves and Trends Detection
the use of topic modeling, sentiment analysis, and wavelet
transforms). To understand the correlation between key offline events and
online activities, we studied the trends of online activities such
III. M ETHODOLOGY as comment volume and sentiment. We also studied the raw
Our approach can be summarized as follows: time series representation of comment volume and sentiment
• Data extraction from candidates Facebook pages.
that is noisy and hard to understand. Thus, we utilized the
• Infer latent topics from candidates posts and assign each
following techniques to identify clear key changes in trends:
post to its topic. • Oblique cumulative curves to show more clean and clear

• Analyze sentiment of comments on posts to understand trends compared to the raw ones.
commentator reactions to different topics. • Wavelet transforms to de-noise the data and precisely

• Study the correlation between key offline events and identify the key events trends in resultant oblique curves
online activities using trends detection algorithm Wavelet of comments volume and sentiment.
transforms (WT) to reveal changes in online activities. Oblique cumulative curve techniques are used in the field of
Detailed description of each bullet is presented next. transportation engineering research to track changes in traffic,
speed and traffic flow respectively [25], [26]. Cumulative curve
A. Dataset showing the cumulative counts of comments C(t) and their
Our dataset is comprised of 9,700 posts by the five presi- sentiment at time t is constructed. An oblique cumulative curve
dential candidates: Clinton, Trump, Sanders, Cruz, and Kasich, is then constructed by deducting the background cumulative
and 12,050,595 comments from the public (who we will count co × t∗ at all time t, where t∗ is the elapsed time from

797
Fig. 1: Distribution of Topics

to to t, and co is the average comments per day. The oblique TABLE I: The Two Categories of Topics
cumulative count at time t is defined by the formula: Election Events Policies
FOX Election Coverage Iran Deal, ISIS and Gun Control
Primaries Planned Parenthood
O(t) = C(t) − co × (to − t) (1)
Election Debates Immigration and Obamacare
Caucus Women Rights and Education
Wavelet transforms (WT) are used to de-noise the oblique Election Events Health Care and Social Security
cumulative curve and precisely identify key events during the Campaigning Taxes and Federal Spending
study period where the users have strongly responded to the Electability Wall Street and Middle Class
candidate’s posts. These events are marked by the changes in Campaign Support and Fund raising Climate Change and Income Inequality
Election Voting
the signal (peaks and dips). WT is a powerful spectral analysis
tool used to analyze non-stationary signals as it can provide
both time and frequency information of the signal unlike
Fourier transforms which only give the frequency information.
p
T = 2 × log(n) × mad(ϕϕ
x) (3)
In this study, discrete wavelet transforms (DWT) are used to
de-noise the raw signal and defined by the formula: Where n is the sample size, mad is the median absolute
deviation and ϕϕx is the first level wavelet coefficients. Matlab
t−τ
Z
ϕ 1 version of Simple Peak Finding algorithm was then used to
ϕx (s, τ ) = p O(t)ϕ( )dt (2)
|s| s detect key dates accurately.
IV. E XPERIMENTAL R ESULTS
Where s ∈ 2j , τ = k × s, (j,k) ∈ Z 2 (s6= 0)
A. Resultant Latent Topics
Wavelet coefficients are computed at discrete values of scale We ran LDA on all candidates’ posts. We experimented
s and time t. The objective here is to obtain de-noised signal with different numbers of topics and found K=20, i.e., 20
by eliminating noise in the raw signal which is assumed to topics was the best number in terms of meaningful topics. Each
be Gaussian distributed. This is done by decomposing the raw topic is described by ten different keywords. There are three
signal (in this case, oblique cumulative curve) into different topics that did not show any relevant information; thus we
frequency components and removing the noise from the high considered them as noise and removed them. The distribution
frequency part of signal by applying certain thresholds on the of the remaining 17 topics is presented in Fig. 1, where the X
wavelet coefficients and then reconstruct the de-noised signal axis represents the topics and Y axis represents the number of
using inverse transforms. More details on this methodology posts associated with the topic. There are two main categories
can be found in [27]. In our study noise is estimated using of the resultant topics: Election Events related topics with
a threshold which is based on median absolute deviation as 4,771 posts and Policies related topics with 3,668 posts. Table
shown below: I shows the topics of each category.

798
Fig. 2: Distribution of Topics per Candidates

Among the most discussed topics in Election Events cat-


egory is the Fox News election coverage where Trump com-
plained that the channel treated him badly and Megan Kelly
insulted him at the first GOP debate on August 3rd , 2015.
His boycotting of Fox News’ hosted sparked discussions as
he continued to blame Fox for leaning toward Cruz, especially
on the Bill O’Reilly TV program. The most discussed topic in
the Policies category is national security in terms of Iran deal,
ISIS, and gun control. Within Policies category there are other
important topics such Planned Parenthood, women’s rights,
education, and taxes/federal spending. Unlike other elections
in 2008 and 2012, the least discussed topics on this election
are climate change and income inequality. Our focus in this
study is Policies related topics.
Fig. 3: Sentiment and Comments Volume Curves for Cruz
The distribution of posts by topic for each candidate is
presented in Fig. 2. The numbers on each topic band are the
total number of posts for that topic, and numbers on each book posts; Cruz discussed topics immigration, Obamacare,
candidate band is his/her total number of posts. Both Demo- Planned Parenthood, the Iran Deal, ISIS, and gun control, and
cratic candidates focus on Healthcare and Social Security; Kasich mainly discussed taxes and federal spending.
Clinton also discussed the topics relating to women’s rights
and education on her Facebook posts while Sanders discussed B. Oblique Cumulative Curves and Trends Detection
the topics of Wall Street and the middle class. Conversely, the Fig. 3 shows the raw comments and sentiment data for pres-
Republican candidates focused on various topics on their Face- idential candidate Cruz. Due to its noisy nature, it is difficult

799
(the peak on August 4th , 2015), his cumulative sentiment curve
dropped due to his poor performance in the debate. He gained
high average cumulative positive sentiment of 11.01 when
the Republican presidential candidate George Pataki dropped
out of the race on January, 2nd 2016. Fig 5 presents Sanders
key trends identified by wavelet transforms. When he posted
his announcement (April 30th , 2015) about his campaign for
president, he received higher comments volume of 42,513
and more positive sentiment (cumulative average of 4.38).
He gained slightly more positive sentiment when Democrats
presidential candidates Jim Webb and Lincoln Chafee dropped
out of the race on October 19th and 24th , 2015. Similarly,
he received higher comments (36,303) and positive sentiment
(5.25) after the Jewish Coalition Forum on December 4th ,
2015.

C. Sentiment Analysis of Reactions


Fig. 4: De-noised Oblique Curves with Identified Trends for
Cruz We further analyzed topics distribution and reactions of
commentators on each candidate’s post in terms of senti-
ment, as given in Fig. 6 (next page). Republican candidates’
campaigns have focused on four major topics: “Iran deal,
ISIS, gun control”, “Planned Parenthood”, “Immigration and
Obamacare”, and “taxes and federal spending”. For all three
candidates, the reactions on the Iran deal, ISIS and gun control
are negative which reflect the commentators’ outrage on the
topic. On the topic of Planned Parenthood, all candidates have
positive reactions as commentators supported their candidate
whether they were for or against the proposed plan. On
immigration and Obamacare, both Cruz and Trump proposed
to stop illegal immigrants and build a wall along U.S.-Mexico
border. Such posts sparked anger from commentators, whether
it was anger on the proposed plan or anger on the immigrants
and other minorities in general. On taxes and federal spending,
Trump posted a few posts on the topic and received positive
reactions; whereas, Kasich was considerably more active on
this topic and received noticeable negative reaction.
Fig. 5: De-noised Oblique Curves with Identified Trends for Democratic candidates focused on three major topics
Sanders namely: “Women rights and Education”, “Wall street and
Middle class”, and “Healthcare and Social security”. In Fig.
7 (last page), Sanders is active on issues that pertain to
to understand trends. Therefore, we constructed cumulative reigning in Wall Street and growing the middle class, for which
oblique curves, presented in Fig 4 for Cruz. In that figure, he received very positive comments. However, he received
curves were de-noised using wavelet transforms and key events negative comments on his next major topic healthcare and
trends were identified as well. Key events trends represent Social Security. Clinton is active on women’s rights and
peaks in sentiment (positive) and comment volume are shown education topics and her commentators reacted positively on
by circles (dots). Due to space limitation, we included the her posts. Clinton is also active on healthcare/Social Security.
analysis of Cruz and Sanders only. She supported the Affordable Care Act and promised to
When analyzing the patterns of trends for both candidates, enhance and protect Social Security and Medicare.
we noticed that changes in trends are driven by three factors: By studying sentiments and the level of positive or negative
1) popular post, 2) offline debate, and 3) candidate’s dropping emotions, we can get estimates of how commentators view
out. Cruz’s popular post on April 10th , 2015 about his support a given candidate’s ability to lead on a given issue. For
for soldiers and the Army caused his cumulative curve of example Sanders clearly has a lead on Wall Street and middle
comments volume to grow and reached a total of 41,949 class issue as estimated by sentiment scores, and if one
comments on that day. However, after the first GOP debate compared the Republican candidates on issues such as Planned

800
Fig. 6: Sentiment Analysis on Topics for Republican Presidential Candidates

Parenthood, Cruz fares better versus Trump’s taxes and federal on Twitter is most often quoted in the media. Second, our
spending plan which gets a lot more positive sentiment. analysis, while insightful, has the benefit of hindsight. Further
work is required to build real-time predictive models that can
V. C ONCLUSION A ND F UTURE W ORK
infer reactions to online activities based on posts by candidates
This study examined the dynamics between candidate posts and offline activities.
and comments they received on Facebook. The paper also We intend to expand on this study to build predictive
explored how online activity might be connected to offline models focused on understanding the effects of online cam-
events. Republican candidates have focused on four major paign activity. Moreover, we can analyze the data to study
topics: “Iran ISIS, gun control”, “Planned Parenthood”, “im- other interesting questions that impact the design of social
migration and Obamacare”, and “taxes and federal spending”. media campaigns. Furthermore, we will study the evolutionary
Democratic candidates have focused on three major topics development of each candidate networks of interactions to
namely: “women’s rights and education”, “Wall Street and understand patterns of growth and scatteredness.
middle class”, and “healthcare and social security”. Detecting
trends in comments and sentiment analysis revealed the corre- R EFERENCES
lation between offline events and candidates’ posts. When an- [1] R. Effing, J. van Hillegersberg, and T. Huibers, “Social media and
alyzing sentiment of commentators on Republican candidates’ political participation: are facebook, twitter and youtube democratizing
posts related to Iran deal, ISIS, and gun control topics, we our political systems?” in Electronic participation. Springer, 2011, pp.
25–35.
found that they reacted with angry comments represented by [2] “Social media user demographics,” http://www.pewinternet.org/data-
negative sentiment. For Democratic candidates, commentators trend/social-media/social-media-user-demographics/, 2014, accessed:
on Wall Street/middle class topics reacted positively, showing 2016-05-20.
[3] J. Gotttfried and M. Barthel, “Among millennials engaged in primaries,
their support to the proposed reform plan. However, the dems more likely to learn about the election from social media,”
significance (strong/weak) and nature (positive/negative) of http://www.pewresearch.org/fact-tank/2016/02/09/among-millennials-
sentiments varied between candidates within political parties engaged-in-primaries-dems-more-likely-to-learn-about-the-election-
from-social-media/, 2016, accessed: 2016-05-20.
based on perceived credibility of the candidate’s degree of [4] S. P. Robertson and R. K. Vatrapu, “Digital government,” Annual review
credibility on a given issue. We employed topic modeling, sen- of information science and technology, vol. 44, no. 1, pp. 317–364, 2010.
timent analysis, and trends detection using wavelet transforms [5] G. S. Enli and E. Skogerbø, “Personalized campaigns in party-centred
politics: Twitter and facebook as arenas for political communication,”
to discover topics, trends, and reactions Information, Communication & Society, vol. 16, no. 5, pp. 757–774,
This study has several limitations that need to be acknowl- 2013.
edged. First, we only analyzed one platform and so it is not [6] W. L. Bennett, “The personalization of politics political identity, social
media, and changing patterns of participation,” The ANNALS of the
fully reflective of the entire social media campaign for the American Academy of Political and Social Science, vol. 644, no. 1,
candidates. Anecdotally we know from the press that activity pp. 20–39, 2012.

801
Fig. 7: Sentiment Analysis on Topics for Democratic Presidential Candidates

[7] P. N. Howard, A. Duffy, D. Freelon, M. M. Hussain, W. Mari, and [21] R. K. Gibson and I. McAllister, “Does cyber-campaigning win votes?
M. Mazaid, “Opening closed regimes: what was the role of social media online communication in the 2004 australian election,” Journal of
during the arab spring?” Available at SSRN 2595096, 2011. Elections, Public Opinion and Parties, vol. 16, no. 3, pp. 243–263, 2006.
[8] H. G. de Zúñiga and S. Valenzuela, “The mediating path to a stronger [22] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent dirichlet allocation,”
citizenship: Online and offline networks, weak ties, and civic engage- J. Mach. Learn. Res., vol. 3, pp. 993–1022, Mar. 2003. [Online].
ment,” Communication Research, vol. 38, no. 3, pp. 397–421, 2011. Available: http://dl.acm.org/citation.cfm?id=944919.944937
[9] A. Kavanaugh, J. M. Carroll, M. B. Rosson, T. T. Zin, and D. D. [23] P. Sojka, “Software framework for topic modelling with large corpora,”
Reese, “Community networks: Where offline communities meet online,” in In Proceedings of the LREC 2010 Workshop on New Challenges for
Journal of Computer-Mediated Communication, vol. 10, no. 4, pp. 00– NLP Frameworks. Citeseer, 2010.
00, 2005. [24] R. Socher, A. Perelygin, J. Y. Wu, J. Chuang, C. D. Manning, A. Y. Ng,
[10] B. Wellman, “An electronic group is virtually a social network,” Culture and C. Potts, “Recursive deep models for semantic compositionality over
of the Internet, vol. 4, pp. 179–205, 1997. a sentiment treebank,” in Proceedings of the conference on empirical
methods in natural language processing (EMNLP), vol. 1631. Citeseer,
[11] C. Williams and G. Gulati, “What is a social network worth? facebook
2013, p. 1642.
and vote share in the 2008 presidential primaries.” American Political
[25] M. J. Cassidy and R. L. Bertini, “Some traffic features at freeway
Science Association, 2008.
bottlenecks,” Transportation Research Part B: Methodological, vol. 33,
[12] C. B. Williams and G. J. Gulati, “Facebook grows up: An empirical no. 1, pp. 25–42, 1999.
assessment of its role in the 2008 congressional elections,” Proceedings [26] ——, “Observations at a freeway bottleneck,” Transportation and Traffic
from Midwest Political Science Association, Chicago, vol. 32, p. 53, Theory, pp. 107–146, 1999.
2009. [27] S. S. Kandala, “Analysis of freeway bottlenecks,” Ph.D. dissertation,
[13] J. E. Carlisle and R. C. Patton, “Is social media changing how we Arizona State University, 2014.
understand political engagement? an analysis of facebook and the 2008
presidential election,” Political Research Quarterly, vol. 66, no. 4, pp.
883–895, 2013.
[14] J. Véronis, “Citations dans la presse et résultats du premier tour de la
présidentielle 2007,” Retrieved December, vol. 15, p. 2009, 2007.
[15] S. P. Robertson, R. K. Vatrapu, and R. Medina, “Off the wall political
discourse: Facebook use in the 2008 us presidential election,” Informa-
tion Polity, vol. 15, no. 1, 2, pp. 11–31, 2010.
[16] K. Strandberg, “A social media revolution or just a case of history
repeating itself? the use of social media in the 2011 finnish parliamentary
elections,” New Media & Society, p. 1461444812470612, 2013.
[17] S. Hong and D. Nadler, “Which candidates do the public discuss online
in an election campaign?: The use of social media by 2012 presidential
candidates and its impact on candidate salience,” Government Informa-
tion Quarterly, vol. 29, no. 4, pp. 455–461, 2012.
[18] B. O’Connor, R. Balasubramanyan, B. R. Routledge, and N. A. Smith,
“From tweets to polls: Linking text sentiment to public opinion time
series.” ICWSM, vol. 11, no. 122-129, pp. 1–2, 2010.
[19] T. Metaxas and E. Mustafaraj, “Social media and the elections,” 2012.
[20] A. Tumasjan, T. O. Sprenger, P. G. Sandner, and I. M. Welpe, “Predict-
ing elections with twitter: What 140 characters reveal about political
sentiment.” ICWSM, vol. 10, pp. 178–185, 2010.

802

Anda mungkin juga menyukai