CHARACTERIZING POLITICALLY ENGAGED USERS' BEHAVIOR DURING THE 2016 US PRESIDENTIAL CAMPAIGN Josemar Alves Caetano, Jussara Almeida, Humberto Torres Marques-Neto
ASONAM 2018 2 Social networks and political campaings
ASONAM 2018 3 Reach of the candidates on Twitter (election day) 17 million followers 12 million followers 35 thousand published tweets 9 thousand published tweets
ASONAM 2018 4 Political biases on social networks
ASONAM 2018 5 Political biases on social networks Advocates
ASONAM 2018 6 Other political groups Political Bots Regular Users
ASONAM 2018 7 Main objective Characterize users in an online social network taking into account political biases and therefore different behaviors
ASONAM 2018 8 Characterizations Which features Language Patterns Popular users Mood Variation of each group Analysis highlight each group Analysis
ASONAM 2018 9 Feature characterization Which features Language Patterns Popular users Mood Variation of each group Analysis highlight each group Analysis
ASONAM 2018 10 Language characterization Which features Language Patterns Popular users Mood Variation of each group Analysis highlight each group Analysis
ASONAM 2018 11 Profile characterization Which features Language Patterns Popular users Mood Variation of each group Analysis highlight each group Analysis
ASONAM 2018 12 Mood characterization Which features Language Patterns Popular users Mood Variation of each group Analysis highlight each group Analysis
ASONAM 2018 13 To perform these characterizations… Which features Language Patterns Popular users Mood Variation of each group Analysis highlight each group Analysis
ASONAM 2018 14 Methodology Identifying Identifying Tweet Mood Collecting Politically Political Sentiment Variation Twitter Data Engaged Tweets Analysis Analysis User Groups
ASONAM 2018 15 Collecting Twitter data Identifying Identifying Tweet Mood Collecting Politically Political Sentiment Variation Twitter Data Engaged Tweets Analysis Analysis User Groups
ASONAM 2018 16 Data collection process
ASONAM 2018 17 Data collection period • Data collected over 122 days ( August 1 st to November 30 th 2016) August 1 st October 9 th November 8 th Data Second collection televised Election day start debate October September 19 th November 30 th 26 th Second Data collection First televised televised end debate debate
ASONAM 2018 18 Dataset # of tweets 23 mi # of users 115 k # of relationships 1.8 mi
ASONAM 2018 19 Identifying political tweets Identifying Identifying Tweet Mood Collecting Politically Political Sentiment Variation Twitter Data Engaged Tweets Analysis Analysis User Groups
ASONAM 2018 20 Candidates references considered Donald Trump Hillary Clinton @realDonaldTrump @HillaryClinton Trump Hillary DT HC
ASONAM 2018 21 Political hashtags Donald Trump Hillary Clinton 1 #Trump #ImWithHer 2 #MAGA #NeverTrump 3 #TrumpTrain #Hillary 4 #TrumpPence16 #HillaryClinton 5 #DrainTheSwamp #Hillary2016 6 #tcot #UniteBlue 7 #Trump2016 #VoteBlue 8 #GOP #HillaryBecause 9 #PJNET #OHHillYes 10 #cco #HillYes
ASONAM 2018 22 Tweet sentiment analysis Identifying Identifying Tweet Mood Collecting Politically Political Sentiment Variation Twitter Data Engaged Tweets Analysis Analysis User Groups
ASONAM 2018 23 How sentiment analysis works? • SentiStrength tool • Dictionary containing emotional words -5 -1 1 5
ASONAM 2018 24 Political sentiment analysis I love Hillary Clinton and her ideas. I hate Hillary Clinton and her ideas.
ASONAM 2018 25 Political sentiment analysis problem I love Hillary Clinton but I hate Donald Trump. I hate Hillary Clinton but I love Donald Trump.
ASONAM 2018 26 Sentiment analysis approaches Non-political Whole text Tweets Sentiment Analysis Tweets about Whole text one candidate Political Tweets Tweets about Words both associated with candidates each candidate
ASONAM 2018 27 Non-political tweets Non-political Whole text Tweets Sentiment Analysis Tweets about Whole text one candidate Political Tweets Tweets about Words both associated with candidates each candidate
ASONAM 2018 28 Political tweets about one candidate Non-political Whole text Tweets Sentiment Analysis Tweets about Whole text one candidate Political Tweets Tweets about Words both associated with candidates each candidate
ASONAM 2018 29 Political tweets about both candidates Non-political Whole text Tweets Sentiment Analysis Tweets about Whole text one candidate Political Tweets Tweets about Words both associated with candidates each candidate
ASONAM 2018 30 Identifying words related to candidates • Stanford Parser tool • Natural language processor
ASONAM 2018 31 Identifying politically engaged user groups Identifying Identifying Tweet Mood Collecting Politically Political Sentiment Variation Twitter Data Engaged Tweets Analysis Analysis User Groups Hillary’s Trump’s Political Regular Advocates Advocates Bots Users
ASONAM 2018 32 Data mining process Features
ASONAM 2018 33 Removing outliers Features Eliminated users that did not have political tweets
ASONAM 2018 34 Identifying political bots Features BotOrNot Users with BotOrNot score >= 0.75
ASONAM 2018 35 Feature set engineering Features 44 Features Syntax Political Bias Sentiment Analysis User Metadata 6 10 11 17
ASONAM 2018 36 Identifying Regular Users, Trump’s Advocates, and Hillary’s Advocates Features Silhouette K-Means Index 3 1 Greedy 2 Selection
ASONAM 2018 37 Two steps clustering Non-political Bots Regular Advocates Users Hillary’s Trump’s Advocates Advocates
ASONAM 2018 38 Identifying regular users and advocates Non-political Bots Regular Advocates Users Hillary’s Trump’s Advocates Advocates
ASONAM 2018 39 Identifying Trump’s Advocates and Hillary’s Advocates Non-political Bots Regular Advocates Users Hillary’s Trump’s Advocates Advocates
ASONAM 2018 40 Mood variation analysis Identifying Identifying Tweet Mood Collecting Politically Political Sentiment Variation Twitter Data Engaged Tweets Analysis Analysis User Groups Tweet Tweet Tweet Retweet Tweet Tweet Tweet 𝑻 𝒄 𝒗 𝑻 𝒃 𝒗
Graduate Program in Informatics 41 Subjective Well-Being definition 𝑂 𝑞 𝑣 𝑢 1 ,𝑢 2 − 𝑂 𝑜 𝑣 𝑢 1 ,𝑢 2 𝑇 𝑣 𝑢 1 , 𝑢 2 = 𝑂 𝑞𝑣 𝑢 1 ,𝑢 2 + 𝑂 𝑜𝑣 𝑢 1 ,𝑢 2 • 𝑂 𝑞 𝑣 𝑢 1 , 𝑢 2 : positive tweets total • 𝑂 𝑜 𝑣 𝑢 1 , 𝑢 2 : negative tweets total • 𝑇 𝑣 𝑢 1 , 𝑢 2 : −1 ≤ 𝑇 𝑣 ≤ 1
Graduate Program in Informatics 42 Mood variation definition ∆𝑇 𝑣 = 𝑇 𝑣 𝑢, 𝑢 + 𝜀 - 𝑇 𝑣 𝑢, 𝑢 − 𝜀 • 𝑇 𝑣 𝑢, 𝑢 + 𝜀 : SWB after retweet • 𝑇 𝑣 𝑢, 𝑢 − 𝜀 : SWB before retweet • ∆ 𝑇 𝑣 values: −2 ≤ ∆ 𝑇 𝑣 ≤ 2
ASONAM 2018 43 What does it mean? 2 1 -2 -1 0
ASONAM 2018 44 Results Language Patterns Which features Popular users of Mood Variation Analysis highlight each group each group Analysis
ASONAM 2018 45 Clustering Regular Users and Advocates Regular Users Advocates (70,290) (40,003) σ σ µ µ political discourse 0.0871 0.4083 0.4614 1.5802 avg number of political hashtags -0.0005 0.0088 -0.0080 0.0297 related to Trump per tweet avg number of political hashtags -0.0066 0.0141 -0.0318 0.0385 related to Hillary per tweet positive/negative bias towards Trump 0.0759 0.0617 0.3431 0.1050 positive/negative bias towards Hillary 0.0833 0.4276 0.6592 2.1534 Sillhouette index: 0.81
ASONAM 2018 46 Clustering Hillary’s Advocates and Trump’s Advocates Hillary’s Trump’s Advocates Advocates (26,230) (13,733) σ σ µ µ # hashtags in user's description 0.4030 0.1494 0.3516 0.1886 avg number of words per tweet 0.2787 0.1934 0.3429 0.1961 % tweets with some reference to Trump 0.5578 0.2349 0.7702 0.2532 % tweets with some reference to Hillary 0.8355 0.1864 0.6504 0.2624 std of the sentiment score of tweets with 3.7241 4.8296 7.8192 5.2273 some reference to Trump std of the sentiment score of tweets with 0.4692 1.4341 0.7009 1.7545 some reference to Hillary Sillhouette index: 0.72
ASONAM 2018 47 Language patterns Hillary’s Advocates Trump’s Advocates Political Bots Regular Users
ASONAM 2018 48 Top 5 Hillary’s Advocates 1 2 3 4 5
ASONAM 2018 49 Top 5 Trump’s Advocates 1 2 3 4 5
ASONAM 2018 50 Top 5 Political Bots Twitter suspended the top 10 Political Bots accounts
ASONAM 2018 51 Top 5 Regular Users 1 2 3 4 5
ASONAM 2018 52 Mood variation – Hillary’s tweets
ASONAM 2018 53 Mood variation – Trump’s tweets
Recommend
More recommend
