Studying the Dark Triad of Personality through Twitter Behavior - PowerPoint PPT Presentation

Studying the Dark Triad of Personality through Twitter Behavior Daniel Preot ¸iuc-Pietro Jordan Carpenter, Salvatore Giorgi, Lyle Ungar Positive Psychology Center Computer and Information Science University of Pennsylvania October 26, 2016

Motivation Online spaces are a medium for self-expression and social communication. There is a concern that these o ff er a medium for expressing darker traits of human personality such as: ◮ Self-promotion ◮ Vanity ◮ Anti-social behavior ◮ Alteration of the truth ◮ Self-interest

The Dark Triad The standard model in psychology for malevolent human personality traits . ◮ Coined in (Paulhus & Williams, 2002) Assessed through questionnaires. ◮ Similar to the ‘Big Five’ personality traits Psychological studies on self-reported behaviors, not data-driven exploration. ◮ Social media o ff ers a unique window into how people that demonstrate these behaviors think and act

User Profiling User profiling automatically quantifying traits from a user’s online footprints: ◮ Text ◮ Images ◮ Platform usage ◮ Likes ◮ Social network ◮ ...

User Profiling Two sides of the problem: 1. Measurement ◮ Goal: build models to predict traits of unknown users ◮ Predictive setup (regression / classification) ◮ Using large scale Machine Learning 2. Insight ◮ Goal: gain a better understanding of group di ff erences ◮ Interpretable features ◮ Use domain experts in analysis

Narcissism Narcissism: ◮ Vanity ◮ Entitlement ◮ Self-su ffi ciency ◮ Superiority ◮ Authority ◮ Exhibitionism ◮ Exploitativeness Sample Items: ◮ I tend to want others to admire me. ◮ I tend to expect special favors from others.

Narcissism Miranda Priestly – The Devil Wears Prada

Psychopathy Psychopathy: ◮ Lack of remorse ◮ Lack of empathy ◮ Pathological lying ◮ Need for stimulation ◮ Superficial charm ◮ Grandiose self-worth Sample Items: ◮ I tend to lack remorse. ◮ I tend to not be too concerned with morality or the morality of my actions.

Psychopathy Anton Chigurh – No Country for Old Men

Machiavellianism Machiavellianism: ◮ Duplicitous ◮ Ends justify the means ◮ Rarely reveal their true intentions ◮ Manipulate to get ahead ◮ Money and power over relationships ◮ Flattery ◮ Cynical view of human nature Sample Items: ◮ I have used deceit or lied to get my way. ◮ I tend to exploit others towards my own end.

Machiavellianism Frank Underwood – House of Cards

Data Set Collected through a study on Amazon Mechanical Turk. 863 Twitter users with public profiles. 491 Twitter users posted > 500 tokens. Collected all their tweets ( < 3200), their profile picture and profile information.

Dark Triad Score Completed the ’Dirty Dozen’ questionnaire: ◮ 12 questions; ◮ 1–5 scale; ◮ 4 questions / trait. Reported age and gender. We use the log of the traits for the rest of the experiments.

Trait Inter-Correlation ◮ Treats are moderately inter-correlated – as expected; ◮ We compute an additional ‘Dark Triad’ score as the average of the three in accordance to previous work; ◮ In our analysis of each trait, we control for the other two traits in addition to age and gender using partial correlation to isolate distinctive behaviors.

Features – Text ◮ Unigrams: ◮ Single tokens used by at least 10% of users (N = 6,491) ◮ LIWC: ◮ Manually constructed word categories (Pennebaker et al, 2001) ◮ Include parts-of-speech, topical categories, emotions (N = 64) ◮ Topics: ◮ Obtained by using spectral clustering over word2vec word representations (Preot ¸iuc-Pietro et al, 2015) ◮ Words that appear in similar contexts (N = 200) ◮ Sentiment & Emotions: ◮ Messages tagged with either sentiment or discrete emotions (Mohammad et al. 2010) ◮ Each user is assigned its average message emotion scores (N = 10)

Features – Profile Image ◮ Color features: ◮ Grayscale, Brightness, Contrast, Saturation, Sharpness, Blur ◮ Facial features: ◮ Type of image: default, # faces, one face, multiple faces (Face ++ ) ◮ Facial presentation: ratio, glasses, posture, smile

Features – Platform Usage ◮ Profile features: ◮ No. tweets, tweets / day ◮ # friends, #followers, follower–friend ratio, #listed ◮ Default background, geo-enabled ◮ Proportion and count of tweets that were retweeted or liked ◮ Shallow features: ◮ # characters, # tokens per tweet ◮ Retweets or duplicate messages ◮ Proportion of messages which use hashtags, @-replies, @-mentions, URLs or ask for followers

‘Core’ Dark Triad Word2Vec Topics R = .126 R = .126 R = .117 R = .152 Posting about work and addresses. Topics significant at p < .01 (two-tailed t-test), controlled for Age and Gender.

‘Core’ Dark Triad LIWC Categories SWEAR PRESENT SPACE ANGER R = .127 R = .106 R = .123 R = .119 Related to present activities. Topics significant at p < .01 (two-tailed t-test), controlled for Age and Gender.

‘Core’ Dark Triad Emotions Trust Negative Disgust R = .093 R = .108 R = .102 Overall negative emotions, but also trust. Topics significant at p < .01 (two-tailed t-test), controlled for Age and Gender.

‘Core’ Dark Triad Image: ◮ less likely to be Grayscale ◮ lower sharpness Profile: ◮ – Shallow: ◮ Fewer characters per tweet ◮ Fewer retweets performed ◮ Fewer tweets with hashtags and URLs All correlations significant at p < .05; controlled for age and gender.

Narcissism Word2Vec Topics R = .110 R = .111 R = .104 R = .119 Positive face to the world. Support causes, celebrities, TV shows. Post about their mundane activities on Twitter (which they think others are interested in). Topics significant at p < .01 (two-tailed t-test), controlled for Age, Gender, Machiavellianism and Psychopathy.

Narcissism Emotions R = .130 R = .104 Trust Positive Positive face to the world. Positive emotions overlap in most frequent words. Topics significant at p < .01 (two-tailed t-test), controlled for Age, Gender, Machiavellianism and Psychopathy.

Narcissism Image: ◮ Not grayscale ◮ Prefer one face in profile image and not multiple faces ◮ Smiling Profile: ◮ Not default background ◮ Geo-enabled ◮ More tweets that are favorited Shallow: ◮ Fewer duplicate tweets (content curation) ◮ Less tweets with hashtags and @-mentions All correlations significant at p < .05; controlled for age, gender, psychopathy and Machiavellianism.

Psychopathy Word2Vec Topics R = .144 R = .123 R = .123 R = .142 R = .110 R = .116 R = .108 R = .110 Interested in news about violent activities and news (including ‘Positive’ aggression), emergencies, issues.

Psychopathy LIWC Categories R = .153 R = .138 R = .101 R = .110 DEATH ANGER BODY NEGEMO Topics significant at p < .01 (two-tailed t-test), controlled for Age, Gender, Machiavellianism and Psychopathy.

Psychopathy Emotions R = .189 R = .177 R = .174 R = .173 Negative Fear Disgust Anger The entire spectrum of negative emotions. Topics significant at p < .01 (two-tailed t-test), controlled for Age, Gender, Machiavellianism and Psychopathy.

Psychopathy Image: ◮ Less saturated Profile: ◮ – Shallow: ◮ Fewer URLs ◮ Not asking for followers All correlations significant at p < .05; controlled for age, gender, Machiavellianism and narcissism.

Machiavellianism Text: ◮ – Image: ◮ – Profile: ◮ Fewer retweets ◮ Fewer tweets with URLs Shallow: ◮ – All correlations significant at p < .05; controlled for age, gender, psychopathy and narcissism.

Prediction .25 .20 .15 .10 .10 .04 .04 .05 .01 .00 Image Narc Psyc Mach DT Linear Regression, Pearson correlation between predictions and log-scored traits, 10-fold cross-validation.

Prediction .25 .20 .15 .10 .10 .09 .05 .04 .04 .05 .01 .01 .00 .00 Image Profile Narc Psyc Mach DT Linear Regression, Pearson correlation between predictions and log-scored traits, 10-fold cross-validation.

Prediction .25 .20 .15 .14 .12 .11 .10 .10 .09 .05 .04 .04 .05 .02 .01 .01 .00 .00 Image Profile Shallow Narc Psyc Mach DT Linear Regression, Pearson correlation between predictions and log-scored traits, 10-fold cross-validation.

Prediction .25 .20 .20 .16 .16 .15 .14 .12 .11 .10 .10 .09 .05 .04 .04 .05 .02 .02 .01 .01 .00 .00 Image Profile Shallow Emotions Narc Psyc Mach DT Linear Regression, Pearson correlation between predictions and log-scored traits, 10-fold cross-validation.

Prediction .25 .20 .20 .16 .16 .16 .16 .15 .14 .15 .14 .12 .11 .10 .10 .09 .05 .04 .04 .05 .02 .02 .01 .01 .00 .00 Image Profile Shallow Emotions Unigrams Narc Psyc Mach DT Linear Regression, Pearson correlation between predictions and log-scored traits, 10-fold cross-validation.

Prediction .25 .25 .20 .20 .18 .16 .16 .16 .16 .15 .15 .14 .15 .14 .12 .11 .10 .09 .10 .09 .05 .04 .04 .05 .02 .02 .01 .01 .00 .00 Image Profile Shallow Emotions Unigrams LIWC Narc Psyc Mach DT Linear Regression, Pearson correlation between predictions and log-scored traits, 10-fold cross-validation.