Mining User Interests to Predict Perceived Psycho-Demographic Traits on Twitter

We analyze the relation between user interests and their perceived psychodemographic attributes using Twitter data, training models for predicting various personal traits of users. In contrast to existing work, which bases predictions on the textual tweets produced by users, we leverage the fact that users are embedded in the Twitter social network. We examine the accounts that our users follow, and use them to determine the high-level interests of these users, then use these areas of interest as features for predicting perceived personal traits. We cover target attributes such as gender, age, educational background, political stand and personality. We evaluate our technique on a dataset of over 4,000 Twitter user profiles. We use crowdsourcing to annotate these user profiles with perceptions regarding their personal traits, and correlate these with user interests, as captured by the accounts they follow and their classification in the Twitter “Who To Follow” hierarchy. We compare the accuracy of our personal trait prediction methods with the state-of-the-art approaches that solely rely on user tweets, and discuss the correlations between perceived user demographics and interests.