The aim of this study is to analyze a broad range of interpretable image features from Twitter profile pictures, such as colors, aesthetics, facial presentation and emotions. Working with these features, we uncover their relationships with personality traits from the Big Five model. With our analysis, we aim to present a procedure that can scale up psychological profiling without requiring users to undertake costly questionnaires and that better matches their online persona. We also test the predictive performance of our interpretable features in held-out data prediction.
Recent advances in computer science and a wider availability of inexpensive user generated data have made automatic personality detection an important research topic. In this study, we focus on static images, and in particular on self-selected profile pictures from social media. Although users can post other photos, studying profile pictures is particularly interesting as these reflect the impressions that the users want to convey to others. Previous work on predicting personality from images has mainly focused on predictive performance.
We use two Twitter data sets in our experiments which differ in size and the set of available user traits. For each user, we have collected up to 3,200 most recent tweets using the Twitter REST API. We use posted tweets from an account as a different modality compared to images in order to predict user attributes and demographics. Text-based prediction methods have been successfully used to predict a wide range of traits including age and gender.
(long)
Auto137