Machine prediction of personality from Facebook profiles

An increasing number of Americans use social networking sites such as Facebook, but few fully appreciate the amount of information they share with the world as a result. Although studies exist on the sharing of specific types of information (photos, posts, etc.), one area that has been less explored is how Facebook profiles can share personality information in a broad, machine-readable fashion. In this study, we apply data-mining and machine learning techniques to predict users´ personality traits (specifically, the traits of the Big Five personality model) using only demographic and text-based attributes extracted from their profiles. We then use these predictions to rank individuals in terms of the five traits, predicting which users will appear in the top or bottom 5% or 10% of these traits. Our results show that when using certain models, we can find the top 10% most Open individuals with nearly 75% accuracy, and across all traits and directions, we can predict the top 10% with at least 34.5% accuracy (exceeding 21.8%, which is the best accuracy when using just the best-performing profile attribute). These results have privacy implications in terms of allowing advertisers and other groups to focus on a specific subset of individuals based on their personality traits.