Show me your friends, and I will tell you whom you vote for: Predicting voting behavior in social networks

Abstract

Increasing use of social media in campaigns raises the question of whether one can predict the voting behavior of social-network users who do not disclose their political preferences in their online profiles. Prior work on this task only considered users who generate politically oriented content or voluntarily disclose their political preferences online. We avoid this bias by using a novel Bayesian-network model that combines demographic, behavioral, and social features; we apply this novel approach to the 2016 U.S. Presidential election. Our model is highly extensible and facilitates the use of incomplete datasets. Furthermore, our work is the first to apply a semi-supervised approach for this task: Using the EM algorithm, we combine labeled survey data with unlabeled Facebook data, thus obtaining larger datasets as well as addressing self-selection bias.