A Machine Learning Approach to Detecting Polarization in Social Media Discourse

• In Event: New Approaches for Prejudice Reduction and Depolarization

Abstract

Beyond social network analysis, political scientists have limited tools to measure polarization in online communities. We offer a new method that combines network analysis and supervised machine learning to detect changes in mass discourse on social media. We use our method to examine the effect of high salience news events on relevant political discussion on Twitter. We find that in day-to-day discourse, Twitter users across the ideological spectrum express diverse opinions on a range of political topics. Following news events salient to those topics, discourse becomes more extreme and partisan. Additionally, we find that the ideological tilt of discourse for users situated in centrist networks remains stable relative to users situated in more partisan networks. We demonstrate the potential of our supervised learning approach to address open questions in political science about who is polarized and when, and then discuss our method’s limitations and practical challenges.

Authors

• Laura Jakli, Harvard University

• Paul Gill