Polarization on Social Media

Abstract

Social media and the web have provided a foundation where users can easily access diverse information from around the world. However, over the years, various factors, such as user homophily (social network structure), and algorithmic filtering (e.g., news feeds and recommendations) have narrowed the breadth of content that a user consumes. This has lead to an ever-increasing cycle where users on social media only consume content that agrees with their beliefs and hence are recommended more such content, ultimately leading to a polarized society where diverse opinions are not encouraged. This thesis provides a broad overview of polarization on social media, along with algorithmic techniques to identify polarized topics, understanding their properties over time, and finally, to reduce polarization. First, we provide methods to identify polarized topics automatically from social-media streams. Our methods are mainly based on interaction networks, i.e., networks of social media users, connected through certain types of interactions. We first show that polarized topics have a special bi-clustered structure in their retweet network and propose an algorithm to quantify the degree of polarization by using a random walk on this network. We then make use of sub-graph patterns (motifs) in the reply network of users to show that we can easily identify polarized topics using such patterns. Since our analysis does not use content, our methods are able to generalize to any topic, domain and language. Next, we study the dynamic aspects of the process of polarization. We understand what happens to the interaction networks defined above in case of a sudden increase in interest of users on the topic. We then address the question on whether polarization on Twitter has increased over the last 8 years and find evidence to support that it does. Finally, given these findings, we design algorithms to reduce polarization. We propose two approaches. In the first approach, we propose connecting users with opposing viewpoints in order to reduce polarization. Our method takes into account the users' interests and their current level of polarization to help them get connected to the people they feel comfortable in doing so. In the second approach, we take an information-diffusion route. We pose the problem of reducing polarization as a task of spreading information that reaches both sides of the polarized topic.