2024 Theses Doctoral
Opinion Dynamics in Social Networks: Fairness, Radicalization, and Polarization
Social media sites have been perceived as a "common digital town square" in which opinions are exchanged at an unprecedented scale and speed. People not only share their own opinions on controversial issues but also consume news and assimilate opinions shared by friends in their social circles. The process by which individuals update their opinions in a social network is called opinion dynamics.
This dissertation focuses on the negative consequences that arise from opinion dynamics, namely unfairness, polarization, and radicalization. We leverage techniques from social network modeling, graph theory, and supervised machine learning to formalize the notions of fairness in opinion dynamics over social networks, diagnose which and when algorithms exacerbate unfairness in networks, and design mitigation algorithms to counter radicalization.
In the first project, we formalize two aspects of fairness in opinion dynamics, namely procedural fairness and distributive fairness. Through theoretical analysis and simulations on real-world social networks, we show how the combined effect of homophilous and reinforcing dynamics plays a special role in both types of fairness.
In the second project, we study radicalization pathways that lead users from moderate to more extreme content and propose algorithms to mitigate radicalization. In particular, we study and propose the concept of gateway entities, i.e., non-problematic entities that are nevertheless associated with a higher likelihood of future engagement with radicalized content. We show, via a real-world application on Facebook groups, that a simple definition of gateway entities can be leveraged to reduce exposure to radicalized content without adversely impacting user engagement metrics.
Through offline experiments, we show that survival analysis-based methods are effective at identifying individuals at risk. In the third project, we study the interplay between algorithms over social networks and fairness in opinion dynamics. We propose a framework that allows us to study the joint effects of algorithms over social networks and the opinion dynamic process.
Through extensive simulations, we demonstrate that people-recommender algorithms do not exacerbate procedure unfairness but have a significant impact on distributive fairness, and polarization reduction algorithms have no significant impact on fairness. We close by discussing the limitations of our work and directions for future research.
Subjects
Files
This item is currently under embargo. It will be available starting 2026-10-14.
More About This Work
- Academic Units
- Computer Science
- Thesis Advisors
- Chaintreau, Augustin
- Degree
- Ph.D., Columbia University
- Published Here
- October 30, 2024