Date of Award

Spring 6-13-2021

Document Type

Thesis (Undergraduate)


Department of Computer Science

First Advisor

Prof. Soroush Vosoughi


In this work, we model users' behavior on Twitter in discussion of the COVID-19 outbreak using inverse reinforcement learning to better understand the underlying forces that drive the observed pattern of polarization. In doing so, we address the largely untapped potential of inverse reinforcement learning to model users' behavior on social media, and contribute to the body of sociology, psychology, and communication research seeking to elucidate the causes of socio-cultural polarization. We hypothesize that structural characteristics of each week's retweet network as well as COVID-19 data on cases, hospitalizations, and outcomes are related to the Twitter users' reward function which leads to polarized discussion of COVID-19 on the platform. To derive the state space of our inverse reinforcement learning model, we compute the relative modularity of retweet networks formed from retweets about COVID-19. The action space is determined by the distribution of mask-wearing sentiment in tweets about COVID-19. We build a fine-tune a BERT text classifier to determine mask-wearing sentiment in tweet. We design state features which reflect both structural characteristics of the retweet networks and COVID-19 data on cases, hospitalizations, and outcomes. Our results indicate that polarized Twitter discussion about COVID-19 weighs more heavily on features relating to the severity of the COVID-19 outbreak and less heavily on features relating to the structure of retweet networks. Overall, our results demonstrate the aptitude of inverse reinforcement learning in helping understand user behavior on social media.