Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Using recommendations to reduce opinion polarization in social networks

Chochlakis Ioannis

Full record

---

Summary

In recent years, social media platforms have become a battleground for sociopolitical discourse. Once a herald of global connection, social networks are now a key factor in amplifying political and ideological divides around the globe. Although the internet offers an abundance of information, from diverse perspectives, individuals often gravitate towards and connect with others who share similar opinions. This ``natural'' segregation of social networks is made worse by malicious actors, such as bots and misinformation agents, exploiting users' psychological tendencies (confirmation bias) in order to further their own goals. The personalization of recommenders also played a key role in furthering this divide among online communities. Individuals are funneled into insulated groups or “echo chambers” where preexisting beliefs get reinforced and diverse viewpoints get limited exposure, with the goal of maximizing engagement on the platform and thus profit. The consequences of this algorithmic bias extend beyond the social network space and pose a threat to our society as a whole, dividing us slowly and steadily. In this thesis, we focus on the role of the recommender inside of the social network and attempt to create a recommendation system that values not only user engagement, but also the reduction of polarization and the disruption of echo chambers inside the network. First, we provide a detailed explanation of how we create a polarized social network space, modeled after the social platform X, where each user inside the network has an opinion along with a set of users they follow. Next, we introduce our model of asynchronous recommendation-driven opinion evolution, explaining how suggested content can influence users' opinions in our network. Additionally, we explain how in this model, a group of users may be left completely segregated, unable to be influenced by other users. Next, we define our proposed metric for measuring polarization that will be used in order to evaluate our proposed system’s success. We then explain how we will be adapting this problem to a Deep Reinforcement Learning framework in order to properly train and create agents that will act as recommenders, without any knowledge of the model parameters. This means our implementation can be used with other opinion evolution models and polarization metrics, as long as the necessary Reinforcement Learning modeling parameters are properly defined. We then define our state space, action space, the reward function and the terminal states for our proposed Reinforcement Learning Model. Finally, we conduct a set of experiments, on small polarized network communities, to evaluate if our proposed solution can indeed learn to reduce polarization. Our results confirm the effectiveness of our approach.

Available Files

Services