Mitigating radicalization in recommender systems by rewiring graph with deep reinforcement learning

Recommender systems play a crucial role in enhancing user experiences by suggesting content based on users consumption histories. However, a significant challenge they encounter is managing the radicalized contents spreading and preventing users from becoming trapped in radicalized pathways. This paper address the radicalization problem in recommendation systems (RS) by proposing a graph-based approach called Deep Reinforcement Learning Graph Rewiring (DRLGR). First, we measure the radicalization score (Rad(G)) for the recommendation graph by assessing the extent of users’ exposure to radical content. Second, we develop a Reinforcement Learning (RL) method, which learns over time which edges among many possible ones should be rewired to reduce the Rad(G). The experimental results on video and news recommendation datasets show that DRLGR consistently reduces the radicalization score and demonstrates more sustained improvements over time, particularly in more complex graphs compared to baseline methods and heuristic approach such as HEU that may reduce radicalization more rapidly in the early stages with fewer interventions but plateau over time.