Recommender System Design via Online Feedback Optimization

Abstract

Conventional recommender systems enhance user engagement through personalized content. However, personalization usually induces significant side effects on opinion formation, such as polarization and echo chambers that need to be prevented. With this motivation, we design a recommender system algorithm that addresses user engagement maximization and opinion polarization mitigation by operating in feedback with the social platform. The recommender is agnostic about real-time opinions, network topology, and users' clicking behaviour, all estimated online. We numerically verify the efficacy of the designed recommender on synthetic data. We show that by providing network-aware recommendations to the users as opposed to users' tailored content, we significantly reduce polarization effects without sacrificing user engagement.