Towards long-term depolarized interactive recommendations

Abstract

Personalization is a prominent process in today’s recommender systems (RS) that enhances user satisfaction and platform profitability. However, recent studies suggest that over-personalization may lead to polarized user preferences, which can result in filter bubbles and echo-chamber effects. These effects have usually been mitigated by focusing on short-term recommendation goals using immediate polarization solutions in static RS settings. In this work, we explore the problem of long-term user polarization resulting from over-personalized multi-step interactive recommendations. We propose a framework to measure and limit the polarization of user preferences, based on item categories consumed over continuous $T-$step recommendations. In this framework, we developed three recommendation approaches based on Deep Q-Networks (DQN), each one incorporating distinct polarization constraining and training techniques. First, we proposed I-CDQN, an instantaneously constrained DQN algorithm in which user polarization is forced to remain below a certain threshold at each recommendation step. Second, we proposed RP-DQN, a DQN-based method that incorporates polarization penalization terms into the reward and DQN loss function. Third, we introduced RC-DQN with a double DQN architecture, which constrains user polarization at the category-level using the first DQN, then trains the second unconstrained DQN using items from restricted category-related action spaces. The proposed methods differ in the way they apply polarization constraints, which can significantly impact their performance and suitability for specific application use cases. We conducted extensive experiments on real world datasets using cold- and warm-start scenarios for $T-$step interactive recommendations. Interestingly, RC-DQN outperforms both I-CDQN and RP-DQN, demonstrating the best balance between user polarization and personalization, and achieving significant improvement in personalization results when compared to the best performing baseline methods across all experiments, e.g., about 3.6% for $T=30$ steps.