Personalizing Fairness: Adaptive RL with User Diversity Preference for Recommender Systems

Luana Guedes Barros Martins, Bryan Lincoln Marques de Oliveira, Bruno Brandão, Telma Woerle de Lima Soares, Marlesson Santana

Keywords: Recommender Systems, Reinforcement Learning, Fairness, Exposure Fairness, Personalization, User Diversity Preference, Actor-Critic

TL;DR: This paper introduces A2Fair, a reinforcement learning framework that personalizes recommendations by dynamically balancing relevance and supplier exposure fairness, adapting to individual user preferences for content diversity.

Abstract: Reinforcement learning is increasingly applied to optimize recommender systems for long-term user engagement and system objectives. However, a significant challenge remains in ensuring fair supplier exposure alongside user relevance, as traditional methods often lead to popularity bias. Addressing this challenge by adaptively balancing relevance and fairness can lead to more sustainable, equitable digital platforms and improved long-term user engagement. We introduce A2Fair, a RL framework that personalizes recommendations by dynamically balancing relevance and exposure fairness through an adaptive reward function that considers individual user diversity preferences and a rich state representation.

Submission Number: 21