Algorithmic bias in recommendation systems can lead to unfair treatment of users or underrepresentation of certain groups, which has become a critical concern in online platforms, potentially resulting in significant societal impacts such as discrimination and commercial losses due to reduced user trust and platform engagement. In this paper, we propose FAIR-RL, a fairness-aware deep reinforcement learning framework that mitigates bias while maintaining recommendation quality. Unlike traditional recommendation approaches that focus solely on predictive accuracy, our method incorporates a fairness-sensitive reward function that balances user engagement with group-level fairness. In contrast to existing fairness-aware recommendation methods, which typically address fairness in isolation, FAIR-RL combines fairness and performance in a unified, dynamic framework, ensuring better adaptability to evolving user preferences. The framework employs a dual-policy network: one policy optimizes individual user satisfaction, while the other enforces fairness constraints across diverse user groups. Additionally, dynamic reward adjustment allows the system to adaptively reduce bias in real-time recommendation scenarios. Experimental results on benchmark datasets demonstrate that FAIR-RL significantly improves fairness metrics without sacrificing recommendation performance, providing an effective solution for equitable personalized recommendations.
Zhi et al. (Fri,) studied this question.