A Multi-Side Fairness-Aware Recommendation System Based on a Pareto-Efficient Perspective

Abstract

In this paper proposed was a method to solve the multi-side fairness of the recommender system from the perspective of Pareto. It eliminates the sensitive attribute information in the user embedding through the adversarial regularizer, and adopts the negative sampling strategy based on exposure to improve the accuracy of the recommender system, so as to achieve Pareto optimality. In addition, the exposure-based negative sampling strategy solves the problem of item exposure bias to a certain extent, ensures the fairness of item side, and realizes the multi-side fairness of users and items. Experimental results show that the method effectively improves the fairness of users and items while ensuring the accuracy of recommendation.

Keywords

• fairness-aware recommendation
• multi-side fairness-aware recommendation
• pareto efficient
• adversarial learning
• negative sample