On User-side Fairness in Negative Sampling for Recommender Systems

Abstract: Recommender systems are usually trained to discern between positive and negative instances for each user. Negative sampling plays an important role in selecting informative negative items. Since positive data is disproportionately contributed by a minority of active users, negative samplers might be affected by data imbalance thus choosing more informative negative items for active users. Consequently, users with low participation are further underrepresented in the training data, potentially causing subpar treatment from recommenders. In this paper we demonstrate empirically that active users receive more accurate recommendation than inactive users for state-of-the-art negative sampling strategies, and the degree of data imbalance influences the severity of performance disparities. We further show that the performance gain brought by sampling more negative instances for each positive item is unequally distributed across user groups. Generally, active users benefit from performance gain whereas inactive users might suffer from performance degradation. To address these shortcomings, we propose a group-wise negative ratio setup where we use the appropriate smaller negative ratio for inactive users and a bigger ratio for active users. Comprehensive experiments show our proposed group-wise ratio outperforms a single global ratio in user-side fairness and performance improvement.