Providing Item-side Individual Fairness for Deep Recommender Systems

Recent advent of deep learning techniques have reinforced the development of new recommender systems. Although these systems have been demonstrated as efficient and effective, the issue of item popularity bias in these recommender systems has raised serious concerns. While most of the existing works focus on group fairness at item side, individual fairness at item side is left largely unexplored. To address this issue, in this paper, first, we define a new notion of individual fairness from the perspective of items, namely (α, β)-fairness, to deal with item popularity bias in recommendations. In particular, (α, β)-fairness requires that similar items should receive similar coverage in the recommendations, where α and β control item similarity and coverage similarity respectively, and both item and coverage similarity metrics are defined as task specific for deep recommender systems. Next, we design two bias mitigation methods, namely embedding-based re-ranking (ER) and greedy substitution (GS), for deep recommender systems. ER is an in-processing mitigation method that equips (α, β)-fairness as a constraint to the objective function of the recommendation algorithm, while GS is a post-processing approach that accepts the biased recommendations as the input, and substitutes high-coverage items with low-coverage ones in the recommendations to satisfy (α, β)-fairness. We evaluate the performance of both mitigation algorithms on two real-world datasets and a set of state-of-the-art deep recommender systems. Our results demonstrate that both ER and GS outperform the existing minimum-coverage (MC) mitigation solutions [Koutsopoulos and Halkidi 2018; Patro et al. 2020] in terms of both fairness and accuracy of recommendations. Furthermore, ER delivers the best trade-off between fairness and recommendation accuracy among a set of alternative mitigation methods, including GS, the hybrid of ER and GS, and the existing MC solutions.