Enhancing recommender systems with provider fairness through preference distribution awareness

Abstract

Going beyond recommendations’ effectiveness, by ensuring properties such as unbiased and fair results, is an aspect that is receiving more and more attention in the literature. This means not only providing accurate recommendations but also ensuring that the visibility of providers aligns with user preferences and demographic representation, which has been identified as a key aspect of fairness in recommender systems. In particular, provider fairness enables the generation of results which are equitable for different (groups of) providers. In this paper, we raise the problem of how recommendations are distributed when enabling provider fairness. Indeed, on the one hand, users have clear preferences with respect to which providers they choose (e.g., Italian users mostly buy Italian food), so recommendations should reflect these preferences. On the other hand, content providers should be able to reach a diverse audience, and be visible across the different user groups that expressed a preference for them. Specifically, we consider demographic groups based on their continent of origin for both users and providers, and assess how the preferences of the user groups are distributed across the provider groups. We first show that the state-of-the-art models and the existing approaches that enable provider fairness do not reflect the original distribution of the user preferences. To enable this property, we propose a re-ranking approach that, thanks to the use of buckets associating users and items, favors what we call preference distribution-aware provider fairness. Results on two real-world datasets (i.e., the Book-Crossing and COCO) show that our approach can enable provider fairness and tailor the recommendations to the original distribution of the user preferences, with negligible losses in effectiveness. In particular, in the Books dataset, our approach obtains an overall disparity that is around 6%. On the other hand, in the case of the COCO dataset, the disparities are reduced to 2%.