Interplay between upsampling and regularization for provider fairness in recommender systems

Considering the impact of recommendations on item providers is one duties multi-sided recommender systems. Item are key stakeholders in online platforms, and their earnings plans influenced by exposure items receive recommended lists. Prior work showed that certain minority groups providers, characterized a common sensitive attribute (e.g., gender or race), being disproportionately affected indirect unintentional discrimination. Our study this paper handles situation where ($i$) same provider associated with multiple list suggested to user, ($ii$) an created more than jointly, ($iii$) predicted user-item relevance scores biasedly estimated for groups. Under scenario, we assess disparities relevance, visibility, exposure, simulating diverse representations group catalog interactions. Based emerged unfair outcomes, devise treatment combines observation upsampling loss regularization, while learning scores. Experiments real-world data demonstrate our leads lower disparate relevance. The resulting lists show fairer visibility higher coverage, negligible recommendation utility.