<i>EquiRate</i>: balanced rating injection approach for popularity bias mitigation in recommender systems

Recommender systems often suffer from popularity bias problem, favoring popular items and overshadowing less known or niche content, which limits recommendation diversity and content exposure. The root reason for this issue is the imbalances in the rating distribution; a few popular items receive a disproportionately large share of interactions, while the vast majority garner relatively few. In this study, we propose the <i>EquiRate</i> method as a pre-processing approach, addressing this problem by injecting synthetic ratings into less popular items to make the dataset regarding rating distribution more balanced. More specifically, this method utilizes several synthetic rating injection and synthetic rating generation strategies: (<i>i</i>) the first ones focus on determining which items to inject synthetic ratings into and calculating the total number of these ratings, while (<i>ii</i>) the second ones concentrate on computing the concrete values of the ratings to be included. We also introduce a holistic and highly efficient evaluation metric, <i>i.e</i>., the <i>FusionIndex</i>, concurrently measuring accuracy and several beyond-accuracy aspects of recommendations. The experiments realized on three benchmark datasets conclude that several <i>EquiRate</i>'s variants, with proper parameter-tuning, effectively reduce popularity bias and enhance recommendation diversity. We also observe that some prominent popularity-debiasing methods, when assessed using the <i>FusionIndex</i>, often fail to balance the referrals' accuracy and beyond-accuracy factors. On the other hand, our best-performing <i>EquiRate</i> variants significantly outperform the existing methods regarding the <i>FusionIndex</i>, and their superiority is more apparent for the high-dimension data collections, which are more realistic for real-world scenarios.