Enhancing Calibration and Reducing Popularity Bias in Recommender Systems

The recent literature highlights that recommendation systems are significantly influenced by popularity bias. This phenomenon has far-reaching implications for the fairness and accuracy of recommendations. This bias often results in some users finding their preferences inadequately reflected in their recommendations, while others benefit from more consistent suggestions. Nevertheless, despite the current state-of-art efforts in this field that primarily aim to provide fairer recommendations, a crucial aspect has been overlooked: the impact of popularity bias on the long tail effect, which leads to a decline in the visibility of less popular items in recommendations. To address this research gap, the present study introduces a calibration approach designed to cater to the diverse interests of users across various levels of item popularity. To achieve this objective, we propose a post-processing system that is independent of any specific recommendation algorithm. Building upon the foundational idea presented by [20], we evaluate the efficacy of our proposed system using an additional dataset from the domain of music. The performance assessment of our system encompasses a range of metrics that consider aspects related to popularity, accuracy, and fairness. Additionally, four recommendation algorithms and two distinct baselines are employed. As a result, the proposed technique mitigates popularity bias, augmenting diversity and fairness within the considered datasets.