A Two-Stage Calibration Approach for Mitigating Bias and Fairness in Recommender Systems

Popularity bias and unfairness are problems caused by the lack of calibration in recommender systems. Works that intend to reduce the effect of popularity bias do not consider the distribution of item genres/categories in the users' profiles. Other studies aim to calibrate the system to generate fair recommendations according to users' profiles, but usually are still biased towards popularity. We propose a system calibration approach based on users' preferences for different levels of popularity of items and their genres. The proposed approach works in the post-processing stage and can be combined with different recommendation models. We evaluated the system with offline experiments using one state-of-the-art dataset, three recommender algorithms, six baselines, and different metrics for popularity, fairness, and accuracy. The results indicate reduced popularity bias and improved fairness.