SUPER: Smart User-Centric Popularity Exposure Reduction for Fair and Diverse Recommendations

Abstract

We present SUPER, a novel approach to mitigating popularity bias in recommender systems by explicitly modeling user-specific preferences for item popularity. Unlike existing methods that apply uniform adjustments to reduce the dominance of popular items, SUPER leverages each individual’s inclination toward mainstream versus niche content. By combining a Pareto-based item partitioning with personalized preference calibration, we incorporate both popular and tail subsets into the training and recommendation refinement processes. Experimental evaluations on multiple benchmark datasets demonstrate that SUPER substantially improves diversity and fairness while maintaining robust accuracy. Furthermore, our findings indicate that aligning the recommendations with each user’s interaction-derived popularity tendency promotes a broader exposure to underrepresented items and produces more satisfying recommendation experiences. These insights underscore the potential of personalized debiasing strategies to promote equitable and context-sensitive content curation in modern recommendation systems.

Keywords