Enhancing Uniformity Loss via Popularity Weighting

Abstract

Current research efforts are dedicated to improving loss functions to enhance the quality of representations in a recommendation system. Among them, the utilization of the alignment-uniformity loss has demonstrated remarkable performance. Previous studies have explored strategies for assigning different weights to alignment losses, but studies on assigning different weights per item within uniformity losses are lacking. This study begins by noting a difference in uniformity between popular and unpopular item groups when trained with the current alignment-uniformity loss. Based on this observation, we introduce a new methodology that includes weight reduction associated with popular items for uniformity loss. Empirical results show that our method mitigates the uniformity difference between item groups with varying levels of popularity and significantly improves the performance of the recommender system.