Improvement research of Invariant Collaborative Filtering

Abstract

The rapid expansion of e-commerce has led to product oversaturation on online platforms. To help users find their preferred products in a more efficient and accurate way, many e-commerce platforms have introduced personalized recommendation systems. Collaborative filtering is one of the most successful techniques, while its improvement, Invariant Collaborative Filtering (Inv-CF), aims to address the popularity bias problem of traditional CF models by capturing unbiased preferences that remain constant despite the change in popularity distributions. However, Inv-CF model still experiences some problems such as ignoring the influence of attention, causing performance less effective when analyzing the representation of implicit feedback. This paper explores the enhancement of Inv-CF, a recommendation system model designed to mitigate the influence of popularity bias. We defined the experimental framework and evaluated the performance of the improved Inv-CF on two benchmark datasets, Yahoo! R3 and COAT. And the results demonstrate significant performance gains over the original Inv-CF, highlighting the effectiveness of the proposed enhancements. In conclusion, this paper presents improvements to the Inv-CF model's loss functions, addressing issues related to popularity bias in collaborative filtering.