Causal inference and attribute correlation consistency for eliminating popularity bias in recommendation system

Abstract

Popularity bias has long been a persistent issue in recommendation systems, leading to misleading results and significant problems such as the Matthew Effect and the Information Cocoon Room. Existing studies have primarily focused on the elevation of long-tailed items, overlooking the crucial connection between users and items. Drawing inspiration from causal graphs, this paper introduces a novel framework called CIACC (Causal Inference and Attribute Correlation Consistency) to tackle the challenges posed by popularity bias. The framework leverages causal graphs to evaluate the compatibility between users and items and to gauge the influence of item popularity on rankings. It employs counterfactual inference to estimate the impact of item popularity on rankings and adheres to the consistent principle of attribute correlation to enhance the feature representation of long-tailed items. Through rigorous experiments conducted on three public datasets, we demonstrate that our CIACC framework outperforms state-of-the-art methods.