Item Attribute-aware Graph Collaborative Filtering

Abstract

Collaborative filtering (CF) is a widely used technique in recommender systems. While many CF methods primarily focus on collaborative signals derived from user–item interactions, they often overlook other types of similarities (e.g., user similarity in loving actors, movie similarity in sharing genres, etc.). These similarities offer the fine-grained knowledge to understand users and items, which can complement the collaborative signal. In this paper, we introduce a model that harnesses these additional similarities for improved recommendations. Our approach involves constructing both a user and an item relational graph, based on multiple item attributes. These graphs reflect semantic similarities from various perspectives among users and items. Then, we develop a recommendation framework that employs a dual graph neural network, integrating these graphs into the recommendation process. We name our approach Item Attribute-aware Graph Collaborative Filtering (IAGCF). Through evaluations on six real-world datasets, we found that IAGCF surpasses several state-of-the-art recommenders.