Super-Node Generation for GNN-Based Recommender Systems: Enhancing Distant Node Integration via Graph Coarsening

Shasha Hu, Chao Wang, Chuan Qin, Hengshu Zhu, Hui Xiong

Abstract: In the realm of recommender systems, Graph Neural Network based collaborative filtering has emerged as a leading approach. This method represents user-item interactions as a bipartite graph, facilitating the extraction of user and item embeddings by aggregating neighborhood data and capturing potential user preferences. However, a glaring limitation persists: GNNs, by design, primarily concentrate on closely located nodes, often sidelining those situated farther apart. This issue becomes more pronounced due to the “long-tail” effect prevalent in recommendation systems. Attempts to enhance GNNs, either by adding more layers or integrating contrastive learning, have only seen limited success. To overcome these hurdles, we present the General Graph Coarsening Recommendation Framework (GGCRF). This novel methodology utilizes graph coarsening algorithms to forge crucial connections between user/item nodes and newly created super-nodes. A key aspect of GGCRF is the division of user-item graphs into smaller segments using community detection, leading to the formation of super-nodes that encapsulate groups of similar nodes. GGCRF also introduces a unique indirect connectivity loss tailored for collaborative filtering, enhancing its adaptability to a range of GNN-based applications. Our extensive evaluations of GGCRF across five real-world datasets confirm its effectiveness and adaptability, underscoring its capability for generalization in various GNN-based models.