Fusing micro-preserved representations and item-homogeneous relationships for graph convolution collaborative filtering

Tao Wang, Jiwei Qin, Aahua Gao, Chao Ma

Abstract: Due to the success of Graph Convolutional Networks (GCNs) in efficiently learning node representations and topological structures in the user-item bipartite graph, GCNs have widely been employed in Collaborative Filtering (CF) to enhance recommendation performance. However, existing GCN-based CF models still suffer from two issues: (1) insufficient utilization of initial user (or item) representations, resulting in suboptimal representation learning; (2) neglect of the homogeneous relationship between items, limiting the effectiveness of graph representation learning. To address these issues, we propose a novel recommendation model named Fusing micro-preserved representations and Item-homogeneous relationships for Graph Convolution Collaborative Filtering (FIGCCF). First, FIGCCF devises a micro-preserved representations learning module. It can retain the information of initial representations and prevent information loss by integrating partial low-order representations of users (or items) into each propagation layer. Second, FIGCCF leverages items’ high-order representations and similarities to construct an item-homogeneous relationship graph, aggregating representations from homogeneous neighbors to improve the capability of graph representation learning. Finally, extensive experiments on six real recommendation datasets demonstrate that FIGCCF outperforms several mainstream graph collaborative filtering models in performance. We also analyze the effectiveness and plausibility of components of FIGCCF.

External IDs:dblp:journals/mlc/WangQGM25