RPGCN: A relation-preserving graph convolutional network for enhanced node feature prediction in recommender systems

Abstract

Recommender systems (RSs) increasingly rely on graph-based models to capture the intricate relationships between users and items. However, existing approaches often struggle to preserve the full spectrum of relational information—particularly in sparse and heterogeneous datasets—leading to degraded recommendation quality. To bridge this gap, we introduce RPGCN (Relation-Preserving Graph Convolutional Network), a novel hybrid framework that unifies early and intermediate fusion strategies with multi-branch Graph Attention Networks (GATs) for enhanced relation modeling. Unlike conventional GCN-based methods that suffer from oversmoothing and relational dilution, RPGCN explicitly preserves both local and global user-item interactions through specialized User-GCN and Item-GCN pathways. Additionally, an auxiliary self-supervised learning task strengthens the model’s robustness in sparse scenarios. Extensive experiments conducted on five benchmark datasets—ML-100 K, ML-1 M, ModCloth, RentTheRunway, and Epinions—demonstrate that RPGCN consistently surpasses state-of-the-art baselines across predictive and ranking metrics, including AUC, RMSE, MAE, NDCG, and HR@10. These results highlight RPGCN’s ability to retain fine-grained relational structure and its superior performance in complex recommendation environments, establishing a compelling alternative to existing graph-based RS architectures.