Multi-view graph contrastive learning for recommendations with sparse user-item interactions

Graph-based recommendation systems often suffer from severe performance degradation under sparse user–item interactions. To address this challenge, we propose a multi-view graph contrastive learning framework that jointly leverages user-, item-, and global-level signals to learn robust representations for recommendations under sparse user-item interactions. Specifically, on the user side, we refine preference modeling through contrastive learning based on embedding propagation and structural perturbation in a graph neural network. On the item side, we construct two complementary views—a structural view from a knowledge graph and a semantic view derived from inter-item similarity—to enforce semantic coherence. At the global level, we introduce a novel bidirectional contrastive objective that captures high-order collaborative signals via user–item and item–user contrastive pairs, enhanced by an asymmetric Gaussian weighting strategy that suppresses noisy or uninformative negatives, thereby mitigating gradient instability and improving discriminability. Extensive experiments on three benchmark datasets—Book-Crossing, Last.FM, and MovieLens-1M—with varying sparsity levels demonstrate that MVGCL consistently outperforms 16 state-of-the-art baselines. A case study further confirms its enhanced semantic consistency, accuracy, and interpretability under extreme data sparsity.