A graph contrastive learning recommendation algorithm based on variational inference

Abstract

Recommender systems are essential tools for alleviating information overload by providing personalized suggestions to users. Graph neural networks (GNNs), capable of capturing complex node relationships and higher-order interactions, have emerged as a prominent research direction in recommendation tasks. In this work, we propose SGICL, a novel recommendation algorithm that integrates graph contrastive learning with variational inference. The graph contrastive learning module applies feature augmentation to mitigate data sparsity and noise, while the variational inference module models the latent distribution of node embeddings, capturing uncertainty and improving representation quality. A crosslayer contrast mechanism further strengthens the discriminative power of node embeddings. We conduct extensive experiments on three benchmark datasets: Yelp2018, Douban-Book, and Alibaba-iFashion. The results show that SGICL consistently outperforms state-of-the-art models, including LightGCN, MixGCF, SimGCL, and SEPT. Specifically, on Yelp2018, SGICL achieves a Hit Rate of 0.0653 and NDCG of 0.0612, surpassing the best baseline by over 3%; on Douban-Book, it attains a Hit Rate of 0.1371 and NDCG of 0.1618; and on Alibaba-iFashion, it records a Hit Rate of 0.1146 and NDCG of 0.0574, demonstrating notable improvements across all four evaluation metrics. These results confirm that SGICL effectively enhances recommendation accuracy, robustness, and the capacity to learn rich latent graph representations, providing a practical and scalable framework for graph-based recommender systems.