Dual-Channel Variational Graph Autoencoder with Hierarchical Contrastive Learning for Recommendation

Abstract In recent years, Graph Neural Networks (GNNs) have been widely utilized in Recommendation. Although current GNNs-based recommendation algorithms have improved performance, they are still plagued by the data sparsity problem. Contrastive learning, which extracts self-supervised signals from a large amount of unlabeled data through auxiliary tasks, is one of the effective means to alleviate this problem. However, existing contrastive learning based on graph structure augmentation may discard important nodes or edges, resulting in the bias of semantic information; while contrastive learning based on embedding augmentation adds random noise to nodes, ignoring the differences between them. To address this problem, we propose a Dual Channel Variational Graph Autoencoder with Hierarchical Contrastive Learning for Recommendation. First, we use variational graph autoencoder to learn the node representation. Second, the collaborative neighbor graph is constructed through the similarity between users (items), which further models the higher-order signals while preserving the information to achieve global-level contrastive learning. Finally, in each variational graph autoencoder, we sample the embeddings from the Gaussian distribution of the intermediate and the final layer according to its estimator variance, respectively, to realize local-level contrastive learning. Extensive experiments on two real-world datasets show that the model achieves state-of-the-art performance, validating its effectiveness.