Enhanced Contrastive Hypergraph Information Bottleneck Learning for Item Recommendation

Existing works on item recommendation have shown the advantages of using graph neural networks to enhance performance. However, some challenges still need to be addressed: 1) Most existing works tend to directly maximize the mutual information between views in the contrastive learning process and cannot ensure that all the information relevant to the recommendation task is captured; 2) User behavior modeling based on graph neural networks is always disturbed by the noise from interaction behavior. Therefore, we propose a novel Information bottleneck principle-enhanced denoising hypergraph convolutional contrastive recommendation method for item recommendation, named DHCLIB. Specifically, we first use a pretraining strategy with ranked sampling to generate a new hypergraph adjacency matrix, well mitigating the noise interference in the user-item interaction. Then, in the contrastive learning framework based on the new augmented hypergraph view and the original user-item hypergraph view, we introduce the information bottleneck principle to guide the minimization of the mutual information process, accurately capturing the collaborative information and implicit features with different semantics among different views. Our model demonstrates effectiveness and efficiency through extensive experiments on public datasets.