Hyperbolic Multi-view Contrastive Learning for Bundle Recommendation

Bundle recommendation aims at recommending a set of related items (bundles) to users. To address the limitations of existing methods in capturing the hierarchical structure of interaction graphs and integrating multi-view information, this paper proposes a hyperbolic multi-view contrastive learning framework for bundle recommendation (HMCBR). The model embeds entities into hyperbolic space and leverages a hyperbolic graph convolutional network to learn user and bundle representations across different views. Additionally, a hyperbolic self-attention mechanism is introduced to adaptively allocate view weights, optimizing multi-view information fusion. Moreover, both intra-view and inter-view contrastive learning are incorporated to enhance feature consistency and multi-view information interaction. Experimental results demonstrate that HMCBR outperforms baseline models on three benchmark datasets, effectively improving recommendation performance.