Multi-view Neighbor-Enriched Contrastive Learning Framework for Bundle Recommendation

Bundle recommendation aims to recommend a group of items with a similar theme to users. The previous methods devoted to alleviating the data sparsity problem. However, they either modeled the intuitive interaction between users and items (bundles) or randomly sampled the negative samples during the training process. It is far from enough to learn the user and bundle embeddings because of insufficient modeling of collaborative information. We propose a Multi-view Neighbor-enriched Contrastive learning framework for Bundle Recommendation (MNCBR). MNCBR learns representations of users and bundles from two separate views (i.e. item and bundle view). Meanwhile, different contrastive learning strategies are applied to each view respectively. Specifically, the item-view contrastive mechanism jointly learns the high-order relations of users and bundles, and obtains the global preferences of users. The bundle-view contrastive mechanism explores the collaborative information via structural neighbors on the interaction graph. Extensive experiments on two public datasets show the proposed MNCBR outperforms the state-of-the-art methods.