MVCBRec: Multi-View Contrastive Learning for Bundle Recommendation

Since bundled recommendation can meet various demands of users in one stop, it has always been a research hotspot in the recommendation field. Recent methods usually construct bundle view and item view based on user-bundle interaction and user-item interaction information, and learn representations for users and bundles from these two views. However, they all ignore the affiliation information between bundles and items, which will lead to serious information loss. To address this, we propose MVCBRec, a novel Multi-View Contrastive learning Bundle Recommendation model. It first learns the comprehensive representations of users and bundles on three views by constructing user-bundle (U-B) graph, user-item (U-I) graph and bundle-item (B-I) graph respectively. Multi-view contrastive learning is then used to model the cooperative associations between the three views. Multi-view contrastive learning encourages the alignment of the same user/bundle between different views can extract complementary information, and increases the dispersion of different users/bundles can enhance the self-discrimination ability. Extensive experiments on three public datasets show that our method outperforms SOTA baseline by 4.04% ~ 21.14% on recall. In addition, various ablation studies are performed to unveil the mystery of the key components.