Multi-view Contrastive Learning Network for Recommendation

Knowledge graphs (KGs) are being introduced into recommender systems in more and more scenarios. However, the supervised signals of the existing KG-aware recommendation models only come from the historical interactions between users and items, which will lead to the sparse supervised signal problem. Inspired by self-supervised learning, which can mine supervised signals from the data itself, we apply its contrastive learning framework to KG-aware recommendation, and propose a novel model named Multi-view Contrastive Learning Network (MCLN). Unlike previous contrastive learning methods that usually generate different views by ruining graph nodes, MCLN comprehensively considers four different views, including collaborative knowledge graph (CKG), user-item interaction graph (UIIG), and user-user graph (UUG) and item-item graph (IIG). We treat the CKG as a global-level structural view, and the other three views as local-level collaborative views. Therefore, MCLN performs contrastive learning between the four views at the local and global levels, aiming to mine the collaborative signals between users and items, between users, and between items, and the global structural information. Besides, in the construction of UUG and IIG, a receptive field is designed to capture important user-user and item-item collaborative signals. Extensive experiments on three datasets show that MCLN significantly outperforms state-of-the-art baselines.