Enhancing knowledge-aware recommendation with a cross-view contrastive learning

Knowledge graph (KG) can effectively address the sparsity issue in recommendation systems. In recent years, graph neural networks (GNNs) have gained popularity in knowledge-aware recommendation (KGR) due to their powerful capability in modeling graph structures. However, existing GNN-based methods have the following limitations: (1) Primarily focusing on extracting collaborative signals between items from the user-item graph and neglecting the influence of preferences between different users. (2) The sparsity of user interaction data is a problem, and mining item knowledge associations on the KG with limited interaction data as a supervision signal will lead to limited performance improvement. In this paper, we propose an efficient cross-view contrastive learning method, KECL, to address the above challenges. Specifically, we first construct a user-user social graph based on the user-item graph to capture potential social connections among users. Then, we selectively utilize the entity information demanded by users and items to construct user-entity and item-entity graphs. Based on this, we design two contrastive loss tasks to perform contrastive learning on the above four graph views from both user and knowledge levels. This approach enables us to model the influence of preferences between users with social connections. It also allows us to efficiently mine item knowledge associations through a self-supervised paradigm, thereby learning high-quality node representations. Experimental results on three publicly available datasets demonstrate that our KECL outperforms state-of-the-art methods.