Knowledge Graph Cross-View Contrastive Learning for Recommendation

Zeyuan Meng, Iadh Ounis, Craig Macdonald, Zixuan Yi

Abstract: Knowledge Graphs (KGs) are useful side information that help recommendation systems improve recommendation quality by providing rich semantic information about entities and items. Recently, models based on graph neural networks (GNNs) have adopted knowledge graphs to capture further high-order structural information, such as shared preferences between users and similarities between items. However, existing GNN-based methods suffer from two challenges: (1) Sparse supervisory signal, where a large amount of information in the knowledge graph is non-relevant to recommendation, and the training labels are insufficient, thereby limiting the recommendation performance of the trained model; (2) Valuable information is discarded whereby the use by the existing models of edge or node dropout strategies to obtain augmented views during self-supervised learning could lead to valuable information being discarded in recommendation. These two challenges limit the effective representation of users and items by existing methods. Inspired by self-supervised learning to mine supervision signals from data, in this paper, we focus on exploring contrastive learning based on knowledge graph enhancement, and propose a new model named Knowledge Graph Cross-view Contrastive Learning for Recommendation (KGCCL) to address the two challenges. Specifically, to address supervision sparseness, we perform contrastive learning between graph views at different levels and mine graph feature information in a self-supervised learning manner. In addition, we use noise augmentation to enhance the representation of users and items, while retaining all triplet information in the knowledge graph to address the challenge of valuable information being discarded. Experimental results on three public datasets show that our proposed KGCCL model outperforms existing state-of-the-art methods. In particular, our model outperforms the best baseline performance by 10.65% on the MIND dataset.