Knowledge-aware Recommender System with Cross-views Contrastive Learning

The knowledge-aware recommendation (KGR) domain commonly suffers from the problem of supervised signal sparsity, and contrast learning methods are increasingly studied to address this issue. However, existing contrast learning-based KGR models still have the following limitations. First, existing methods failed to suppress the interference information of unnecessary neighbouring nodes in the knowledge graph because graph convolution is used to directly aggregate all neighbouring nodes; Second, focusing only on the global information would lead to ignoring the fine-grained local features, causing over-smooth issues. In this work, a Knowledge-aware Recommender System with Cross-Views Contrastive Learning (KRSCCL) is proposed to address the aforementioned issues. In the KRSCCL, a relational graph attention network is proposed to construct a global view, including user, item and entity nodes. A lightweight graph convolutional network is designed to construct a local view, including user and item nodes, in which local features are emphasized to effectively mitigate the over-smooth problem. Finally, the contrastive learning mechanism is performed between intra- and inter-graph node pairs of the two views to fully extract KG signals and further optimize the user and item representations. Experimental results on three public datasets from different domains demonstrate that the proposed KRSCCL achieves expected performance improvement on all the three datasets over selective baselines, F1 score improvement on Movielens-1M, Last.FM and Book-crossing are 2.0%, 0.3% and 5.1%, respectively. Most importantly, the relational graph attention network can effectively exclude the noise during the feature aggregation of complex networks, the local views can optimize the generation of the node representation and alleviate the over-smooth problem. © 2024 Editorial Department of Journal of Sichuan University. All rights reserved.