Bi-knowledge views recommendation based on user-oriented contrastive learning

In recommender system, knowledge graph (KG) is usually leveraged as side information to enhance representation ability, and has been proven to mitigate the cold-start and data sparsity issues. However, due to the complexity of KG construction, it inevitably brings a large amount of noise, thus simply introducing KG into recommender system may hurt the performance of models. In addition, the current KG-based recommendation models mainly include the following issues: (1) The rich facts and semantic knowledge contained in KG are not fully explored. (2) The useless noise in KG is not effectively filtered, and the representation obtained by neighborhood aggregation shows poor quality. (3) Nodes with long-tail distribution are easily ignored and the models fail to balance the attention between popular and unpopular items. Therefore, we propose a Bi-Knowledge Views Recommendation Based on User-Oriented Contrastive Learning architecture (BUCL) to improve the representation quality and alleviate the long-tail distribution of entities. In particular, different graph embedding methods are applied to fully extract the rich facts and semantic knowledge in the KG to obtain multiple views of nodes. Based on the different representation views, a user-oriented item quality estimation method is proposed to guide the model to generate multiple augmented subgraphs. Each node provides enough negative samples to ensure that the model discriminates the same node from other nodes in differentiated subgraphs with contrastive learning. Experiments on three benchmark datasets show that BUCL consistently outperforms state-of-the-art models, alleviating the long-tail distribution problem and reducing the impact of noise.