Multi-Frequency Filter Based Contrastive Learning for Knowledge-Aware Recommendation

The introduction of knowledge graph (KG) into recommender system has proven to be successful in further expressing user preferences. However, KGs with fruitful entities lead to noise issue and the data sparsity by the long-tail distribution. Recently, contrastive learning has been a successful method in removing noise. In this paper, we propose a multi-frequency filter based contrastive learning algorithm (MFCL) for knowledgeaware recommendation to effectively alleviate the noise and the data sparsity problem. Firstly, we design multi-frequency filters, which retain different frequency signals with different proportions simultaneously. In frequency domain perspective, our filters divide long-tail data of KG and user-item interaction graph (U-I graph). Then GNNs are used to obtain embedding representations of augmented U-I graph and KGs. In addition, we propose a dual-group cross-view contrastive learning strategy, which captures valuable knowledge for interaction information to enrich user preference. The strategy can optimize the model performance during iterative training. Extensive experiments on three real-world datasets demonstrate that our proposed MFCL outperforms the-state-of-art recommendation methods.