Recommendation Algorithm Based on Refined Knowledge Graphs and Contrastive Learning

Knowledge graphs are a vital tool for improving recommendation performance and interpretability since they are rich in auxiliary data. Two obstacles must be overcome in the recommendation process, though: 1) Real-world knowledge graphs are frequently noisy and contain connections unrelated to items and entities; 2) Data sparsity is a problem due to the long-tail distribution of user-item interactions. To address the aforementioned issues, a recommendation algorithm based on refined knowledge graphs and contrastive learning (RKGCL) is proposed. Firstly, the algorithm utilizes a graph pruning strategy to trim task-irrelevant knowledge associations, obtaining a high-quality knowledge graph. Secondly, it employs a graph convolutional network based on attention mechanisms to learn item embeddings. Subsequently, a simple yet effective noise-based embedding enhancement is applied for cross-layer contrastive learning, thereby alleviating the data sparsity issue. Finally, a message propagation strategy is employed to obtain user and item embeddings for recommendation prediction. Experimental results on three public datasets - Yelp, amazon-book, and Last-FM - show that our model outperforms other benchmark models on Recall@K and NDCG@K.