RsDiff: Rational score based knowledge graph diffusion for recommendation

Abstract

The Knowledge graph (KG) provides auxiliary information to improve the recommendation system performance. However, the knowledge graph includes a large number of triplets that have nothing to do with the recommendation task, leading to suboptimal results. To address this challenge, we propose a knowledge graph diffusion model based on rationality score for recommendation, called RsDiff. Firstly, we design a rational scoring mechanism for the knowledge graph triplets. Then, we propose a knowledge graph diffusion model based on rational scores to mitigate the impact of noise. Finally, we employ cross-view contrastive learning to align collaborative signals across different graphs. Experiments show that our proposed RsDiff outperforms the most advanced recommendation models in terms of NDCG@20 and Recall@20 indicators in the Last-FM, Alibaba-iFashion, and MIND datasets.