Recommendation actively selects information for users, yet it persistently face data sparsity and cold-start problems. The incorporation of knowledge graph as side information has demonstrated effectiveness in mitigating these issues, leading to the development of knowledge-aware recommendation. Existing methods often use graph augmentation by constructing other knowledge views from the original knowledge graph (KG) to address external noise (e.g., erroneous triplets) and information overload (e.g., redundant data), yet they seldom consider the knowledge incompleteness (e.g., inherent missing facts and long-tail relation sparsity), which leads to decreased recommendation performance. Unlike noise (which introduces distortions) and information overload (which causes selection inefficiency), knowledge incompleteness stems from structural gaps in the graph that hinder semantic connectivity. To address this challenge, we propose RKGRec, a relation-guided conditional diffusion framework that generates a relation-centric auxiliary KG to alleviate knowledge incompleteness for recommendation. The model primarily consists of three core modules tailored for knowledge-aware recommendation: (1) a relation-attention network that captures multi-hop entity-relation patterns to obtain knowledge embeddings. (2) a relation-guided conditional diffusion model that strategically refines knowledge graph through controlled noise injection by the forward process and relation-guided denoising by the reverse process. (3) a joint prediction and optimization module that jointly trains recommendation and knowledge graph generation. Experimental results show RKGRec outperforms baselines across multiple datasets, particularly achieving both comprehensive leading predictive accuracy and competitive predictive diversity. The model also demonstrates robust performance in cold-start users, long-tail items, interaction noise, and knowledge graph noise or sparsity conditions.
原文链接
分享
引用
