Knowledge graphs (KGs) have been increasingly integrated in recommendation systems. They capture relations among items and attributes, offering supplementary knowledge beyond observed user-item interactions. However, KG-enhanced recommendation still faces the challenge of sparsity, given that most entities in real-world KGs are associated with only a limited number of attributes. This scarcity impedes information propagation across KG and undermines the quality of item representations. To this end, we propose DiffKGR, a diffusion-based framework that generates virtual edges to alleviate sparsity in KGs and enhance downstream recommendation performance. DiffKGR consists of two key components: diffusion-based virtual edge generation and relevance-aware KG aggregation. The generation module employs a structure-guided diffusion process to produce structurally consistent virtual edges. This effectively mitigates the sparsity of the knowledge graph. The KG aggregation module introduces relevance scores to evaluate task-specific relevance of edges, which are then used to adjust edges’ weights during aggregation. This mechanism suppresses the influence of irrelevant information during message passing. To evaluate the effectiveness of DiffKGR, we conduct experiments on three public benchmark datasets. Experimental results show that our method outperforms state-of-the-art recommendation models, achieving up to 4.74% improvement. The implementations of DiffKGR are available at https://github.com/WLW8991/DiffKGR .