Knowledge-Stability-Guided Dual-Graph Contrastive Learning for Recommendation

Abstract

Knowledge-graph-enhanced recommendation leverages external knowledge to characterize item semantics and alleviate the limitations of representation learning under sparse user–item interactions. However, existing methods inadequately model the correspondence between item-level collaborative signals and knowledge semantics, and most graph augmentation strategies rely on random perturbations that fail to distinguish edge-specific retention values. To address these limitations, we propose Knowledge-Stability-Guided Dual-Graph Contrastive Learning for Recommendation (KSDGCL). KSDGCL learns collaborative representations from the user–item interaction graph and knowledge-semantic representations from the knowledge graph, and introduces a cross-graph semantic alignment objective to strengthen the semantic correspondence between the collaborative and knowledge-semantic representations of the same item. To construct informative augmented views, KSDGCL defines edge-level knowledge stability by measuring the consistency of preference matching for the same user–item interaction edge across knowledge-perturbed views. The resulting stability scores are converted into edge retention probabilities to guide augmented interaction graph construction, thereby preserving valuable collaborative relations. Finally, multi-view representations from the original interaction graph, the knowledge graph, and the augmented interaction graphs are fused into comprehensive representations for recommendation. Experiments on Amazon-Book and LastFM show Recall@20 gains of 3.7% and 5.6% over the strongest baseline, respectively, demonstrating the effectiveness of KSDGCL in improving recommendation performance.