DPRec: A dual-path preference evolution framework with semantic alignment for knowledge graph recommendation

Although recent advances in knowledge graph-based recommendation have achieved remarkable progress, several key challenges remain. First, most existing methods treat user preferences as static, failing to capture their semantic evolution over time. Second, many graph neural network models rely on uniform relation aggregation, overlooking differences in semantic importance and increasing the risk of over-smoothing. Furthermore, existing contrastive learning approaches primarily emphasize structural consistency by generating samples through random perturbations, which neglect preference hierarchies and can lead to semantic inconsistency. To address these limitations, this paper proposes DPRec, a dual-path preference evolution framework with semantic alignment for knowledge graph recommendation. DPRec jointly captures multi-hop user interests through direct and indirect preference path diffusion, incorporates a preference evolution attention mechanism for adaptive semantic aggregation, and employs a multi-level contrastive learning framework to align preference-level and path-level representations. Together, these components address three key aspects of knowledge graph recommendation, namely user preference modeling, dynamic preference evolution, and semantic alignment, thereby overcoming the limitations of existing approaches. Comprehensive experiments on multiple public datasets demonstrate that DPRec consistently outperforms existing baselines in terms of AUC, F1, Hit@K, and NDCG@K. These results demonstrate that dual-path preference reasoning produces more accurate, semantically consistent, and robust knowledge-aware recommendations.