Enhancing Sequential Recommendation with Knowledge Graph-Based Intent Network

Sequential recommendation models have achieved strong performance by capturing users’ temporal preferences from historical interaction sequences. However, they often lack the ability to leverage rich semantic information about items, limiting both their accuracy and explainability. In this work, we propose a novel framework that enhances sequential recommendation by integrating a Knowledge Graph-based Intent Network (KGIN) through a late-fusion strategy. KGIN models latent user intents as attentive combinations of knowledge graph relations and performs relational path-aware aggregation to encode long-range semantic dependencies into user and item representations. By combining the sequential model’s output with KGIN’s knowledge-aware reasoning at the score level, our framework enriches recommendation signals without modifying the underlying architecture. Experiments on three benchmark datasets—MovieLens, Amazon-Book, and Last-FM—demonstrate that KGIN-enhanced models consistently outperform their original counterparts in terms of Recall@K, NDCG@K, and MRR@K.