知识图谱与大语言模型的语义增强推荐方法

A Semantic-Enhanced Recommendation Method Integrating Knowledge Graphs and Large Language Models

摘要: 针对传统知识图谱推荐方法在异构知识融合困难、局部语义挖掘不足和复杂关联关系建模能力有限等问题，提出一种知识图谱与大语言模型的语义增强推荐方法LLM-KGRec。首先，利用大语言模型对多源异构知识图谱进行语义标准化，统一实体表达与关系表示。其次，围绕候选物品构建局部知识子图，并借助大语言模型挖掘局部知识的深层语义信息。进一步，引入全局跨源语义检索机制，补充候选物品的外部语义上下文。最后，融合交互特征、知识图谱结构特征与多粒度语义特征，实现推荐预测。实验表明，所提方法在多个公开数据集上的NDCG@10, Recall@10等指标均优于对比模型，在数据稀疏和复杂语义场景下表现出较好的鲁棒性与泛化能力。

Abstract: To address the limitations of traditional knowledge graph-based recommendation methods in heterogeneous knowledge fusion, local semantic mining, and complex relational modeling, a semantic-enhanced recommendation method integrating knowledge graphs and large language models, named LLM-KGRec, is proposed. First, a large language model is employed to perform semantic standardization on multi-source heterogeneous knowledge graphs, so as to unify entity expressions and relation representations. Second, local knowledge subgraphs are constructed around candidate items, and the large language model is used to capture the deep semantic information embedded in local structures. Furthermore, a global cross-source semantic retrieval mechanism is introduced to supplement external semantic context for candidate items. Finally, interaction features, structural features of the knowledge graph, and multi-granularity semantic features are fused to achieve recommendation prediction. Experimental results on multiple public datasets show that the proposed method outperforms baseline models in terms of NDCG@10, Recall@10, and other evaluation metrics, and demonstrates better robustness and generalization ability in sparse-data, cold-start, and complex semantic scenarios.