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

Heterogeneous KG fusion difficulty: Traditional knowledge-graph-based recommendation methods have difficulty fusing heterogeneous knowledge.

Insufficient local semantic mining: Traditional knowledge-graph-based recommendation methods insufficiently mine local semantic information.

Limited complex relation modeling: Traditional knowledge-graph-based recommendation methods have limited ability to model complex relational associations.

Sparse and cold-start recommendation robustness: Recommendation methods need to perform robustly in sparse-data, cold-start, and complex semantic scenarios.

Heterogeneous KG fusion difficulty: Traditional knowledge-graph-based recommendation methods have difficulty fusing heterogeneous knowledge.

LLM-KGRec semantic-enhanced KG recommendation: The paper proposes LLM-KGRec, a semantic-enhanced recommendation method integrating knowledge graphs and large language models.

LLM-based heterogeneous KG semantic standardization: The method uses a large language model to semantically standardize multi-source heterogeneous knowledge graphs and unify entity and relation representations.

LLM-enhanced local knowledge subgraphs: The method constructs local knowledge subgraphs around candidate items and uses an LLM to capture deep semantic information in local structures.

Global cross-source semantic retrieval: The method introduces a global cross-source semantic retrieval mechanism to add external semantic context for candidate items.

Multi-feature recommendation prediction fusion: The method fuses interaction features, knowledge-graph structural features, and multi-granularity semantic features for recommendation prediction.

Improved ranking accuracy over baselines:

Better robustness and generalization:

Improved ranking accuracy over baselines: