.png){kind=icon}
.png){kind=icon}
Return
Information entropy-guided knowledge graph recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.115539.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge graphs (KGs) are a crucial auxiliary technique in modern recommender systems for incorporating rich external information and capturing complex entity relationships. However, their inherent noise can obscure key semantic information, diminishing the effectiveness of KG-based recommendations. Additionally, effectively modeling and leveraging the varying importance of different relation types remains a significant research challenge. Information entropy, as a measure of uncertainty or redundancy, can quantify the diversity of entity connections as well as the semantic information capacity it carries, facilitating the identification of structurally uncertain or connections with low semantic information. Motivated by this, we propose an Information Entropy-guided Knowledge Graph Recommendation (IEKGR) framework. Specifically, we employ information entropy to quantify the semantic richness of entities and their connections within the KG. By combining user behavior data, we adaptively generate personalized weights to guide a conditional diffusion model for denoising the original knowledge graph. This guidance mechanism enables the model to selectively identify and filter out noisy edges, while reinforcing connections indicative of user interests or item attributes, thus producing a recommendation-relevant denoised knowledge subgraph. Furthermore, to fully capture the diverse and complex relational information in the KG, we design a Relation-aware Graph Attention Network (RGAN). RGAN dynamically assigns attention weights according to the types of relations between entities, thereby enabling more accurate learning of entity and relation representations for recommendation tasks. Experimental results demonstrate that the proposed IEKGR framework outperforms several state-of-the-art baselines on public datasets including MIND, Last-FM, and Alibaba-iFashion.
Keywords:
Knowledge graph
Information entropy
Recommender systems
Graph denoising
Relation-aware attention