.png){kind=icon}
.png){kind=icon}
Return
Noisy or structural? Knowledge graph-based dual-view diffusion for recommendation systems
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.115607.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge graph-based (KG-based) recommendation systems aim to extract task-relevant information from the knowledge graph (KG) and enhance the effectiveness of recommendations. Diffusion models denoise irrelevant triples from KG and have shown impressive performance in KG-based recommendation. However, diffusion models may lead to excessive elimination of some task-relevant triples. To address this problem, we propose a KG-based Dual-view Diffusion model for Recommendation systems (KDDRec). Specifically, we generate a structured noise that is similar to the KG distribution by a Variational Autoencoder. It perturbs KG with random noise from a different perspective. Additionally, we inject random noise and structured noise into two independent diffusion models, respectively. Then, we apply these two diffusion models to diffuse and denoise KG, respectively. Finally, merging the triples from these two denoising processes effectively prevents the over-denoising issue. Experimental evaluations demonstrate that KDDRec performs excellently in overcoming these challenges. Compared to the existing state-of-the-art recommendation model, KDDRec improves NDCG@20 by 8.99% on the Last-FM dataset and by 9.44% on the Alibaba-iFashion dataset. The code is available at https://github.com/HubuKG/KDDRec .
Keywords:
Knowledge graph
Diffusion model
Recommendation system
Dual-view
Over-denoising