.png){kind=icon}
Return
IDiffKG: Deterministic and implicit knowledge graph diffusion for recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116660.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge-aware recommendation benefits from external semantics, yet its effectiveness is often limited by noisy and task-irrelevant knowledge graph (KG) structure. Existing diffusion-based approaches provide a principled denoising perspective, but their stochastic reverse trajectories can introduce additional variance and weaken the preservation of recommendation-relevant relational patterns during KG reconstruction. We propose IDiffKG, a deterministic and implicit KG diffusion framework that addresses recommendation-oriented KG refinement through a non-Markovian reverse process. IDiffKG preserves the standard diffusion training objective while replacing stochastic reverse sampling with controllable deterministic reconstruction, calibrating reconstructed relations with collaborative signals, and aligning the original and reconstructed graphs through cross-view contrastive learning. This design makes diffusion explicitly recommendation-oriented rather than merely generative. Experiments on three public benchmarks from multiple domains demonstrate consistent improvements over representative baselines, with relative Recall@20 and NDCG@20 gains of up to 7.71% and 8.15%, respectively. Ablation and sensitivity analyses further indicate that deterministic reconstruction, collaborative calibration, and cross-view alignment play complementary roles, while the proposed formulation offers a controllable accuracy-efficiency trade-off. These results suggest that controllable, behavior-grounded KG refinement is an effective use of implicit diffusion for KG-based recommendation.
Keywords:
Recommendation systems
Knowledge graph
Implicit diffusion
Contrastive learning
Journal
K
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available