GCCM: Enhancing Graph Prediction with Contrastive Consistency Model
A new method called GCCM (Graph Contrastive Consistency Model) addresses shortcut solutions in consistency training for graph prediction. Existing diffusion-based methods require expensive iterative denoising, while consistency training can collapse into deterministic predictors by ignoring noisy targets. GCCM mitigates this by incorporating contrastive learning to maintain sensitivity to noise, improving stability and efficiency. The approach is validated on benchmark datasets, showing competitive results with fewer inference steps.
Key facts
- GCCM stands for Graph Contrastive Consistency Model.
- It targets shortcut solutions in consistency training for graph prediction.
- Diffusion-based methods require expensive iterative denoising.
- Consistency training can collapse into deterministic predictors.
- GCCM uses contrastive learning to maintain noise sensitivity.
- The method improves stability and efficiency.
- Validated on benchmark datasets.
- Achieves competitive results with fewer inference steps.
Entities
Institutions
- arXiv