IMPRESS: Hyperbolic Space and Diffusion for Graph Few-Shot Learning
A new framework called IMPRESS (IMproving graPh few-shot learning with hypeRbolic spacE and denoiSing diffuSion) addresses limitations in graph few-shot learning. Existing methods perform node representation learning in Euclidean space, failing to capture hierarchical structures in real-world graph data. Additionally, during meta-testing, they fit an empirical target distribution from few support samples, which may deviate from the true distribution. IMPRESS uses hyperbolic space to better model hierarchies and denoising diffusion to refine the target distribution. The approach is detailed in arXiv:2604.27462v1.
Key facts
- IMPRESS is a framework for graph few-shot learning.
- It uses hyperbolic space to capture hierarchical structures.
- It employs denoising diffusion to improve target distribution fitting.
- Existing methods suffer from limitations in Euclidean space.
- The paper is available on arXiv with ID 2604.27462v1.
- Graph few-shot learning focuses on learning from few labeled nodes.
- Meta-training typically uses Euclidean space for node representations.
- Meta-testing often uses empirical distributions from few support samples.
Entities
Institutions
- arXiv