Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs
The Graph Geometry-aware MLP (G2MLP) framework addresses spectral failure modes when distilling knowledge from Graph Neural Networks (GNNs) to Multi-Layer Perceptrons (MLPs). It uses Ollivier–Ricci curvature to guide supervision, allocating training effort based on local graph geometry. This results in improved performance across benchmarks while maintaining the fast, graph-free inference speed of a standard MLP.