AUGKD: Adaptive Universal Graph Knowledge Distillation

Abstract

Graph data distillation frameworks have shown great potential in compressing graph representations and improving learning efficiency. Nevertheless, their robustness and generalization remain limited when applied to different graph neural network architectures. To address these limitations, we propose Adaptive Universal Graph Knowledge Distillation Framework (AUGKD). AUGKD adaptively enhances node representations by adapting to dataset heterogeneity and designing an importance-based node feature enhancement strategy. Then we introduce a label propagation–guided feature extraction module within the multilayer perceptron to mitigate over-smoothing, and employs a weight-adaptive mechanism that enables robust generalization to heterophilic graphs through adaptive and negative propagation coefficients. Experiments on multiple benchmark datasets demonstrate that AUGKD achieves robust and effective performance on both homophilic and heterophilic graphs benchmark datasets, significantly outperforming the baselines.