Hierarchical Restructuring of Graph Neural Networks for Robust Learning under Edge Perturbations

Zhengdao Li, Kai Hwang

Keywords: Hierarchical Variational Inference, graph neural networks, graph structural learning

Abstract: Graph Neural Networks (GNNs) are central to graph representation learning, yet their robustness is challenged by structural perturbations in graphs, leading to suboptimal analytical outcomes. These perturbations, common in real-world graphs due to factors like adversarial interferences, result in noisy and incomplete data. Addressing this issue, we propose the hierarchical restructuring (HR) framework, utilizing a hierarchical Bayesian model to capture these latent disruptions. Our framework is uniquely adaptable as a plug-in for various GNN variants, optimizing a hierarchical variational lower bound alongside downstream task training. The HR-enhanced models show superior performance in node-level classification, graph classification, and spatial-temporal graph classification tasks. The results indicate accuracy gains in a range of 3% to 21% under a 90% perturbation ratio for node classification tasks and up to 38\% for graph classification tasks under a 50% perturbation ratio. These findings underscore the effectiveness of our framework in enhancing the robustness and accuracy of GNNs in the presence of structural perturbations.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 5908