Community detection robustness of graph neural networks

2025-09-29

AI Summary

This study systematically evaluates the robustness of six mainstream graph neural networks (GNNs)—GCN, GAT, GraphSAGE, DiffPool, MinCUT, and DMoN—in attributed network community detection under three threat categories: node attribute perturbation, topological disruption, and adversarial attacks. Experiments reveal that supervised GNNs achieve high accuracy but exhibit poor robustness, whereas the unsupervised model DMoN demonstrates superior resilience under adversarial perturbations. Community strength emerges as a critical determinant of robustness, and joint attribute-structure perturbations induce the most severe performance degradation. The work provides the first empirical characterization of differential impacts across perturbation types on community recovery capability and validates element-level similarity as an effective metric for robustness assessment. These findings establish an empirical benchmark for robust community detection and yield actionable design principles for developing resilient GNN-based community detection methods.

Technology Category

• Data Mining & Knowledge Management: Graph Mining, Social Network Analysis & Community
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Social Networks and Social Media: Influence propagation, information diffusion, and the prediction on networks

Abstract

Graph neural networks (GNNs) are increasingly widely used for community detection in attributed networks. They combine structural topology with node attributes through message passing and pooling. However, their robustness or lack of thereof with respect to different perturbations and targeted attacks in conjunction with community detection tasks is not well understood. To shed light into latent mechanisms behind GNN sensitivity on community detection tasks, we conduct a systematic computational evaluation of six widely adopted GNN architectures: GCN, GAT, Graph- SAGE, DiffPool, MinCUT, and DMoN. The analysis covers three perturbation categories: node attribute manipulations, edge topology distortions, and adversarial attacks. We use element-centric similarity as the evaluation metric on synthetic benchmarks and real-world citation networks. Our findings indicate that supervised GNNs tend to achieve higher baseline accuracy, while unsupervised methods, particularly DMoN, maintain stronger resilience under targeted and adversarial pertur- bations. Furthermore, robustness appears to be strongly influenced by community strength, with well-defined communities reducing performance loss. Across all models, node attribute perturba- tions associated with targeted edge deletions and shift in attribute distributions tend to cause the largest degradation in community recovery. These findings highlight important trade-offs between accuracy and robustness in GNN-based community detection and offer new insights into selecting architectures resilient to noise and adversarial attacks.

Problem

• Evaluating GNN robustness against perturbations in community detection tasks
• Analyzing how different attacks affect GNN-based community identification performance
• Identifying trade-offs between accuracy and resilience in graph neural networks

Innovation

• Evaluated six GNN architectures systematically
• Tested robustness under three perturbation categories
• Used element-centric similarity as evaluation metric