Revisiting Edge Perturbation for Graph Neural Network in Graph Data Augmentation and Attack

TL;DR

This work reveals that edge perturbation for graph neural networks can be cast as a single optimization problem that subsumes both augmentation (Gaug) and attack (Gatk) under different constraints. The authors introduce Edge Priority Detector (EPD), a modular framework with two solutions, to compute edge perturbation priorities and enable flexible augmentation or attack with reduced computational overhead. Theoretical unification is complemented by extensive experiments across multiple GNN backbones and real-world graphs, showing that EPD achieves comparable or superior performance to existing methods while significantly lowering time costs. Overall, the approach provides a principled boundary between augmentation and attack, and offers a practical tool for controlled graph perturbations and future method design.

Abstract

Edge perturbation is a basic method to modify graph structures. It can be categorized into two veins based on their effects on the performance of graph neural networks (GNNs), i.e., graph data augmentation and attack. Surprisingly, both veins of edge perturbation methods employ the same operations, yet yield opposite effects on GNNs' accuracy. A distinct boundary between these methods in using edge perturbation has never been clearly defined. Consequently, inappropriate perturbations may lead to undesirable outcomes, necessitating precise adjustments to achieve desired effects. Therefore, questions of ``why edge perturbation has a two-faced effect?'' and ``what makes edge perturbation flexible and effective?'' still remain unanswered. In this paper, we will answer these questions by proposing a unified formulation and establishing a clear boundary between two categories of edge perturbation methods. Specifically, we conduct experiments to elucidate the differences and similarities between these methods and theoretically unify the workflow of these methods by casting it to one optimization problem. Then, we devise Edge Priority Detector (EPD) to generate a novel priority metric, bridging these methods up in the workflow. Experiments show that EPD can make augmentation or attack flexibly and achieve comparable or superior performance to other counterparts with less time overhead.

Figures (10)

• Figure 1: Illustrations of edge perturbation methods in GNNs.
• Figure 2: Variations of test accuracy when applying Gaug and Gatk methods on three datasets under different EPR.
• Figure 3: Statistics on the maximal reported EPR among typical methods. The first four from left to right are Gatk methods, and the rest are Gaug methods.
• Figure 4: Comparisons of the accuracy of GNN models (2-layer GNN and 8-layer GNN) with different Gaug and Gatk methods applied. For each dataset, the only variate is the number of layers of GNN models as we keep all other parameters regarding GNN models and training configurations the same, including GNN backbones, learning rate, EPR, etc.
• Figure 5: Gaug and Gatk methods are bridged theoretically and practically by proposing unified priority metrics. The metrics will be leveraged in EPD to efficiently perturb edges, enabling flexible effects of attack or augmentation.