ADEdgeDrop: Adversarial Edge Dropping for Robust Graph Neural Networks

TL;DR

The paper tackles the robustness and generalization gaps of graph neural networks when trained on noisy or redundant graphs. It presents ADEdgeDrop, a novel framework that couples a downstream GNN with a trainable edge predictor operating on the line graph to adversarially drop edges using a min–max objective and a perturbation bound $ orm{m{ abla}}_p leq eta$, guided by a retention threshold $m{ au}$. By transforming the graph to its line graph, the method yields interpretable edge-level decisions and can be integrated with multiple GNN backbones, showing superior performance and resilience to edge-removal/addition attacks across eight datasets. The empirical results, ablation studies, and convergence analyses demonstrate that ADEdgeDrop learns a sparser, higher-quality graph structure that preserves critical connections, improving both generalization and robustness in practice. These findings suggest a practical, scalable route to robust graph representation learning through adversarial, edge-aware augmentation.

Abstract

Although Graph Neural Networks (GNNs) have exhibited the powerful ability to gather graph-structured information from neighborhood nodes via various message-passing mechanisms, the performance of GNNs is limited by poor generalization and fragile robustness caused by noisy and redundant graph data. As a prominent solution, Graph Augmentation Learning (GAL) has recently received increasing attention. Among prior GAL approaches, edge-dropping methods that randomly remove edges from a graph during training are effective techniques to improve the robustness of GNNs. However, randomly dropping edges often results in bypassing critical edges, consequently weakening the effectiveness of message passing. In this paper, we propose a novel adversarial edge-dropping method (ADEdgeDrop) that leverages an adversarial edge predictor guiding the removal of edges, which can be flexibly incorporated into diverse GNN backbones. Employing an adversarial training framework, the edge predictor utilizes the line graph transformed from the original graph to estimate the edges to be dropped, which improves the interpretability of the edge-dropping method. The proposed ADEdgeDrop is optimized alternately by stochastic gradient descent and projected gradient descent. Comprehensive experiments on six graph benchmark datasets demonstrate that the proposed ADEdgeDrop outperforms state-of-the-art baselines across various GNN backbones, demonstrating improved generalization and robustness.

Figures (7)

• Figure 1: Left (existing methods): random edge dropping with probability $\rho$. Right (our work): edge dropping by an adversarial edge predictor $\xi(\cdot)$.
• Figure 2: The framework of the proposed ADEdgeDrop, which consists of a basic GNN related to the downstream task, and a GNN-based adversarial edge predictor for the edge-dropping process.
• Figure 3: GCN classification accuracy of different GAL methods with varying rates of edge-addition attacks.
• Figure 4: GCN classification accuracy of different GAL methods with varying rates of edge-removal attacks.
• Figure 5: Visualization of retained edges generated by DropEdge, GAUG-O and ADEdgeDrop on the toy example dataset, where pink edges indicate noisy connections.

Theorems & Definitions (2)

• Definition 1
• Definition 2