RobGC: Towards Robust Graph Condensation

TL;DR

RobGC addresses the vulnerability of graph condensation to structural noise by coupling condensation with a training‑free denoising procedure that uses the condensed graph as a denoising signal. It introduces reliable graph structure modeling and label‑propagation guided alternating optimization to refine both the condensed graph and the training graph, plus test‑time denoising for inductive nodes. The method is compatible with multiple GC baselines, improves robustness under random and adversarial noise, and generalizes across several GNN architectures, while maintaining practical runtime. This data‑centric robustness approach enhances GC applicability in real‑world, noisy graph environments and offers a scalable path for reliable inductive graph inference.

Abstract

Graph neural networks (GNNs) have attracted widespread attention for their impressive capability of graph representation learning. However, the increasing prevalence of large-scale graphs presents a significant challenge for GNN training due to their computational demands, limiting the applicability of GNNs in various scenarios. In response to this challenge, graph condensation (GC) is proposed as a promising acceleration solution, focusing on generating an informative compact graph that enables efficient training of GNNs while retaining performance. Despite the potential to accelerate GNN training, existing GC methods overlook the quality of large training graphs during both the training and inference stages. They indiscriminately emulate the training graph distributions, making the condensed graphs susceptible to noises within the training graph and significantly impeding the application of GC in intricate real-world scenarios. To address this issue, we propose robust graph condensation (RobGC), a plug-and-play approach for GC to extend the robustness and applicability of condensed graphs in noisy graph structure environments. Specifically, RobGC leverages the condensed graph as a feedback signal to guide the denoising process on the original training graph. A label propagation-based alternating optimization strategy is in place for the condensation and denoising processes, contributing to the mutual purification of the condensed graph and training graph. Additionally, as a GC method designed for inductive graph inference, RobGC facilitates test-time graph denoising by leveraging the noise-free condensed graph to calibrate the structure of the test graph. Extensive experiments show that RobGC is compatible with various GC methods, significantly boosting their robustness under different types and levels of graph structural noises.

Figures (6)

• Figure 1: The graph condensation process and inductive inference scenario within the noisy environment. (a) In the training procedure, the noisy training graph is condensed and then the condensed graph is leveraged to train multiple GNN models with different architectures. (b) In the inference stage, new graph is integrated into the noisy training graph, and GNNs are then deployed on this expanded noisy graph for inference.
• Figure 2: The test accuracy of GCN DBLP:conf/iclr/KipfW17 trained on the condensed graphs. GCond jin2022graph is adopted to condense the graph and random noise is applied. (a) The performance of the GCN trained on condensed graphs derived from the training graph with different levels of noise and tested on a clean graph. (b) The results of the GCN trained on a clean condensed graph and tested on graphs containing varying levels of noise.
• Figure 3: The training and test procedures of RobGC. During the training stage, the noisy training graph is condensed, and the condensed graph is utilized as a denoising signal to optimize the structure of the training graph. In the inference stage, the noisy test graph is denoised by referring the condensed graph. Features are omitted in test-time denoising for simplicity.
• Figure 4: The accuracy of the model trained on the original and condensed graphs throughout our optimization procedure. The GCond and random noise are applied and the noise level is 100%. The condensation ratios are 5.80%, 3.30%, 0.17%, and 0.50% for datasets, respectively.
• Figure 5: The hyper-parameter sensitivity analysis for the number of selected nearest neighbors, hops, and searching candidate on Citeseer dataset.