Enhancing GNN learning with node augmentation.

Abstract

Graph Neural Networks (GNNs) have achieved remarkable success across numerous applications. However, their performance often suffers from overfitting and poor generalization when trained on limited or low-diversity graph datasets. Graph data augmentation has been explored as a solution, but most existing approaches rely on static transformations and do not increase the number of training samples. To address this limitation, we propose Node-based Augmentation (NAug), a novel framework that augments the training set by synthesizing new data samples. NAug introduces two key components, an augmented node feature generator and a link placement predictor, jointly trained to produce node features and edges that are semantically and structurally consistent with the local graph context. Unlike prior methods, NAug employs a learning-based generation process and supports multiple downstream graph learning tasks. We evaluate NAug across eleven benchmark datasets spanning node classification, link prediction, and graph classification. Experimental results demonstrate that NAug consistently outperforms state-of-the-art augmentation methods, while also improving model performance on ML safety measures including calibration, robustness to noise, consistency and robustness to adversarial attacks. These findings highlight the effectiveness and versatility of NAug as a general-purpose augmentation framework for GNNs. The code is available at https://github.com/MariaMarrium/NAug.