SGDA: Class Augmentation for Imbalanced Graph Learning at Scale

Graph Neural Networks (GNNs) are widely applied in node classification tasks across social networks, bioinformatics, and healthcare. However, datasets in these applications typically exhibit class imbalance, where the majority classes contain abundant labeled samples while the minority classes have limited annotations. During the training process, nodes from the majority classes are more likely to be labeled. In contrast, nodes from minority classes are more difficult to annotate. This labeling bias causes GNNs to favor majority classes while under-representing minority classes, leading to poor classification performance. This paper proposes Self-supervised Graph Data Augmentation (SGDA) to improve accuracy in imbalanced learning. First, we utilize graph partitioning to identify nodes that require augmentation and generate new nodes with edges based on their neighborhoods, ensuring topological balance. Second, to address the processing of large-scale graphs, we propose mini-batch training where nodes in a batch are selected based on their label distribution. Experimental results show that SGDA achieves up to 0.38 points improvement in balanced accuracy compared to GraphSMOTE and ReNode. The mini-batch training achieves 1.5 \(\times \) speedup over GCN.