Structural mixup for graph neural networks

Abstract

Recent mixup-based augmentations have shown promising results in various domains, including computer vision. However, its application in the graph domain remains a challenge due to the inherent structural information in graphs. While recent studies have attempted to address such scenarios, they mostly concentrate on graph classification tasks. The research in node classification, on the other hand, is still under-explored. In this paper, we propose a novel mixup augmentation especially designed for the node classification in the graph domain called Structural Mixup (S-Mixup). The core idea is to take into account the structural information while mixing nodes. S-Mixup first passes through a classic Graph Neural Network (GNN) classifier to obtain pseudo-labels along with their prediction confidence. These serve as the criteria for the composition of the mixup pool for both inter and intra-class mixups. Furthermore, we utilize the edge gradient obtained from GNN training to suggest an edge gradient-based neighborhood edge selection for the newly generated nodes. Through extensive experiments on real-world benchmark datasets, we demonstrate the effectiveness of our mixup method evaluated on the node classification task. We observe that S-Mixup enhances the robustness and generalization performance of GNNs, especially in heterophilous situations, making it a valuable tool for graph-based learning tasks.