Graph Neural Network with Adaptive Node-Edge Augmentation Strategies for Imbalanced Node Classification

Graph neural networks (GNNs) play an increasingly significant role in node classification. Most GNNs have been conducted on balanced datasets. However, real-world graphs often reflect imbalanced distributions, with some classes in the graph having a much larger number of nodes than others, leading to accuracy degradation remarkably when applying GNNs. Although some methods attempt to use the interpolation strategy to improve imbalance node classification, they still lack flexibility. This paper presents a graph neural network with adaptive node-edge augmentation strategies (GraphANAS), effectively promoting the performance of imbalanced node classification. Specifically, all nodes in the target minority class are considered to enhance the diversity of node features. Then, we construct an adaptive node generator using the neural network's learnable ability to adaptively produce new nodes with a more diverse set of features. Further, an adaptive edge generator is devised to discriminate whether there is a connection between two nodes and form the final adjacency matrix of the augmentation graph. Experiments on several datasets reveal that GraphANAS has a significant improvement over baselines. GraphANAS offers a plug-and-play adaptive node-edge augmentation strategy for imbalanced node classification.