GraphSR: A Data Augmentation Algorithm for Imbalanced Node Classification

Graph neural networks (GNNs) have achieved great success in node classification tasks. However, existing GNNs naturally bias towards the majority classes with more labelled data and ignore those minority relatively few ones. The traditional techniques often resort over-sampling methods, but they may cause overfitting problem. More recently, some works propose to synthesize additional nodes for from nodes, however, there is no any guarantee if generated really stand corresponding classes. In fact, improperly synthesized result insufficient generalization of algorithm. To resolve problem, this paper we seek automatically augment massive unlabelled graph. Specifically, \textit{GraphSR}, a novel self-training strategy significant diversity which based on Similarity-based selection module Reinforcement Learning(RL) module. first finds subset are most similar second one further determines representative reliable via RL technique. Furthermore, RL-based can adaptively determine sampling scale according current training data. This general be easily combined different models. Our experiments demonstrate proposed approach outperforms state-of-the-art baselines various class-imbalanced datasets.