A Dual-branch Graph Convolutional Network on Imbalanced Node Classification

Abstract

Graph convolutional neural networks (GCNs) have attracted much attention in dealing with various node classification tasks on graphs. Some real-world node classification tasks face the situation that the number of minority class nodes is significantly less than that of majority class nodes. This makes us more concerned about how to effectively solve the problem of imbalanced node classification based on GCNs. To solve this problem, we propose a Dual-branch Graph Convolutional Network framework (D-GCN), which can reduce the dominant effect of majority class on topology aggregation and the negative impact of information differences caused by graph structure reconstruction. This framework achieves the goal of decreasing the possibility of misrecognizing the minority class nodes as majority class and improving the classification performance of minority class nodes. Experiments on several graph datasets demonstrate that D-GCN outperforms representative baselines in solving imbalanced node classification tasks.