Attention and Cost-Sensitive Graph Neural Network for Imbalanced Node Classification

Abstract

Semi-supervised node classification of imbalanced graphs is one of the important tasks in the field of graph neural network (GNN). Most of the current methods focus on how to aggregate feature information from neighbor nodes, but they do not distinguish the importance of minority class and majority class samples in the process of aggregation. To this end, this paper introduces an attention mechanism in the process of aggregating feature information, which flexibly assigns individualized weights to minority and majority class samples. At the same time, we improve the loss function using cost-sensitive techniques to increase the minority class misclassification cost. Finally, we design experiments to verify the effectiveness of the proposed method.