Effective-aggregation Graph Convolutional Network for Imbalanced Classification

Abstract

Classification is a common task and can be achieved by learning a predictive model from a labeled training dataset. However, the imbalanced data distribution makes the model tend to favor the majority class, which reduces the classification performance. Unlike traditional classification models, graph convolutional networks (GCNs) can extract useful feature information from unlabeled data. In this paper, a novel framework for imbalanced classification named effective-aggregation graph convolutional network (EGCN) is proposed. First, a graph generator constructs graph-structured data using both labeled and unlabeled data. Then, an aggregation control unit (ACU) is performed to improve the effectiveness of aggregation. ACU uses local estimation density to limit the aggregation of inter-class edges from a local perspective, and it enhances the aggregation of the minority class from a global perspective based on the imbalance ratio. Finally, the prediction results are obtained by a graph convolutional network. Experimental results on several real-world datasets show that EGCN has promising performance.