Expanding Training Set for Graph-Based Semi-supervised Classification

Abstract

Graph Convolutional Networks (GCNs) have made significant improvements in semi-supervised learning for graph structured data and have been successfully used in node classification tasks in network data mining. So far, there have been many methods that can improve GCNs, but only a few works improved it by expanding the training set. Some existing methods try to expand the label sets by using a random walk that only considers the structural relationships or selecting the most confident predictions for each class by comparing the softmax scores. However, the spatial relationships in low-dimensional feature space between nodes is ignored. In this paper, we propose a method to expand the training set by considering the spatial relationships in low-dimensional feature space between nodes. Firstly, we use existing classification methods to predict the pseudo-label information of nodes, and use such information to compute the category center of nodes which has the same pseudo label. Then, we select the k nearest nodes of the category center to expand the training set. At last, we use the expanded training set to reclassify the nodes. In order to further verify our proposed method, we randomly select the same number of nodes to expand the training set, and use the expanded training set to reclassify nodes. Comprehensive experiments conducted on several public data sets demonstrate effectiveness of the proposed method over the state-of-art methods.