Effects of Random Errors on Graph Convolutional Networks

Abstract

The use of Graph Convolutional Networks (GCN) has been an emerging trend in the network science research community. While GCN achieves excellent performance in several tasks, there exists an open issue in applying GCN to real-world applications. The issue is the effects of network errors on GCN. Since real-world network data contain several types of noises and errors, GCN is desirable to be less affected by such errors. However, the effects have not been sufficiently evaluated before. In this paper, we analyze the effects of random errors on GCN through extensive experiments. The results show that the node classification accuracy of GCN is decreased only 5% even when 50% of the edges are randomly increased or decreased. Moreover, in terms of false labels, the accuracy of node classification is decreased only 10% even when 20% of the labels are changed.