How graph features from message passing affect graph classification and regression?

Abstract

Graph neural networks (GNNs) have been applied to various graph domains. However, GNNs based on the message passing scheme, which iteratively aggregates information from neighboring nodes, have difficulty learning to represent larger subgraph structures because of the nature of the scheme. We investigate the prediction performance of GNNs when the number of message passing iteration increases to capture larger subgraph structures on classification and regression tasks using various real-world graph datasets. Our empirical results show that the averaged features over nodes obtained by the message passing scheme in GNNs are likely to converge to a certain value, which significantly deteriorates the resulting prediction performance. This is in contrast to the state-of-the-art Weisfeiler–Lehman graph kernel, which has been used actively in machine learning for graphs, as it can comparably learn the large subgraph structures and its performance does not usually drop significantly drop from the first couple of rounds of iterations. Moreover, we report that when we apply node features obtained via GNNs to SVMs, the performance of the Weisfeiler-Lehman kernel can be superior to that of the graph convolutional model, which is a typically employed approach in GNNs.