Can higher-order structural features improve the performance of graph neural networks for graph classification?

Abstract

Graph classification is a problem with applications in many different domains, which classifies a collection of graphs with categorical labels. One of the increasingly popular approaches to classify graphs is to use graph neural networks (GNNs), which capture the dependence of graph elements via message passing between the nodes. The key idea is to represent graphs in low-dimensional vectors by collectively aggregating node information guided by the graph structure. There are two types of information that can be used for graph classification. One is the textual features associated with each node in a graph, such as the keywords of a publication in a citation network. The other is the structural features that capture the higher-order dependencies between graph nodes. In this paper, we present a GNN-based graph classification framework that utilizes both textual and structural features, where the structural features of each node is calculated based on a set of small induced subgraphs (i.e., graphlets). We carry out experiments on several well-known graph-structured data sets, i.e., DD, MUTAG, NCI1, ENZYMES, and PROTEINS. By comparing with the state-of-the-art graph convolutional networks (GCNs), i.e., the spectral-based GCNs, the graph attention networks (GAT), and the transformer-based GCNs, we evaluate the effectiveness of involving graphlet-based structural features on the task of graph classification. The results also show that the transformer-based GCN, which integrates higher-order structural features as input, can significantly improve the accuracy of graph classification.