Edge utilization in graph convolutional networks for graph classification

Graph convolutional neural networks are designed to apply convolutional operations directly on non-Euclidean structure graph data, generating orderly arranged matrix representations of graphs. However, only node features are fully exploited even though edge features may also play an important role in some domains such as chemoinformatics. In this paper, we proposed two new approaches of utilizing edge features on graph convolutional neural networks, Feature embedding adjacent matrix and Reverse graph. Methodologies of basic graph convolutional neural networks only tend to propagate node features to neighbor nodes along edges by convolutional operations. By applying Feature embedding adjacent matrix, edge features are synthesized into node features and also propagated to neighbor nodes during propagation process. Reverse graph approach builds a special auxiliary graph to propagate edge features to neighbor edges. Therefore, a synthetical presentation including both edge features and node features is built. Experiments demonstrated our new approaches improve graph classification accuracies, especially on data sets with low accuracies on basic GCNs.