Multi-scale Heat Kernel Graph Network for Graph Classification

Graph neural networks (GNNs) have been shown to be useful in a variety of graph classification tasks, from bioinformatics to social networks. However, most GNNs represent the graph using local neighbourhood aggregation. This mechanism is inherently difficult to learn about the global structure of a graph and does not have enough expressive power to distinguish simple non-isomorphic graphs. To overcome the limitation, here we propose multi-head heat kernel convolution for graph representation. Unlike the conventional approach of aggregating local information from neighbours using an adjacency matrix, the proposed method uses multiple heat kernels to learn the local information and the global structure simultaneously. The proposed algorithm outperforms the competing methods in most benchmark datasets or at least shows comparable performance.