Graph Neural Networks with Stability and Discernability

Abstract

Graph neural networks (GNNs) are popular for solving graph-related tasks in fields ranging from bioinformatics to social networks. The essential mechanism of a GNN model is local-neighborhood aggregation. However, it is innately difficult to aggregate the global information of a graph even though such information plays a key role in graph-classification tasks such as molecular-function prediction and community detection in social networks. Moreover, the local-aggregation mechanism suffers from limited expressive power, making the model unstable and difficult to distinguish non-isomorphic graphs. To overcome these limitations, we propose a GNN equipped with heat kernel convolution for gathering global information while ensuring stability and heat kernel trace normalization for improving the expressive power. We validated our model against previous methods using various benchmark graph datasets, and the experimental results demonstrate that the proposed GNN's performance is comparable to those of state-of-the-art methods.