Abstract
Deep learning's success has been widely recognized in a variety of machinelearning tasks, including image classification, audio recognition, and naturallanguage processing. As an extension of deep learning beyond these domains,graph neural networks (GNNs) are designed to handle the non-Euclideangraph-structure which is intractable to previous deep learning techniques.Existing GNNs are presented using various techniques, making direct comparisonand cross-reference more complex. Although existing studies categorize GNNsinto spatial-based and spectral-based techniques, there hasn't been a thoroughexamination of their relationship. To close this gap, this study presents asingle framework that systematically incorporates most GNNs. We organizeexisting GNNs into spatial and spectral domains, as well as expose theconnections within each domain. A review of spectral graph theory andapproximation theory builds a strong relationship across the spatial andspectral domains in further investigation.