Geometric and topological representations in graph neural networks

Abstract

Graph Neural Networks (GNNs) show impressive performance in link-prediction analysis and node classification problems as compared to other neural network approaches. In this paper, the geometric and topological structures of various kinds of node embedding GNNs such as basic GNN, Graph Convolutional Network (GCN), Graph SAmple and aggreGatE (Graph SAGE), and Gated GNN are investigated. Interpretation and comparison between these models are made to provide better comprehension. Sub-graph embedding which is a relatively recent approach is also mentioned in the paper. In particular, two GCN models, i.e. Decagon and convolution spatial graph embedding network (C-SGEN), are studied. In order to enhance the models mentioned, some future works are suggested.