Introduction
A large number of recent studies have allowed us to see the powerful potential of the graph neural network model (GNN), and many research teams are constantly improving and building basic modules. But most research uses very small data sets, such as Cora and TU. In this case, even the performance of the non-graph neural network is considerable. If you make further comparisons and use a medium-sized data set, the advantages of graph neural networks can be revealed.
After Stanford graph neural network Daniel Jure and others released "Open Graph Benchmark", another study aimed at building "ImageNet of graph neural network" appeared. Recently, papers from Nanyang Technological University, Loyola Marymount University, University of Montreal, and MILA were submitted to the preprint platform of the paper. In this study, the author introduced six medium-sized benchmark data sets at once. (12k-70k graph, 8-500 nodes), and tested some representative graph neural networks. In addition to the baseline model that only uses node features, graph neural networks are divided into two categories with or without opposite-edge attention. The GNN research community has been seeking a common benchmark to evaluate the capabilities of new models. This tool may allow us to achieve our goals.
