A systematic analysis and guidelines of graph neural networks for practical applications

A graph neural network (GNN) draws attention to deal with many problems in social networks and bioinformatics, as graph data proliferate in a wide variety of applications. Despite the large amount of investigation, it is still difficult to choose the most suitable method for a given problem due to the lack of a thorough analysis on the feasible methods. An anatomical comparison of GNNs would help to devise a prospective method for better solution to real-world problems. In order to give guidelines to make full use of the GNN for graph classification, this paper attempts to analyze the state-of-the-art methods of the GNN and provide practicable guidelines for applications. The representative methods are described with a systematic scheme in four phases for GNN: 1) preprocessing, 2) aggregation, 3) readout, and 4) classification with graph embedding, resulting in a large coverage of more than 1300 methods. The 13 well-known benchmark datasets are categorized into three types with respect to the properties of graph data such as connectivity. In total, more than 3600 runs are executed to systematically analyze and compare the GNN models while changing only one method for each phase. Experimental reproducibility and replicability are also verified by comparing the results with the performance from the literature. Finally, five guidelines for an appropriate model are deduced according to the graph characteristics such as complexity on connectivity and node feature.

Automated summary

An anatomical comparison of GNNs provides insight into devising better solutions. By systematically analyzing and comparing thousands of runs, researchers deduced five guidelines for an appropriate model. The study aims to guide the use of GNN for graph classification and validation of experimental reproducibility and replicability.