Training Sensitivity in Graph Isomorphism Network

TL;DR

This paper investigates how different training configurations affect the performance of Graph Isomorphism Networks (GINs) across various datasets, revealing that common techniques may not always be optimal.

Contribution

It provides an empirical analysis of training sensitivities in GINs, highlighting the impact of different functions and configurations on their ability to learn graph structures.

Findings

Certain training techniques outperform others depending on graph structure

Commonly used methods do not always capture graph features effectively

Performance varies significantly across different benchmark datasets

Abstract

Graph neural network (GNN) is a popular tool to learn the lower-dimensional representation of a graph. It facilitates the applicability of machine learning tasks on graphs by incorporating domain-specific features. There are various options for underlying procedures (such as optimization functions, activation functions, etc.) that can be considered in the implementation of GNN. However, most of the existing tools are confined to one approach without any analysis. Thus, this emerging field lacks a robust implementation ignoring the highly irregular structure of the real-world graphs. In this paper, we attempt to fill this gap by studying various alternative functions for a respective module using a diverse set of benchmark datasets. Our empirical results suggest that the generally used underlying techniques do not always perform well to capture the overall structure from a set of graphs.