Automatic Graph Learning with Evolutionary Algorithms: An Experimental Study

In recent years, automated machine learning (AutoML) has received widespread attention from academia and industry owing to its ability to significantly reduce the threshold and labor cost of machine learning. It has demonstrated its powerful functions in hyperparameter optimization, model selection, neural network search, and feature engineering. Most AutoML frameworks are not specifically designed to process graph data. That is, in most AutoML tools, only traditional neural networks are integrated without using a graph neural network (GNN). Although traditional neural networks have achieved great success, GNNs have more advantages in processing non-Euclidean data (e.g., graph data) and have gained popularity in recent years. However, to the best of our knowledge, there is currently only one open-source AutoML framework for graph learning, i.e., AutoGL. For the AutoGL framework, traditional AutoML optimization algorithms such as grid search, random search, and Bayesian optimization are used to optimize the hyperparameters. Because each type of traditional optimization algorithm has its own advantages and disadvantages, more options are required. This study analyzes the performance of different evolutionary algorithms (EAs) on AutoGL through experiments. The experimental results show that EAs could be an effective alternative to the hyperparameter optimization of GNN.