Fitness Landscape Analysis of Graph Neural Network Architecture Search Spaces

Neural Architecture Search (NAS) is the name given to a set of methods designed to automatically configure the layout of neural networks. Their success on Convolutional Neural Networks inspired its use on optimizing other types of neural network architectures, including Graph Neural Networks (GNNs). GNNs have been extensively applied over several collections of real-world data, achieving state-of-the-art results in tasks such as circuit design, molecular structure generation and anomaly detection. Many GNN models have been recently proposed, and choosing the best model for each problem has become a cumbersome and error-prone task. Aiming to alleviate this problem, recent works have proposed strategies for applying NAS to GNN models. However, different search methods converge relatively fast in the search for a good architecture, which raises questions about the structure of the problem. In this work we use Fitness Landscape Analysis (FLA) measures to characterize the search space explored by NAS methods for GNNs. We sample almost 90k different architectures that cover most of the fitness range, and represent them using both a one-hot encoding and an embedding representation. Results of the fitness distance correlation and dispersion metrics show the fitness landscape is easy to be explored, and presents low neutrality.