Genetic-GNN: Evolutionary architecture search for Graph Neural Networks

Neural architecture search (NAS) has seen significant attention throughout the computational intelligence research community and pushed forward state-of-the-art of many neural models to address grid-like data such as texts images. However, little work been done on Graph Network (GNN) dedicated unstructured network data. Given huge number choices combinations components aggregators activation functions, determining suitable GNN model for a specific problem normally necessitates tremendous expert knowledge laborious trials. In addition, moderate change hyperparameters learning rate dropout would dramatically impact capacity model. this paper, we propose novel framework through evolution individual in large searching space. Instead simply optimizing structures, an alternating process is performed between structures dynamically approach optimal fit each other. Experiments validations demonstrate that evolutionary NAS capable matching existing reinforcement methods both transductive inductive graph representation node classification.