基于神经网络预测器的图神经网络架构搜索

Neural Architecture Search (NAS) has proven successful in discovering promising architectures for vision or language modeling tasks, and was also recently introduced to search Graph Neural Networks (GNNs). Despite initial success, the search strategy efficiency needs to be improved. In this paper, the search strategy of the neural network predictor (Neural predictor) in the neural network architecture search is introduced into the GNN architecture search. Qualitative graph embedding methods get architectural embeddings and predict their accuracy. Experiments have been carried out on multiple graph data sets of different types and scales with node classification downstream tasks, and the results have proved to be significantly improved.