NAVIGATOR-D3: Neural Architecture Search Using VarIational Graph Auto-encoder Toward Optimal aRchitecture Design for Diverse Datasets

Neural architecture search (NAS) is an automated machine learning method that optimizes neural network architectures depending on the dataset or its purpose. With the advances in NAS, high-accuracy neural network architectures can be built for a specific dataset without any expert skills. However, NAS is an expensive, time-consuming, and resource-intensive technique. Therefore, searching for the optimal architecture from scratch for each new dataset is inefficient. To accommodate the expected future increase in datasets, a technique is required that directly predicts the optimized architecture for unknown datasets. Therefore, we propose a framework that generates architectures for unknown datasets by mapping adequate architectures for existing datasets into the latent feature space. A variational graph autoencoder (VGAE) is utilized for latent feature mapping. Our experimental results indicate that the architecture generated by the proposed method from the information of previously obtained high-accuracy architectures performs effectively for new datasets.