Clustering Method based on VGAE Improving Interpretability of Search Results of Neural Architecture Search

Abstract

Neural Architecture Search (NAS) is a method of AutoML that the optimization of neural network models according to the given data and objectives. NAS needs a lot of time to search, and it is not efficient to search the network models from scratch for each new task. In other words, the improvement of NAS efficiency can be achieved by searching the architecture of similar models. The clustering and visualization of models based on their features are considered viable methods to search for similar network models. However, the quantitative classification of models based on their features is a challenge, and there is no widely known approach to visualize the features of models searched by NAS. Therefore, this study aims to obtain latent features of network models through a machine learning approach incorporating Variational Graph Auto-Encoders (VGAE), which is one of the graph neural network methods. As a result, the proposed method enables the search for similar network models and the comparisons between models.