Training data augmentation: An empirical study using generative adversarial net-based approach with normalizing flow models for materials informatics

Abstract We address the issue of small data size for training models for regression problems, which is a significant issue in materials science. Many density estimators that use generative models based on deep neural networks have been proposed. With generative models, normalizing flows can provide exact density estimations. Using normalizing flows, we address training data augmentation issue, where we use a real-valued non-volume preserving model (real-NVP) as the normalizing flow. A generative adversarial net (GAN)-based training method is applied to improve real-NVP training using real-NVP as the generator. Using kernel ridge regression trained by generated data, generalization performance was measured for evaluating the models. Experiments were conducted with seven benchmark datasets and a dataset of ionic conductivity of materials to compare the GAN-based real-NVP to state-of-the-art models, such as real-NVP and masked autoregressive flows. The experimental results demonstrated that the GAN-based real-NVP was comparable to state-of-the-art models and implied that the data sampled by the GAN-based real-NVP were available as new training data.