Training data augmentation using generative models with statistical guarantees for materials informatics

doi:10.1007/s00500-021-06533-3

Return

Training data augmentation using generative models with statistical guarantees for materials informatics

2021-12-01

PRE

Hiroshi Ohno *

DOI：10.1007/s00500-021-06533-3

Original

Original request for help

Save

Abstract

En 中文

In materials double dagger science, the small data size problem occurs frequently when applying machine learning algorithms. This issue is primarily due to, for example, real experimental measurement cost and time-consuming simulations based on materials models, e.g., density functional theory. We address the small data issue by generating training data using generative models. The proposed training data augmentation method can generate data using kernel density estimation models while maintaining statistical guarantees using unbiased estimators of linear regression, which has low computational costs and easy hyper-parameter tuning compared to deep neural networks. In addition, we derive an upper bound for the L-1 loss of the proposed method relative to probability density functions. Experiments were conducted with four benchmark and three materials (i.e., binary compounds, oxide ionic conductivity, and phosphorescent materials) datasets. Regarding the generated data, we examined training and generalization performances using kernel ridge regression compared to those of generative adversarial networks and real-NVPs. For the materials datasets, we analyzed influential factors regarding material properties (conductivity and emission intensity). The experimental results demonstrate that the proposed method can generate data that can be used as new training data.

Keywords:

Linear regression model

Density estimation

Data augmentation

Generative models

Materials informatics