CSIML: a cost-sensitive and iterative machine-learning method for small and imbalanced materials data sets

Abstract

Materials science research benefits from the powerful machine-learning (ML) surrogate models, but it is also limited by the implicit requirement for sufficiently big and balanced data distribution for ML. In this paper, we propose a model to obtain more credible results for small and imbalanced materials data sets as well as chemical knowledge. Taking 2 bandgaps imbalanced data sets as instances, we demonstrate the usability and performance of our model compared with common ML models with normal sampling and resampling methods.