A. Shokurov

Data Augmentation for Machine Learning on Missing Data

It is proved that drop-and-replace augmentation allows one to maximize balanced accuracy when learning on small linearly separable datasets containing missing values, and is more efficient when learning on missing data.