Dynamic imputation for improved training of neural network with missing values

To train a neural network with an incomplete dataset containing missing values, the dataset is required to be completed in advance. The conventional approach applies missing value imputation before training, and consistently uses the same imputations to update the neural network. However, some inaccurate imputations would negatively affect the generalization performance of the neural network. In this study, we propose a dynamic imputation method to improve the training of a neural network in the presence of missing values. We use a dynamic imputer that can provide different imputations for the same missing value. At each training epoch, imputations for all missing values are newly obtained using the dynamic imputer, and are then used to update the neural network. By diversifying the imputations with the dynamic imputer during the training, the neural network is expected to exhibit more robustness to imputation inaccuracies, thereby improving the generalization performance. The effectiveness of the proposed method was demonstrated on benchmark datasets with various missing rates.