Imputation of missing data with neural networks for classification

Abstract We propose a mechanism to use data with missing values for designing classifiers which is different from predicting missing values for classification. Our imputation method uses an auto-encoder neural network. We make an innovative use of the training data without missing values to train the auto-encoder so that it is better equipped to predict missing values. It is a two-stage training scheme. Unlike most of the existing auto-encoder based methods which use a bottleneck layer for missing data handling, we justify and use a latent space of much higher dimension than that of the input. Now to design a classifier using a training set with missing values, we use the trained auto-encoder to predict missing values based on the hypothesis that a good choice for a missing value would be the one which can reconstruct itself via the auto-encoder. For this we make an initial guess of the missing value using the nearest neighbor rule and then refine the missing value minimizing the reconstruction error. We train several classifiers using the union of the imputed instances and the remaining training instances without missing values. We also train another classifier of the same type with the same configuration using the corresponding complete dataset. The performances of these classifiers are compared. We compare the proposed method with eight state-of-the-art imputation techniques using fourteen datasets and eight classification strategies.