Imputations of missing values using a tracking-removed autoencoder trained with incomplete data

摘要

The presence of missing values in incomplete datasets increases the difficulty of data mining. In this paper, we use the autoencoder (AE) to model the incomplete data for imputations of missing values, which reduces the complexity of data modeling effectively. In order to strengthen the dependence of missing values on known attribute values for each incomplete record, we propose an architecture named tracking-removed autoencoder (TRAE) by redesigning the input structure of hidden neurons in a dynamic way on the basis of the traditional AE. Moreover, a training scheme for TRAE which treats missing values as variables and allows them to participate in the network training has been designed with the consideration of the data incompleteness. The imputation will be completed at the end of the training process. The proposed method makes full use of the complete data and builds the correlation of attributes in incomplete data for imputations in complicated missing patterns. The experimental results obtained from several UCI datasets validate the effectiveness of the proposed method. (C) 2019 Elsevier B.V. All rights reserved.