Imputation Using a Correlation-Enhanced Auto-Associative Neural Network with Dynamic Processing of Missing Values

The missing value is a common phenomenon in real-world datasets, which makes the analysis of incomplete data become an active research area. In this paper, a correlation-enhanced auto-associative neural network (CE-AANN) is proposed for imputations of missing values. We design correlation-enhanced hidden neurons and combine them with traditional hidden neurons organically, thereby constructing CE-AANN. Compared with the traditional auto-associative neural network (AANN), the improved architecture can mine cross-correlations among attributes more effectively. The introduction of correlation-enhanced hidden neurons keeps the network from learning a meaningless identity mapping. Moreover, a training scheme named MVPT is used for network training. Missing values are regarded as variables of the loss function and adjusted dynamically based on optimization algorithms. The dynamic processing mechanism takes account of the incompleteness of data during training, which makes the imputation accuracy increase as the training goes further. Experiments validate the effectiveness of the proposed method.