Neural network inversion techniques for EM training and testing of incomplete data

The expectation-maximization (EM) algorithm is a successful statistical approach for maximum likelihood estimation of incomplete-data problems. The performance of an EM algorithm highly depends on assumptions made about the probability density function (commonly, the multivariate Gaussian) of the multivariate data. When the EM algorithm is used for classification applications, it is commonly done by replacing the missing values based on the estimated probability density function of the same class for getting the maximum likelihood labeling without jointly considering the discrimination among classes. In this paper, the authors propose an EM procedure based on a neural network inversion technique for improving the training accuracy using incomplete data sets and the classification accuracy in testing new incomplete data. The authors´ approach relaxes the assumption made about the probability density function, and more importantly, the missing value replacements take into account the discrimination among classes