Mixture Models for Maximum Likelihood Estimation from Incomplete Values

In this paper we consider methods for the analysis of the relationship between input and output variables when missing values occur in the input data. In such situations the incomplete cases cannot be suppressed and then the missing values must be estimated on the basis of some suitable statistical model. This problem is here approached by means of mixture distributions in which the parameters are estimated using likelihood-based methods. Applications in neural network training from incomplete data are discussed and the results are compared with those obtained using the mean imputation method. These results lead to some practical criteria for the use of either method in the learning of neural network from incomplete data.