Recovering Missing Data with Functional and Bayesian Networks

Abstract

The paper presents some methods for recovering missing data using functional and Bayesian networks. In the case of a small set of missing data one can consider the missing data as variables and learn them together with the model parameters in the minimization process. If on the contrary, the missing data set is large, one can learn the functional or neural network from complete data and use them to learn the missing data, one case at a time. Finally, some examples of application to illus- trate the methodology are presented. They show how the missing data recovery degenerates as the number of missing data per case increases us- ing an adimensional error measure that allows a direct comparison with the case of all missing data. In addition, the Bayesian network approach seems to give better results than the functional network.