An Out of Sample Version of the EM Algorithm for Imputing Missing Values in Classification.

Finding real-world applications whose records contain missing values is not uncommon. As many data analysis algorithms are not designed to work with missing data, a frequent approach is to remove all variables associated with such records from the analysis. A much better alternative is to employ data imputation techniques to estimate the missing values using statistical relationships among the variables. The Expectation Maximization (EM) algorithm is a classic method to deal with missing data, but is not designed to work in typical Machine Learning settings that have training set and testing set. In this work we present an extension of the EM algorithm that can deal with this problem. We test the algorithm with ADNI (Alzheimer’s Disease Neuroimaging Initiative) data set, where about 80% of the sample has missing values. Our extension of EM achieved higher accuracy and robustness in the classification performance. It was evaluated using three different classifiers and showed a significant improvement with regard to similar approaches proposed in the literature.