Feature selection by block addition and block deletion

Abstract

In our previous work, we have developed methods for selecting input variables for function approximation based on block addition and block deletion. In this paper, we extend these methods to feature selection. To avoid random tie breaking for a small sample size problem with a large number of features, we introduce the weighted sum of the recognition error rate and the average of margin errors as the feature selection and feature ranking criteria. In our methods, starting from the empty set of features, we add several features at a time until a stopping condition is satisfied. Then we search deletable features by block deletion. To further speedup feature selection, we use a linear programming support vector machine (LP SVM) as a preselector. By computer experiments using benchmark data sets we show that the addition of the average of margin errors is effective for small sample size problems with large numbers of features in realizing high generalization ability. © 2012 Springer-Verlag.