Soft feature selection by using a histogram-based classifier

Abstract

Proposed is a histogram approach for feature selection and classification. The axes are divided into equally-spaced intervals, but the division numbers are different among axes. The main difference from similar approaches is that feature selection mechanism is embedded in the method. The optimal division is determined by an MDL criterion, so that the classifier is guaranteed to converge to the Bayes optimal classifier. We also introduce the concept of "soft feature selection" that is carried out by this method as an extension of traditional "feature selection." © 2008 Springer Berlin Heidelberg.