Two new metrics for feature selection in pattern recognition

Abstract

The purpose of this paper is to discuss about feature selection methods. We present two common feature selection approaches: statistical methods and artificial intelligence approach. Statistical methods are exposed as antecedents of classification methods with specific techniques for choice of variables because we pretend to try the feature selection techniques in classification problems. We show the artificial intelligence approaches from different points of view. We also present the use of the information theory to build decision trees. Instead of using Quinlan's Gain we discuss others alternatives to build decision trees. We introduce two new feature selection measures: MLRelevance formula and the PRelevance. These criteria maximize the heterogeneity among elements that belong to different classes and the homogeneity among elements that belong to the same class. Finally, we compare different feature selection methods by means of the classification of two medical data sets. © Springer-Verlag Berlin Heidelberg 2003.