Feature selection SDA method in ensemble nearest neighbor classifier

Abstract

The curse of dimensionality is still a big problem in the pattern recognition field. Feature extraction and feature selection have been presented as two general solutions for this problem. In this paper, a new approach based on combination of these methods has been proposed to classify different classes in large dimensional problems. Among the vast variety of search strategies,Tabu Search (TB) is chosen here as a core for feature selection. Filters & Wrappersare the two traditional types of objective functions for feature selection which both applied in this study. Following a feature extraction approach is considered. Subclass Discriminant Analysis (SDA), as a successful strategy in feature extraction has been employed. the result of each objective function on the standard UCI datasets are shown. Finally a revised nearest neighbor classifier has been used to classify the patterns in the new feature space on the UCI data sets and the results show the supremacy of our combinatorial approach in comparison with the traditional methods. © 2008 Springer-Verlag.