Integrating liknon feature selection and committee training

Abstract

We address a practical application of feature selection and training a committee of classifiers in high dimensional classification problems. Embedded Liknon feature selection method is integrated into the training of a committee of classifiers via external K-fold crossvalidation with an inner loop. In problems, characterized by nonlinear class separation, the Liknon-selected feature profiles, optimal for linear class separation, also identify the relevant feature subspaces. The capabilities of the proposed approach are illustrated in a benchmark of NIPS2003 feature selection challenge. © 2009 Springer-Verlag Berlin Heidelberg.