A Composite Wrapper for Feature Selection

In many bioinformatics problems the number of features is very relevant. Especially in classification, feature selection plays an essential role and it improves classification results. In this paper we propose a new method for feature selection based on the ensemble approach. Our method is a composite wrapper: it simultaneously takes care of the sets of features selected by different classifiers. By a cross validation strategy, each learner produces a feature ranking, which we subsequently merge to determine the ideal subset. We report classification results on some biological datasets considering many state of the art classifiers and the features selected by our composite method. Experiments demonstrate that the proposed method is often effective and overall results are encouraging.