Ant system-based feature set partitioning algorithm for classifier ensemble construction

Ensemble method is considered as a new direction in pattern classification. Accuracy and diversity in a set of classifiers are two important things to be considered in constructing classifier ensemble.Several approaches have been proposed to construct the classifier ensemble. All of these approaches attempt to generate diversity in the ensemble.However, classifier ensemble construction still remains a problem because there is no standard guideline in constructing a set of accurate and diverse classifiers. In this study, Ant system-based feature set partitioning algorithm for classifier ensemble construction is proposed.The Ant System Algorithm is used to form an optimal feature set partition of the original training set which represents the number of classifiers.Experiments were carried out to construct several homogeneous classifier ensembles using nearest mean classifier, naive Bayes classifier, k-nearest neighbor and linear discriminant analysis as base classifier and majority voting technique as combiner. Experimental results on several datasets from University of California, Irvine have shown that the proposed algorithm has successfully constructed better classifier ensembles.