Classifier Ensemble Framework Based on Clustering Method

Abstract

This paper proposes an innovative combinational method how to select the number of clusters in the Classifier Selection by Clustering (CSC) to improve the performance of classifier ensembles both in stabilities of their results and in their accuracies as much as possible. The CSC uses bagging as the generator of base classifiers. Base classifiers are kept fixed as either decision trees or multilayer perceptron during the creation of the ensemble. Then it partitions the classifiers using a clustering algorithm. After that by selecting one classifier per each cluster, it produces the final ensemble. The weighted majority vote is taken as consensus function of the ensemble. Here it is probed how the cluster number affects the performance of the CSC method and how we can switch to a well approximation option for a dataset adaptively. We expand our studies on a large number of real datasets of UCI repository to reach a well conclusion.