On the performance of ensembles of classifiers based on kernel density estimation

A combination of classification rules (classifiers) is known as an Ensemble, and in general it is more accurate than the individual classifiers used to build it. Two popular methods to construct an Ensemble are Bagging introduced by Breiman, (1996) and Boosting (Freund and Schapire, 1996). Both method rely on resampling techniques to obtain different training sets for each of the classifiers. Previous work has shown that Bagging as well as Boosting are very effective for unstable classifiers. In this paper we present experimental results of application of both combining techniques using classifiers where the class conditional density is estimated using kernel density estimators. The effect of sequential forward selection on the performance of the Ensemble also is considered.