Adaptive Selection of Classifier Ensemble Based on GMDH

Abstract

In multiple classifiers combination, how to choose an effective combination method is a very critical issue. This article introduces group method of data handing (GMDH) theory into multiple classifiers combination, and proposes a novel algorithm GAES for classifier ensemble selection. It is able to select an appropriate ensemble from the classifier pool adaptively, determine the combination weights among base classifiers, and complete the combination process automatically. We experimentally test GAES over 16 UCI data sets and 4 ELENA data sets. The results show that compared with the commonly used combination methods, GAES is statistically superior to Bayesian (Kittler et al., 1998), Linear (Benediktsson et al., 1997) and ESNN (Lipnickas and Korbicz, 2004) methods, and achieves a comparable classification rate than MAJ (Xu et al., 1992) and Genetic (Cho, 1999).