Classifier combination with kernelized eigenclassifiers

Classifier combination has become a very important topic, because it is possible to train many classifiers using different feature, instance subsets or different types of classifiers. Classifier diversity and accuracy are two competing requirements for classifier combination. In this paper, we study classifier combination using the kernelized eigenclassifiers The eigenclassifier method, tries to handle the linear correlations among classifier outputs by applying PCA to uncorrelate them before fusing with a second classifier, is introduced by Ulas et al. (2012). Our contribution is to adapt the kernel PCA in this method to handle non-linear correlations among classifier outputs and we compared the eigen and kernelized eigenclassifiers to SVM based stacking algorithms both for linear and rbf kernels and simple average method to see the performance of these methods. Our experiments on the 38 datasets used by (Ulas et al. 2009) show that the kernelized eigenclassifiers method performs better than the other methods in terms of test accuracy.