Robust kernel fisher discriminant analysis with weighted kernels
Nadja Louw · South African Statistical Journal · 2012
Kernel Fisher discriminant analysis (KFDA) is a very popular kernel classification technique that performs well in situations where linear classifiers fail. The performance of the KFD classifier is however adversely affected by outliers or noise in the data. In this paper we propose a more robust KFD classifier by making use of weighted kernels. The performance of the proposed classifiers is compared to that of the KFD classifier with a Gaussian kernel in Monte Carlo simulation studies as well as on several benchmark data sets. Based on the results we conclude that the proposed weighted kernels are successful in achieving a lower error rate than the Gaussian kernel when used in a KFD classifier.