Abstract
The Mahalanobis-Taguchi (MT) method is one of statistical methods for analyzing asymmetric discriminant data in the field of quality control and medical diagnosis. However, some researchers point out that the accuracy of the MT method is deteriorated when (i) the data generating process of a given normal group is not based on an elliptical distribution, and (ii) some of variables are highly correlated with each other. In order to solve these problems, we propose a new asymmetric discriminant analysis "the Kernel MT method," which is formulated by introducing the kernel technique used in the field of machine learning into the MT divergences proposed by Miyakawa and Nagata (2003). Through the numerical experiments and the application of the kernel MT method to the sensory evaluation problem of the wine quality, we show that the kernel MT method is superior to the MT method.
