Abstract
Mahalanobis-Taguchi system, composed of Mahalanobis distance and Taguchi's robust engineering, is a multivariate technique proposed for diagnosis, forecasting, and feature selection. Although MTS has been widely studied and successfully applied in binary-class problems, very little research has been conducted for MTS to deal with multi-class problems. This is unfortunate because multi-class problems frequently arise in real application. By constructing the multiple Mahalanobis spaces, this study develops a multi-class MTS to accomplish the multi-class data classification and feature selection. Besides, in order to enhance the robustness and performance of MMTS, the Gram-Schmidt orthogonalization process is employed to eliminate the multicollinearity between features, and a weighted Mahalanobis distance which takes the contribution of the features for classification into account is proposed to be the distance metric. The numerical experiment results show that MMTS outperforms other well-known algorithms not only on classification accuracy but also on feature selection efficiency.