The Research And Application Of Classification Method On MTS Based On Ridge Estimaiton

Multicollinearity is often among variables in the multi-dimensional data, which will affect the performence of parameter estimation, make parameters extremely sensitive on slight variable's perturbation. Mahalanobis-Taguchi System (MTS) is a new methodology of multi-dimensional pattern recognition whose measure scale is based on the mahalanobis distance(MD), combining the orthogonal array and signal-to-noise ratio to reduce the system dimension. As a technique of data mining, it can classify or predict the class of multi-dimensional samples with the simplest system. However, the performance must be affected by the defects of MTS itself and the unavoidable negative effect of multicollinearity. From the perspective of overcoming multicollinearity, this paper try to improve MTS in measure scale, dimension reduction and classifier reconstruction to decrease the negative effect, enhance the calculation robustness of MD and the classification performance of MTS.This paper introduces the idea of ridge regression to MD to form a new measure scale through researching the negative effect of multicollinearity to MD and MTS. In order to draw the ridge trace to determine the ridge parameter, three indexes of condition and partial derivative are generated by performing the sensitivity analysis. Then a multi-objectives optimization model for MTS binary classification is established, and present an adaptive multi-objectives genetic algorithm combine the strategy of external elite reservation with internal random selection to achieve the pareto non-dominated solutions, the optimal solution is selected based on the idea of TOPSIS. At last, for the purpose of improving the MTS classifier, increase the system accuracy and the category explanatory of samples. A fuzzy threshold classifier is presented based in the fuzzy theory because of the unreliability of hard threshold classifier in MTS, the membership function is determined by assignment method and the parameter is achieved by fuzzy statistic.This paper introduces the idea of biased estimation to covariance to decrease the over sensitive of MD, implements the dimension reduction based in the optimal theory and constructs the fuzzy threshold classifier to perform the classification. The research can not only make MTS deal with the data with multicollinearity well, but also highly enhance the applied field and explanatory ability.