Ensemble Pruning of MT Method

Abstract

The Mahalanobis Taguchi(MT) method is used for pattern recognition and anomaly detection. It defines apopulation homogeneous to the objective as the unit space and uses theMahalanobis distance from its center to discriminate. However, a problem withthis method is that the number of samples must be greater than the number ofvariables in the data. This problem is caused by the calculation of theMahalanobis distance. To solve this problem, the MT -bagging method, whichapplies feature bagging to the MT method was proposed. This study proposes twomethods that apply pruning to MT bagging. One is based on ordering pruning andthe other, on clustering pruning. In the ordering-based method, thesignal-to-noise ratio of the training abnormality data are calculated for eachweak learner. Only the weak learners with high signal-to-noise ratios areensembled. In the clustering-based method, weak learners are clustered usingthe K-means method with the Mahalanobis distance of the training anomaly data,and the centers of the clusters are ensemble learned. Breast cancer and drybean datasets are used to verify the performance of the proposed methods. Bothmethods outperformed the MT and MT-bagging methods in terms of abnormalitydiscrimination accuracy, suggesting that ensemble pruning is effective in somesituations.