A Multi-Classification Method Based on Optimized Binary Tree Mahalanobis-Taguchi System for Imbalanced Data

Abstract

Data imbalance is a common problem in classification tasks. The Mahalanobis-Taguchi system (MTS) has proven to be promising due to its lack of requirements for data distribution. The MTS is a binary classifier. However, multi-classification problems are more common in real life and the diversity of categories may further aggravate the difficulty of classifying imbalanced data. Imbalanced multi-classification has become an important research topic. To improve the performance of MTS in imbalanced multi-classification, we propose an algorithm called optimized binary tree MTS (Optimized BT-MTS). Mahalanobis space (MS) construction, feature selection, and threshold determination are incorporated in a unified classification framework, and joint optimization is carried out according to the principles of maximizing separability, signal-to-noise ratio, dimensionality reduction, and minimizing misclassification cost. Experimental results on several datasets show that the method can significantly reduce the overall misclassification cost and improve the performance of imbalanced data multi-classification.