Progressive two-class decision classifier for optimization of class discriminations

Abstract

A progressive two-class decision classifier (pTCDC) is presented in this article. This multilayer procedure converts a multiclass classification problem into a several independent two-class separations. It not only provides the general advantages of hierarchical classification schemes over single-stage classification but it is also free of the need for hierarchical structure design and offers an optimal class pair discrimination environment. At each decision node, only one class pair is considered. Data processing aimed at maximising individual class pair separations, such as feature selection, classification algorithm selection and data source selection or data transformation, becomes more reliable and efficient. Experiments carried out using an AVIRIS data set are presented and the results demonstrate that fewer features are needed and classification accuracy is improved with the new procedure compared with single-stage classification.