Title: Multicategory classification via forward-backward support vector machine
Authors: Donglin Zeng - University of North Carolina at Chapel Hill (United States) [presenting]
Yuanjia Wang - Columbia University (United States)
Xuan Zhou - University of North Carolina (United States)
Abstract: A novel and computationally efficient learning algorithm, namely forward-backward support vector machine (FB-SVM), is proposed to perform multicategory learning. The new method is based on a sequential binary classification algorithms: we first classify a particular class by excluding the possibility of labeling as any other classes using a forward procedure of sequential SVM; we then exclude already classified classes and repeat the same learning for the remaining classes in a backward way. The proposed algorithm relies on support vector machines for each binary classification and utilizes only feasible data in each step of this procedure; therefore, the method guarantees convergence and computation burden is little. Furthermore, we show that the derived rule from FB-SVM is Fisher consistent and we obtain the risk bound for the predicted misclassification rate. We conduct extensive simulation studies, including benchmark examples from existing methods, to demonstrate that the proposed method has superior performance in terms of small misclassification rates and significantly improved computation speed. Finally, we apply the proposed method to analyze real data for further illustration.