Multiclass support vector machines with GenSVM

Abstract

Binary support vector machines (SVM) have become a standard tool for supervised machine learning. Attractive features of the SVM are that its solution only depends on badly fitting observations, it combats overfitting through regularization, can handle high dimensionality (thus many predictors), allows for nonlinear predictions, and is robust against outliers. Much less attention has been given to classification problems with more than two classes. In the machine learning literature, such multiclass problems tend to be solved by repeatedly applying binary SVMs, for example, through one-versus-one (OvO) or one-versus-all (OvA). Although such approaches are generally fast, they can lead to regions that are inconclusive in their prediction. As an alternative, the present authors have proposed a single machine classifier, called GenSVM (see, [1]). In this paper, we present its main properties and discuss examples of its implementations in the R and Python packages.