ABSTRACT

The natural approach to use Support Vector Machines (SVMs) in multiclass problems is to reduce the problem to several binary classification problems. The most used methods are the One-versus-Rest (OvR) and the One-versus-One (OvO) based methods. Although these classical methods give good results in many problems, all of them require a large computational complexity to solve problems with large number of classes. This paper proposes a new approach, which we named Truth Table Fitting Multiclass SVM (TTF-MCSVM), that uses considerably smaller complexity and have a performance comparable to the classical methods. It is based on the overlapping of the decision functions of all used binary SVMs, constructed of different class groups. The application in synthetic problems showed results similar to the OvR and OvO methods, despite this results are very dependent of the way the classes are grouped.