TI-QSSVM: Two Independent Quarter Sphere Support Vector Machine for binary classification

Abstract

One of the important extensions of SVM is TWSVM which uses two hyperplanes to classify two classes of data. Since one hyperplane cannot efficiently model one class of data so the better choice is employing one hypersphere which covers as many data points in the corresponding class as possible and can better depict the characteristics of that class. Quarter sphere SVM uses a minimum radius centered hypersphere to describe data points such that it covers the majority of data and makes the outliers lied out of this hypersphere. In this paper inspired by the merit of QSSVM algorithm, we proposed a new two independent quarter sphere SVM (TI-QSSVM) to classify two classes of data. TI-QSSVM generates two quarter sphere with the minimum radiuses for two classes such that each one centered at the mean point of the corresponding class and covers as many data points in that class as possible. TI-QSSVM obtains these two quarter sphere by solving two linear programming problems. As can be seen in the experiment section, TI-QSSVM has significant advantages in terms of the learning speed and generalization performance compared with the other algorithms.