PROXIMAL AND TWIN SUPPORT VECTOR MACHINE FOR BINARY CLASSIFICATION

Abstract

Support vector machines (SVMs) have become the standard choice for classification tasks such as pattern recognition and regression problems over the last two decades. The main advantage of support vector machines is that it results in a convex programming problem which yields a global optimal solution. However, SVM also has some limitations when applied to classification problems. SVM is sensitive to outliers and noises and it requires solving large size quadratic programming problem (QPP) to train the algorithm, which is computationally expensive. Several variants of SVM have been developed to overcome these limitations of SVM. In this study, we have explored SVMs, and its two important variants: proximal SVM and twin SVM. In the proximal support vector machine, only a system of linear equations has to be solved instead of QPP which speeds up the computation steps involved in the algorithm, compared to SVM. However, in twin support vector machines we obtained the planes which need not be parallel. Unlike the classical SVM, the twin SVM solves two smaller size QPP, which makes it four times faster than SVM. Several artificial datasets have also been constructed for proximal and twin SVMs using python code to see the effectiveness of these variants.