Projection Support Vector Machines

Abstract

Large-scale classification is a very active research line in data mining. It can be applied to problems like credit card fraud detection or content-based document browsing. In recent years, several efficient algorithms for this area have been proposed by Mangasarian and Musicant. These approaches, based on quadratic problems, are: Successive OverRelaxation (SOR), Active Support Vector Machines (ASVM) and Lagrangian Support Vector Machines (LSVM). These algorithms have solved linear classification problems with millions of points. ASVM is perhaps the fastest and more scalable among them. This paper presents a projection-based SVM algorithm that outperforms ASVM on a 50,000 point data set generated by means of NDC (Normally Distributed Clusters), which has become a common tool in large-scales SVM research.