Fast SVM-based One-Class Classification in Large Training Sets

Abstract

SVM is one of the popular methods to solve One-Class classification problem. However, it is time and space-consuming. This fact makes it hard or even impossible to apply SVM for large training sets. In this paper we propose fast method (One-Class Kernel-based Mean Decision Rule method, OC-KMDR) to find an approximate decision of One-Class SVM problem. The main advantages of the proposed approach are: 1) the obtained decision is near exact (and in a number of cases the method can outperform the original SVM in quality); 2) the obtained decision has absolutely the same structure as the original SVM; 3) the absence of theoretical restriction for the training set size; 4) it can be realized in an iterative manner but without inter-iteration data dependencies, and, as a result, provides the possibility for effective parallel computing. Experimental study of the proposed OC-KMDR-method in series of simulated large data sets shows that it outperforms existing methods for solving One-Class SVM problem in a computing time or(and) in a decision quality.