A New Approach to Multi-class SVM Learning
Based on OC-SVM for Huge Databases
Djeffal Abdelhamid1 , Babahenini Mohamed Chaouki1 ,
and Taleb-Ahmed Abdelmalik2
1
Computer Science Department, LESIA Laboratory,
Biskra University, Algeria
2
LAMIH Laboratory FRE CNRS 3304 UVHC,
Valenciennes University, France
,
,
Abstract. In this paper, we propose a new learning method for multi-
class support vector machines based on single class SVM learning method.
Unlike the methods 1vs1 and 1vsR, used in the literature and mainly
based on binary SVM method, our method learns a classifier for each
class from only its samples and then uses these classifiers to obtain a
multiclass decision model. To enhance the accuracy of our method, we
build from the obtained hyperplanes new hyperplanes, similar to those of
the 1vsR method, for use in classification. Our method represents a con-
siderable improvement in the speed of training and classification as well
the decision model size while maintaining the same accuracy as other
methods.
Keywords: Support vector machine, Multiclass SVM, One-class SVM,
1vs1, 1vsR.
1 Introduction and Related Work
The support vector machine (SVM) method is, in its origin, binary. It is based on
the principle of separation between the samples of tow classes, one positive and
one negative. In this form, the SVM method is very successful in several areas
of application given the precision it offers. In practice, we find more applications
with multi-class problems, hence the need to extend the binary model to meet
multi-class problems. Existing methods currently try mainly to optimize two
phases: a training phase and a classification phase. The first phase constructs
the hyperplane, and the second uses it. The evaluation of the methods is based
on the evaluation of the performances of the two phases. Among the well known
methods, there are methods for direct solution without using the binary SVM
model as Weston & Watkins model [1],but which suffers, always, from some
slowness and weak accuracy. The widely used methods are based essentially
on the extension of the binary model, namely, the one-against-rest (1vsR) and
A. Abd Manaf et al. (Eds.): ICIEIS 2011, Part II, CCIS 252, pp. 677–690, 2011.
c Springer-Verlag Berlin Heidelberg 2011
Based on OC-SVM for Huge Databases
Djeffal Abdelhamid1 , Babahenini Mohamed Chaouki1 ,
and Taleb-Ahmed Abdelmalik2
1
Computer Science Department, LESIA Laboratory,
Biskra University, Algeria
2
LAMIH Laboratory FRE CNRS 3304 UVHC,
Valenciennes University, France
,
,
Abstract. In this paper, we propose a new learning method for multi-
class support vector machines based on single class SVM learning method.
Unlike the methods 1vs1 and 1vsR, used in the literature and mainly
based on binary SVM method, our method learns a classifier for each
class from only its samples and then uses these classifiers to obtain a
multiclass decision model. To enhance the accuracy of our method, we
build from the obtained hyperplanes new hyperplanes, similar to those of
the 1vsR method, for use in classification. Our method represents a con-
siderable improvement in the speed of training and classification as well
the decision model size while maintaining the same accuracy as other
methods.
Keywords: Support vector machine, Multiclass SVM, One-class SVM,
1vs1, 1vsR.
1 Introduction and Related Work
The support vector machine (SVM) method is, in its origin, binary. It is based on
the principle of separation between the samples of tow classes, one positive and
one negative. In this form, the SVM method is very successful in several areas
of application given the precision it offers. In practice, we find more applications
with multi-class problems, hence the need to extend the binary model to meet
multi-class problems. Existing methods currently try mainly to optimize two
phases: a training phase and a classification phase. The first phase constructs
the hyperplane, and the second uses it. The evaluation of the methods is based
on the evaluation of the performances of the two phases. Among the well known
methods, there are methods for direct solution without using the binary SVM
model as Weston & Watkins model [1],but which suffers, always, from some
slowness and weak accuracy. The widely used methods are based essentially
on the extension of the binary model, namely, the one-against-rest (1vsR) and
A. Abd Manaf et al. (Eds.): ICIEIS 2011, Part II, CCIS 252, pp. 677–690, 2011.
c Springer-Verlag Berlin Heidelberg 2011
