A Fast Multi-label Classification Algorithm Based on Double Label Support Vector Machine

Combing one-versus-one decomposition strategy with support vector machine has become an efficient means for multi-label classification problem. But how to speed up its training and test procedures is still a challenging issue. In this paper, we generalize the primary binary support vector machine to construct a double label support vector machine through locating double label instances at marginal region between positive and negative instances, and then design a fast multi-label classification algorithm using the voting rule. Experiments on benchmark datasets Yeast and Scene illustrate that our novel method can be comparable with some existing methods according to some widely used evaluation criteria, and can run faster 17% averagely than the current corresponding method in training procedure.

1. Introduction

The multi-label problem is a special classification issue where some instances belong to many classes or include many labels at the same time [1]–[3]. Nowadays, there are mainly two kinds of methods to deal with such a problem. One is to directly generalize binary or multi-class classification algorithms to design multi-label ones based on one optimization problem, for example, rank-SVM [4], ML-KNN [5] [6], ML-BP [7], and so on. The other is to divide a training set into many binary class subsets, to solve all of them using popular binary algorithms, and to integrate all sub-classifiers together. Many researchers have paid more attention to such a kind of decomposition methods since they run very fast in training procedures. Currently two data decomposition strategies are widely adopted: one-versus-other and one-versus-one.