Fast Extended One-Versus-Rest Multi-Label Support Vector Machine Using Approximate Extreme Points

Abstract

Existing extended one-versus-rest multi-label support vector machine (OVR-ESVM) adopting non-linear kernel is seriously restricted by excessive training time when it is applied to large-scale data set. In order to overcome this problem, we improve the OVR-ESVM by introducing the principle of approximate extreme points and new approximate ranking loss to construct a novel extended OVR-ESVM using approximate extreme points (AEML-ESVM). By optimizing only on the representative set which can be acquired via adopting the approximate extreme points method, the AEML-ESVM classification algorithm can substantially shorten the training time and its classification performance is comparable to that of the OVR-ESVM classification algorithm. And it uses the new approximate ranking loss as empirical loss term to exploit label correlation of individual instance directly. Experimental study on three benchmark large-scale data sets illustrates that AEML-ESVM classification algorithm can reduce training time greatly and achieve comparable classification performance with OVR-ESVM classification algorithm. And it is also superior to the existing fast multi-label SVM classification algorithms in terms of classification performance and training time.