Learning the boundary of One-Class-Classifier globally and locally

Abstract

The one class classification problem aims to distinguish a target class from outliers. Two popular algorithms, one-class SVM (OCSVM) and single-class MPM (SCMPM), solve this problem by finding a hyperplane with the maximum distance to the origin. Their essential difference is that OCSVM focuses on the support vectors (SV) in a local manner while SCMPM emphasizes the whole datapsilas distribution using global information. In fact, these two seemingly different yet complementary characteristics are all important prior knowledge for the one-class-classifier (OCC) design. In this paper, we propose a novel OCC called global & local (GLocal) OCC, which incorporates the global and local information in a unified framework. Through embedding the samplespsila distribution information into the original OCSVM, the GLocal OCC provides a general way to extend the present SVM algorithm to consider global information. Moreover, the optimization problem of the GLocal OCC can be solved using the standard SVM approach similar to OCSVM, and preserves all the advantages of SVM. Experiment results on benchmark data sets show that the GLocal OCC really has better generalization compared with OCSVM and SCMPM.