Online Training Framework of SVM Within Locally-discriminant Neighborhood

Abstract

SVM is characterized of excellent behaviors in diverse classification communities. But its expensive training cost dependent on the size of training data prevents its wider application. For that this paper presents an Online Training Framework for SVM (OTF). On the observation of a new data, OTF updates the decision model through optimizing a global objective within a locally-discriminant neighborhood of the new data. A novel of OTF is that based on the new data, the hyperplane of SVM is explored to derive the discriminant information, which is used to define metric and consequently the new data’s neighborhood. Experiments on real datasets verify the performance and efficiency of OTF when compared with state of the arts.