Oversample Based Large Scale Support Vector Machine for Online Class Imbalance Problem

Abstract

Dealing with online class imbalance from evolving stream is a critical issue than the conventional class imbalance problem. Usually, the class imbalance problem occurs when one class of data severely outnumbers the other classes of data, thus leads to skewed class boundaries. In the case of online class imbalance problem, the degree of class imbalance changes over time and the present state of imbalance is not known a prior to the learner. To address such problem, in this paper, we present an Oversampling based Online Large Scale Support Vector Machine (OOLASVM) algorithm which is a hybrid of active sample selection and over sampling of Support Vectors and thereby both oversampling and under sampling coexists while learning the new boundary. Further, OOLASVM maintains the balanced boundary throughout the learning process. Results on simulated and real world datasets demonstrate that proposed OOLASVM yields better performance than existing approaches such as Generalized Oversampling based Online Imbalanced Learners and Over Online Bagging.