New SV selection strategy and local-global regularisation method for improving online SVM learning

During the online learning process of support vector machines (SVMs), when a newly added sample is violating the Karush-Kuhn-Tucker) conditions, the new sample should be a new SV and transfer the old samples between the SVs and the non-SVs. Normally, the performance of an SVM model is decided by the SVs, and the model should be updated by the newly added SVs; therefore, the selection of high-quality candidate SVs will lead to a better learning accuracy, whereas low-quality candidate SVs may result in low learning efficiency and unnecessary updating. A new strategy is proposed to select the candidate SVs. SVs are selected according to two new criteria: the importance and the informativeness criteria. Furthermore, a mixed local-global regularisation method is applied during the online learning process to improve the penalty coefficients. Experiment results show that the proposed algorithm can achieve a better performance with a faster speed and a higher accuracy when compared with traditional methods.