A New Incremental Learning Support Vector Machine
Yingchun Zhang, Guosheng Hu, Feng-feng Zhu, Jin-Lian Yu · 2009
A new incremental learning method for support vector machine (SVM) is proposed, which train SVM quickly and incrementally. In this paper, we first choose the violating KKT samples which maybe be new support vector candidates. Then for a given new-added sample, the proposed training method validate whether they are border vectors. If true, we add them to training sample set to retrain support vector machine, otherwise omit it. Hence, the training samples can be reduced and training complexity be lessened. Finally, an incremental algorithm is presented to train SVM by using the selected samples violating KKT conditions. Experiment results show that the test error and support vector number of the proposed algorithm is almost same as those of SMO algorithm, however, the training speed of the new incremental algorithm are more quickly that of SMO method.