A training algorithm of incremental supprot vector machine with recombining method [supprot read support]

Support vector machine (SVM) has become a popular tool of pattern recognition in recent years for its outstanding learning performance. When dealing with large-scale learning problems, incremental SVM framework is generally used because SVM can summarize the data space in a concise way. This paper proposes a training algorithm of incremental SVM with recombining method. Considering the differences of data distribution and the impact of new training data on history data, the history training dataset and the new training one are divided into independent groups and are recombined to train a classifier. In fact, this method can be implemented in a parallel structure for the actions of dividing may decrease the computation complexity of training a SVM. Meanwhile, the actions of recombining may weaken the potential impact caused by the difference of data distribution. The experiment results on a text dataset show that this training algorithm is effective and the classification accuracy of proposed incremental algorithm is superior to that using batch SVM model.