Abstract: We propose a parallel training strategy, which is an improved parallel algorithm of support vector machine (SVM), to shorten the training time based on SVM's classification. The strategy uses the master-salve mode and divides the whole training task into several subtasks, each sub-task computed by a node. The master node collects the training results from slave nodes to produce the classifiable model. Performance of this algorithm is analyzed and evaluated with sparse and dense dataset on a high-performance computer ZQ3000 cluster. The results indicate that the proposed method can ensure high precision in the original multi-classification and reduce training time.

Key words: decision-function, parallel training, support vector machine (SVM), parallel computing