An approach of Support Vector Machine to improve the training speed

When training a Support Vector Machine (SVM) we need to solve a very large convex optimization problems, typically Quadratic Programs (QP) that will largely reduce the solving speed of SVM. An improved algorithm to speed up training of SVM is proposed. The algorithm uses the fact that the solution of the QP is same if you remove the training samples that correspond to zero Lagrange multipliers. The training samples of zero Lagrange multipliers are nonsupport vectors. The samples located in the center of the training samples are nonsupport vectors. We can identify the samples in the center of the training samples and remove them using variance in statistics to decrease training samples as much as possible. In the paper, the balance factor that is the key parameter is analyzed and its value range is provided. Simulation experiments are performed using the databases including an artificial and an UCI real data. Simulation experiments indicate that the improved algorithm can cut down the training time of SVM by 80% while the ability of SVM to classification is unaffected. This approach is practicable.