The Optimization of Large Scale Multiple Kernel SVM Based on K-Means Clustering in Kernel Space

The generalization ability of multiple kernel support vector machines is better than the single kernel ones. If the training datasets are large scale, solving the optimal multiple kernels´ combination coefficients with semidefinite programming method is difficult, and the time-consuming is large. We use K-means Clustering algorithm in kernel space to reduce the scale of SVM´s training datasets, then the scale of the corresponding semidefinite programming is reduced. Our experimental results show that: the new method received more than several times faster than the old one in solving the semidefinite programming problem of SVM, and does not reduce the classification accuracy of multi-kernel SVM model.