Optimization ClusterSVM using improved nonlinear kernel

ClusterSVM exploiting the distributional properties of training data accelerates the training process with large-scale data set, and especially benefits two-class problem with a large number of boundary support vectors. The algorithm first partition the training data into disjoint clusters, then train an initial SVM using representatives of these clusters. This initial SVM gives us a global picture of the solution. The initial SVM can approximately identify the support vectors and non-support vectors. The training process is accelerated by replacing non-support vectors with few data. The initial SVM of cluster is the key of training ClusterSVM. This paper proposed an improved nonlinear kernel to generate a nonlinear separating surface which depends on expanding the reduced set incrementally according to information criterion. It uses this kernel to train the initial SVM of clusters. Computational results indicate computational times and the number of training data are much smaller for improved method than that of the conventional ClusterSVM.