Design of a Two Layers Support Vector Machine for Classification

The idea of support vector machine (SVM) is to project the primal data which are not separable in the primal space into a new high dimensional feature space by a nonlinear function, so that the data can be separated in the new space correctly. But sometimes the result is not satisfied. In order to improve the accuracy, a two-layer SVM is put forward in the paper. Through mapping the primal data into a much higher dimensional space by nonlinear function for two times, the data could be separated in the final feature space as most as possible. A new kernel function is also deduced from two layers SVM which can also satisfy the Mercer theorem. The algorithm is deducted, proved and simulated in detail. The complexity of SVM is not increased, but the classification accuracy can be improved by this means.