The epsilon-SVR / nu-SVR formulation of the kernel matrix: the kernel matrix is expanded to twice of the sample size so that the two slack variables of each sample can be handled at the same time.