train structur svm kernel sampl cut chun nam john yu thorsten joachim discrimin train structur output found increas applic area natur languag process bioinformat inform retriev comput vision focus larg margin method gener term loss function model structur train algorithm date base cut plane approach algorithm effici linear model train complex quadrat number exampl kernel overcom bottleneck propos train algorithm approxim cut plane random sampl enabl effici train kernel prove algorithm improv time complex provid approxim guarante empir evalu algorithm produc solut train test error rate close exact solver binari classif problem highli optim convent train method exist svm light method order magnitud faster convent train method larg dataset remain competit speed dataset medium size 