A geometric approach to train SVM on very large data sets | IEEE Conference Publication | IEEE Xplore