A criterion for choosing between full-sample and hold-out classifier design | IEEE Conference Publication | IEEE Xplore