Traditional　SVM　classification　model　constructs　linear　discriminant　function　by　maximizing　the　margin　between　two　classes,　and　the　weight　vector　of　the　discriminant　function　is　only　related　to　a　small　number　of　support　vectors　near　the　decision　boundary.　The　small　amount　of　support　vectors　is　hard　to　describe　the　global　distributive　information　when　the　distributions　of　the　samples　are　nonlinear　manifolds　structure.　To　solve　this　problem,　the　graph　regularization　term　with　discrimination　information　is　introduced　into　the　objective　function　of　SVM　model.　Experimental　results　on　public　data　sets　show　that　the　classification　accuracy　of　this　method　has　improved　significantly　compared　to　traditional　SVM　models.　©　2015　IEEE.