This　paper　introduced　an　improved-LDA　to　overcome　the　drawbacks　existing　in　traditional　linear　discriminant　analysis　method.　It　redefined　the　characteristic　matrix　by　adding　a　weight　vector　which　is　determined　by　the　posterior　classification　rate　of　each　feature.　Therefore　it　can　discriminate　different　classes　of　samples　in　the　projection　space　more　effectively　than　traditional　methods.　The　numerical　experiments　based　on　UCI　data　sets　show　that　this　method　can　reduce　the　within-class　scatter　and　increase　the　recognition　accuracy　rate　of　the　support　vector　machine.