There　is　still　a　problem,　lack　of　enough　generalization　ability,　with　existing　feature　selection　methods.　To　solve　this　problem,　a　supervised　feature　selection　method　base　on　support　vector　machine　is　proposed　in　view　of　generalization　ability　of　support　vector　machine　for　small　sample　set　and　ability　of　processing　high-dimensional　data　of　kernel　function.　The　new　method　introduces　the　categoryseparability　criterion　in　terms　of　minimum　coverage　hypersphere　of　samples,　and　uses　the　criterion　as　the　feature　assessment　index　to　feature　sorting　and　feature　selection.　The　experimental　results　show　that　this　method　can　obtain　a　reasonable　feature　sorting,　eliminate　unrelated　feature　in　the　data　set　effectively.　©　2009　IEEE.