The　classification　of　different　cancer　subtypes　and　feature　subset　selection　is　of　great　importance　in　cancer　diagnosis　and　has　recently　received　a　great　deal　of　attention　in　the　field　of　bioinformatics.　The　purpose　of　this　study　is　to　develop　a　method　of　classifying　tumors　to　specific　categories　and　select　a　small　subset　of　genes　for　classification　based　on　gene　expression　profiles.　Firstly,　a　new　metric　for　class　separability　was　proposed　in　order　to　remove　the　genes　irrelevant　to　the　classification　task,　and　then　a　support　vector　machine　was　applied　to　distinguish　different　cancer　types.　The　feature　subset　selection　process　is　performed　by　pair-wise　redundancy　analysis　and　the　sequential　floating　forward　search　method　after　irrelevant　genes　have　been　removed.　We　analyzed　the　gene　expression　profiles　of　human　acute　leukemia　as　a　test　case,　and　the　results　showed　the　feasibility　and　effectiveness　of　the　method　proposed.