Feature　selection　is　an　important　process　to　choose　a　subset　of　features　relevant　to　a　particular　application　in　text　classification.　Based　on　the　mutual　information　method,　we　designed　variance-mean　based　feature　selection　(VM).　After　computing　and　ranking　the　variance　of　class　discrimination　value　vector　for　each　word,　we　can　choose　the　most　distinguishable　features.　This　method　has　advantages　in　the　case　of　choosing　smaller　number　of　features,　especially　for　classes　with　small　number　of　training　documents.　It　keeps　the　best　features,　and　thus　improves　the　final　performance　of　the　classification　system.　The　experiment　results　indicate　the　effectiveness　of　the　proposed　feature　selection　method　in　a　text　classification.