In　recent　years,　feature　selection　(FS)　methods　based　on　sparsity　have　been　extensively　investigated　due　to　their　high　performance.　These　methods　solve　the　FS　problem　mainly　by　introducing　some　kinds　of　sparsity　regularization　terms.　However,　recent　existing　feature　selection　algorithms　combine　sparsity　regularization　with　simple　linear　loss　function,　which　may　lead　to　deficient　in　performance.　To　this　end,　we　propose　a　fresh　and　robust　feature　selection　method　that　combines　the　structured　sparsity　regularization,　i.e.,　2,　0　-norm　regularization,　with　the　Softmax　model　to　find　a　stable　row-sparse　solution,　where　we　can　select　the　features　in　group　according　to　the　solution　of　the　projected　matrix,　and　the　classification　performance　can　be　improved　by　Softmax.　Extensive　experiments　on　six　different　datasets　indicate　that　our　method　can　obtain　better　or　comparable　classification　performance　by　using　fewer　features　compared　with　other　advanced　sparsity-based　FS　methods.　©　2022,　Springer　Nature　Singapore　Pte　Ltd.