Feature　selection　technique　has　become　a　basic　but　desired　technique　when　analyzing　high-throughput　biological　data　due　to　its　nature　of　large　p　and　small　n.　In　recent　years,　ensemble　learning　based　feature　selection　methods　have　been　widely　proposed　and　studied.　Ensemble　methods　employ　multiple　learning　algorithms　to　obtain　better　predictive　performance　than　any　of　the　constituent　learning　algorithms　separately.　Also,　the　feature　selected　by　ensemble　classifiers　can　yield　more　accurate　classification　performance　and　more　robust　results.　In　our　work,　the　bagging　algorithm　Random　Forest　(RF),　and　the　boosting　algorithms　Gradient　Boosting　Decision　Tree　(GBDT),　Extreme　Gradient　Boosting　(XGBoost)　are　the　main　research　objects.　We　compared　the　accuracy　and　robustness　of　three　algorithms　on　six　different　datasets　from　TCGA　database.　Also,　the　three　feature　selection　algorithms　are　further　ensembled　using　a　bagging　procedure　for　the　purpose　of　comparison　with　the　original　single　classifier.　The　results　of　our　work　indicated　that　for　single　base　feature　selectors,　boosting　algorithms　all　outperform　than　bagging　one　in　both　performance　and　robustness.　By　applying　the　bagging-based　feature　selection　procedure,　the　robustness　of　three　single　base　feature　selectors　is　improved　significantly,　but　the　accuracy　of　them　is　slightly　reduced.　GBDT　with　bagging-based　feature　selection　procedure　achieved　the　best　performance　using　our　proposed　comprehensive　metric　which　balances　equally　accuracy　and　robustness.　©　2020　ACM.