Feature　selection　has　become　a　critical　process　in　training　models　with　high　throughput　biological　data.　One　of　the　most　critical　categories　of　feature　selection　techniques　is　penalty-based　approaches　because　of　the　sparsity　of　selected　features.　Penalty-based　methods　automatically　set　small　estimated　coefficients　to　zero　to　reduce　model　complexity.　There　are　many　penalty-based　methods　with　different　benefits,　drawbacks　and　statistical　property.　The　choosing　of　these　penalty-based　methods　under　different　situations　has　become　a　problem.　So,　in　this　paper,　we　mainly　focus　on　the　comparison　and　evaluation　of　four　popular　penalty-based　methods　by　evaluating　the　three　metrics　which　are　accuracy,　robustness　and　the　robustness-performance　trade-off　(RPT)　for　each　method.　Since　each　of　them　has　its　statistical　properties,　our　comparison　may　be　helpful　for　researchers　when　making　choices　when　dealing　with　high　throughput　data.　The　result　shows　that　LASSO　achieves　the　best　robustness　and　accuracy　among　the　four　feature　selection　methods　when　dealing　with　high　throughput　TCGA　datasets.　©　2020　ACM.