A Simple Extension of Stability Feature Selection

摘要: Stability selection [9] is a general principle for performing feature selection. It functions as meta-layer on top of “baseline” method, and consists in repeatedly applying the baseline to random data subsamples half-size, finally outputting features with frequency larger than fixed threshold. In present work, we suggest study simple extension original stability method submatrices matrix X given size returning those having largest frequency. We analyze from theoretical point view effect this subsampling selected variables, particular influence subsample size. report experimental results large-dimension artificial real identify which settings be recommended.