Kernel Knockoffs Selection for Nonparametric Additive Models

Thanks to its fine balance between model flexibility and interpretability, the nonparametric additive has been widely used, variable selection for this type of frequently studied. However, none existing solutions can control false discovery rate (FDR) unless sample size tends infinity. The knockoff framework is a recent proposal that address issue, but few are directly applicable models. In article, we propose novel kernel knockoffs procedure model. We integrate three key components: knockoffs, subsampling stability, random feature mapping function approximation. show proposed method guaranteed FDR any size, achieves power approaches one as demonstrate efficacy our through intensive simulations comparisons with alternative solutions. Our thus, makes useful contributions methodology selection, FDR-based inference, well knockoffs. Supplementary materials article available online.