Powerful knockoffs via minimizing reconstructability

Model-X knockoffs (J. R. Stat. Soc. Ser. B. Methodol. 80 (2018) 551–577) allows analysts to perform feature selection using almost any machine learning algorithm while provably controlling the expected proportion of false discoveries. This procedure involves constructing synthetic variables, called knockoffs, which effectively act as controls during selection. The gold standard for has been minimize mean absolute correlation (MAC) between features and their but, surprisingly, we prove this can be powerless in extremely easy settings, including Gaussian linear models with correlated exchangeable features. key problem is that minimizing MAC creates joint dependencies allow algorithms reconstruct effect on response knockoffs. To improve power, propose generating reconstructability (MRC) features, demonstrate our proposal by showing it computationally efficient, robust, powerful. We also certain MRC a notion estimation error models. Through extensive simulations, show often dramatically outperform MAC-minimizing find no settings more than slight margin. implement methods many others from literature new python package knockpy.