Title:The Holdout Randomization Test: Principled and Easy Black Box Feature Selection

Abstract:We consider the problem of feature selection using black box predictive models. For example, high-throughput devices in science are routinely used to gather thousands of features for each sample in an experiment. The scientist must then sift through the many candidate features to find explanatory signals in the data, such as which genes are associated with sensitivity to a prospective therapy. Often, predictive models are used for this task: the model is fit, error on held out data is measured, and strong performing models are assumed to have discovered some fundamental properties of the system. A model-specific heuristic is then used to inspect the model parameters and rank important features, with top features reported as "discoveries." However, such heuristics provide no statistical guarantees and can produce unreliable results. We propose the holdout randomization test (HRT) as a principled approach to feature selection using black box predictive models. The HRT is model agnostic and produces a valid p-value for each feature, enabling control over the false discovery rate (or Type I error) for any predictive model. Further, the HRT is computationally efficient and, in simulations, has greater power than a competing knockoffs-based approach. Code is available at this https URL.