A1267
Title: Response-guided knockoffs for directional FDR control in linear models
Authors: Jack Freestone - Macquarie University (Australia) [presenting]
Garth Tarr - University of Sydney (Australia)
Samuel Muller - Macquarie University (Australia)
Uri Keich - University of Sydney (Australia)
Abstract: In many scientific applications feature selection is inherently one-sided: the analyst cares only about discoveries whose effect acts in a pre-specified direction. Finite-sample control of the directional FDR (FDR) in linear models under this one-sided regime is studied. Existing knockoff filters control the directional FDR but treat positive and negative effects symmetrically, and their constructions are entirely response-agnostic, allocating resources uniformly across features regardless of their likely sign. Two procedures address these limitations. The first, the signed-knockoff filter, adapts the standard knockoff filter to target a pre-specified direction with provable directional FDR control. The second, the response-guided knockoff filter, uses a noise-perturbed copy of the response to steer the knockoff construction toward features whose effects are likely to match the target sign. A simple correction condition on the knockoffs overcomes the apparent double-use of the response and preserves finite-sample FDR guarantees. The construction also provides a reduction in the sample-size requirement of the fixed-X knockoff filter and extends directly to directional FDX and k-FWER control. Simulations and HIV drug-resistance experiments demonstrate substantial power gains over existing knockoff-based methods.
