Bayesian knockoff filter for controlling false discovery rate.

The Bayesian knockoff filter (BKF) controls false discoveries by using a tight posterior null bound and an e-value extension that guarantees frequentist FDR control.

• Why it matters: Many feature-selection problems require finding true effects among many candidates while keeping the false discovery proportion controlled. Existing knockoff and Bayesian selection approaches do not fully combine Bayesian side-information handling with reliable FDR guarantees.
• What they did: BKF builds on a Bayesian “flip-sign” property to bound posterior null probabilities, and it regenerates knockoff variables within each Markov chain Monte Carlo iteration. It also introduces an e-value BKF method to provide theoretical frequentist FDR control.
• The result: Experiments on synthetic and real data show BKF has comparable power, lower FDR, and more stable selection sets than existing knockoff methods and Bayesian variable selection approaches.

AI summary of the abstract · may contain errors · not medical advice · read the paper.