Calibrated automatic relevance determination: finite-sample false discovery rate control for variable selection in Gaussian process regression
Abstract Automatic relevance determination (ARD) endows Gaussian process (GP) regression with a per-input relevance score—the reciprocal lengthscale—widely used to rank and select explanatory variables. In practice a variable is declared relevant by thresholding these scores with an informal rule, often calibrated against “inert” reference variables known to be inactive; such rules carry no guarantee on the proportion of false selections. We formalize the inert-reference device into a procedure with provable finite-sample false discovery rate (FDR) control. The central observation is an exchangeability argument: under a symmetric (independent-column) design, the relevance scores of genuinely null inputs are exchangeable with the scores of synthetic reference variables, yielding valid conformal-style p-values and e-values from a single GP fit. Feeding these e-values to the e-Benjamini–Hochberg procedure controls FDR under arbitrary dependence; the Benjamini–Hochberg procedure applies when references are scored by independent per-input fits. A salient feature is robustness to non-identifiability: because the guarantee rests on column symmetry of the relevance map rather than consistent estimation of the lengthscales, it sidesteps their well-known weak identifiability, which affects power but not validity. The procedure adds a calibration layer to one GP fit; we establish its exchangeability lemma, FDR theorems, a derandomized variant and selection consistency, and analyse its computational cost, scalability and implementation. Simulations confirm finite-sample control where the uncorrected reference rule fails, and a model-X knockoff variant handles correlated designs. An application to National Health and Nutrition Examination Survey data, with planted nulls and a semi-synthetic response, illustrates error control and power. MSC Classification: 62J15 , 62G08 , 62F15 , 60G15
Authors
- Jia Cai
Institutions
- George Mason University (US)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11319814/v1
- Primary Topic
- Statistical Methods and Inference
- Type
- preprint