Conformal Selection for Multivariate Data

Selecting high-quality candidates is crucial in drug discovery and precision medicine. While Conformal Selection (CS) ensures uncertainty quantification, it is limited to univariate responses. We propose Multivariate Conformal Selection (mCS), extending CS to multivariate settings using regional monotonicity and multivariate nonconformity scores for conformal p-values, ensuring finite-sample False Discovery Rate (FDR) control. We introduce two variants: one using distance-based scores and another optimizing scores via differentiable learning. Experiments on simulated and real-world data show mCS enhances selection power while maintaining FDR control, making it a robust tool for multivariate selection.

About Archer Yang

Archer Yang is an Associate Professor in the Department of Mathematics and Statistics at McGill University, as well as an associate member of the School of Computer Science and the Quantitative Life Sciences program. He is also an associate academic member of Mila – Quebec AI Institute.  His research focuses on statistical machine learning, statistical computing, and high-dimensional statistics, with applications in biomedical and biochemical sciences, industrial data science, and drug discovery.