Date: Friday, 26 June 2026, at 12.30pm (CET)

Dual format: ESSEC Paris La Défense (CNIT), Room 209
and via Zoom, please click here

Uncertainty quantification is a fundamental problem to solve for adopting modern machine learning models in critical real-world applications. In a standard supervised regression or classification setting, this is often translated into constructing prediction sets. A meaningful prediction set would contain the unknown label of the given test point with high probability. Conformal prediction offers a simple mechanism to construct prediction sets with finite-sample guarantees, independently of the choice of the underlying models and estimation methods. While the method is well adapted under the exchangeability assumption on the data-generating process, it fails under the perturbations such as noisy labels or missing data. In this work, we revisit the problem of conformal prediction under missing data scenarios. It has been noted that, due to heterogeneity induced by missingness, it would be more useful to have a prediction set that adapts to the pattern of the missingness, thus giving rise to the notion of mask-conditional validity. Existing methods are known to be limited to Missing-Complete-At-Random setting only. We develop a new framework valid for general missing mechanisms. Noting that imputation is widely used as a pre-processing step, we propose a novel preimpute-mask-then-correct framework that is applicable in conjunction with imputation. Our proposed methods are shown to give valid prediction sets without any assumptions on the underlying missing mechanism. Practical implications are illustrated with simulated and real data examples, including ICU length-of-stay forecasting.

Kind regards,
Pierre Alquier, Marie Kratz, Roberto Reno, and Riada Djebbar (Singapore Actuarial Society - ERM)