BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Conformal Prediction with Missing Values
DTSTART:20231024T091500Z
DTEND:20231024T101500Z
DTSTAMP:20260504T062300Z
UID:indico-event-10245@indico.math.cnrs.fr
DESCRIPTION:Speakers: Margaux Zaffran (INRIA\, CMAP)\n\nUncertainty quanti
 fication of predictive models is crucial in decision-making problems. Conf
 ormal Prediction (CP) is a theoretically grounded framework for constructi
 ng prediction intervals with finite sample distribution-free marginal cove
 rage guarantee for any underlying machine learning model. The presence of 
 missing values in real data brings additional challenges to uncertainty qu
 antification. Despite an abundant literature on missing data\, as far as w
 e know\, there is no work studying the quantification of uncertainty in pr
 edictive models. In this talk\, we will first introduce in details CP\, a
 long with its limitations and current active research directions. Then\, w
 e will study conformal prediction with missing covariates. We first show t
 hat the marginal coverage guarantee of conformal prediction holds on imput
 ed data for any missingness distribution and almost all imputation functio
 ns. However\, we emphasize that the average coverage varies depending on t
 he pattern of missing values: conformal methods tend to construct predicti
 on intervals that under-cover the response conditionally to some missing p
 atterns. This motivates our novel generalized conformalized quantile regre
 ssion framework\, missing data augmentation\, which yields prediction inte
 rvals that are valid conditionally to the patterns of missing values\, des
 pite their exponential number. Using synthetic and data from critical medi
 cal care\, we corroborate our theory and report improved performance of ou
 r methods.\n\nhttps://indico.math.cnrs.fr/event/10245/
LOCATION:Salle K. Johnson\, 1er étage (1R3)
URL:https://indico.math.cnrs.fr/event/10245/
END:VEVENT
END:VCALENDAR