A1963
Title: A unified framework for multicalibration boosting with application to survival prediction
Authors: Hanxuan Ye - University of Pennsylvania (United States) [presenting]
Hongzhe Li - University of Pennsylvania (United States)
Abstract: Reliable risk prediction across diverse subpopulations is a central challenge when training and target populations differ. Multicalibration addresses this by strengthening classical calibration to require predictions to be unbiased across a rich class of functions encompassing prediction slices and subpopulations. A post-processing boosting algorithm for censored survival data produces multicalibrated predictors of survival probability and restricted mean survival time. Building on pseudo-observations and a functional delta-method analysis, the method is competitive with, and often outperforms, inverse propensity score weighting in unlabeled target domains. A cardiovascular risk study using two large prospective cohorts demonstrates its value across heterogeneous patient subgroups. The mechanism and theoretical foundations of multicalibration boosting are further explored, unifying variants such as multiaccuracy, BatchGCP, and BatchMVP under weaker conditions. The boosting limit is characterized as the Bregman projection of the population-optimal predictor onto the cumulative span of the audit class, identifying the function space on which multicalibration is achieved. Convergence rates, finite-sample guarantees, and a principled early-stopping rule controlling a calibration risk trade-off are derived. Universal adaptability is further extended, clarifying when multicalibrated predictors generalize across domains.
