From Empirical Success to Theoretical Foundations: Multicalibration Gradient Boosting at Scale

Abstract: Multicalibration requires predictions to be calibrated simultaneously across a vast collection of potentially overlapping subpopulations, and has emerged as a rigorous criterion for trustworthy machine learning with applications ranging from algorithmic fairness and beyond. Despite its promise, multicalibration has seen limited industrial adoption, hampered by the need for manual group specification, poor scalability, and the risk of harming model performance. In this talk, I will present two complementary contributions. First, I will introduce MCGrad, a scalable multicalibration algorithm deployed in production at Meta across hundreds of models. MCGrad eliminates the needfor pre-specified protected groups by leveraging recursive gradientboosted decision trees over an augmented feature space, with logitrescaling and early stopping ensuring safe deployment. Second, I willaddress a fundamental open question about whether this algorithm classactually converges. By modelling multicalibration gradient boosting as a deterministic dynamical system, I will present the first finite-time convergence guarantees for this setting, showing that empirical multicalibration error decays at O(1/√T), improving to linear convergence under smoothness conditions on the weak learners. Together, these results bridge the gap between the empirical success of multicalibration gradient boosting and its theoretical foundations.Joint work with Pavlos Athanasios Apostolopoulos, Daniel Haimovich, DimaKaramshuk, Fridolin Linder, Nastaran Okati, Lorenzo Perini, and Niek Tax.