Conformalized Super Learner

A new statistical method called the Conformalized Super Learner (CSL) bridges ensemble machine learning with rigorous uncertainty quantification. Developed by researchers Zhanli Wu, Fabrizio Leisen, Miguel-Angel Luque-Fernandez, and F. Javier Rubio, the method integrates conformal prediction (CP)—a framework that provides prediction intervals with finite-sample coverage guarantees under mild assumptions—into the Super Learner (SL) ensemble. The SL combines predictions from a library of learners based on their performance, but traditional interval methods rely on asymptotic arguments or computationally expensive bootstrapping. CSL mirrors the original SL framework by using individual learner weights and combining learner-specific conformity scores via a weighted majority vote, enabling interval predictions that are both valid and efficient.

In a comprehensive simulation study, CSL demonstrated valid finite-sample coverage across settings including exchangeability, violations of exchangeability, heteroscedasticity, sparsity, and other distributional heterogeneities. The method's practical utility was showcased in an application predicting creatinine levels using socio-demographic, biometric, and laboratory measurements. This example highlighted the benefits of a carefully selected ensemble of learners designed to capture non-linear effects, interactions, sparsity, heteroscedasticity, and robustness to outliers. The work is published on arXiv and includes R code and data, making it accessible for practitioners seeking reliable uncertainty quantification in complex regression tasks.