Boosted Conformal Prediction Intervals

Authors: Ran Xie, Rina Barber, Emmanuel Candes

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Systematic experiments demonstrate that starting from conventional conformal methods, our boosted procedure achieves substantial improvements in reducing interval length and decreasing deviation from target conditional coverage.
Researcher Affiliation Academia Ran Xie Department of Statistics Stanford University ranxie@stanford.edu Rina Foygel Barber Department of Statistics University of Chicago rina@uchicago.edu Emmanuel J. Cand es Department of Statistics Department of Mathematics Stanford University candes@stanford.edu
Pseudocode Yes Algorithm 1 Boosting stage
Open Source Code Yes Source code for implementing the boosted conformal procedure is available online at https://github.com/ran-xie/boosted-conformal.
Open Datasets Yes All datasets, except for the meps and star data sets, are licensed under CC-BY 4.0. The Medical Expenditure Panel Survey (meps) data is subject to copyright and usage rules. The licensing status of the star dataset could not be determined.
Dataset Splits Yes The number τ is calculated using k-fold cross-validation on the training dataset, selecting τ from potential values up to a predefined maximum T (e.g., 500). We partition the dataset into k folds and for each j = 1, . . . , k, we hold out fold j for sub-calibration and the remaining k 1 folds for sub-training.
Hardware Specification Yes All experiments were conducted on a dual-socket AMD EPYC 7502 32-Core Processor system, utilizing 8 of its 128 CPUs each time.
Software Dependencies No The paper mentions 'Python s scikit-learn package' and 'Python s XGBoost package' but does not specify exact version numbers for these or other software dependencies.
Experiment Setup Yes We set the hyper-parameters τ1, τ2 in the approximated loss (17) to 50. The approximated loss is then passed to the Gradient Boosting Machine from Python s XGBoost package along with a base conformity score. We set the maximum tree depth to 1 to avoid overfitting and perform cross-validation for the number of boosting rounds, as outlined in Section 3.