Reproduce “Conformal Prediction with Missing Values” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Margaux Zaffran, Aymeric Dieuleveut, Julie Josse, Yaniv Romano
This paper demonstrates that while standard conformal prediction methods preserve marginal coverage on imputed data, they fail to provide reliable coverage conditional on specific missing data patterns due to induced heteroskedasticity. Researchers from EDF R&D, INRIA, École polytechnique, and Technion propose Conformal Prediction with Missing Data Augmentation (CP-MDA) methods that adjust the calibration process based on missingness, ensuring statistically valid prediction intervals even when input features are incomplete.
Key insights
- Applying standard split conformal prediction on imputed data preserves marginal coverage, but fails to provide valid coverage conditional on specific missing data patterns.
- Missing values inherently induce heteroskedasticity, meaning the optimal prediction interval length varies depending on which features are missing.
- Adjusting the conformal calibration process to explicitly account for and match the missingness pattern of the test point is crucial for achieving mask-conditional validity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.