Reproduce “Conformal prediction beyond exchangeability” on OpenResearch

About this paper

Rina Foygel Barber, Emmanuel J. Candes, Aaditya Ramdas, Ryan J. Tibshirani

This work by Rina Foygel Barber and collaborators extends conformal prediction to reliably operate with non-exchangeable data and non-symmetric machine learning algorithms. The methods provide theoretical robustness guarantees on coverage probability and demonstrate empirically that they maintain target coverage while potentially achieving narrower prediction intervals in settings with distribution drift or correlated data.

Key insights

• The coverage degradation under non-exchangeability is bounded by the total variation distance between residual vectors, not the raw data distributions, highlighting the importance of the nonconformity score.
• The proposed generalized methods retain exact coverage at the target level if the data is, in fact, exchangeable, even when using arbitrary fixed weights.
• By allowing non-symmetric, weighted algorithms within the conformal framework, it is possible to achieve tighter prediction intervals while maintaining statistical coverage guarantees.

Summary by alphaXiv.