Set-Preserving Calibration from Conformal P-Values to E-Values
Researchers have developed a new method for converting conformal p-values into e-values, addressing limitations in existing approaches. This novel P2E calibrator ensures that prediction sets remain unchanged while potentially improving statistical efficiency. The new technique is demonstrated to satisfy coverage guarantees and enhance efficiency in applications like cross-conformal prediction and conformal aggregation, opening new avenues for uncertainty quantification. AI
IMPACT Enhances uncertainty quantification methods applicable to machine learning models.