Configure CV+, jackknife+ and cross-conformal prediction for conformal over a model trained with outer resampling.
setup_CVPlus(alpha = 0.1, score = NULL, seed = NULL)CVPlusConfig object.
Numeric (0, 1): Miscoverage rate. The region covers with
probability at least 1 - 2 * alpha in the worst case.
Optional Character {"absolute", "LAC", "APS"}: Nonconformity score. NULL resolves to "absolute" for a regression and "APS" for a classification. See setup_SplitConformal for the trade between the two set-valued scores, and for why APS draws.
Optional Integer [0, Inf): Random seed for the APS draw. NULL draws one and records it.
EDG
CV+ (Barber, Candes, Ramdas and Tibshirani, 2021) spends no data on
calibration: every case is out-of-fold exactly once, so every case both
trains and calibrates. A SupervisedRes stores the per-fold models and their
out-of-fold predictions, so the computation is predicting newdata under
each fold's model and taking two order statistics.
It delivers 1 - 2 * alpha in the worst case, not 1 - alpha. The bound
is conservative and observed coverage usually sits near 1 - alpha; the
region states the guarantee it carries.
The folds must partition the cases, each case out-of-fold exactly once.
setup_KFold() and setup_LOOCV() do; the subsampling and
bootstrap types do not, and conformal() refuses them by checking the
indices.
Which relative runs follows from the object: leave-one-out folds make it jackknife+, and a classification outcome makes it cross-conformal (Vovk, 2015), CV+ being defined for intervals. The result records which.
setup_CVPlus()
setup_CVPlus(alpha = 0.2)
Run the code above in your browser using DataLab