File size: 2,067 Bytes
fc329a3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 | """Approximate jackknife+ for simplex-valued prediction tasks.
This implementation uses leave-one-out local score normalization on the
calibration sample as a practical surrogate when full model retraining is
unavailable. It is closer in spirit to jackknife+ than split CP, but it is
still an approximation in this project because the underlying predictor is not
re-fit for each leave-one-out fold.
"""
import numpy as np
from ._knn_sigma import knn_sigma_hat, knn_sigma_leave_one_out
from ._split_quantile import split_conformal_quantile
from .base import ConformalResult
def jackknife_plus_conformal(
R_cal: np.ndarray,
R_test: np.ndarray,
alpha: float,
U_cal: np.ndarray | None = None,
U_test: np.ndarray | None = None,
loo_scores: np.ndarray | None = None,
k: int = 20,
) -> ConformalResult:
"""Approximate jackknife+ using leave-one-out calibration scores.
Args:
R_cal: calibration residuals (n_cal,)
R_test: test residuals (n_test,)
alpha: miscoverage level
U_cal: calibration predictions (n_cal, K), optional
U_test: test predictions (n_test, K), optional
loo_scores: pre-computed leave-one-out scores, optional
k: kNN neighbors for local scale estimation when U inputs are provided
Returns:
ConformalResult with a global or locally-rescaled radius.
"""
if loo_scores is None:
if U_cal is not None and U_test is not None:
sigma_loo = knn_sigma_leave_one_out(U_cal, R_cal, k=k)
loo_scores = R_cal / sigma_loo
else:
loo_scores = np.asarray(R_cal, dtype=float)
else:
loo_scores = np.asarray(loo_scores, dtype=float)
q = split_conformal_quantile(loo_scores, alpha)
if U_cal is not None and U_test is not None:
sigma_test = knn_sigma_hat(U_cal, R_cal, U_test, k=k)
radius = sigma_test * q
else:
radius = np.full_like(R_test, q, dtype=float)
covered = R_test <= radius
return ConformalResult(covered=covered, radius=radius, threshold=q)
|