API: Calibration

Split-conformal calibration normalized by a DEUP uncertainty scale.

Parameters:

Name	Type	Description	Default
method	ConformalMethod	"normalized" (default), "mondrian", or "cqr".	'normalized'
alpha	float	Miscoverage level; target coverage is 1 - alpha.	0.1
eps	float	Floor added to the uncertainty scale to avoid division by zero.	1e-09

Notes

Fit on a held-out calibration set that the base model and g did not train on, otherwise coverage guarantees do not hold. The DEUP estimators handle this by calibrating on out-of-fold predictions.

fit(y_true, y_pred, uncertainty=None, *, groups=None, y_lower=None, y_upper=None)

Calibrate on a held-out set.

Parameters:

Name	Type	Description	Default
y_true	ArrayLike	Calibration targets and base-model point predictions.	required
y_pred	ArrayLike	Calibration targets and base-model point predictions.	required
uncertainty	ArrayLike | None	The DEUP scale g(x) on the calibration set (normalized/mondrian).	None
groups	ArrayLike | None	Per-row group labels (mondrian only).	None
y_lower	ArrayLike | None	Pre-fit quantile predictions on the calibration set (cqr only).	None
y_upper	ArrayLike | None	Pre-fit quantile predictions on the calibration set (cqr only).	None

predict_interval(y_pred, uncertainty=None, *, groups=None, y_lower=None, y_upper=None)

Return calibrated (lower, upper) interval bounds for new points.

Lower/upper prediction-interval bounds.

Return a callable X -> g(x) for use as a MAPIE-style normalization function.

MAPIE's locally adaptive conformal methods accept a per-point scale; passing a fitted DEUP estimator's epistemic estimate as that scale yields DEUP-normalized conformal intervals inside MAPIE. The returned callable exposes predict so it can stand in as a residual-scaling model.

Example

from deup import DEUPRegressor from deup.calibration import deup_normalizer model = DEUPRegressor().fit(X_train, y_train) normalizer = deup_normalizer(model) scale = normalizer.predict(X_cal) # == model.predict_epistemic(X_cal)