Ce Calibrated Predict

CE Calibrated Predict

You are obtaining calibrated predictions without explanations from a CE explainer. This is the lightest entry point — no explanation rules are generated, just the calibrated prediction and (optionally) its uncertainty interval.

A fully fit + calibrated WrapCalibratedExplainer is required.

Quick reference

Classification — predict

# Point prediction (returns class labels for classification)
y_hat = explainer.predict(x_test)

# With uncertainty interval (returns tuple)
y_hat, (low, high) = explainer.predict(x_test, uq_interval=True)
# Interpretation: low ≤ p(class) ≤ high (Venn-Abers probability bounds)
#   Binary: p(class=1) bounds (always positive class)
#   Multiclass: p(class=y_hat) bounds (predicted class)

# Force uncalibrated predictions (bypasses calibration; warns if calibrated)
y_uncal = explainer.predict(x_test, calibrated=False)

Classification — predict_proba

# Calibrated probability array (shape: [n, n_classes])
proba = explainer.predict_proba(x_test)

# With uncertainty interval (tuple of arrays)
proba, (low, high) = explainer.predict_proba(x_test, uq_interval=True)
#   Binary: 1D probability arrays (low_1, high_1) for positive class 1
#   Multiclass: 2D arrays (n_samples, n_classes) for all classes

# Uncalibrated (uses underlying learner's predict_proba directly)
proba_uncal = explainer.predict_proba(x_test, calibrated=False)

Regression — predict

# Calibrated point prediction (CPS median by default)
y_hat = explainer.predict(x_test)

# With 90% conformal interval (target variable space)
# Interpretation: low ≤ y_hat ≤ high (e.g., in dollars or years)
y_hat, (low, high) = explainer.predict(x_test, uq_interval=True)

# Custom percentile interval
y_hat, (low, high) = explainer.predict(
    x_test,
    uq_interval=True,
    low_high_percentiles=(10, 90),
)

Regression — predict_proba (thresholded)

# P(y ≤ 50.0) for each instance — scalar threshold
proba = explainer.predict_proba(x_test, threshold=50.0)

# With uncertainty interval
proba, (low, high) = explainer.predict_proba(
    x_test, threshold=50.0, uq_interval=True
)
# Interpretation: low ≤ proba ≤ high are probability bounds (not y-space)

# Two-sided window: P(40.0 < y ≤ 60.0)
proba = explainer.predict_proba(x_test, threshold=(40.0, 60.0))

Conditional predictions (Mondrian bins)

If the explainer was calibrated with mc= or bins=, pass the same group assignments at predict time:

# Predict with group-specific calibration
y_hat = explainer.predict(x_test, bins=group_labels_test)
proba = explainer.predict_proba(x_test, bins=group_labels_test)

See ce-mondrian-conditional for step-by-step group calibration setup.

Reject-aware predictions

from calibrated_explanations.core.reject.policy import RejectPolicy

# Returns RejectResult envelope instead of plain array
result = explainer.predict(x_test, reject_policy=RejectPolicy.FLAG)
y_hat = result.prediction
rejected = result.rejected  # bool mask

See ce-reject-policy for full policy documentation.

Uncalibrated model — expected warnings

Calling predict / predict_proba with calibrated=True (default) on an uncalibrated explainer emits a UserWarning and falls back to the underlying learner's predict method. This is intentional (fallback visibility policy).

# Before calibrate() — will warn:
with pytest.warns(UserWarning):
    y_hat = explainer.predict(x_test)   # warns + returns uncalibrated

Do not suppress this warning — it is a mandatory fallback signal (see ce-fallback-impl).

Choosing between predict and explain_factual

predict and predict_proba do not generate explanation rules; they are significantly cheaper than the explain* entry points.

Evaluation Checklist

• fitted=True and calibrated=True confirmed before calling predict.
• Return type identified: array (no uq_interval) vs tuple (with uq_interval).
• For regression with uq_interval=True: invariant low ≤ predict ≤ high verified.
• For regression predict_proba(threshold=t): result is a probability (0–1), not y-space.
• Mondrian bins= passed at predict time if calibrated with mc=.
• Uncalibrated fallback warning handled (not suppressed) in tests.