Quick API

This page is a compact method map for WrapCalibratedExplainer.

Mode-specific semantics notes

Classification

  • Calibration prerequisites: fit on proper split and calibrate on held-out calibration split.

  • Mode-specific guarantees: Venn-Abers calibrated class probabilities with intervals.

  • Assumptions: exchangeability or calibration-deployment distribution match.

  • Explicit non-guarantees: no guarantee under drift or regime shift.

  • Explanation-envelope limits: feature-level intervals are model-response summaries, not causal claims.

  • Formal semantics: Calibrated interval semantics.

Percentile or interval regression

  • Calibration prerequisites: fit regression model and calibrate CPS on held-out calibration split.

  • Mode-specific guarantees: CPS percentile intervals for requested low_high_percentiles.

  • Assumptions: exchangeability or calibration-deployment distribution match.

  • Explicit non-guarantees: no guarantee under drift or fixed interval width across subpopulations.

  • Explanation-envelope limits: interval effects on explanations summarize model behavior under perturbation.

  • Formal semantics: Calibrated interval semantics.

Probabilistic or thresholded regression

  • Calibration prerequisites: fit regression model and calibrate before threshold queries.

  • Mode-specific guarantees: threshold events use CPS with Venn-Abers calibrated probabilities.

  • Assumptions: exchangeability or calibration-deployment distribution match.

  • Explicit non-guarantees: no guarantee under drift and no causal guarantee from threshold probabilities.

  • Explanation-envelope limits: feature-level probability shifts are model-response summaries.

  • Formal semantics: Calibrated interval semantics.

Core methods

pred = explainer.predict(X_query)
pred, (low, high) = explainer.predict(X_query, uq_interval=True)
probs = explainer.predict_proba(X_query)
probs, (low, high) = explainer.predict_proba(X_query, uq_interval=True)
factual = explainer.explain_factual(X_query)
alternatives = explainer.explore_alternatives(X_query)

Reject-aware prediction note

When predict(...) or predict_proba(...) is called with a non-NONE reject_policy, the return value is a RejectResult envelope. For ONLY_ACCEPTED and ONLY_REJECTED, result.prediction remains full-batch; use result.metadata["source_indices"] and result.metadata["matched_count"] to identify the matched subset. Reject-enabled explain_factual(...) and explore_alternatives(...) return reject-aware collections (RejectCalibratedExplanations / RejectAlternativeExplanations), not a RejectResult envelope.

Classification

explainer.fit(x_proper, y_proper)
explainer.calibrate(x_cal, y_cal, feature_names=feature_names)
probs, (low, high) = explainer.predict_proba(X_sample, uq_interval=True)

Percentile or interval regression

pred, (low, high) = explainer.predict(
    X_sample,
    uq_interval=True,
    low_high_percentiles=(5, 95),
)

Probabilistic or thresholded regression

p = explainer.predict_proba(X_sample, threshold=120.0)
p, (plo, phi) = explainer.predict_proba(X_sample, uq_interval=True, threshold=120.0)

Entry-point tier: Tier 2.