CalibratedClassifierCV
Post-hoc probability calibration that maps a base classifier's scores onto well-calibrated probabilities using sigmoid (Platt) or isotonic regression fit by cross-validation.
The examples/calibration/calibrated_classifier_cv.cpp program renders a reliability curve comparing predicted positive-class probabilities with observed positive fractions:
Algorithm
For each CV fold the base estimator is fit on the training part and a calibrator (1-D logistic regression for sigmoid, PAVA for isotonic) is fit on the held-out scores. Predictions average the per-fold calibrators. v1.1.0 covers binary classification.
Constructor
Skigen::CalibratedClassifierCV<Base, Scalar> model(Base est, Method = Sigmoid, int cv = 5, int n_jobs = 1, bool ensemble = true);
Parameters
| Parameter | Default | Description |
|---|---|---|
estimator | — | Base classifier to calibrate. |
method | Sigmoid | Sigmoid (Platt) or Isotonic. |
cv | 5 | Cross-validation folds. |
ensemble | true | Average per-fold calibrators. |
Methods
| Method | Description |
|---|---|
fit(X, y) | Fit base estimators and calibrators. |
predict(X) | Calibrated class labels. |
predict_proba(X) | Calibrated probabilities. |
Fitted Attributes
| Accessor | Description |
|---|---|
n_classes() | Number of classes. |
n_estimators_fitted() | Calibrated base estimators held. |
Example
Skigen::GaussianNB<double> nb;
Skigen::CalibratedClassifierCV cc(nb, Skigen::CalibrationMethod::Sigmoid, 5);
cc.fit(X, y);
This estimator is checked by the parity suite. See the generator tests/parity/generate_calibration_reference.py and the reference fixtures in tests/parity/data/calibrated_classifier_cv/, exercised by tests/parity/parity_calibration.cpp.
For full signatures see the CalibratedClassifierCV API Reference.