"""Methods for calibrating predicted probabilities.""" # Authors: The scikit-learn developers # SPDX-License-Identifier: BSD-3-Clause import warnings from functools import partial from inspect import signature from math import log from numbers import Integral, Real import numpy as np from scipy.optimize import minimize, minimize_scalar from scipy.special import expit from sklearn._loss import HalfBinomialLoss, HalfMultinomialLoss from sklearn.base import ( BaseEstimator, ClassifierMixin, MetaEstimatorMixin, RegressorMixin, _fit_context, clone, ) from sklearn.externals import array_api_extra as xpx from sklearn.frozen import FrozenEstimator from sklearn.isotonic import IsotonicRegression from sklearn.model_selection import LeaveOneOut, check_cv, cross_val_predict from sklearn.preprocessing import LabelEncoder, label_binarize from sklearn.svm import LinearSVC from sklearn.utils import Bunch, _safe_indexing, column_or_1d, get_tags, indexable from sklearn.utils._array_api import ( _convert_to_numpy, _half_multinomial_loss, _is_numpy_namespace, get_namespace, get_namespace_and_device, move_to, ) from sklearn.utils._param_validation import ( HasMethods, Interval, StrOptions, validate_params, ) from sklearn.utils._plotting import ( _BinaryClassifierCurveDisplayMixin, _validate_style_kwargs, ) from sklearn.utils._response import _get_response_values, _process_predict_proba from sklearn.utils.extmath import softmax from sklearn.utils.metadata_routing import ( MetadataRouter, MethodMapping, _routing_enabled, process_routing, ) from sklearn.utils.multiclass import check_classification_targets from sklearn.utils.parallel import Parallel, delayed from sklearn.utils.validation import ( _check_method_params, _check_pos_label_consistency, _check_response_method, _check_sample_weight, _num_samples, check_array, check_consistent_length, check_is_fitted, ) class CalibratedClassifierCV(ClassifierMixin, MetaEstimatorMixin, BaseEstimator): """Calibrate probabilities using isotonic, sigmoid, or temperature scaling. This class uses cross-validation to both estimate the parameters of a classifier and subsequently calibrate a classifier. With `ensemble=True`, for each cv split it fits a copy of the base estimator to the training subset, and calibrates it using the testing subset. For prediction, predicted probabilities are averaged across these individual calibrated classifiers. When `ensemble=False`, cross-validation is used to obtain unbiased predictions, via :func:`~sklearn.model_selection.cross_val_predict`, which are then used for calibration. For prediction, the base estimator, trained using all the data, is used. This is the prediction method implemented when `probabilities=True` for :class:`~sklearn.svm.SVC` and :class:`~sklearn.svm.NuSVC` estimators (see :ref:`User Guide ` for details). Already fitted classifiers can be calibrated by wrapping the model in a :class:`~sklearn.frozen.FrozenEstimator`. In this case all provided data is used for calibration. The user has to take care manually that data for model fitting and calibration are disjoint. The calibration is based on the :term:`decision_function` method of the `estimator` if it exists, else on :term:`predict_proba`. Read more in the :ref:`User Guide `. In order to learn more on the CalibratedClassifierCV class, see the following calibration examples: :ref:`sphx_glr_auto_examples_calibration_plot_calibration.py`, :ref:`sphx_glr_auto_examples_calibration_plot_calibration_curve.py`, and :ref:`sphx_glr_auto_examples_calibration_plot_calibration_multiclass.py`. Parameters ---------- estimator : estimator instance, default=None The classifier whose output need to be calibrated to provide more accurate `predict_proba` outputs. The default classifier is a :class:`~sklearn.svm.LinearSVC`. .. versionadded:: 1.2 method : {'sigmoid', 'isotonic', 'temperature'}, default='sigmoid' The method to use for calibration. Can be: - 'sigmoid', which corresponds to Platt's method (i.e. a binary logistic regression model). - 'isotonic', which is a non-parametric approach. - 'temperature', temperature scaling. Sigmoid and isotonic calibration methods natively support only binary classifiers and extend to multi-class classification using a One-vs-Rest (OvR) strategy with post-hoc renormalization, i.e., adjusting the probabilities after calibration to ensure they sum up to 1. In contrast, temperature scaling naturally supports multi-class calibration by applying `softmax(classifier_logits/T)` with a value of `T` (temperature) that optimizes the log loss. For very uncalibrated classifiers on very imbalanced datasets, sigmoid calibration might be preferred because it fits an additional intercept parameter. This helps shift decision boundaries appropriately when the classifier being calibrated is biased towards the majority class. Isotonic calibration is not recommended when the number of calibration samples is too low ``(≪1000)`` since it then tends to overfit. .. versionchanged:: 1.8 Added option 'temperature'. cv : int, cross-validation generator, or iterable, default=None Determines the cross-validation splitting strategy. Possible inputs for cv are: - None, to use the default 5-fold cross-validation, - integer, to specify the number of folds. - :term:`CV splitter`, - An iterable yielding (train, test) splits as arrays of indices. For integer/None inputs, if ``y`` is binary or multiclass, :class:`~sklearn.model_selection.StratifiedKFold` is used. If ``y`` is neither binary nor multiclass, :class:`~sklearn.model_selection.KFold` is used. Refer to the :ref:`User Guide ` for the various cross-validation strategies that can be used here. .. versionchanged:: 0.22 ``cv`` default value if None changed from 3-fold to 5-fold. n_jobs : int, default=None Number of jobs to run in parallel. ``None`` means 1 unless in a :obj:`joblib.parallel_backend` context. ``-1`` means using all processors. Base estimator clones are fitted in parallel across cross-validation iterations. See :term:`Glossary ` for more details. .. versionadded:: 0.24 ensemble : bool, or "auto", default="auto" Determines how the calibrator is fitted. "auto" will use `False` if the `estimator` is a :class:`~sklearn.frozen.FrozenEstimator`, and `True` otherwise. If `True`, the `estimator` is fitted using training data, and calibrated using testing data, for each `cv` fold. The final estimator is an ensemble of `n_cv` fitted classifier and calibrator pairs, where `n_cv` is the number of cross-validation folds. The output is the average predicted probabilities of all pairs. If `False`, `cv` is used to compute unbiased predictions, via :func:`~sklearn.model_selection.cross_val_predict`, which are then used for calibration. At prediction time, the classifier used is the `estimator` trained on all the data. Note that this method is also internally implemented in :mod:`sklearn.svm` estimators with the `probabilities=True` parameter. .. versionadded:: 0.24 .. versionchanged:: 1.6 `"auto"` option is added and is the default. Attributes ---------- classes_ : ndarray of shape (n_classes,) The class labels. n_features_in_ : int Number of features seen during :term:`fit`. Only defined if the underlying estimator exposes such an attribute when fit. .. versionadded:: 0.24 feature_names_in_ : ndarray of shape (`n_features_in_`,) Names of features seen during :term:`fit`. Only defined if the underlying estimator exposes such an attribute when fit. .. versionadded:: 1.0 calibrated_classifiers_ : list (len() equal to cv or 1 if `ensemble=False`) The list of classifier and calibrator pairs. - When `ensemble=True`, `n_cv` fitted `estimator` and calibrator pairs. `n_cv` is the number of cross-validation folds. - When `ensemble=False`, the `estimator`, fitted on all the data, and fitted calibrator. .. versionchanged:: 0.24 Single calibrated classifier case when `ensemble=False`. See Also -------- calibration_curve : Compute true and predicted probabilities for a calibration curve. References ---------- .. [1] B. Zadrozny & C. Elkan. `Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers `_, ICML 2001. .. [2] B. Zadrozny & C. Elkan. `Transforming Classifier Scores into Accurate Multiclass Probability Estimates `_, KDD 2002. .. [3] J. Platt. `Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods `_, 1999. .. [4] A. Niculescu-Mizil & R. Caruana. `Predicting Good Probabilities with Supervised Learning `_, ICML 2005. .. [5] Chuan Guo, Geoff Pleiss, Yu Sun, Kilian Q. Weinberger. :doi:`On Calibration of Modern Neural Networks<10.48550/arXiv.1706.04599>`. Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1321-1330, 2017. Examples -------- >>> from sklearn.datasets import make_classification >>> from sklearn.naive_bayes import GaussianNB >>> from sklearn.calibration import CalibratedClassifierCV >>> X, y = make_classification(n_samples=100, n_features=2, ... n_redundant=0, random_state=42) >>> base_clf = GaussianNB() >>> calibrated_clf = CalibratedClassifierCV(base_clf, cv=3) >>> calibrated_clf.fit(X, y) CalibratedClassifierCV(...) >>> len(calibrated_clf.calibrated_classifiers_) 3 >>> calibrated_clf.predict_proba(X)[:5, :] array([[0.110, 0.889], [0.072, 0.927], [0.928, 0.072], [0.928, 0.072], [0.072, 0.928]]) >>> from sklearn.model_selection import train_test_split >>> X, y = make_classification(n_samples=100, n_features=2, ... n_redundant=0, random_state=42) >>> X_train, X_calib, y_train, y_calib = train_test_split( ... X, y, random_state=42 ... ) >>> base_clf = GaussianNB() >>> base_clf.fit(X_train, y_train) GaussianNB() >>> from sklearn.frozen import FrozenEstimator >>> calibrated_clf = CalibratedClassifierCV(FrozenEstimator(base_clf)) >>> calibrated_clf.fit(X_calib, y_calib) CalibratedClassifierCV(...) >>> len(calibrated_clf.calibrated_classifiers_) 1 >>> calibrated_clf.predict_proba([[-0.5, 0.5]]) array([[0.936, 0.063]]) """ _parameter_constraints: dict = { "estimator": [ HasMethods(["fit", "predict_proba"]), HasMethods(["fit", "decision_function"]), None, ], "method": [StrOptions({"isotonic", "sigmoid", "temperature"})], "cv": ["cv_object"], "n_jobs": [Integral, None], "ensemble": ["boolean", StrOptions({"auto"})], } def __init__( self, estimator=None, *, method="sigmoid", cv=None, n_jobs=None, ensemble="auto", ): self.estimator = estimator self.method = method self.cv = cv self.n_jobs = n_jobs self.ensemble = ensemble def _get_estimator(self): """Resolve which estimator to return (default is LinearSVC)""" if self.estimator is None: # we want all classifiers that don't expose a random_state # to be deterministic (and we don't want to expose this one). estimator = LinearSVC(random_state=0) if _routing_enabled(): estimator.set_fit_request(sample_weight=True) else: estimator = self.estimator return estimator @_fit_context( # CalibratedClassifierCV.estimator is not validated yet prefer_skip_nested_validation=False ) def fit(self, X, y, sample_weight=None, **fit_params): """Fit the calibrated model. Parameters ---------- X : array-like of shape (n_samples, n_features) Training data. y : array-like of shape (n_samples,) Target values. sample_weight : array-like of shape (n_samples,), default=None Sample weights. If None, then samples are equally weighted. **fit_params : dict Parameters to pass to the `fit` method of the underlying classifier. Returns ------- self : object Returns an instance of self. """ check_classification_targets(y) X, y = indexable(X, y) estimator = self._get_estimator() _ensemble = self.ensemble if _ensemble == "auto": _ensemble = not isinstance(estimator, FrozenEstimator) self.calibrated_classifiers_ = [] # Set `classes_` using all `y` label_encoder_ = LabelEncoder().fit(y) self.classes_ = label_encoder_.classes_ if self.method == "temperature" and isinstance(y[0], str): # for temperature scaling if `y` contains strings then encode it # right here to avoid fitting LabelEncoder again within the # `_fit_calibrator` function. y = label_encoder_.transform(y=y) if _routing_enabled(): routed_params = process_routing( self, "fit", sample_weight=sample_weight, **fit_params, ) else: # sample_weight checks fit_parameters = signature(estimator.fit).parameters supports_sw = "sample_weight" in fit_parameters if sample_weight is not None and not supports_sw: estimator_name = type(estimator).__name__ warnings.warn( f"Since {estimator_name} does not appear to accept" " sample_weight, sample weights will only be used for the" " calibration itself. This can be caused by a limitation of" " the current scikit-learn API. See the following issue for" " more details:" " https://github.com/scikit-learn/scikit-learn/issues/21134." " Be warned that the result of the calibration is likely to be" " incorrect." ) routed_params = Bunch() routed_params.splitter = Bunch(split={}) # no routing for splitter routed_params.estimator = Bunch(fit=fit_params) if sample_weight is not None and supports_sw: routed_params.estimator.fit["sample_weight"] = sample_weight xp, is_array_api, device_ = get_namespace_and_device(X) if is_array_api: y, sample_weight = move_to(y, sample_weight, xp=xp, device=device_) # Check that each cross-validation fold can have at least one # example per class if isinstance(self.cv, int): n_folds = self.cv elif hasattr(self.cv, "n_splits"): n_folds = self.cv.n_splits else: n_folds = None if n_folds and xp.any(xp.unique_counts(y)[1] < n_folds): raise ValueError( f"Requesting {n_folds}-fold " "cross-validation but provided less than " f"{n_folds} examples for at least one class." ) if isinstance(self.cv, LeaveOneOut): raise ValueError( "LeaveOneOut cross-validation does not allow" "all classes to be present in test splits. " "Please use a cross-validation generator that allows " "all classes to appear in every test and train split." ) cv = check_cv(self.cv, y, classifier=True) if _ensemble: parallel = Parallel(n_jobs=self.n_jobs) self.calibrated_classifiers_ = parallel( delayed(_fit_classifier_calibrator_pair)( clone(estimator), X, y, train=train, test=test, method=self.method, classes=self.classes_, xp=xp, sample_weight=sample_weight, fit_params=routed_params.estimator.fit, ) for train, test in cv.split(X, y, **routed_params.splitter.split) ) else: this_estimator = clone(estimator) method_name = _check_response_method( this_estimator, ["decision_function", "predict_proba"], ).__name__ predictions = cross_val_predict( estimator=this_estimator, X=X, y=y, cv=cv, method=method_name, n_jobs=self.n_jobs, params=routed_params.estimator.fit, ) if self.classes_.shape[0] == 2: # Ensure shape (n_samples, 1) in the binary case if method_name == "predict_proba": # Select the probability column of the positive class predictions = _process_predict_proba( y_pred=predictions, target_type="binary", classes=self.classes_, pos_label=self.classes_[1], ) predictions = predictions.reshape(-1, 1) if sample_weight is not None: # Check that the sample_weight dtype is consistent with the # predictions to avoid unintentional upcasts. sample_weight = _check_sample_weight( sample_weight, predictions, dtype=predictions.dtype ) this_estimator.fit(X, y, **routed_params.estimator.fit) # Note: Here we don't pass on fit_params because the supported # calibrators don't support fit_params anyway calibrated_classifier = _fit_calibrator( this_estimator, predictions, y, self.classes_, self.method, xp=xp, sample_weight=sample_weight, ) self.calibrated_classifiers_.append(calibrated_classifier) first_clf = self.calibrated_classifiers_[0].estimator if hasattr(first_clf, "n_features_in_"): self.n_features_in_ = first_clf.n_features_in_ if hasattr(first_clf, "feature_names_in_"): self.feature_names_in_ = first_clf.feature_names_in_ return self def predict_proba(self, X): """Calibrated probabilities of classification. This function returns calibrated probabilities of classification according to each class on an array of test vectors X. Parameters ---------- X : array-like of shape (n_samples, n_features) The samples, as accepted by `estimator.predict_proba`. Returns ------- C : ndarray of shape (n_samples, n_classes) The predicted probas. """ check_is_fitted(self) # Compute the arithmetic mean of the predictions of the calibrated # classifiers xp, _, device_ = get_namespace_and_device(X) mean_proba = xp.zeros((_num_samples(X), self.classes_.shape[0]), device=device_) for calibrated_classifier in self.calibrated_classifiers_: proba = calibrated_classifier.predict_proba(X) mean_proba += proba mean_proba /= len(self.calibrated_classifiers_) return mean_proba def predict(self, X): """Predict the target of new samples. The predicted class is the class that has the highest probability, and can thus be different from the prediction of the uncalibrated classifier. Parameters ---------- X : array-like of shape (n_samples, n_features) The samples, as accepted by `estimator.predict`. Returns ------- C : ndarray of shape (n_samples,) The predicted class. """ xp, _ = get_namespace(X) check_is_fitted(self) class_indices = xp.argmax(self.predict_proba(X), axis=1) if isinstance(self.classes_[0], str): class_indices = _convert_to_numpy(class_indices, xp=xp) return self.classes_[class_indices] def get_metadata_routing(self): """Get metadata routing of this object. Please check :ref:`User Guide ` on how the routing mechanism works. Returns ------- routing : MetadataRouter A :class:`~sklearn.utils.metadata_routing.MetadataRouter` encapsulating routing information. """ router = ( MetadataRouter(owner=self) .add_self_request(self) .add( estimator=self._get_estimator(), method_mapping=MethodMapping().add(caller="fit", callee="fit"), ) .add( splitter=self.cv, method_mapping=MethodMapping().add(caller="fit", callee="split"), ) ) return router def __sklearn_tags__(self): tags = super().__sklearn_tags__() estimator_tags = get_tags(self._get_estimator()) tags.input_tags.sparse = estimator_tags.input_tags.sparse tags.array_api_support = ( estimator_tags.array_api_support and self.method == "temperature" ) return tags def _fit_classifier_calibrator_pair( estimator, X, y, train, test, method, classes, xp, sample_weight=None, fit_params=None, ): """Fit a classifier/calibration pair on a given train/test split. Fit the classifier on the train set, compute its predictions on the test set and use the predictions as input to fit the calibrator along with the test labels. Parameters ---------- estimator : estimator instance Cloned base estimator. X : array-like, shape (n_samples, n_features) Sample data. y : array-like, shape (n_samples,) Targets. train : ndarray, shape (n_train_indices,) Indices of the training subset. test : ndarray, shape (n_test_indices,) Indices of the testing subset. method : {'sigmoid', 'isotonic', 'temperature'} Method to use for calibration. classes : ndarray, shape (n_classes,) The target classes. xp : namespace Array API namespace. sample_weight : array-like, default=None Sample weights for `X`. fit_params : dict, default=None Parameters to pass to the `fit` method of the underlying classifier. Returns ------- calibrated_classifier : _CalibratedClassifier instance """ fit_params_train = _check_method_params(X, params=fit_params, indices=train) X_train, y_train = _safe_indexing(X, train), _safe_indexing(y, train) X_test, y_test = _safe_indexing(X, test), _safe_indexing(y, test) estimator.fit(X_train, y_train, **fit_params_train) predictions, _ = _get_response_values( estimator, X_test, response_method=["decision_function", "predict_proba"], ) if predictions.ndim == 1: # Reshape binary output from `(n_samples,)` to `(n_samples, 1)` predictions = predictions.reshape(-1, 1) if sample_weight is not None: # Check that the sample_weight dtype is consistent with the predictions # to avoid unintentional upcasts. sample_weight = _check_sample_weight(sample_weight, X, dtype=predictions.dtype) sw_test = _safe_indexing(sample_weight, test) else: sw_test = None calibrated_classifier = _fit_calibrator( estimator, predictions, y_test, classes, method, xp=xp, sample_weight=sw_test, ) return calibrated_classifier def _fit_calibrator(clf, predictions, y, classes, method, xp, sample_weight=None): """Fit calibrator(s) and return a `_CalibratedClassifier` instance. A separate calibrator is fitted for each of the `n_classes` (i.e. `len(clf.classes_)`). However, if `n_classes` is 2 or if `method` is 'temperature', only one calibrator is fitted. Parameters ---------- clf : estimator instance Fitted classifier. predictions : array-like, shape (n_samples, n_classes) or (n_samples, 1) \ when binary. Raw predictions returned by the un-calibrated base classifier. y : array-like, shape (n_samples,) The targets. For `method="temperature"`, `y` needs to be label encoded. classes : ndarray, shape (n_classes,) All the prediction classes. method : {'sigmoid', 'isotonic', 'temperature'} The method to use for calibration. xp : namespace Array API namespace. sample_weight : ndarray, shape (n_samples,), default=None Sample weights. If None, then samples are equally weighted. Returns ------- pipeline : _CalibratedClassifier instance """ calibrators = [] if method in ("isotonic", "sigmoid"): Y = label_binarize(y, classes=classes) label_encoder = LabelEncoder().fit(classes) pos_class_indices = label_encoder.transform(clf.classes_) for class_idx, this_pred in zip(pos_class_indices, predictions.T): if method == "isotonic": calibrator = IsotonicRegression(out_of_bounds="clip") else: # "sigmoid" calibrator = _SigmoidCalibration() calibrator.fit(this_pred, Y[:, class_idx], sample_weight) calibrators.append(calibrator) elif method == "temperature": if classes.shape[0] == 2 and predictions.shape[-1] == 1: response_method_name = _check_response_method( clf, ["decision_function", "predict_proba"], ).__name__ if response_method_name == "predict_proba": predictions = xp.concat([1 - predictions, predictions], axis=1) calibrator = _TemperatureScaling() calibrator.fit(predictions, y, sample_weight) calibrators.append(calibrator) pipeline = _CalibratedClassifier(clf, calibrators, method=method, classes=classes) return pipeline class _CalibratedClassifier: """Pipeline-like chaining a fitted classifier and its fitted calibrators. Parameters ---------- estimator : estimator instance Fitted classifier. calibrators : list of fitted estimator instances List of fitted calibrators (either 'IsotonicRegression' or '_SigmoidCalibration'). The number of calibrators equals the number of classes. However, if there are 2 classes, the list contains only one fitted calibrator. classes : array-like of shape (n_classes,) All the prediction classes. method : {'sigmoid', 'isotonic'}, default='sigmoid' The method to use for calibration. Can be 'sigmoid' which corresponds to Platt's method or 'isotonic' which is a non-parametric approach based on isotonic regression. """ def __init__(self, estimator, calibrators, *, classes, method="sigmoid"): self.estimator = estimator self.calibrators = calibrators self.classes = classes self.method = method def predict_proba(self, X): """Calculate calibrated probabilities. Calculates classification calibrated probabilities for each class, in a one-vs-all manner, for `X`. Parameters ---------- X : ndarray of shape (n_samples, n_features) The sample data. Returns ------- proba : array, shape (n_samples, n_classes) The predicted probabilities. Can be exact zeros. """ predictions, _ = _get_response_values( self.estimator, X, response_method=["decision_function", "predict_proba"], ) if predictions.ndim == 1: # Reshape binary output from `(n_samples,)` to `(n_samples, 1)` predictions = predictions.reshape(-1, 1) n_classes = self.classes.shape[0] proba = np.zeros((_num_samples(X), n_classes)) if self.method in ("sigmoid", "isotonic"): label_encoder = LabelEncoder().fit(self.classes) pos_class_indices = label_encoder.transform(self.estimator.classes_) for class_idx, this_pred, calibrator in zip( pos_class_indices, predictions.T, self.calibrators ): if n_classes == 2: # When binary, `predictions` consists only of predictions for # clf.classes_[1] but `pos_class_indices` = 0 class_idx += 1 proba[:, class_idx] = calibrator.predict(this_pred) # Normalize the probabilities if n_classes == 2: proba[:, 0] = 1.0 - proba[:, 1] else: denominator = np.sum(proba, axis=1)[:, np.newaxis] # In the edge case where for each class calibrator returns a zero # probability for a given sample, use the uniform distribution # instead. uniform_proba = np.full_like(proba, 1 / n_classes) proba = np.divide( proba, denominator, out=uniform_proba, where=denominator != 0 ) elif self.method == "temperature": xp, _ = get_namespace(predictions) if n_classes == 2 and predictions.shape[-1] == 1: response_method_name = _check_response_method( self.estimator, ["decision_function", "predict_proba"], ).__name__ if response_method_name == "predict_proba": predictions = xp.concat([1 - predictions, predictions], axis=1) proba = self.calibrators[0].predict(predictions) # Deal with cases where the predicted probability minimally exceeds 1.0 proba[(1.0 < proba) & (proba <= 1.0 + 1e-5)] = 1.0 return proba # The max_abs_prediction_threshold was approximated using # logit(np.finfo(np.float64).eps) which is about -36 def _sigmoid_calibration( predictions, y, sample_weight=None, max_abs_prediction_threshold=30 ): """Probability Calibration with sigmoid method (Platt 2000) Parameters ---------- predictions : ndarray of shape (n_samples,) The decision function or predict proba for the samples. y : ndarray of shape (n_samples,) The targets. sample_weight : array-like of shape (n_samples,), default=None Sample weights. If None, then samples are equally weighted. Returns ------- a : float The slope. b : float The intercept. References ---------- Platt, "Probabilistic Outputs for Support Vector Machines" """ predictions = column_or_1d(predictions) y = column_or_1d(y) F = predictions # F follows Platt's notations scale_constant = 1.0 max_prediction = np.max(np.abs(F)) # If the predictions have large values we scale them in order to bring # them within a suitable range. This has no effect on the final # (prediction) result because linear models like Logisitic Regression # without a penalty are invariant to multiplying the features by a # constant. if max_prediction >= max_abs_prediction_threshold: scale_constant = max_prediction # We rescale the features in a copy: inplace rescaling could confuse # the caller and make the code harder to reason about. F = F / scale_constant # Bayesian priors (see Platt end of section 2.2): # It corresponds to the number of samples, taking into account the # `sample_weight`. mask_negative_samples = y <= 0 if sample_weight is not None: prior0 = (sample_weight[mask_negative_samples]).sum() prior1 = (sample_weight[~mask_negative_samples]).sum() else: prior0 = float(np.sum(mask_negative_samples)) prior1 = y.shape[0] - prior0 T = np.zeros_like(y, dtype=predictions.dtype) T[y > 0] = (prior1 + 1.0) / (prior1 + 2.0) T[y <= 0] = 1.0 / (prior0 + 2.0) bin_loss = HalfBinomialLoss() def loss_grad(AB): # .astype below is needed to ensure y_true and raw_prediction have the # same dtype. With result = np.float64(0) * np.array([1, 2], dtype=np.float32) # - in Numpy 2, result.dtype is float64 # - in Numpy<2, result.dtype is float32 raw_prediction = -(AB[0] * F + AB[1]).astype(dtype=predictions.dtype) l, g = bin_loss.loss_gradient( y_true=T, raw_prediction=raw_prediction, sample_weight=sample_weight, ) loss = l.sum() # TODO: Remove casting to np.float64 when minimum supported SciPy is 1.11.2 # With SciPy >= 1.11.2, the LBFGS implementation will cast to float64 # https://github.com/scipy/scipy/pull/18825. # Here we cast to float64 to support SciPy < 1.11.2 grad = np.asarray([-g @ F, -g.sum()], dtype=np.float64) return loss, grad AB0 = np.array([0.0, log((prior0 + 1.0) / (prior1 + 1.0))]) opt_result = minimize( loss_grad, AB0, method="L-BFGS-B", jac=True, options={ "gtol": 1e-6, "ftol": 64 * np.finfo(float).eps, }, ) AB_ = opt_result.x # The tuned multiplicative parameter is converted back to the original # input feature scale. The offset parameter does not need rescaling since # we did not rescale the outcome variable. return AB_[0] / scale_constant, AB_[1] def _convert_to_logits(decision_values, eps=1e-12, xp=None): """Convert decision_function values to 2D and predict_proba values to logits. This function ensures that the output of `decision_function` is converted into a (n_samples, n_classes) array. For binary classification, each row contains logits for the negative and positive classes as (-x, x). If `predict_proba` is provided instead, it is converted into log-probabilities using `numpy.log`. Parameters ---------- decision_values : array-like of shape (n_samples,) or (n_samples, 1) \ or (n_samples, n_classes). The decision function values or probability estimates. - If shape is (n_samples,), converts to (n_samples, 2) with (-x, x). - If shape is (n_samples, 1), converts to (n_samples, 2) with (-x, x). - If shape is (n_samples, n_classes), returns unchanged. - For probability estimates, returns `numpy.log(decision_values + eps)`. eps : float Small positive value added to avoid log(0). Returns ------- logits : ndarray of shape (n_samples, n_classes) """ xp, _, device_ = get_namespace_and_device(decision_values, xp=xp) decision_values = check_array( decision_values, dtype=[xp.float64, xp.float32], ensure_2d=False ) if (decision_values.ndim == 2) and (decision_values.shape[1] > 1): # Check if it is the output of predict_proba entries_zero_to_one = xp.all((decision_values >= 0) & (decision_values <= 1)) # TODO: simplify once upstream issue is addressed # https://github.com/data-apis/array-api-extra/issues/478 row_sums_to_one = xp.all( xpx.isclose( xp.sum(decision_values, axis=1), xp.asarray(1.0, device=device_, dtype=decision_values.dtype), ) ) if entries_zero_to_one and row_sums_to_one: logits = xp.log(decision_values + eps) else: logits = decision_values elif (decision_values.ndim == 2) and (decision_values.shape[1] == 1): logits = xp.concat([-decision_values, decision_values], axis=1) elif decision_values.ndim == 1: decision_values = xp.reshape(decision_values, (-1, 1)) logits = xp.concat([-decision_values, decision_values], axis=1) return logits class _SigmoidCalibration(RegressorMixin, BaseEstimator): """Sigmoid regression model. Attributes ---------- a_ : float The slope. b_ : float The intercept. """ def fit(self, X, y, sample_weight=None): """Fit the model using X, y as training data. Parameters ---------- X : array-like of shape (n_samples,) Training data. y : array-like of shape (n_samples,) Training target. sample_weight : array-like of shape (n_samples,), default=None Sample weights. If None, then samples are equally weighted. Returns ------- self : object Returns an instance of self. """ X = column_or_1d(X) y = column_or_1d(y) X, y = indexable(X, y) self.a_, self.b_ = _sigmoid_calibration(X, y, sample_weight) return self def predict(self, T): """Predict new data by linear interpolation. Parameters ---------- T : array-like of shape (n_samples,) Data to predict from. Returns ------- T_ : ndarray of shape (n_samples,) The predicted data. """ T = column_or_1d(T) return expit(-(self.a_ * T + self.b_)) class _TemperatureScaling(RegressorMixin, BaseEstimator): """Temperature scaling model. Attributes ---------- beta_ : float The optimized inverse temperature. """ def fit(self, X, y, sample_weight=None): """Fit the model using X, y as training data. Parameters ---------- X : ndarray of shape (n_samples,) or (n_samples, n_classes) Training data. This should be the output of `decision_function` or `predict_proba`. If the input appears to be probabilities (i.e., values between 0 and 1 that sum to 1 across classes), it will be converted to logits using `np.log(p + eps)`. Binary decision function outputs (1D) will be converted to two-class logits of the form (-x, x). For shapes of the form (n_samples, 1), the same process applies. y : array-like of shape (n_samples,) Training target. sample_weight : array-like of shape (n_samples,), default=None Sample weights. If None, then samples are equally weighted. Returns ------- self : object Returns an instance of self. """ xp, _, xp_device = get_namespace_and_device(X, y) X, y = indexable(X, y) check_consistent_length(X, y) logits = _convert_to_logits(X) # guarantees xp.float64 or xp.float32 dtype_ = logits.dtype labels = column_or_1d(y, dtype=dtype_) if sample_weight is not None: sample_weight = _check_sample_weight(sample_weight, labels, dtype=dtype_) if _is_numpy_namespace(xp): multinomial_loss = HalfMultinomialLoss(n_classes=logits.shape[1]) else: multinomial_loss = partial(_half_multinomial_loss, xp=xp) def log_loss(log_beta=0.0): """Compute the log loss as a parameter of the inverse temperature (beta). Parameters ---------- log_beta : float The current logarithm of the inverse temperature value during optimisation. Returns ------- negative_log_likelihood_loss : float The negative log likelihood loss. """ # TODO: numpy 2.0 # Ensure raw_prediction has the same dtype as labels using .astype(). # Without this, dtype promotion rules differ across NumPy versions: # # beta = np.float64(0) # logits = np.array([1, 2], dtype=np.float32) # # result = beta * logits # - NumPy < 2: result.dtype is float32 # - NumPy 2+: result.dtype is float64 # # This can cause dtype mismatch errors downstream (e.g., buffer dtype). log_beta = xp.asarray(log_beta, dtype=dtype_, device=xp_device) raw_prediction = xp.exp(log_beta) * logits return multinomial_loss(labels, raw_prediction, sample_weight) xatol = 64 * xp.finfo(dtype_).eps log_beta_minimizer = minimize_scalar( log_loss, bounds=(-10.0, 10.0), options={ "xatol": xatol, }, ) if not log_beta_minimizer.success: # pragma: no cover raise RuntimeError( "Temperature scaling fails to optimize during calibration. " "Reason from `scipy.optimize.minimize_scalar`: " f"{log_beta_minimizer.message}" ) self.beta_ = xp.exp( xp.asarray(log_beta_minimizer.x, dtype=dtype_, device=xp_device) ) return self def predict(self, X): """Predict new data by linear interpolation. Parameters ---------- X : ndarray of shape (n_samples,) or (n_samples, n_classes) Data to predict from. This should be the output of `decision_function` or `predict_proba`. If the input appears to be probabilities (i.e., values between 0 and 1 that sum to 1 across classes), it will be converted to logits using `np.log(p + eps)`. Binary decision function outputs (1D) will be converted to two-class logits of the form (-x, x). For shapes of the form (n_samples, 1), the same process applies. Returns ------- X_ : ndarray of shape (n_samples, n_classes) The predicted data. """ logits = _convert_to_logits(X) return softmax(self.beta_ * logits) def __sklearn_tags__(self): tags = super().__sklearn_tags__() tags.input_tags.one_d_array = True tags.input_tags.two_d_array = False return tags @validate_params( { "y_true": ["array-like"], "y_prob": ["array-like"], "pos_label": [Real, str, "boolean", None], "n_bins": [Interval(Integral, 1, None, closed="left")], "strategy": [StrOptions({"uniform", "quantile"})], }, prefer_skip_nested_validation=True, ) def calibration_curve( y_true, y_prob, *, pos_label=None, n_bins=5, strategy="uniform", ): """Compute true and predicted probabilities for a calibration curve. The method assumes the inputs come from a binary classifier, and discretize the [0, 1] interval into bins. Calibration curves may also be referred to as reliability diagrams. Read more in the :ref:`User Guide `. Parameters ---------- y_true : array-like of shape (n_samples,) True targets. y_prob : array-like of shape (n_samples,) Probabilities of the positive class. pos_label : int, float, bool or str, default=None The label of the positive class. .. versionadded:: 1.1 n_bins : int, default=5 Number of bins to discretize the [0, 1] interval. A bigger number requires more data. Bins with no samples (i.e. without corresponding values in `y_prob`) will not be returned, thus the returned arrays may have less than `n_bins` values. strategy : {'uniform', 'quantile'}, default='uniform' Strategy used to define the widths of the bins. uniform The bins have identical widths. quantile The bins have the same number of samples and depend on `y_prob`. Returns ------- prob_true : ndarray of shape (n_bins,) or smaller The proportion of samples whose class is the positive class, in each bin (fraction of positives). prob_pred : ndarray of shape (n_bins,) or smaller The mean predicted probability in each bin. See Also -------- CalibrationDisplay.from_predictions : Plot calibration curve using true and predicted labels. CalibrationDisplay.from_estimator : Plot calibration curve using an estimator and data. References ---------- Alexandru Niculescu-Mizil and Rich Caruana (2005) Predicting Good Probabilities With Supervised Learning, in Proceedings of the 22nd International Conference on Machine Learning (ICML). See section 4 (Qualitative Analysis of Predictions). Examples -------- >>> import numpy as np >>> from sklearn.calibration import calibration_curve >>> y_true = np.array([0, 0, 0, 0, 1, 1, 1, 1, 1]) >>> y_pred = np.array([0.1, 0.2, 0.3, 0.4, 0.65, 0.7, 0.8, 0.9, 1.]) >>> prob_true, prob_pred = calibration_curve(y_true, y_pred, n_bins=3) >>> prob_true array([0. , 0.5, 1. ]) >>> prob_pred array([0.2 , 0.525, 0.85 ]) """ y_true = column_or_1d(y_true) y_prob = column_or_1d(y_prob) check_consistent_length(y_true, y_prob) pos_label = _check_pos_label_consistency(pos_label, y_true) if y_prob.min() < 0 or y_prob.max() > 1: raise ValueError("y_prob has values outside [0, 1].") labels = np.unique(y_true) if len(labels) > 2: raise ValueError( f"Only binary classification is supported. Provided labels {labels}." ) y_true = y_true == pos_label if strategy == "quantile": # Determine bin edges by distribution of data quantiles = np.linspace(0, 1, n_bins + 1) bins = np.percentile(y_prob, quantiles * 100) elif strategy == "uniform": bins = np.linspace(0.0, 1.0, n_bins + 1) else: raise ValueError( "Invalid entry to 'strategy' input. Strategy " "must be either 'quantile' or 'uniform'." ) binids = np.searchsorted(bins[1:-1], y_prob) bin_sums = np.bincount(binids, weights=y_prob, minlength=len(bins)) bin_true = np.bincount(binids, weights=y_true, minlength=len(bins)) bin_total = np.bincount(binids, minlength=len(bins)) nonzero = bin_total != 0 prob_true = bin_true[nonzero] / bin_total[nonzero] prob_pred = bin_sums[nonzero] / bin_total[nonzero] return prob_true, prob_pred class CalibrationDisplay(_BinaryClassifierCurveDisplayMixin): """Calibration curve (also known as reliability diagram) visualization. It is recommended to use :func:`~sklearn.calibration.CalibrationDisplay.from_estimator` or :func:`~sklearn.calibration.CalibrationDisplay.from_predictions` to create a `CalibrationDisplay`. All parameters are stored as attributes. Read more about calibration in the :ref:`User Guide ` and more about the scikit-learn visualization API in :ref:`visualizations`. For an example on how to use the visualization, see :ref:`sphx_glr_auto_examples_calibration_plot_calibration_curve.py`. .. versionadded:: 1.0 Parameters ---------- prob_true : ndarray of shape (n_bins,) The proportion of samples whose class is the positive class (fraction of positives), in each bin. prob_pred : ndarray of shape (n_bins,) The mean predicted probability in each bin. y_prob : ndarray of shape (n_samples,) Probability estimates for the positive class, for each sample. estimator_name : str, default=None Name of estimator. If None, the estimator name is not shown. pos_label : int, float, bool or str, default=None The positive class when calibration curve computed. If not `None`, this value is displayed in the x- and y-axes labels. .. versionadded:: 1.1 Attributes ---------- line_ : matplotlib Artist Calibration curve. ax_ : matplotlib Axes Axes with calibration curve. figure_ : matplotlib Figure Figure containing the curve. See Also -------- calibration_curve : Compute true and predicted probabilities for a calibration curve. CalibrationDisplay.from_predictions : Plot calibration curve using true and predicted labels. CalibrationDisplay.from_estimator : Plot calibration curve using an estimator and data. Examples -------- >>> from sklearn.datasets import make_classification >>> from sklearn.model_selection import train_test_split >>> from sklearn.linear_model import LogisticRegression >>> from sklearn.calibration import calibration_curve, CalibrationDisplay >>> X, y = make_classification(random_state=0) >>> X_train, X_test, y_train, y_test = train_test_split( ... X, y, random_state=0) >>> clf = LogisticRegression(random_state=0) >>> clf.fit(X_train, y_train) LogisticRegression(random_state=0) >>> y_prob = clf.predict_proba(X_test)[:, 1] >>> prob_true, prob_pred = calibration_curve(y_test, y_prob, n_bins=10) >>> disp = CalibrationDisplay(prob_true, prob_pred, y_prob) >>> disp.plot() <...> """ def __init__( self, prob_true, prob_pred, y_prob, *, estimator_name=None, pos_label=None ): self.prob_true = prob_true self.prob_pred = prob_pred self.y_prob = y_prob self.estimator_name = estimator_name self.pos_label = pos_label def plot(self, *, ax=None, name=None, ref_line=True, **kwargs): """Plot visualization. Extra keyword arguments will be passed to :func:`matplotlib.pyplot.plot`. Parameters ---------- ax : Matplotlib Axes, default=None Axes object to plot on. If `None`, a new figure and axes is created. name : str, default=None Name for labeling curve. If `None`, use `estimator_name` if not `None`, otherwise no labeling is shown. ref_line : bool, default=True If `True`, plots a reference line representing a perfectly calibrated classifier. **kwargs : dict Keyword arguments to be passed to :func:`matplotlib.pyplot.plot`. Returns ------- display : :class:`~sklearn.calibration.CalibrationDisplay` Object that stores computed values. """ self.ax_, self.figure_, name = self._validate_plot_params(ax=ax, name=name) info_pos_label = ( f"(Positive class: {self.pos_label})" if self.pos_label is not None else "" ) default_line_kwargs = {"marker": "s", "linestyle": "-"} if name is not None: default_line_kwargs["label"] = name line_kwargs = _validate_style_kwargs(default_line_kwargs, kwargs) ref_line_label = "Perfectly calibrated" existing_ref_line = ref_line_label in self.ax_.get_legend_handles_labels()[1] if ref_line and not existing_ref_line: self.ax_.plot([0, 1], [0, 1], "k:", label=ref_line_label) self.line_ = self.ax_.plot(self.prob_pred, self.prob_true, **line_kwargs)[0] # We always have to show the legend for at least the reference line self.ax_.legend(loc="lower right") xlabel = f"Mean predicted probability {info_pos_label}" ylabel = f"Fraction of positives {info_pos_label}" self.ax_.set(xlabel=xlabel, ylabel=ylabel) return self @classmethod def from_estimator( cls, estimator, X, y, *, n_bins=5, strategy="uniform", pos_label=None, name=None, ax=None, ref_line=True, **kwargs, ): """Plot calibration curve using a binary classifier and data. A calibration curve, also known as a reliability diagram, uses inputs from a binary classifier and plots the average predicted probability for each bin against the fraction of positive classes, on the y-axis. Extra keyword arguments will be passed to :func:`matplotlib.pyplot.plot`. Read more about calibration in the :ref:`User Guide ` and more about the scikit-learn visualization API in :ref:`visualizations`. .. versionadded:: 1.0 Parameters ---------- estimator : estimator instance Fitted classifier or a fitted :class:`~sklearn.pipeline.Pipeline` in which the last estimator is a classifier. The classifier must have a :term:`predict_proba` method. X : {array-like, sparse matrix} of shape (n_samples, n_features) Input values. y : array-like of shape (n_samples,) Binary target values. n_bins : int, default=5 Number of bins to discretize the [0, 1] interval into when calculating the calibration curve. A bigger number requires more data. strategy : {'uniform', 'quantile'}, default='uniform' Strategy used to define the widths of the bins. - `'uniform'`: The bins have identical widths. - `'quantile'`: The bins have the same number of samples and depend on predicted probabilities. pos_label : int, float, bool or str, default=None The positive class when computing the calibration curve. By default, `estimators.classes_[1]` is considered as the positive class. .. versionadded:: 1.1 name : str, default=None Name for labeling curve. If `None`, the name of the estimator is used. ax : matplotlib axes, default=None Axes object to plot on. If `None`, a new figure and axes is created. ref_line : bool, default=True If `True`, plots a reference line representing a perfectly calibrated classifier. **kwargs : dict Keyword arguments to be passed to :func:`matplotlib.pyplot.plot`. Returns ------- display : :class:`~sklearn.calibration.CalibrationDisplay`. Object that stores computed values. See Also -------- CalibrationDisplay.from_predictions : Plot calibration curve using true and predicted labels. Examples -------- >>> import matplotlib.pyplot as plt >>> from sklearn.datasets import make_classification >>> from sklearn.model_selection import train_test_split >>> from sklearn.linear_model import LogisticRegression >>> from sklearn.calibration import CalibrationDisplay >>> X, y = make_classification(random_state=0) >>> X_train, X_test, y_train, y_test = train_test_split( ... X, y, random_state=0) >>> clf = LogisticRegression(random_state=0) >>> clf.fit(X_train, y_train) LogisticRegression(random_state=0) >>> disp = CalibrationDisplay.from_estimator(clf, X_test, y_test) >>> plt.show() """ y_prob, pos_label, name = cls._validate_and_get_response_values( estimator, X, y, response_method="predict_proba", pos_label=pos_label, name=name, ) return cls.from_predictions( y, y_prob, n_bins=n_bins, strategy=strategy, pos_label=pos_label, name=name, ref_line=ref_line, ax=ax, **kwargs, ) @classmethod def from_predictions( cls, y_true, y_prob, *, n_bins=5, strategy="uniform", pos_label=None, name=None, ax=None, ref_line=True, **kwargs, ): """Plot calibration curve using true labels and predicted probabilities. Calibration curve, also known as reliability diagram, uses inputs from a binary classifier and plots the average predicted probability for each bin against the fraction of positive classes, on the y-axis. Extra keyword arguments will be passed to :func:`matplotlib.pyplot.plot`. Read more about calibration in the :ref:`User Guide ` and more about the scikit-learn visualization API in :ref:`visualizations`. .. versionadded:: 1.0 Parameters ---------- y_true : array-like of shape (n_samples,) True labels. y_prob : array-like of shape (n_samples,) The predicted probabilities of the positive class. n_bins : int, default=5 Number of bins to discretize the [0, 1] interval into when calculating the calibration curve. A bigger number requires more data. strategy : {'uniform', 'quantile'}, default='uniform' Strategy used to define the widths of the bins. - `'uniform'`: The bins have identical widths. - `'quantile'`: The bins have the same number of samples and depend on predicted probabilities. pos_label : int, float, bool or str, default=None The positive class when computing the calibration curve. When `pos_label=None`, if `y_true` is in {-1, 1} or {0, 1}, `pos_label` is set to 1, otherwise an error will be raised. .. versionadded:: 1.1 name : str, default=None Name for labeling curve. ax : matplotlib axes, default=None Axes object to plot on. If `None`, a new figure and axes is created. ref_line : bool, default=True If `True`, plots a reference line representing a perfectly calibrated classifier. **kwargs : dict Keyword arguments to be passed to :func:`matplotlib.pyplot.plot`. Returns ------- display : :class:`~sklearn.calibration.CalibrationDisplay`. Object that stores computed values. See Also -------- CalibrationDisplay.from_estimator : Plot calibration curve using an estimator and data. Examples -------- >>> import matplotlib.pyplot as plt >>> from sklearn.datasets import make_classification >>> from sklearn.model_selection import train_test_split >>> from sklearn.linear_model import LogisticRegression >>> from sklearn.calibration import CalibrationDisplay >>> X, y = make_classification(random_state=0) >>> X_train, X_test, y_train, y_test = train_test_split( ... X, y, random_state=0) >>> clf = LogisticRegression(random_state=0) >>> clf.fit(X_train, y_train) LogisticRegression(random_state=0) >>> y_prob = clf.predict_proba(X_test)[:, 1] >>> disp = CalibrationDisplay.from_predictions(y_test, y_prob) >>> plt.show() """ pos_label_validated, name = cls._validate_from_predictions_params( y_true, y_prob, sample_weight=None, pos_label=pos_label, name=name ) prob_true, prob_pred = calibration_curve( y_true, y_prob, n_bins=n_bins, strategy=strategy, pos_label=pos_label ) disp = cls( prob_true=prob_true, prob_pred=prob_pred, y_prob=y_prob, estimator_name=name, pos_label=pos_label_validated, ) return disp.plot(ax=ax, ref_line=ref_line, **kwargs)