.. DO NOT EDIT.
.. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY.
.. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE:
.. "auto_examples/calibration/plot_calibration_curve.py"
.. LINE NUMBERS ARE GIVEN BELOW.

.. only:: html

    .. note::
        :class: sphx-glr-download-link-note

        :ref:`Go to the end <sphx_glr_download_auto_examples_calibration_plot_calibration_curve.py>`
        to download the full example code. or to run this example in your browser via JupyterLite or Binder

.. rst-class:: sphx-glr-example-title

.. _sphx_glr_auto_examples_calibration_plot_calibration_curve.py:


==============================
Probability Calibration curves
==============================

When performing classification one often wants to predict not only the class
label, but also the associated probability. This probability gives some
kind of confidence on the prediction. This example demonstrates how to
visualize how well calibrated the predicted probabilities are using calibration
curves, also known as reliability diagrams. Calibration of an uncalibrated
classifier will also be demonstrated.

.. GENERATED FROM PYTHON SOURCE LINES 14-18

.. code-block:: Python


    # Authors: The scikit-learn developers
    # SPDX-License-Identifier: BSD-3-Clause








.. GENERATED FROM PYTHON SOURCE LINES 19-27

Dataset
-------

We will use a synthetic binary classification dataset with 100,000 samples
and 20 features. Of the 20 features, only 2 are informative, 10 are
redundant (random combinations of the informative features) and the
remaining 8 are uninformative (random numbers). Of the 100,000 samples, 1,000
will be used for model fitting and the rest for testing.

.. GENERATED FROM PYTHON SOURCE LINES 27-39

.. code-block:: Python


    from sklearn.datasets import make_classification
    from sklearn.model_selection import train_test_split

    X, y = make_classification(
        n_samples=100_000, n_features=20, n_informative=2, n_redundant=10, random_state=42
    )

    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.99, random_state=42
    )








.. GENERATED FROM PYTHON SOURCE LINES 40-58

Calibration curves
------------------

Gaussian Naive Bayes
^^^^^^^^^^^^^^^^^^^^

First, we will compare:

* :class:`~sklearn.linear_model.LogisticRegression` (used as baseline
  since very often, properly regularized logistic regression is well
  calibrated by default thanks to the use of the log-loss)
* Uncalibrated :class:`~sklearn.naive_bayes.GaussianNB`
* :class:`~sklearn.naive_bayes.GaussianNB` with isotonic and sigmoid
  calibration (see :ref:`User Guide <calibration>`)

Calibration curves for all 4 conditions are plotted below, with the average
predicted probability for each bin on the x-axis and the fraction of positive
classes in each bin on the y-axis.

.. GENERATED FROM PYTHON SOURCE LINES 58-78

.. code-block:: Python


    import matplotlib.pyplot as plt
    from matplotlib.gridspec import GridSpec

    from sklearn.calibration import CalibratedClassifierCV, CalibrationDisplay
    from sklearn.linear_model import LogisticRegression
    from sklearn.naive_bayes import GaussianNB

    lr = LogisticRegression(C=1.0)
    gnb = GaussianNB()
    gnb_isotonic = CalibratedClassifierCV(gnb, cv=2, method="isotonic")
    gnb_sigmoid = CalibratedClassifierCV(gnb, cv=2, method="sigmoid")

    clf_list = [
        (lr, "Logistic"),
        (gnb, "Naive Bayes"),
        (gnb_isotonic, "Naive Bayes + Isotonic"),
        (gnb_sigmoid, "Naive Bayes + Sigmoid"),
    ]








.. GENERATED FROM PYTHON SOURCE LINES 79-119

.. code-block:: Python

    fig = plt.figure(figsize=(10, 10))
    gs = GridSpec(4, 2)
    colors = plt.get_cmap("Dark2")

    ax_calibration_curve = fig.add_subplot(gs[:2, :2])
    calibration_displays = {}
    for i, (clf, name) in enumerate(clf_list):
        clf.fit(X_train, y_train)
        display = CalibrationDisplay.from_estimator(
            clf,
            X_test,
            y_test,
            n_bins=10,
            name=name,
            ax=ax_calibration_curve,
            color=colors(i),
        )
        calibration_displays[name] = display

    ax_calibration_curve.grid()
    ax_calibration_curve.set_title("Calibration plots (Naive Bayes)")

    # Add histogram
    grid_positions = [(2, 0), (2, 1), (3, 0), (3, 1)]
    for i, (_, name) in enumerate(clf_list):
        row, col = grid_positions[i]
        ax = fig.add_subplot(gs[row, col])

        ax.hist(
            calibration_displays[name].y_prob,
            range=(0, 1),
            bins=10,
            label=name,
            color=colors(i),
        )
        ax.set(title=name, xlabel="Mean predicted probability", ylabel="Count")

    plt.tight_layout()
    plt.show()




.. image-sg:: /auto_examples/calibration/images/sphx_glr_plot_calibration_curve_001.png
   :alt: Calibration plots (Naive Bayes), Logistic, Naive Bayes, Naive Bayes + Isotonic, Naive Bayes + Sigmoid
   :srcset: /auto_examples/calibration/images/sphx_glr_plot_calibration_curve_001.png
   :class: sphx-glr-single-img





.. GENERATED FROM PYTHON SOURCE LINES 120-137

Uncalibrated :class:`~sklearn.naive_bayes.GaussianNB` is poorly calibrated
because of
the redundant features which violate the assumption of feature-independence
and result in an overly confident classifier, which is indicated by the
typical transposed-sigmoid curve. Calibration of the probabilities of
:class:`~sklearn.naive_bayes.GaussianNB` with :ref:`isotonic` can fix
this issue as can be seen from the nearly diagonal calibration curve.
:ref:`Sigmoid regression <sigmoid_regressor>` also improves calibration
slightly,
albeit not as strongly as the non-parametric isotonic regression. This can be
attributed to the fact that we have plenty of calibration data such that the
greater flexibility of the non-parametric model can be exploited.

Below we will make a quantitative analysis considering several classification
metrics: :ref:`brier_score_loss`, :ref:`log_loss`,
:ref:`precision, recall, F1 score <precision_recall_f_measure_metrics>` and
:ref:`ROC AUC <roc_metrics>`.

.. GENERATED FROM PYTHON SOURCE LINES 137-171

.. code-block:: Python


    from collections import defaultdict

    import pandas as pd

    from sklearn.metrics import (
        brier_score_loss,
        f1_score,
        log_loss,
        precision_score,
        recall_score,
        roc_auc_score,
    )

    scores = defaultdict(list)
    for i, (clf, name) in enumerate(clf_list):
        clf.fit(X_train, y_train)
        y_prob = clf.predict_proba(X_test)
        y_pred = clf.predict(X_test)
        scores["Classifier"].append(name)

        for metric in [brier_score_loss, log_loss, roc_auc_score]:
            score_name = metric.__name__.replace("_", " ").replace("score", "").capitalize()
            scores[score_name].append(metric(y_test, y_prob[:, 1]))

        for metric in [precision_score, recall_score, f1_score]:
            score_name = metric.__name__.replace("_", " ").replace("score", "").capitalize()
            scores[score_name].append(metric(y_test, y_pred))

        score_df = pd.DataFrame(scores).set_index("Classifier")
        score_df.round(decimals=3)

    score_df






.. raw:: html

    <div class="output_subarea output_html rendered_html output_result">
    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }

        .dataframe tbody tr th {
            vertical-align: top;
        }

        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th>Brier  loss</th>
          <th>Log loss</th>
          <th>Roc auc</th>
          <th>Precision</th>
          <th>Recall</th>
          <th>F1</th>
        </tr>
        <tr>
          <th>Classifier</th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th>Logistic</th>
          <td>0.098932</td>
          <td>0.323200</td>
          <td>0.937443</td>
          <td>0.871965</td>
          <td>0.851348</td>
          <td>0.861533</td>
        </tr>
        <tr>
          <th>Naive Bayes</th>
          <td>0.117608</td>
          <td>0.782755</td>
          <td>0.940374</td>
          <td>0.857400</td>
          <td>0.875941</td>
          <td>0.866571</td>
        </tr>
        <tr>
          <th>Naive Bayes + Isotonic</th>
          <td>0.098332</td>
          <td>0.370738</td>
          <td>0.938613</td>
          <td>0.883065</td>
          <td>0.836224</td>
          <td>0.859007</td>
        </tr>
        <tr>
          <th>Naive Bayes + Sigmoid</th>
          <td>0.108880</td>
          <td>0.368896</td>
          <td>0.940201</td>
          <td>0.861106</td>
          <td>0.871277</td>
          <td>0.866161</td>
        </tr>
      </tbody>
    </table>
    </div>
    </div>
    <br />
    <br />

.. GENERATED FROM PYTHON SOURCE LINES 172-196

Notice that although calibration improves the :ref:`brier_score_loss` (a
metric composed
of calibration term and refinement term) and :ref:`log_loss`, it does not
significantly alter the prediction accuracy measures (precision, recall and
F1 score).
This is because calibration should not significantly change prediction
probabilities at the location of the decision threshold (at x = 0.5 on the
graph). Calibration should however, make the predicted probabilities more
accurate and thus more useful for making allocation decisions under
uncertainty.
Further, ROC AUC, should not change at all because calibration is a
monotonic transformation. Indeed, no rank metrics are affected by
calibration.

Linear support vector classifier
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Next, we will compare:

* :class:`~sklearn.linear_model.LogisticRegression` (baseline)
* Uncalibrated :class:`~sklearn.svm.LinearSVC`. Since SVC does not output
  probabilities by default, we naively scale the output of the
  :term:`decision_function` into [0, 1] by applying min-max scaling.
* :class:`~sklearn.svm.LinearSVC` with isotonic and sigmoid
  calibration (see :ref:`User Guide <calibration>`)

.. GENERATED FROM PYTHON SOURCE LINES 196-222

.. code-block:: Python


    import numpy as np

    from sklearn.svm import LinearSVC


    class NaivelyCalibratedLinearSVC(LinearSVC):
        """LinearSVC with `predict_proba` method that naively scales
        `decision_function` output for binary classification."""

        def fit(self, X, y):
            super().fit(X, y)
            df = self.decision_function(X)
            self.df_min_ = df.min()
            self.df_max_ = df.max()

        def predict_proba(self, X):
            """Min-max scale output of `decision_function` to [0, 1]."""
            df = self.decision_function(X)
            calibrated_df = (df - self.df_min_) / (self.df_max_ - self.df_min_)
            proba_pos_class = np.clip(calibrated_df, 0, 1)
            proba_neg_class = 1 - proba_pos_class
            proba = np.c_[proba_neg_class, proba_pos_class]
            return proba









.. GENERATED FROM PYTHON SOURCE LINES 223-236

.. code-block:: Python


    lr = LogisticRegression(C=1.0)
    svc = NaivelyCalibratedLinearSVC(max_iter=10_000)
    svc_isotonic = CalibratedClassifierCV(svc, cv=2, method="isotonic")
    svc_sigmoid = CalibratedClassifierCV(svc, cv=2, method="sigmoid")

    clf_list = [
        (lr, "Logistic"),
        (svc, "SVC"),
        (svc_isotonic, "SVC + Isotonic"),
        (svc_sigmoid, "SVC + Sigmoid"),
    ]








.. GENERATED FROM PYTHON SOURCE LINES 237-276

.. code-block:: Python

    fig = plt.figure(figsize=(10, 10))
    gs = GridSpec(4, 2)

    ax_calibration_curve = fig.add_subplot(gs[:2, :2])
    calibration_displays = {}
    for i, (clf, name) in enumerate(clf_list):
        clf.fit(X_train, y_train)
        display = CalibrationDisplay.from_estimator(
            clf,
            X_test,
            y_test,
            n_bins=10,
            name=name,
            ax=ax_calibration_curve,
            color=colors(i),
        )
        calibration_displays[name] = display

    ax_calibration_curve.grid()
    ax_calibration_curve.set_title("Calibration plots (SVC)")

    # Add histogram
    grid_positions = [(2, 0), (2, 1), (3, 0), (3, 1)]
    for i, (_, name) in enumerate(clf_list):
        row, col = grid_positions[i]
        ax = fig.add_subplot(gs[row, col])

        ax.hist(
            calibration_displays[name].y_prob,
            range=(0, 1),
            bins=10,
            label=name,
            color=colors(i),
        )
        ax.set(title=name, xlabel="Mean predicted probability", ylabel="Count")

    plt.tight_layout()
    plt.show()




.. image-sg:: /auto_examples/calibration/images/sphx_glr_plot_calibration_curve_002.png
   :alt: Calibration plots (SVC), Logistic, SVC, SVC + Isotonic, SVC + Sigmoid
   :srcset: /auto_examples/calibration/images/sphx_glr_plot_calibration_curve_002.png
   :class: sphx-glr-single-img





.. GENERATED FROM PYTHON SOURCE LINES 277-295

:class:`~sklearn.svm.LinearSVC` shows the opposite
behavior to :class:`~sklearn.naive_bayes.GaussianNB`; the calibration
curve has a sigmoid shape, which is typical for an under-confident
classifier. In the case of :class:`~sklearn.svm.LinearSVC`, this is caused
by the margin property of the hinge loss, which focuses on samples that are
close to the decision boundary (support vectors). Samples that are far
away from the decision boundary do not impact the hinge loss. It thus makes
sense that :class:`~sklearn.svm.LinearSVC` does not try to separate samples
in the high confidence region regions. This leads to flatter calibration
curves near 0 and 1 and is empirically shown with a variety of datasets
in Niculescu-Mizil & Caruana [1]_.

Both kinds of calibration (sigmoid and isotonic) can fix this issue and
yield similar results.

As before, we show the :ref:`brier_score_loss`, :ref:`log_loss`,
:ref:`precision, recall, F1 score <precision_recall_f_measure_metrics>` and
:ref:`ROC AUC <roc_metrics>`.

.. GENERATED FROM PYTHON SOURCE LINES 295-316

.. code-block:: Python


    scores = defaultdict(list)
    for i, (clf, name) in enumerate(clf_list):
        clf.fit(X_train, y_train)
        y_prob = clf.predict_proba(X_test)
        y_pred = clf.predict(X_test)
        scores["Classifier"].append(name)

        for metric in [brier_score_loss, log_loss, roc_auc_score]:
            score_name = metric.__name__.replace("_", " ").replace("score", "").capitalize()
            scores[score_name].append(metric(y_test, y_prob[:, 1]))

        for metric in [precision_score, recall_score, f1_score]:
            score_name = metric.__name__.replace("_", " ").replace("score", "").capitalize()
            scores[score_name].append(metric(y_test, y_pred))

        score_df = pd.DataFrame(scores).set_index("Classifier")
        score_df.round(decimals=3)

    score_df






.. raw:: html

    <div class="output_subarea output_html rendered_html output_result">
    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }

        .dataframe tbody tr th {
            vertical-align: top;
        }

        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th>Brier  loss</th>
          <th>Log loss</th>
          <th>Roc auc</th>
          <th>Precision</th>
          <th>Recall</th>
          <th>F1</th>
        </tr>
        <tr>
          <th>Classifier</th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th>Logistic</th>
          <td>0.098932</td>
          <td>0.323200</td>
          <td>0.937443</td>
          <td>0.871965</td>
          <td>0.851348</td>
          <td>0.861533</td>
        </tr>
        <tr>
          <th>SVC</th>
          <td>0.144943</td>
          <td>0.465660</td>
          <td>0.937597</td>
          <td>0.872186</td>
          <td>0.851792</td>
          <td>0.861868</td>
        </tr>
        <tr>
          <th>SVC + Isotonic</th>
          <td>0.099820</td>
          <td>0.376999</td>
          <td>0.936480</td>
          <td>0.853174</td>
          <td>0.877981</td>
          <td>0.865400</td>
        </tr>
        <tr>
          <th>SVC + Sigmoid</th>
          <td>0.098758</td>
          <td>0.321301</td>
          <td>0.937532</td>
          <td>0.873724</td>
          <td>0.848743</td>
          <td>0.861053</td>
        </tr>
      </tbody>
    </table>
    </div>
    </div>
    <br />
    <br />

.. GENERATED FROM PYTHON SOURCE LINES 317-337

As with :class:`~sklearn.naive_bayes.GaussianNB` above, calibration improves
both :ref:`brier_score_loss` and :ref:`log_loss` but does not alter the
prediction accuracy measures (precision, recall and F1 score) much.

Summary
-------

Parametric sigmoid calibration can deal with situations where the calibration
curve of the base classifier is sigmoid (e.g., for
:class:`~sklearn.svm.LinearSVC`) but not where it is transposed-sigmoid
(e.g., :class:`~sklearn.naive_bayes.GaussianNB`). Non-parametric
isotonic calibration can deal with both situations but may require more
data to produce good results.

References
----------

.. [1] `Predicting Good Probabilities with Supervised Learning
       <https://dl.acm.org/doi/pdf/10.1145/1102351.1102430>`_,
       A. Niculescu-Mizil & R. Caruana, ICML 2005


.. rst-class:: sphx-glr-timing

   **Total running time of the script:** (0 minutes 2.168 seconds)


.. _sphx_glr_download_auto_examples_calibration_plot_calibration_curve.py:

.. only:: html

  .. container:: sphx-glr-footer sphx-glr-footer-example

    .. container:: binder-badge

      .. image:: images/binder_badge_logo.svg
        :target: https://mybinder.org/v2/gh/scikit-learn/scikit-learn/1.6.X?urlpath=lab/tree/notebooks/auto_examples/calibration/plot_calibration_curve.ipynb
        :alt: Launch binder
        :width: 150 px

    .. container:: lite-badge

      .. image:: images/jupyterlite_badge_logo.svg
        :target: ../../lite/lab/index.html?path=auto_examples/calibration/plot_calibration_curve.ipynb
        :alt: Launch JupyterLite
        :width: 150 px

    .. container:: sphx-glr-download sphx-glr-download-jupyter

      :download:`Download Jupyter notebook: plot_calibration_curve.ipynb <plot_calibration_curve.ipynb>`

    .. container:: sphx-glr-download sphx-glr-download-python

      :download:`Download Python source code: plot_calibration_curve.py <plot_calibration_curve.py>`

    .. container:: sphx-glr-download sphx-glr-download-zip

      :download:`Download zipped: plot_calibration_curve.zip <plot_calibration_curve.zip>`


.. include:: plot_calibration_curve.recommendations


.. only:: html

 .. rst-class:: sphx-glr-signature

    `Gallery generated by Sphinx-Gallery <https://sphinx-gallery.github.io>`_