# Probability calibration

As a follow-up of my previous post on reliability diagrams, I have worked jointly with Alexandre Gramfort, Mathieu Blondel and Balazs Kegl (with reviews by the whole team, in particular Olivier Grisel) on adding probability calibration and reliability diagrams to scikit-learn. Those have been added in the recent 0.16 release of scikit-learn as CalibratedClassifierCV and calibration_curve.

This post contains an interactive version of the documentation in the form of an IPython notebook; parts of the text/code are thus due to my coauthors.

Note that the 0.16 release of scikit-learn contains a bug in IsotonicRegression, which has been fixed in the 0.16.1 release. For obtaining correct results with this notebook, you need to use 0.16.1 or any later version.

## Reliability curves¶

In [1]:
# <!-- collapse=True -->
import numpy as np
np.random.seed(0)

import matplotlib
matplotlib.use("svg")
import matplotlib.pyplot as plt
from matplotlib import cm
%matplotlib inline

from sklearn import datasets
from sklearn.naive_bayes import GaussianNB
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.svm import LinearSVC
from sklearn.calibration import calibration_curve, CalibratedClassifierCV
from sklearn.metrics import (brier_score_loss, precision_score, recall_score,
f1_score, log_loss)
from sklearn.cross_validation import train_test_split


When performing classification you often want not only to predict the class label, but also obtain a probability of the respective label. This probability gives you some kind of confidence on the prediction. Some models can give you poor estimates of the class probabilities and some even do not not support probability prediction. The calibration module allows you to better calibrate the probabilities of a given model, or to add support for probability prediction.

Well calibrated classifiers are probabilistic classifiers for which the output of the predict_proba method can be directly interpreted as a confidence level. For instance, a well calibrated (binary) classifier should classify the samples such that among the samples to which it gave a predict_proba value close to 0.8, approximately 80% actually belong to the positive class. The following plot compares how well the probabilistic predictions of different classifiers are calibrated:

In [2]:
# <!-- collapse=True -->
X, y = datasets.make_classification(n_samples=100000, n_features=20,
n_informative=2, n_redundant=2)

train_samples = 100  # Samples used for training the models

X_train = X[:train_samples]
X_test = X[train_samples:]
y_train = y[:train_samples]
y_test = y[train_samples:]

# Create classifiers
lr = LogisticRegression()
gnb = GaussianNB()
svc = LinearSVC(C=1.0)
rfc = RandomForestClassifier(n_estimators=100)

In [3]:
# <!-- collapse=True -->
plt.figure(figsize=(9, 9))
ax1 = plt.subplot2grid((3, 1), (0, 0), rowspan=2)
ax2 = plt.subplot2grid((3, 1), (2, 0))

ax1.plot([0, 1], [0, 1], "k:", label="Perfectly calibrated")
for clf, name in [(lr, 'Logistic'),
(gnb, 'Naive Bayes'),
(svc, 'Support Vector Classification'),
(rfc, 'Random Forest')]:
clf.fit(X_train, y_train)
if hasattr(clf, "predict_proba"):
prob_pos = clf.predict_proba(X_test)[:, 1]
else:  # use decision function
prob_pos = clf.decision_function(X_test)
prob_pos = \
(prob_pos - prob_pos.min()) / (prob_pos.max() - prob_pos.min())
fraction_of_positives, mean_predicted_value = \
calibration_curve(y_test, prob_pos, n_bins=10)

ax1.plot(mean_predicted_value, fraction_of_positives, "s-",
label="%s" % (name, ))

ax2.hist(prob_pos, range=(0, 1), bins=10, label=name,
histtype="step", lw=2)

ax1.set_ylabel("Fraction of positives")
ax1.set_ylim([-0.05, 1.05])
ax1.legend(loc="lower right")
ax1.set_title('Calibration plots  (reliability curve)')

ax2.set_xlabel("Mean predicted value")
ax2.set_ylabel("Count")
ax2.legend(loc="upper center", ncol=2)

plt.tight_layout()


LogisticRegression returns well calibrated predictions by default as it directly optimizes log-loss. In contrast, the other methods return biased probabilities; with different biases per method:

• Naive Bayes (GaussianNB) tends to push probabilties to 0 or 1 (note the counts in the histograms). This is mainly because it makes the assumption that features are conditionally independent given the class, which is not the case in this dataset which contains 2 redundant features.

• RandomForestClassifier shows the opposite behavior: the histograms show peaks at approximately 0.2 and 0.9 probability, while probabilities close to 0 or 1 are very rare. An explanation for this is given by Niculescu-Mizil and Caruana [4]: "Methods such as bagging and random forests that average predictions from a base set of models can have difficulty making predictions near 0 and 1 because variance in the underlying base models will bias predictions that should be near zero or one away from these values. Because predictions are restricted to the interval [0,1], errors caused by variance tend to be one-sided near zero and one. For example, if a model should predict p = 0 for a case, the only way bagging can achieve this is if all bagged trees predict zero. If we add noise to the trees that bagging is averaging over, this noise will cause some trees to predict values larger than 0 for this case, thus moving the average prediction of the bagged ensemble away from 0. We observe this effect most strongly with random forests because the base-level trees trained with random forests have relatively high variance due to feature subseting." As a result, the calibration curve shows a characteristic sigmoid shape, indicating that the classifier could trust its "intuition" more and return probabilties closer to 0 or 1 typically.

• Linear Support Vector Classification (LinearSVC) shows an even more sigmoid curve as the RandomForestClassifier, which is typical for maximum-margin methods (compare Niculescu-Mizil and Caruana [4]), which focus on hard samples that are close to the decision boundary (the support vectors).

## Calibration of binary classifiers¶

Two approaches for performing calibration of probabilistic predictions are provided: a parametric approach based on Platt's sigmoid model and a non-parametric approach based on isotonic regression (sklearn.isotonic). Probability calibration should be done on new data not used for model fitting. The class CalibratedClassifierCV uses a cross-validation generator and estimates for each split the model parameter on the train samples and the calibration of the test samples. The probabilities predicted for the folds are then averaged. Already fitted classifiers can be calibrated by CalibratedClassifierCV via the paramter cv="prefit". In this case, the user has to take care manually that data for model fitting and calibration are disjoint.

The following images demonstrate the benefit of probability calibration. The first image present a dataset with 2 classes and 3 blobs of data. The blob in the middle contains random samples of each class. The probability for the samples in this blob should be 0.5.

In [4]:
# <!-- collapse=True -->

n_samples = 50000
n_bins = 3  # use 3 bins for calibration_curve as we have 3 clusters here

# Generate 3 blobs with 2 classes where the second blob contains
# half positive samples and half negative samples. Probability in this
# blob is therefore 0.5.
centers = [(-5, -5), (0, 0), (5, 5)]
X, y = datasets.make_blobs(n_samples=n_samples, n_features=2, cluster_std=1.0,
centers=centers, shuffle=False, random_state=42)

y[:n_samples // 2] = 0
y[n_samples // 2:] = 1
sample_weight = np.random.RandomState(42).rand(y.shape[0])

# split train, test for calibration
X_train, X_test, y_train, y_test, sw_train, sw_test = \
train_test_split(X, y, sample_weight, test_size=0.9, random_state=42)

plt.figure()
y_unique = np.unique(y)
colors = cm.rainbow(np.linspace(0.0, 1.0, y_unique.size))
for this_y, color in zip(y_unique, colors):
this_X = X_train[y_train == this_y]
this_sw = sw_train[y_train == this_y]
plt.scatter(this_X[:, 0], this_X[:, 1], s=this_sw * 50, c=color, alpha=0.5,
label="Class %s" % this_y)
plt.legend(loc="best")
plt.title("Data")

Out[4]:
<matplotlib.text.Text at 0x5b37b10>

The following image shows on the data above the estimated probability using a Gaussian naive Bayes classifier without calibration, with a sigmoid calibration and with a non-parametric isotonic calibration. One can observe that the non-parametric model provides the most accurate probability estimates for samples in the middle, i.e., 0.5.

In [5]:
# <!-- collapse=True -->

# Gaussian Naive-Bayes with no calibration
clf = GaussianNB()
clf.fit(X_train, y_train)  # GaussianNB itself does not support sample-weights
prob_pos_clf = clf.predict_proba(X_test)[:, 1]

# Gaussian Naive-Bayes with isotonic calibration
clf_isotonic = CalibratedClassifierCV(clf, cv=2, method='isotonic')
clf_isotonic.fit(X_train, y_train, sw_train)
prob_pos_isotonic = clf_isotonic.predict_proba(X_test)[:, 1]

# Gaussian Naive-Bayes with sigmoid calibration
clf_sigmoid = CalibratedClassifierCV(clf, cv=2, method='sigmoid')
clf_sigmoid.fit(X_train, y_train, sw_train)
prob_pos_sigmoid = clf_sigmoid.predict_proba(X_test)[:, 1]

print("Brier scores: (the smaller the better)")

clf_score = brier_score_loss(y_test, prob_pos_clf, sw_test)
print("No calibration: %1.3f" % clf_score)

clf_isotonic_score = brier_score_loss(y_test, prob_pos_isotonic, sw_test)
print("With isotonic calibration: %1.3f" % clf_isotonic_score)

clf_sigmoid_score = brier_score_loss(y_test, prob_pos_sigmoid, sw_test)
print("With sigmoid calibration: %1.3f" % clf_sigmoid_score)

Brier scores: (the smaller the better)
No calibration: 0.104
With isotonic calibration: 0.084
With sigmoid calibration: 0.109

In [6]:
# <!-- collapse=True -->

plt.figure()
order = np.lexsort((prob_pos_clf, ))
plt.plot(prob_pos_clf[order], 'r', label='No calibration (%1.3f)' % clf_score)
plt.plot(prob_pos_isotonic[order], 'g', linewidth=3,
label='Isotonic calibration (%1.3f)' % clf_isotonic_score)
plt.plot(prob_pos_sigmoid[order], 'b', linewidth=3,
label='Sigmoid calibration (%1.3f)' % clf_sigmoid_score)
plt.plot(np.linspace(0, y_test.size, 51)[1::2],
y_test[order].reshape(25, -1).mean(1),
'k', linewidth=3, label=r'Empirical')
plt.ylim([-0.05, 1.05])
plt.xlabel("Instances sorted according to predicted probability "
"(uncalibrated GNB)")
plt.ylabel("P(y=1)")
plt.legend(loc="upper left")
plt.title("Gaussian naive Bayes probabilities")

Out[6]:
<matplotlib.text.Text at 0x623ce90>

The following experiment is performed on an artificial dataset for binary classification with 100.000 samples (1.000 of them are used for model fitting) with 20 features. Of the 20 features, only 2 are informative and 10 are redundant. The figure shows the estimated probabilities obtained with logistic regression, a linear support-vector classifier (SVC), and linear SVC with both isotonic calibration and sigmoid calibration. The calibration performance is evaluated with Brier score brier_score_loss, reported in the legend (the smaller the better).

In [7]:
# <!-- collapse=True -->

# Create dataset of classification task with many redundant and few
# informative features
X, y = datasets.make_classification(n_samples=100000, 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)

In [8]:
# <!-- collapse=True -->

def plot_calibration_curve(est, name, fig_index):
"""Plot calibration curve for est w/o and with calibration. """
# Calibrated with isotonic calibration
isotonic = CalibratedClassifierCV(est, cv=2, method='isotonic')

# Calibrated with sigmoid calibration
sigmoid = CalibratedClassifierCV(est, cv=2, method='sigmoid')

# Logistic regression with no calibration as baseline
lr = LogisticRegression(C=1., solver='lbfgs')

fig = plt.figure(fig_index, figsize=(9, 9))
ax1 = plt.subplot2grid((3, 1), (0, 0), rowspan=2)
ax2 = plt.subplot2grid((3, 1), (2, 0))

ax1.plot([0, 1], [0, 1], "k:", label="Perfectly calibrated")
for clf, name in [(lr, 'Logistic'),
(est, name),
(isotonic, name + ' + Isotonic'),
(sigmoid, name + ' + Sigmoid')]:
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
if hasattr(clf, "predict_proba"):
prob_pos = clf.predict_proba(X_test)[:, 1]
else:  # use decision function
prob_pos = clf.decision_function(X_test)
prob_pos = \
(prob_pos - prob_pos.min()) / (prob_pos.max() - prob_pos.min())

clf_score = brier_score_loss(y_test, prob_pos, pos_label=y.max())
print("%s:" % name)
print("\tBrier: %1.3f" % (clf_score))
print("\tPrecision: %1.3f" % precision_score(y_test, y_pred))
print("\tRecall: %1.3f" % recall_score(y_test, y_pred))
print("\tF1: %1.3f\n" % f1_score(y_test, y_pred))

fraction_of_positives, mean_predicted_value = \
calibration_curve(y_test, prob_pos, n_bins=10)

ax1.plot(mean_predicted_value, fraction_of_positives, "s-",
label="%s (%1.3f)" % (name, clf_score))

ax2.hist(prob_pos, range=(0, 1), bins=10, label=name,
histtype="step", lw=2)

ax1.set_ylabel("Fraction of positives")
ax1.set_ylim([-0.05, 1.05])
ax1.legend(loc="lower right")
ax1.set_title('Calibration plots  (reliability curve)')

ax2.set_xlabel("Mean predicted value")
ax2.set_ylabel("Count")
ax2.legend(loc="upper center", ncol=2)

plt.tight_layout()

In [9]:
# Plot calibration cuve for Linear SVC
plot_calibration_curve(LinearSVC(), "SVC", 2)

Logistic:
Brier: 0.099
Precision: 0.872
Recall: 0.851
F1: 0.862

SVC:
Brier: 0.163
Precision: 0.872
Recall: 0.852
F1: 0.862

SVC + Isotonic:
Brier: 0.100
Precision: 0.853
Recall: 0.878
F1: 0.865

SVC + Sigmoid:
Brier: 0.099
Precision: 0.874
Recall: 0.849
F1: 0.861



One can observe here that logistic regression is well calibrated as its curve is nearly diagonal. Linear SVC's calibration curve has a sigmoid curve, which is typical for an under-confident classifier. In the case of LinearSVC, this is caused by the margin property of the hinge loss, which lets the model focus on hard samples that are close to the decision boundary (the support vectors). Both kinds of calibration can fix this issue and yield nearly identical results. The next figure shows the calibration curve of Gaussian naive Bayes on the same data, with both kinds of calibration and also without calibration.

In [10]:
# Plot calibration cuve for Gaussian Naive Bayes
plot_calibration_curve(GaussianNB(), "Naive Bayes", 1)

Logistic:
Brier: 0.099
Precision: 0.872
Recall: 0.851
F1: 0.862

Naive Bayes:
Brier: 0.118
Precision: 0.857
Recall: 0.876
F1: 0.867

Naive Bayes + Isotonic:
Brier: 0.098
Precision: 0.883
Recall: 0.836
F1: 0.859

Naive Bayes + Sigmoid:
Brier: 0.109
Precision: 0.861
Recall: 0.871
F1: 0.866



One can see that Gaussian naive Bayes performs very badly but does so in an other way than linear SVC: While linear SVC exhibited a sigmoid calibration curve, Gaussian naive Bayes' calibration curve has a transposed-sigmoid shape. This is typical for an over-confident classifier. In this case, the classifier's overconfidence is caused by the redundant features which violate the naive Bayes assumption of feature-independence.

Calibration of the probabilities of Gaussian naive Bayes with isotonic regression can fix this issue as can be seen from the nearly diagonal calibration curve. Sigmoid calibration also improves the brier score slightly, albeit not as strongly as the non-parametric isotonic calibration. This is an intrinsic limitation of sigmoid calibration, whose parametric form assumes a sigmoid rather than a transposed-sigmoid curve. The non-parametric isotonic calibration model, however, makes no such strong assumptions and can deal with either shape, provided that there is sufficient calibration data. In general, sigmoid calibration is preferable if the calibration curve is sigmoid and when there is few calibration data while isotonic calibration is preferable for non- sigmoid calibration curves and in situations where many additional data can be used for calibration.

## Multi-class classification¶

CalibratedClassifierCV can also deal with classification tasks that involve more than two classes if the base estimator can do so. In this case, the classifier is calibrated first for each class separately in an one-vs-rest fashion. When predicting probabilities for unseen data, the calibrated probabilities for each class are predicted separately. As those probabilities do not necessarily sum to one, a postprocessing is performed to normalize them.

The next image illustrates how sigmoid calibration changes predicted probabilities for a 3-class classification problem. Illustrated is the standard 2-simplex, where the three corners correspond to the three classes. Arrows point from the probability vectors predicted by an uncalibrated classifier to the probability vectors predicted by the same classifier after sigmoid calibration on a hold-out validation set. Colors indicate the true class of an instance (red: class 1, green: class 2, blue: class 3).

In [11]:
# <!-- collapse=True -->

np.random.seed(0)

# Generate data
X, y = datasets.make_blobs(n_samples=1000, n_features=2, random_state=42,
cluster_std=5.0)
X_train, y_train = X[:600], y[:600]
X_valid, y_valid = X[600:800], y[600:800]
X_train_valid, y_train_valid = X[:800], y[:800]
X_test, y_test = X[800:], y[800:]

In [12]:
# <!-- collapse=True -->

# Train uncalibrated random forest classifier on whole train and validation
# data and evaluate on test data
clf = RandomForestClassifier(n_estimators=25)
clf.fit(X_train_valid, y_train_valid)
clf_probs = clf.predict_proba(X_test)
score = log_loss(y_test, clf_probs)

# Train random forest classifier, calibrate on validation data and evaluate
# on test data
clf = RandomForestClassifier(n_estimators=25)
clf.fit(X_train, y_train)
clf_probs = clf.predict_proba(X_test)
sig_clf = CalibratedClassifierCV(clf, method="sigmoid", cv="prefit")
sig_clf.fit(X_valid, y_valid)
sig_clf_probs = sig_clf.predict_proba(X_test)
sig_score = log_loss(y_test, sig_clf_probs)

In [13]:
# <!-- collapse=True -->

# Plot changes in predicted probabilities via arrows
plt.figure(0, figsize=(10, 8))
colors = ["r", "g", "b"]
for i in range(clf_probs.shape[0]):
plt.arrow(clf_probs[i, 0], clf_probs[i, 1],
sig_clf_probs[i, 0] - clf_probs[i, 0],
sig_clf_probs[i, 1] - clf_probs[i, 1],

# Plot perfect predictions
plt.plot([1.0], [0.0], 'ro', ms=20, label="Class 1")
plt.plot([0.0], [1.0], 'go', ms=20, label="Class 2")
plt.plot([0.0], [0.0], 'bo', ms=20, label="Class 3")

# Plot boundaries of unit simplex
plt.plot([0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0], 'k', label="Simplex")

# Annotate points on the simplex
plt.annotate(r'($\frac{1}{3}$, $\frac{1}{3}$, $\frac{1}{3}$)',
xy=(1.0/3, 1.0/3), xytext=(1.0/3, .23), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.plot([1.0/3], [1.0/3], 'ko', ms=5)
plt.annotate(r'($\frac{1}{2}$, $0$, $\frac{1}{2}$)',
xy=(.5, .0), xytext=(.5, .1), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.annotate(r'($0$, $\frac{1}{2}$, $\frac{1}{2}$)',
xy=(.0, .5), xytext=(.1, .5), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.annotate(r'($\frac{1}{2}$, $\frac{1}{2}$, $0$)',
xy=(.5, .5), xytext=(.6, .6), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.annotate(r'($0$, $0$, $1$)',
xy=(0, 0), xytext=(.1, .1), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.annotate(r'($1$, $0$, $0$)',
xy=(1, 0), xytext=(1, .1), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.annotate(r'($0$, $1$, $0$)',
xy=(0, 1), xytext=(.1, 1), xycoords='data',
arrowprops=dict(facecolor='black', shrink=0.05),
horizontalalignment='center', verticalalignment='center')
plt.grid("off")
for x in [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]:
plt.plot([0, x], [x, 0], 'k', alpha=0.2)
plt.plot([0, 0 + (1-x)/2], [x, x + (1-x)/2], 'k', alpha=0.2)
plt.plot([x, x + (1-x)/2], [0, 0 + (1-x)/2], 'k', alpha=0.2)

plt.title("Change of predicted probabilities after sigmoid calibration")
plt.xlabel("Probability class 1")
plt.ylabel("Probability class 2")
plt.xlim(-0.05, 1.05)
plt.ylim(-0.05, 1.05)
plt.legend(loc="best")

print("Log-loss of")
print(" * uncalibrated classifier trained on 800 datapoints: %.3f "
% score)
print(" * classifier trained on 600 datapoints and calibrated on "
"200 datapoint: %.3f" % sig_score)

Log-loss of
* uncalibrated classifier trained on 800 datapoints: 1.280
* classifier trained on 600 datapoints and calibrated on 200 datapoint: 0.536


The base classifier is a random forest classifier with 25 base estimators (trees). If this classifier is trained on all 800 training datapoints, it is overly confident in its predictions and thus incurs a large log-loss. Calibrating an identical classifier, which was trained on 600 datapoints, with method='sigmoid' on the remaining 200 datapoints reduces the confidence of the predictions, i.e., moves the probability vectors from the edges of the simplex towards the center:

In [14]:
# <!-- collapse=True -->

# Illustrate calibrator
plt.figure(1, figsize=(10, 8))
# generate grid over 2-simplex
p1d = np.linspace(0, 1, 20)
p0, p1 = np.meshgrid(p1d, p1d)
p2 = 1 - p0 - p1
p = np.c_[p0.ravel(), p1.ravel(), p2.ravel()]
p = p[p[:, 2] >= 0]

calibrated_classifier = sig_clf.calibrated_classifiers_[0]
prediction = np.vstack([calibrator.predict(this_p)
for calibrator, this_p in
zip(calibrated_classifier.calibrators_, p.T)]).T
prediction /= prediction.sum(axis=1)[:, None]

# Ploit modifications of calibrator
for i in range(prediction.shape[0]):
plt.arrow(p[i, 0], p[i, 1],
prediction[i, 0] - p[i, 0], prediction[i, 1] - p[i, 1],
# Plot boundaries of unit simplex
plt.plot([0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0], 'k', label="Simplex")

plt.grid("off")
for x in [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]:
plt.plot([0, x], [x, 0], 'k', alpha=0.2)
plt.plot([0, 0 + (1-x)/2], [x, x + (1-x)/2], 'k', alpha=0.2)
plt.plot([x, x + (1-x)/2], [0, 0 + (1-x)/2], 'k', alpha=0.2)

plt.title("Illustration of sigmoid calibrator")
plt.xlabel("Probability class 1")
plt.ylabel("Probability class 2")
plt.xlim(-0.05, 1.05)
plt.ylim(-0.05, 1.05)

Out[14]:
(-0.05, 1.05)

This calibration results in a lower log-loss. Note that an alternative would have been to increase the number of base estimators which would have resulted in a similar decrease in log-loss.

## Summary¶

In summary, the newly added CalibratedClassifierCV allows to improve the quality of predicted class probabilities of binary and multi-class classifiers. It provides both a parametric calibration assuming a sigmoid shape of the calibration curve and a non-parametric calibration based on isotonic regression, which can cope with any shape of the calibration curve, provided that sufficient calibration data exists. The main bottlenecks of the method are the increased computation time due to the internal cross-validation loop and the necessity of additional calibration data, which can be alleviated by cross-validation.

## References¶

[1] Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers, B. Zadrozny & C. Elkan, ICML 2001

[2] Transforming Classifier Scores into Accurate Multiclass Probability Estimates, B. Zadrozny & C. Elkan, (KDD 2002)

[3] Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods, J. Platt, (1999)

[4] Predicting Good Probabilities with Supervised Learning, A. Niculescu-Mizil & R. Caruana, ICML 2005

In [15]:
%load_ext watermark
%watermark -a "Jan Hendrik Metzen" -d -v -m -p numpy

Jan Hendrik Metzen 14/04/2015

CPython 2.7.5+
IPython 2.4.1

numpy 1.9.2

compiler   : GCC 4.8.1
system     : Linux
release    : 3.16.0-28-generic
machine    : x86_64
processor  : x86_64
CPU cores  : 4
interpreter: 64bit


This post was written as an IPython notebook. You can download this notebook.