Python Scikit-learn Perceptron Output Probabilities

Question

I'm using scikit-learn's Perceptron algorithm to do binary classification. When using some of the other algorithms in the library (RandomForestClassifer, LogisticRegression, etc.), I can use model.predict_proba() to have the algorithm output the probability of getting a positive (1) for each example. Is there a way to get a similar output for the Perceptron algorithm?

The closest I've been able to come is model.decision_function(), which outputs a confidence score for the example based on the signed distance to the hyperplane, but I'm not sure how to convert these confidence scores to the probability figures I want.

model.predict() is also only returning binary values.

Answer 1

I think what you want is CalibratedClassifierCV:

from sklearn import linear_model
from sklearn.datasets import make_classification
from sklearn.calibration import CalibratedClassifierCV
# Build a classification task using 3 informative features
X, y = make_classification(n_samples=1000,
                           n_features=10,
                           n_informative=3,
                           n_redundant=0,
                           n_repeated=0,
                           n_classes=2,
                           random_state=0,
                           shuffle=False)


per = linear_model.Perceptron()

clf_isotonic = CalibratedClassifierCV(per, cv=10, method='isotonic')

clf_isotonic.fit(X[:900], y[:900])

preds = clf_isotonic.predict_proba(X[900:])
print preds

[Edit] You can use this also to make other linear_models produce probabilities for classification problems