how to enable probability estimates when using scikitlearn's LinearSVC classifier

问题

How can I get the probability estimates of predictions from a sklearn.svm.LinearSVC model in similar fashion to sklearn.svm.SVC's probability=True option that allows predict_proba() I need to avoid the quadratic fit penalty of the underlying libsvm of SVC as my training set is large.

回答1:

sklearn.svm.LinearSVC does not have predict_proba method as you noticed correctly.

However, you may try the following trick to circumvent this shortcoming:

from sklearn.svm import LinearSVC
from sklearn.calibration import CalibratedClassifierCV
svm = CalibratedClassifierCV(LinearSVC())
svm
CalibratedClassifierCV(base_estimator=LinearSVC(C=1.0, class_weight=None, dual=True, fit_intercept=True,
     intercept_scaling=1, loss='squared_hinge', max_iter=1000,
     multi_class='ovr', penalty='l2', random_state=None, tol=0.0001,
     verbose=0),
            cv=3, method='sigmoid')

The resulting svm model indeed has predict_proba method available.

You may read more about CalibratedClassifierCV

来源：https://stackoverflow.com/questions/54402210/how-to-enable-probability-estimates-when-using-scikitlearns-linearsvc-classifie