Question

我知道sgdclassifier铰链损失不支持概率估计。那么在使用log_loss指标时如何与GridSearchCV一起使用？

clf = SGDClassifier(loss='hinge')

grid_params = {'alpha': [0.0001, 0.001, 0.01]}
grid_search = GridSearchCV(clf, grid_params, scoring='neg_log_loss')
grid_search.fit(X_train, y_train)

它返回：

AttributeError：概率估计不适用于 loss ='铰链'

有什么办法可以使这项工作完成？

Answer 1

将损失从铰链转移到日志正将算法从SVM更改为逻辑回归，因此我认为这是不可能的。

但是，您可以在Scikit-learn的CalibratedClassifierCV中将SGDClassifier设置为基本估计量，这将生成概率估计值。

这是一个例子：

from sklearn.calibration import CalibratedClassifierCV
from sklearn.linear_model import SGDClassifier
from sklearn.model_selection import GridSearchCV, train_test_split
from sklearn.datasets import load_iris

# load some example data
data = load_iris()
X = data['data']
y = data['target']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

clf = SGDClassifier(loss='hinge', max_iter=100)
calibrated_clf = CalibratedClassifierCV(base_estimator=clf, method='sigmoid', cv=3)  # set the SGD classifier as the base estimator


grid_params = {'base_estimator__alpha': [0.0001, 0.001, 0.01]}  # note 'base_estimator__' in the params because you want to change params in the SGDClassifier
grid_search = GridSearchCV(estimator=calibrated_clf, param_grid=grid_params, cv=3)
grid_search.fit(X_train, y_train)

print(grid_search.best_params_)

{'base_estimator__alpha': 0.0001}

现在使校准的分类器具有最佳参数：

calibrated_clf.set_params(**grid_search.best_params_)
calibrated_clf.fit(X_train, y_train)
preds = calibrated_clf.predict_proba(X_test)
print(preds)

# probabilities for each of the 3 classes:

array([[7.62825746e-02, 5.24891243e-01, 3.98826183e-01],
       [9.24810700e-01, 7.50659865e-02, 1.23313813e-04],
       [8.40690799e-01, 1.59138563e-01, 1.70637465e-04],
       [7.10696359e-01, 2.88969750e-01, 3.33891072e-04],
       [7.99360835e-02, 7.83076911e-01, 1.36987006e-01],
       [9.90417693e-03, 7.72846023e-02, 9.12811221e-01],
       [1.07116396e-02, 3.03030985e-01, 6.86257375e-01],
       [1.43944221e-02, 1.17223024e-01, 8.68382554e-01],
       [1.11659634e-01, 7.35051942e-01, 1.53288424e-01],
       [8.30127745e-03, 1.39546231e-01, 8.52152492e-01],
       [2.07825315e-02, 1.56925620e-01, 8.22291849e-01],
       [8.88421387e-01, 1.11384933e-01, 1.93680314e-04],
       [6.90696963e-01, 3.09038629e-01, 2.64408097e-04],
       [1.26043359e-01, 5.78366890e-01, 2.95589750e-01],
       [3.83356263e-03, 4.06197230e-01, 5.89969207e-01],
       [7.78520570e-01, 2.21144460e-01, 3.34969184e-04],
       [5.11227086e-02, 6.32329915e-01, 3.16547377e-01],
       [8.24310445e-01, 1.75412791e-01, 2.76763715e-04],
       [3.50118697e-02, 3.91028064e-01, 5.73960067e-01],
       [1.23034113e-01, 7.32289832e-01, 1.44676055e-01],
       [3.44588463e-01, 5.92799831e-01, 6.26117056e-02],
       [2.67170305e-02, 5.78551461e-01, 3.94731509e-01],
       [5.92943916e-02, 5.57127843e-01, 3.83577765e-01],
       [7.16297083e-01, 2.83282184e-01, 4.20732771e-04],
       [7.82091800e-03, 1.30949377e-01, 8.61229705e-01],
       [1.70781668e-01, 5.47432635e-01, 2.81785697e-01],
       [8.38288358e-01, 1.61495161e-01, 2.16480625e-04],
       [2.11106665e-02, 4.66121567e-01, 5.12767766e-01],
       [9.20496389e-02, 6.29184167e-01, 2.78766194e-01],
       [1.29649784e-02, 2.73576019e-01, 7.13459002e-01]])

如何使用对数损失指标将sgdclassifier铰链损失与Gridsearchcv一起使用？

1 个答案: