Description
Describe the bug
I'm not sure if it is a proper bug, or my lack of understanding of the metadata routing API ;)
When enable_metadata_routing=True, the score method of a LogisticRegressionCV estimator will ignore sample_weight.
set_config(enable_metadata_routing=True)
logreg_cv = LogisticRegressionCV().fit(X, y)
logreg_cv.score(X, y, sample_weight=sw)==logreg_cv.score(X, y) #unweighted accuracyI found it surprising, because the score method works fine when enable_metadata_routing=False, so the same piece of code behaves differently depending on the metadata routing config.
set_config(enable_metadata_routing=False)
logreg_cv = LogisticRegressionCV().fit(X, y)
logreg_cv.score(X, y, sample_weight=sw) #weighted accuracyIf I understood the metadata routing API correctly, to make the score method sample_weight aware we need to explicitly pass a scorer that request it:
set_config(enable_metadata_routing=True)
weighted_accuracy = make_scorer(accuracy_score).set_score_request(sample_weight=True)
logreg_cv = LogisticRegressionCV(scoring=weighted_accuracy).fit(X, y)
logreg_cv.score(X, y, sample_weight=sw) #weighted accuracyIf it's the intended behavior of the metadata routing API, maybe we should warn the user or raise an error in the first case, instead of silently ignoring sample_weight ?
Steps/Code to Reproduce
from sklearn import set_config
from sklearn.metrics import make_scorer, accuracy_score
from sklearn.linear_model import LogisticRegressionCV
import numpy as np
rng = np.random.RandomState(22)
n_samples, n_features = 10, 4
X = rng.rand(n_samples, n_features)
y = rng.randint(0, 2, size=n_samples)
sw = rng.randint(0, 5, size=n_samples)
set_config(enable_metadata_routing=True)
logreg_cv = LogisticRegressionCV()
logreg_cv.fit(X, y)
# sample_weight is silently ignored in logreg_cv.score
assert logreg_cv.score(X, y) == logreg_cv.score(X, y, sample_weight=sw)
assert not logreg_cv.score(X, y, sample_weight=sw)==accuracy_score(logreg_cv.predict(X),y, sample_weight=sw)Expected Results
Either logreg_cv.score(X, y, sample_weight=sw) raises an error/warning or the assertions are false.
Actual Results
The assertions are true.
Versions
sklearn: 1.7.dev0Metadata
Metadata
Assignees
Type
Projects
Status