Description

The fit method of OneVsRestClassifier and other MetaEstimators (e.g. OneVsOneClassifier) only accepts X and y as parameters. Underlying estimators like SGDClassifier accept more optional keyword args to this function which are essential for some tasks (e.g. the sample_weight parameter is the only way to add weights to training samples in multi-label classification problems).

Steps/Code to Reproduce

Here's how I solve a multi-label classification task with a linear SVM

from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.multiclass import OneVsRestClassifier
from sklearn.linear_model import SGDClassifier

X = [[0.0, 0.0], [1.0, 1.0], [1.0, 0.0]]
y = [[0], [1], [0, 1]]
y_mlb = MultiLabelBinarizer().fit_transform(y)
sample_weight = [1.0, 0.5, 0.8]
clf = OneVsRestClassifier(SGDClassifier(loss="hinge"))
clf.fit(X, y_mlb)   # unable to pass `sample_weight`

see also this related question on stackoverflow.

Expected Results

For regular (single-label) classification tasks, I can pass the sample_weight kwarg directly to SGDClassifier.fit, but with OneVsRestClassifier this is not possible

Actual Results

I cannot add weights to my training samples when OneVsRestClassifier (or a similar meta-estimator) is used.

Feature Request

Please let the fit method of OneVsRestClassifier (and similar meta-estimators) accept arbitrary kwargs and pass them on to the fit method of the wrapped estimator. The same may be useful for partial_fit and score, though I'm not sure about that.