FIX pass sample weights to final estimator (scikit-learn#15773)

Author: J-A16

doc/whats_new/v0.23.rst

@@ -58,6 +58,14 @@ Changelog
   :func:`datasets.make_moons` now accept two-element tuple.
   :pr:`15707` by :user:`Maciej J Mikulski <mjmikulski>`
 
+:mod:`sklearn.linear_model`
+...........................
+
+- |Fix| Fixed a bug where if a `sample_weight` parameter was passed to the fit
+  method of :class:`linear_model.RANSACRegressor`, it would not be passed to
+  the wrapped `base_estimator` during the fitting of the final model.
+  :pr:`15573` by :user:`Jeremy Alexandre <J-A16>`.
+
 :mod:`sklearn.preprocessing`
 ............................
 

sklearn/linear_model/_ransac.py

@@ -328,6 +328,7 @@ def fit(self, X, y, sample_weight=None):
         inlier_mask_best = None
         X_inlier_best = None
         y_inlier_best = None
+        inlier_best_idxs_subset = None
         self.n_skips_no_inliers_ = 0
         self.n_skips_invalid_data_ = 0
         self.n_skips_invalid_model_ = 0
@@ -404,6 +405,7 @@ def fit(self, X, y, sample_weight=None):
             inlier_mask_best = inlier_mask_subset
             X_inlier_best = X_inlier_subset
             y_inlier_best = y_inlier_subset
+            inlier_best_idxs_subset = inlier_idxs_subset
 
             max_trials = min(
                 max_trials,
@@ -441,7 +443,13 @@ def fit(self, X, y, sample_weight=None):
                               ConvergenceWarning)
 
         # estimate final model using all inliers
-        base_estimator.fit(X_inlier_best, y_inlier_best)
+        if sample_weight is None:
+            base_estimator.fit(X_inlier_best, y_inlier_best)
+        else:
+            base_estimator.fit(
+                X_inlier_best,
+                y_inlier_best,
+                sample_weight=sample_weight[inlier_best_idxs_subset])
 
         self.estimator_ = base_estimator
         self.inlier_mask_ = inlier_mask_best

sklearn/linear_model/tests/test_ransac.py

@@ -10,6 +10,8 @@
 from sklearn.utils._testing import assert_almost_equal
 from sklearn.utils._testing import assert_raises_regexp
 from sklearn.utils._testing import assert_raises
+from sklearn.utils._testing import assert_allclose
+from sklearn.datasets import make_regression
 from sklearn.linear_model import LinearRegression, RANSACRegressor, Lasso
 from sklearn.linear_model._ransac import _dynamic_max_trials
 from sklearn.exceptions import ConvergenceWarning
@@ -494,3 +496,21 @@ def test_ransac_fit_sample_weight():
     base_estimator = Lasso()
     ransac_estimator = RANSACRegressor(base_estimator)
     assert_raises(ValueError, ransac_estimator.fit, X, y, weights)
+
+
+def test_ransac_final_model_fit_sample_weight():
+    X, y = make_regression(n_samples=1000, random_state=10)
+    rng = check_random_state(42)
+    sample_weight = rng.randint(1, 4, size=y.shape[0])
+    sample_weight = sample_weight / sample_weight.sum()
+    ransac = RANSACRegressor(base_estimator=LinearRegression(), random_state=0)
+    ransac.fit(X, y, sample_weight=sample_weight)
+
+    final_model = LinearRegression()
+    mask_samples = ransac.inlier_mask_
+    final_model.fit(
+        X[mask_samples], y[mask_samples],
+        sample_weight=sample_weight[mask_samples]
+    )
+
+    assert_allclose(ransac.estimator_.coef_, final_model.coef_)