Merge pull request scikit-learn#4448 from pprett/fix-gbrt-min-leaf-weight

Fix gbrt min leaf weight

Author: pprett

sklearn/ensemble/gradient_boosting.py

@@ -1002,12 +1002,19 @@ def _fit_stages(self, X, y, y_pred, sample_weight, random_state,
         n_inbag = max(1, int(self.subsample * n_samples))
         loss_ = self.loss_
 
+        # Set min_weight_leaf from min_weight_fraction_leaf
+        if self.min_weight_fraction_leaf != 0. and sample_weight is not None:
+            min_weight_leaf = (self.min_weight_fraction_leaf *
+                               np.sum(sample_weight))
+        else:
+            min_weight_leaf = 0.
+
         # init criterion and splitter
         criterion = FriedmanMSE(1)
         splitter = PresortBestSplitter(criterion,
                                        self.max_features_,
                                        self.min_samples_leaf,
-                                       self.min_weight_fraction_leaf,
+                                       min_weight_leaf,
                                        random_state)
 
         if self.verbose:

sklearn/ensemble/tests/test_gradient_boosting.py

@@ -976,10 +976,26 @@ def test_non_uniform_weights_toy_edge_case_reg():
     y = [0, 0, 1, 0]
     # ignore the first 2 training samples by setting their weight to 0
     sample_weight = [0, 0, 1, 1]
-    for loss in ('ls', 'huber', 'lad', 'quantile'):
-        gb = GradientBoostingRegressor(n_estimators=5)
+    for loss in ('huber', 'ls', 'lad', 'quantile'):
+        gb = GradientBoostingRegressor(learning_rate=1.0, n_estimators=2, loss=loss)
         gb.fit(X, y, sample_weight=sample_weight)
-        assert_true(gb.predict([[1, 0]])[0] > 0.5)
+        assert_greater(gb.predict([[1, 0]])[0], 0.5)
+
+
+def test_non_uniform_weights_toy_min_weight_leaf():
+    """Regression test for https://github.com/scikit-learn/scikit-learn/issues/4447 """
+    X = [[1, 0],
+         [1, 0],
+         [1, 0],
+         [0, 1],
+        ]
+    y = [0, 0, 1, 0]
+    # ignore the first 2 training samples by setting their weight to 0
+    sample_weight = [0, 0, 1, 1]
+    gb = GradientBoostingRegressor(n_estimators=5, min_weight_fraction_leaf=0.1)
+    gb.fit(X, y, sample_weight=sample_weight)
+    assert_true(gb.predict([[1, 0]])[0] > 0.5)
+    assert_almost_equal(gb.estimators_[0,0].splitter.min_weight_leaf, 0.2)
 
 
 def test_non_uniform_weights_toy_edge_case_clf():