DOC/TST Clarify group order in GroupKFold and LeaveOneGroupOut (#22582)

SamAdamDay · glemaitre · commit fbe2126e5dc4 · 2022-05-19T14:13:33.000+02:00
diff --git a/sklearn/model_selection/_split.py b/sklearn/model_selection/_split.py
@@ -469,6 +469,10 @@ class GroupKFold(_BaseKFold):
         .. versionchanged:: 0.22
             ``n_splits`` default value changed from 3 to 5.
 
+    Notes
+    -----
+    Groups appear in an arbitrary order throughout the folds.
+
     Examples
     --------
     >>> import numpy as np
@@ -1110,6 +1114,12 @@ class LeaveOneGroupOut(BaseCrossValidator):
 
     Read more in the :ref:`User Guide <leave_one_group_out>`.
 
+    Notes
+    -----
+    Splits are ordered according to the index of the group left out. The first
+    split has training set consting of the group whose index in `groups` is
+    lowest, and so on.
+
     Examples
     --------
     >>> import numpy as np
@@ -1137,7 +1147,6 @@ class LeaveOneGroupOut(BaseCrossValidator):
     [[1 2]
      [3 4]] [[5 6]
      [7 8]] [1 2] [1 2]
-
     """
 
     def _iter_test_masks(self, X, y, groups):
diff --git a/sklearn/model_selection/tests/test_split.py b/sklearn/model_selection/tests/test_split.py
@@ -1041,6 +1041,26 @@ def test_leave_group_out_changing_groups():
     assert 3 == LeaveOneGroupOut().get_n_splits(X, y=X, groups=groups)
 
 
+def test_leave_group_out_order_dependence():
+    # Check that LeaveOneGroupOut orders the splits according to the index
+    # of the group left out.
+    groups = np.array([2, 2, 0, 0, 1, 1])
+    X = np.ones(len(groups))
+
+    splits = iter(LeaveOneGroupOut().split(X, groups=groups))
+
+    expected_indices = [
+        ([0, 1, 4, 5], [2, 3]),
+        ([0, 1, 2, 3], [4, 5]),
+        ([2, 3, 4, 5], [0, 1]),
+    ]
+
+    for expected_train, expected_test in expected_indices:
+        train, test = next(splits)
+        assert_array_equal(train, expected_train)
+        assert_array_equal(test, expected_test)
+
+
 def test_leave_one_p_group_out_error_on_fewer_number_of_groups():
     X = y = groups = np.ones(0)
     msg = re.escape("Found array with 0 sample(s)")
