Fix tests with the latest scikit-learn. (#11086)

* Fix tests with the latest scikit-learn.

* dask.

* Remove scikit-learn pin

---------

Co-authored-by: Philip Hyunsu Cho <[email protected]>

ops/conda_env/macos_cpu_test.yml

@@ -11,7 +11,7 @@ dependencies:
 - numpy
 - scipy
 - llvm-openmp
-- scikit-learn>=1.4.1,<=1.5.2
+- scikit-learn>=1.4.1
 - pandas
 - matplotlib
 - dask<=2024.10.0

ops/conda_env/win64_test.yml

@@ -6,7 +6,7 @@ dependencies:
 - numpy
 - scipy
 - matplotlib
-- scikit-learn<=1.5.2
+- scikit-learn
 - pandas
 - pytest
 - boto3

python-package/xgboost/dask/__init__.py

@@ -1616,7 +1616,7 @@ def _client_sync(self, func: Callable, **kwargs: Any) -> Any:
 @xgboost_model_doc(
     """Implementation of the Scikit-Learn API for XGBoost.""", ["estimators", "model"]
 )
-class DaskXGBRegressor(DaskScikitLearnBase, XGBRegressorBase):
+class DaskXGBRegressor(XGBRegressorBase, DaskScikitLearnBase):
     """dummy doc string to workaround pylint, replaced by the decorator."""
 
     async def _fit_async(
@@ -1707,7 +1707,7 @@ def fit(
     "Implementation of the scikit-learn API for XGBoost classification.",
     ["estimators", "model"],
 )
-class DaskXGBClassifier(DaskScikitLearnBase, XGBClassifierBase):
+class DaskXGBClassifier(XGBClassifierBase, DaskScikitLearnBase):
     # pylint: disable=missing-class-docstring
     async def _fit_async(
         self,
@@ -1911,7 +1911,7 @@ def _argmax(x: Any) -> Any:
             For the dask implementation, group is not supported, use qid instead.
 """,
 )
-class DaskXGBRanker(DaskScikitLearnBase, XGBRankerMixIn):
+class DaskXGBRanker(XGBRankerMixIn, DaskScikitLearnBase):
     @_deprecate_positional_args
     def __init__(
         self,

tests/python-gpu/test_gpu_prediction.py

@@ -477,7 +477,7 @@ def test_predict_leaf_dart(self, param: dict, dataset: tm.TestDataset) -> None:
     )
     @settings(deadline=None, max_examples=20, print_blob=True)
     def test_predict_categorical_split(self, df):
-        from sklearn.metrics import mean_squared_error
+        from sklearn.metrics import root_mean_squared_error
 
         df = df.astype("category")
         x0, x1 = df["x0"].to_numpy(), df["x1"].to_numpy()
@@ -504,7 +504,7 @@ def test_predict_categorical_split(self, df):
         )
         bst.set_param({"device": "cuda:0"})
         pred = bst.predict(dtrain)
-        rmse = mean_squared_error(y_true=y, y_pred=pred, squared=False)
+        rmse = root_mean_squared_error(y_true=y, y_pred=pred)
         np.testing.assert_almost_equal(
             rmse, eval_history["train"]["rmse"][-1], decimal=5
         )