stratified group kfold splitter (#899)

* stratified group kfold splitter

* exclude catboost

---------

Co-authored-by: Shaokun <shaokunzhang529@gmail.com>
Co-authored-by: Qingyun Wu <qingyun.wu@psu.edu>

flaml/automl/automl.py

@@ -17,6 +17,7 @@ from sklearn.model_selection import (
     GroupKFold,
     TimeSeriesSplit,
     GroupShuffleSplit,
+    StratifiedGroupKFold,
 )
 from sklearn.utils import shuffle
 from sklearn.base import BaseEstimator
@@ -1575,8 +1576,8 @@ class AutoML(BaseEstimator):
         else:
             # logger.info("Using splitter object")
             self._state.kf = self._split_type
-        if isinstance(self._state.kf, GroupKFold):
+        if isinstance(self._state.kf, (GroupKFold, StratifiedGroupKFold)):
-            # self._split_type is either "group" or a GroupKFold object
+            # self._split_type is either "group", a GroupKFold object, or a StratifiedGroupKFold object
             self._state.kf.groups = self._state.groups_all
 
     def add_learner(self, learner_name, learner_class):

flaml/automl/ml.py

@@ -17,7 +17,12 @@ from sklearn.metrics import (
     mean_absolute_percentage_error,
     ndcg_score,
 )
-from sklearn.model_selection import RepeatedStratifiedKFold, GroupKFold, TimeSeriesSplit
+from sklearn.model_selection import (
+    RepeatedStratifiedKFold,
+    GroupKFold,
+    TimeSeriesSplit,
+    StratifiedGroupKFold,
+)
 from flaml.automl.model import (
     XGBoostSklearnEstimator,
     XGBoost_TS,
@@ -517,7 +522,7 @@ def evaluate_model_CV(
     shuffle = getattr(kf, "shuffle", task not in TS_FORECAST)
     if isinstance(kf, RepeatedStratifiedKFold):
         kf = kf.split(X_train_split, y_train_split)
-    elif isinstance(kf, GroupKFold):
+    elif isinstance(kf, (GroupKFold, StratifiedGroupKFold)):
         groups = kf.groups
         kf = kf.split(X_train_split, y_train_split, groups)
         shuffle = False
@@ -548,8 +553,16 @@ def evaluate_model_CV(
                 weight[val_index],
             )
         if groups is not None:
-            fit_kwargs["groups"] = groups[train_index]
+            fit_kwargs["groups"] = (
-            groups_val = groups[val_index]
+                groups[train_index]
+                if isinstance(groups, np.ndarray)
+                else groups.iloc[train_index]
+            )
+            groups_val = (
+                groups[val_index]
+                if isinstance(groups, np.ndarray)
+                else groups.iloc[val_index]
+            )
         else:
             groups_val = None
         val_loss_i, metric_i, train_time_i, pred_time_i = get_val_loss(

test/automl/test_split.py

@@ -94,6 +94,33 @@ def test_groups():
     automl.fit(X, y, **automl_settings)
 
 
+def test_stratified_groupkfold():
+    from sklearn.model_selection import StratifiedGroupKFold
+    from flaml.data import load_openml_dataset
+
+    X_train, _, y_train, _ = load_openml_dataset(dataset_id=1169, data_dir="test/")
+    splitter = StratifiedGroupKFold(n_splits=5, shuffle=True, random_state=0)
+
+    automl = AutoML()
+    settings = {
+        "time_budget": 6,
+        "metric": "ap",
+        "eval_method": "cv",
+        "split_type": splitter,
+        "groups": X_train["Airline"],
+        "estimator_list": [
+            "lgbm",
+            "rf",
+            "xgboost",
+            "extra_tree",
+            "xgb_limitdepth",
+            "lrl1",
+        ],
+    }
+
+    automl.fit(X_train=X_train, y_train=y_train, **settings)
+
+
 def test_rank():
     from sklearn.externals._arff import ArffException