[pyspark] Filter out the unsupported train parameters (dmlc#8355)

Author: wbo4958

python-package/xgboost/spark/core.py

@@ -126,6 +126,11 @@
     "eval_qid",  # Use spark param `qid_col` instead
 }
 
+_unsupported_train_params = {
+    "evals",  # Supported by spark param validation_indicator_col
+    "evals_result",  # Won't support yet+
+}
+
 _unsupported_predict_params = {
     # for classification, we can use rawPrediction as margin
     "output_margin",
@@ -515,6 +520,7 @@ def setParams(self, **kwargs):  # pylint: disable=invalid-name
                     k in _unsupported_xgb_params
                     or k in _unsupported_fit_params
                     or k in _unsupported_predict_params
+                    or k in _unsupported_train_params
                 ):
                     raise ValueError(f"Unsupported param '{k}'.")
                 _extra_params[k] = v
@@ -620,7 +626,9 @@ def _get_distributed_train_params(self, dataset):
 
     @classmethod
     def _get_xgb_train_call_args(cls, train_params):
-        xgb_train_default_args = _get_default_params_from_func(xgboost.train, {})
+        xgb_train_default_args = _get_default_params_from_func(
+            xgboost.train, _unsupported_train_params
+        )
         booster_params, kwargs_params = {}, {}
         for key, value in train_params.items():
             if key in xgb_train_default_args:

tests/python/test_spark/test_spark_local.py

@@ -1126,3 +1126,7 @@ def test_early_stop_param_validation(self):
         classifier = SparkXGBClassifier(early_stopping_rounds=1)
         with pytest.raises(ValueError, match="early_stopping_rounds"):
             classifier.fit(self.cls_df_train)
+
+    def test_unsupported_params(self):
+        with pytest.raises(ValueError, match="evals_result"):
+            SparkXGBClassifier(evals_result={})