MAINT: Get tests passing with XGBoost 2.0 in test suite (#3257)

* Use "Exact" XGBoost tree method

* Cast feature_names to list for XGBoost

* Use exact tree method

* Remove xgboost version pin

* Fix base_score parameter

* Define base_score in basic xgboost scenario

* Increase tolerance for xgboost margin test

pyproject.toml

@@ -58,7 +58,7 @@ test = [
   "pytest",
   "pytest-mpl",
   "pytest-cov",
-  "xgboost==1.7.6",
+  "xgboost",
   "lightgbm",
   "catboost",
   "gpboost",

tests/explainers/common.py

@@ -19,7 +19,8 @@ def basic_xgboost_scenario(max_samples=None, dataset=shap.datasets.adult):
     X = X.values
 
     # train an XGBoost model (but any other model type would also work)
-    model = xgboost.XGBClassifier()
+    # Specify some hyperparameters for consitency between xgboost v1.X and v2.X
+    model = xgboost.XGBClassifier(tree_method="exact", base_score=0.5)
     model.fit(X, y)
 
     return model, X

tests/explainers/test_tree.py

@@ -463,7 +463,7 @@ def test_provided_background_tree_path_dependent():
     xgboost = pytest.importorskip("xgboost")
 
     X, y = shap.datasets.adult(n_points=100)
-    dtrain = xgboost.DMatrix(X, label=y, feature_names=X.columns)
+    dtrain = xgboost.DMatrix(X, label=y, feature_names=list(X.columns))
 
     params = {
         "booster": "gbtree",
@@ -1285,7 +1285,9 @@ def test_xgboost_classifier_independent_margin(self):
         y = y + abs(min(y))
         y = rs.binomial(n=1, p=y / max(y))
 
-        model = xgboost.XGBClassifier(n_estimators=10, max_depth=5, random_state=random_seed)
+        model = xgboost.XGBClassifier(
+            n_estimators=10, max_depth=5, random_state=random_seed, tree_method="exact"
+        )
         model.fit(X, y)
         predicted = model.predict(X, output_margin=True)
 
@@ -1305,6 +1307,7 @@ def test_xgboost_classifier_independent_margin(self):
         assert np.allclose(
             explanation.values.sum(1) + explanation.base_values,
             predicted,
+            atol=1e-7,
         )
 
     def test_xgboost_classifier_independent_probability(self, random_seed):

tests/plots/conftest.py

@@ -22,7 +22,7 @@ def explainer():
     y = y[:100]
 
     # train an XGBoost model (but any other model type would also work)
-    model = xgboost.XGBClassifier(random_state=0).fit(X, y)
+    model = xgboost.XGBClassifier(random_state=0, tree_method="exact", base_score=0.5).fit(X, y)
 
     # build an Exact explainer and explain the model predictions on the given dataset
     return shap.TreeExplainer(model, X)