BUG: correctly clean up exclude attribute

python/interpret-core/interpret/glassbox/_ebm/_merge_ebms.py

@@ -17,6 +17,7 @@
     EBMModel,
     ExplainableBoostingClassifier,
     ExplainableBoostingRegressor,
+    _clean_exclude,
     is_private,
 )
 from ._utils import (
@@ -450,14 +451,16 @@ def _initialize_ebm(models: List[EBMModel], ebm_type=EBMModel) -> EBMModel:
         # none of the models contains all feature_idxs
         # merged EBM should exclude features included by none of the models
         # -> overlap of all features
-        # following algorithm works only if all models use the same feature names
-        # interactions are probably not handled correctly
-        excluded = {(feat,) for feat in models[0].feature_names_in_}
+        clean_excludes = []
         for model in models:
             if model.exclude == "mains":
-                excluded &= set(models[0].feature_names_in_)
+                clean_excludes.append({(idx,) for idx in range(model.n_features_in_)})
                 continue
-            excluded &= set(model.exclude)
+            feature_map = {
+                name: idx for idx, name in enumerate(model.feature_names_in_)
+            }
+            clean_excludes.append(_clean_exclude(model.exclude, feature_map))
+        excluded = set.intersection(*clean_excludes)
         manual_kdws["exclude"] = list(excluded) if excluded else None
 
     # handle `interactions`
@@ -498,7 +501,6 @@ def monotone(args) -> int:
             for item in zip(*(model.monotone_constraints for model in models))
         ]
 
-    # TODO: treat special cases: exclude, interactions
     for key in kdws:
         values = np.array([getattr(model, key, np.nan) for model in models])
         nan_weight = np.copy(weights)
@@ -544,6 +546,7 @@ def merge_ebms(models):
     if any(not getattr(model, "has_fitted_", False) for model in models):
         msg = "All models must be fitted."
         raise Exception(msg)
+
     ebm = _initialize_ebm(models, ebm_type=ebm_type)
     ebm.has_fitted_ = True
 

python/interpret-core/tests/glassbox/ebm/test_merge_ebms.py

@@ -241,3 +241,54 @@ def test_unfitted():
     # ebm2 is not fitted
     with pytest.raises(Exception, match="All models must be fitted."):
         merge_ebms([ebm1, ebm2])
+
+
+@pytest.mark.skip(reason="Extremely slow")
+def test_merge_monotone():
+    """Check merging of features with `monotone_constraints`."""
+    X, y, names, _ = make_synthetic(classes=2, missing=True, output_type="str")
+    TestEBM = partial(
+        ExplainableBoostingClassifier,
+        feature_names=names,
+        random_state=42,
+        **_fast_kwds,
+    )
+    ebm1 = TestEBM(monotone_constraints=[+0, +0, +0, +1, +1, -1, 0, 0, 0, 0])
+    ebm1.fit(X, y)
+    ebm2 = TestEBM(monotone_constraints=[+0, +1, -1, +1, -1, -1, 0, 0, 0, 0])
+    ebm2.fit(X, y)
+    merged_ebm = merge_ebms([ebm1, ebm2])
+    assert merged_ebm.monotone_constraints == [+0, +0, +0, +1, 0, -1, 0, 0, 0, 0]
+    ebm3 = TestEBM(monotone_constraints=None)
+    ebm3.fit(X, y)
+    merged_ebm = merge_ebms([ebm1, ebm2, ebm3])
+    assert merged_ebm.monotone_constraints is None
+
+
+def test_merge_exclude():
+    """Check merging of features with `exclude`."""
+    X, y, names, _ = make_synthetic(classes=2, missing=True, output_type="str")
+    TestEBM = partial(
+        ExplainableBoostingClassifier,
+        feature_names=names,
+        random_state=42,
+        **_fast_kwds,
+    )
+    ebm1 = TestEBM(exclude=None)
+    ebm1.fit(X, y)
+    ebm2 = TestEBM(exclude=[0, 1, 2])
+    ebm2.fit(X, y)
+    merged_ebm = merge_ebms([ebm1, ebm2])
+    assert merged_ebm.exclude is None
+    ebm1 = TestEBM(exclude=[0, 2])
+    ebm1.fit(X, y)
+    ebm2 = TestEBM(exclude=[0, 1, 2])
+    ebm2.fit(X, y)
+    merged_ebm = merge_ebms([ebm1, ebm2])
+    assert merged_ebm.exclude == [(0,), (2,)]
+    ebm1 = TestEBM(exclude="mains")
+    ebm1.fit(X, y)
+    ebm2 = TestEBM(exclude=[0, 1, 2])
+    ebm2.fit(X, y)
+    merged_ebm = merge_ebms([ebm1, ebm2])
+    assert merged_ebm.exclude == [(0,), (1,), (2,)]