FIX Add validation for FeatureUnion transformer outputs (scikit-learn#31318) (scikit-learn#31559)

Co-authored-by: Jérémie du Boisberranger <jeremie@probabl.ai>

Author: gguiomar

doc/whats_new/upcoming_changes/sklearn.pipeline/31559.fix.rst

@@ -0,0 +1,4 @@
+- :class:`pipeline.FeatureUnion` now validates that all transformers return 2D outputs
+  and raises an informative error when transformers return 1D outputs, preventing
+  silent failures that previously produced meaningless concatenated results.
+  By :user:`gguiomar <gguiomar>`.

sklearn/pipeline.py

@@ -2038,15 +2038,23 @@ def transform(self, X, **params):
         return self._hstack(Xs)
 
     def _hstack(self, Xs):
+        # Check if Xs dimensions are valid
+        for X, (name, _) in zip(Xs, self.transformer_list):
+            if hasattr(X, "shape") and len(X.shape) != 2:
+                raise ValueError(
+                    f"Transformer '{name}' returned an array or dataframe with "
+                    f"{len(X.shape)} dimensions, but expected 2 dimensions "
+                    "(n_samples, n_features)."
+                )
+
         adapter = _get_container_adapter("transform", self)
         if adapter and all(adapter.is_supported_container(X) for X in Xs):
             return adapter.hstack(Xs)
 
         if any(sparse.issparse(f) for f in Xs):
-            Xs = sparse.hstack(Xs).tocsr()
-        else:
-            Xs = np.hstack(Xs)
-        return Xs
+            return sparse.hstack(Xs).tocsr()
+
+        return np.hstack(Xs)
 
     def _update_transformer_list(self, transformers):
         transformers = iter(transformers)

sklearn/tests/test_pipeline.py

@@ -1900,6 +1900,22 @@ def test_feature_union_feature_names_in_():
     assert not hasattr(union, "feature_names_in_")
 
 
+def test_feature_union_1d_output():
+    """Test that FeatureUnion raises error for 1D transformer outputs."""
+    X = np.arange(6).reshape(3, 2)
+
+    with pytest.raises(
+        ValueError,
+        match="Transformer 'b' returned an array or dataframe with 1 dimensions",
+    ):
+        FeatureUnion(
+            [
+                ("a", FunctionTransformer(lambda X: X)),
+                ("b", FunctionTransformer(lambda X: X[:, 1])),
+            ]
+        ).fit_transform(X)
+
+
 # transform_input tests
 # =====================