BUG: columns name retention in groupby methods (pandas-dev#41497)

Author: jbrockmendel

doc/source/whatsnew/v1.3.0.rst

@@ -1143,6 +1143,8 @@ Groupby/resample/rolling
 - Bug in :class:`DataFrameGroupBy` aggregations incorrectly failing to drop columns with invalid dtypes for that aggregation when there are no valid columns (:issue:`41291`)
 - Bug in :meth:`DataFrame.rolling.__iter__` where ``on`` was not assigned to the index of the resulting objects (:issue:`40373`)
 - Bug in :meth:`.DataFrameGroupBy.transform` and :meth:`.DataFrameGroupBy.agg` with ``engine="numba"`` where ``*args`` were being cached with the user passed function (:issue:`41647`)
+- Bug in :class:`DataFrameGroupBy` methods ``agg``, ``transform``, ``sum``, ``bfill``, ``ffill``, ``pad``, ``pct_change``, ``shift``, ``ohlc`` dropping ``.columns.names`` (:issue:`41497`)
+
 
 Reshaping
 ^^^^^^^^^

pandas/core/apply.py

@@ -348,6 +348,7 @@ def agg_list_like(self) -> FrameOrSeriesUnion:
 
         # multiples
         else:
+            indices = []
             for index, col in enumerate(selected_obj):
                 colg = obj._gotitem(col, ndim=1, subset=selected_obj.iloc[:, index])
                 try:
@@ -369,7 +370,9 @@ def agg_list_like(self) -> FrameOrSeriesUnion:
                         raise
                 else:
                     results.append(new_res)
-                    keys.append(col)
+                    indices.append(index)
+
+            keys = selected_obj.columns.take(indices)
 
         # if we are empty
         if not len(results):
@@ -407,6 +410,7 @@ def agg_dict_like(self) -> FrameOrSeriesUnion:
         -------
         Result of aggregation.
         """
+        from pandas import Index
         from pandas.core.reshape.concat import concat
 
         obj = self.obj
@@ -443,8 +447,18 @@ def agg_dict_like(self) -> FrameOrSeriesUnion:
             keys_to_use = [k for k in keys if not results[k].empty]
             # Have to check, if at least one DataFrame is not empty.
             keys_to_use = keys_to_use if keys_to_use != [] else keys
+            if selected_obj.ndim == 2:
+                # keys are columns, so we can preserve names
+                ktu = Index(keys_to_use)
+                ktu._set_names(selected_obj.columns.names)
+                # Incompatible types in assignment (expression has type "Index",
+                # variable has type "List[Hashable]")
+                keys_to_use = ktu  # type: ignore[assignment]
+
             axis = 0 if isinstance(obj, ABCSeries) else 1
-            result = concat({k: results[k] for k in keys_to_use}, axis=axis)
+            result = concat(
+                {k: results[k] for k in keys_to_use}, axis=axis, keys=keys_to_use
+            )
         elif any(is_ndframe):
             # There is a mix of NDFrames and scalars
             raise ValueError(

pandas/core/groupby/generic.py

@@ -1020,13 +1020,15 @@ def aggregate(self, func=None, *args, engine=None, engine_kwargs=None, **kwargs)
 
                     if isinstance(sobj, Series):
                         # GH#35246 test_groupby_as_index_select_column_sum_empty_df
-                        result.columns = [sobj.name]
+                        result.columns = self._obj_with_exclusions.columns.copy()
                     else:
+                        # Retain our column names
+                        result.columns._set_names(
+                            sobj.columns.names, level=list(range(sobj.columns.nlevels))
+                        )
                         # select everything except for the last level, which is the one
                         # containing the name of the function(s), see GH#32040
-                        result.columns = result.columns.rename(
-                            [sobj.columns.name] * result.columns.nlevels
-                        ).droplevel(-1)
+                        result.columns = result.columns.droplevel(-1)
 
         if not self.as_index:
             self._insert_inaxis_grouper_inplace(result)
@@ -1665,7 +1667,7 @@ def _wrap_transformed_output(
             result.columns = self.obj.columns
         else:
             columns = Index(key.label for key in output)
-            columns.name = self.obj.columns.name
+            columns._set_names(self.obj._get_axis(1 - self.axis).names)
             result.columns = columns
 
         result.index = self.obj.index
@@ -1800,7 +1802,6 @@ def nunique(self, dropna: bool = True) -> DataFrame:
         results = self._apply_to_column_groupbys(
             lambda sgb: sgb.nunique(dropna), obj=obj
         )
-        results.columns.names = obj.columns.names  # TODO: do at higher level?
 
         if not self.as_index:
             results.index = Index(range(len(results)))

pandas/core/reshape/concat.py

@@ -362,8 +362,13 @@ def __init__(
                 clean_keys.append(k)
                 clean_objs.append(v)
             objs = clean_objs
-            name = getattr(keys, "name", None)
-            keys = Index(clean_keys, name=name)
+
+            if isinstance(keys, MultiIndex):
+                # TODO: retain levels?
+                keys = type(keys).from_tuples(clean_keys, names=keys.names)
+            else:
+                name = getattr(keys, "name", None)
+                keys = Index(clean_keys, name=name)
 
         if len(objs) == 0:
             raise ValueError("All objects passed were None")

pandas/tests/groupby/aggregate/test_aggregate.py

@@ -300,13 +300,13 @@ def test_agg_multiple_functions_same_name_with_ohlc_present():
     # ohlc expands dimensions, so different test to the above is required.
     df = DataFrame(
         np.random.randn(1000, 3),
-        index=pd.date_range("1/1/2012", freq="S", periods=1000),
-        columns=["A", "B", "C"],
+        index=pd.date_range("1/1/2012", freq="S", periods=1000, name="dti"),
+        columns=Index(["A", "B", "C"], name="alpha"),
     )
     result = df.resample("3T").agg(
         {"A": ["ohlc", partial(np.quantile, q=0.9999), partial(np.quantile, q=0.1111)]}
     )
-    expected_index = pd.date_range("1/1/2012", freq="3T", periods=6)
+    expected_index = pd.date_range("1/1/2012", freq="3T", periods=6, name="dti")
     expected_columns = MultiIndex.from_tuples(
         [
             ("A", "ohlc", "open"),
@@ -315,7 +315,8 @@ def test_agg_multiple_functions_same_name_with_ohlc_present():
             ("A", "ohlc", "close"),
             ("A", "quantile", "A"),
             ("A", "quantile", "A"),
-        ]
+        ],
+        names=["alpha", None, None],
     )
     non_ohlc_expected_values = np.array(
         [df.resample("3T").A.quantile(q=q).values for q in [0.9999, 0.1111]]
@@ -901,14 +902,20 @@ def test_grouby_agg_loses_results_with_as_index_false_relabel_multiindex():
 def test_multiindex_custom_func(func):
     # GH 31777
     data = [[1, 4, 2], [5, 7, 1]]
-    df = DataFrame(data, columns=MultiIndex.from_arrays([[1, 1, 2], [3, 4, 3]]))
+    df = DataFrame(
+        data,
+        columns=MultiIndex.from_arrays(
+            [[1, 1, 2], [3, 4, 3]], names=["Sisko", "Janeway"]
+        ),
+    )
     result = df.groupby(np.array([0, 1])).agg(func)
     expected_dict = {
         (1, 3): {0: 1.0, 1: 5.0},
         (1, 4): {0: 4.0, 1: 7.0},
         (2, 3): {0: 2.0, 1: 1.0},
     }
     expected = DataFrame(expected_dict)
+    expected.columns = df.columns
     tm.assert_frame_equal(result, expected)
 
 

pandas/tests/groupby/test_groupby.py

@@ -637,10 +637,11 @@ def test_as_index_select_column():
 
 def test_groupby_as_index_select_column_sum_empty_df():
     # GH 35246
-    df = DataFrame(columns=["A", "B", "C"])
+    df = DataFrame(columns=Index(["A", "B", "C"], name="alpha"))
     left = df.groupby(by="A", as_index=False)["B"].sum(numeric_only=False)
-    assert type(left) is DataFrame
-    assert left.to_dict() == {"A": {}, "B": {}}
+
+    expected = DataFrame(columns=df.columns[:2], index=range(0))
+    tm.assert_frame_equal(left, expected)
 
 
 def test_groupby_as_index_agg(df):
@@ -1944,8 +1945,8 @@ def test_groupby_agg_ohlc_non_first():
     # GH 21716
     df = DataFrame(
         [[1], [1]],
-        columns=["foo"],
-        index=date_range("2018-01-01", periods=2, freq="D"),
+        columns=Index(["foo"], name="mycols"),
+        index=date_range("2018-01-01", periods=2, freq="D", name="dti"),
     )
 
     expected = DataFrame(
@@ -1957,9 +1958,10 @@ def test_groupby_agg_ohlc_non_first():
                 ("foo", "ohlc", "high"),
                 ("foo", "ohlc", "low"),
                 ("foo", "ohlc", "close"),
-            )
+            ),
+            names=["mycols", None, None],
         ),
-        index=date_range("2018-01-01", periods=2, freq="D"),
+        index=date_range("2018-01-01", periods=2, freq="D", name="dti"),
     )
 
     result = df.groupby(Grouper(freq="D")).agg(["sum", "ohlc"])
@@ -2131,7 +2133,11 @@ def test_groupby_duplicate_index():
 
 
 @pytest.mark.parametrize(
-    "idx", [Index(["a", "a"]), MultiIndex.from_tuples((("a", "a"), ("a", "a")))]
+    "idx",
+    [
+        Index(["a", "a"], name="foo"),
+        MultiIndex.from_tuples((("a", "a"), ("a", "a")), names=["foo", "bar"]),
+    ],
 )
 @pytest.mark.filterwarnings("ignore:tshift is deprecated:FutureWarning")
 def test_dup_labels_output_shape(groupby_func, idx):