FIX handle full NaT columns properly in Random*Sampler (#1059)

Author: glemaitre

doc/whats_new/v0.12.rst

@@ -18,6 +18,11 @@ Bug fixes
   the number of samples in the minority class.
   :pr:`1012` by :user:`Guillaume Lemaitre <glemaitre>`.
 
+- Fix a bug in :class:`~imblearn.under_sampling.RandomUnderSampler` and
+  :class:`~imblearn.over_sampling.RandomOverSampler` where a column containing only
+  NaT was not handled correctly.
+  :pr:`1059` by :user:`Guillaume Lemaitre <glemaitre>`.
+
 Compatibility
 .............
 

imblearn/over_sampling/tests/test_random_over_sampler.py

@@ -287,3 +287,26 @@ def test_random_over_sampling_datetime():
     pd.testing.assert_series_equal(X_res.dtypes, X.dtypes)
     pd.testing.assert_index_equal(X_res.index, y_res.index)
     assert_array_equal(y_res.to_numpy(), np.array([0, 0, 0, 1, 1, 1]))
+
+
+def test_random_over_sampler_full_nat():
+    """Check that we can return timedelta columns full of NaT.
+
+    Non-regression test for:
+    https://github.com/scikit-learn-contrib/imbalanced-learn/issues/1055
+    """
+    pd = pytest.importorskip("pandas")
+
+    X = pd.DataFrame(
+        {
+            "col_str": ["abc", "def", "xyz"],
+            "col_timedelta": pd.to_timedelta([np.nan, np.nan, np.nan]),
+        }
+    )
+    y = np.array([0, 0, 1])
+
+    X_res, y_res = RandomOverSampler().fit_resample(X, y)
+    assert X_res.shape == (4, 2)
+    assert y_res.shape == (4,)
+
+    assert X_res["col_timedelta"].dtype == "timedelta64[ns]"

imblearn/under_sampling/_prototype_selection/tests/test_random_under_sampler.py

@@ -162,3 +162,26 @@ def test_random_under_sampling_datetime():
     pd.testing.assert_series_equal(X_res.dtypes, X.dtypes)
     pd.testing.assert_index_equal(X_res.index, y_res.index)
     assert_array_equal(y_res.to_numpy(), np.array([0, 1]))
+
+
+def test_random_under_sampler_full_nat():
+    """Check that we can return timedelta columns full of NaT.
+
+    Non-regression test for:
+    https://github.com/scikit-learn-contrib/imbalanced-learn/issues/1055
+    """
+    pd = pytest.importorskip("pandas")
+
+    X = pd.DataFrame(
+        {
+            "col_str": ["abc", "def", "xyz"],
+            "col_timedelta": pd.to_timedelta([np.nan, np.nan, np.nan]),
+        }
+    )
+    y = np.array([0, 0, 1])
+
+    X_res, y_res = RandomUnderSampler().fit_resample(X, y)
+    assert X_res.shape == (2, 2)
+    assert y_res.shape == (2,)
+
+    assert X_res["col_timedelta"].dtype == "timedelta64[ns]"

imblearn/utils/_validation.py

@@ -66,7 +66,24 @@ def _transfrom_one(self, array, props):
                 ret = pd.DataFrame.sparse.from_spmatrix(array, columns=props["columns"])
             else:
                 ret = pd.DataFrame(array, columns=props["columns"])
-            ret = ret.astype(props["dtypes"])
+
+            try:
+                ret = ret.astype(props["dtypes"])
+            except TypeError:
+                # We special case the following error:
+                # https://github.com/scikit-learn-contrib/imbalanced-learn/issues/1055
+                # There is no easy way to have a generic workaround. Here, we detect
+                # that we have a column with only null values that is datetime64
+                # (resulting from the np.vstack of the resampling).
+                for col in ret.columns:
+                    if (
+                        ret[col].isnull().all()
+                        and ret[col].dtype == "datetime64[ns]"
+                        and props["dtypes"][col] == "timedelta64[ns]"
+                    ):
+                        ret[col] = pd.to_timedelta(["NaT"] * len(ret[col]))
+                # try again
+                ret = ret.astype(props["dtypes"])
         elif type_ == "series":
             import pandas as pd