Remove nan-likes from category header (#1037)

* Remove nan-likes from category header

Pandas does not accept None/nan as a category (note: of course
it does allow nan-values in the data itself). However outside source
(i.e. ARFF files) do allow nan as a category, so we must filter these.

* Test output of _unpack_categories

Author: PGijsbers

openml/datasets/dataset.py

@@ -639,6 +639,11 @@ def _encode_if_category(column):
 
     @staticmethod
     def _unpack_categories(series, categories):
+        # nan-likes can not be explicitly specified as a category
+        def valid_category(cat):
+            return isinstance(cat, str) or (cat is not None and not np.isnan(cat))
+
+        filtered_categories = [c for c in categories if valid_category(c)]
         col = []
         for x in series:
             try:
@@ -647,7 +652,7 @@ def _unpack_categories(series, categories):
                 col.append(np.nan)
         # We require two lines to create a series of categories as detailed here:
         # https://pandas.pydata.org/pandas-docs/version/0.24/user_guide/categorical.html#series-creation  # noqa E501
-        raw_cat = pd.Categorical(col, ordered=True, categories=categories)
+        raw_cat = pd.Categorical(col, ordered=True, categories=filtered_categories)
         return pd.Series(raw_cat, index=series.index, name=series.name)
 
     def get_data(

tests/test_datasets/test_dataset.py

@@ -50,6 +50,17 @@ def test_init_string_validation(self):
                 name="somename", description="a description", citation="Something by Müller"
             )
 
+    def test__unpack_categories_with_nan_likes(self):
+        # unpack_categories decodes numeric categorical values according to the header
+        # Containing a 'non' category in the header shouldn't lead to failure.
+        categories = ["a", "b", None, float("nan"), np.nan]
+        series = pd.Series([0, 1, None, float("nan"), np.nan, 1, 0])
+        clean_series = OpenMLDataset._unpack_categories(series, categories)
+
+        expected_values = ["a", "b", np.nan, np.nan, np.nan, "b", "a"]
+        self.assertListEqual(list(clean_series.values), expected_values)
+        self.assertListEqual(list(clean_series.cat.categories.values), list("ab"))
+
     def test_get_data_array(self):
         # Basic usage
         rval, _, categorical, attribute_names = self.dataset.get_data(dataset_format="array")