FIX use np.nan instead of None for missing marker in fetch_openml (scikit-learn#26579)

Author: glemaitre

doc/whats_new/v1.3.rst

@@ -283,6 +283,10 @@ Changelog
   the pandas parser. The parameter `read_csv_kwargs` allows to overwrite this behaviour.
   :pr:`26551` by :user:`Guillaume Lemaitre <glemaitre>`.
 
+- |Fix| :func:`dataasets.fetch_openml` will consistenly use `np.nan` as missing marker
+  with both parsers `"pandas"` and `"liac-arff"`.
+  :pr:`26579` by :user:`Guillaume Lemaitre <glemaitre>`.
+
 - |Enhancement| Allows to overwrite the parameters used to open the ARFF file using
   the parameter `read_csv_kwargs` in :func:`datasets.fetch_openml` when using the
   pandas parser.

sklearn/datasets/_arff_parser.py

@@ -204,7 +204,10 @@ def _io_to_generator(gzip_file):
         if len(dfs) >= 2:
             dfs[0] = dfs[0].astype(dfs[1].dtypes)
 
-        frame = pd.concat(dfs, ignore_index=True)
+        # liac-arff parser does not depend on NumPy and uses None to represent
+        # missing values. To be consistent with the pandas parser, we replace
+        # None with np.nan.
+        frame = pd.concat(dfs, ignore_index=True).fillna(value=np.nan)
         del dfs, first_df
 
         # cast the columns frame

sklearn/datasets/tests/test_openml.py

@@ -920,9 +920,7 @@ def datasets_missing_values():
         (1119, "liac-arff", 9, 6, 0),
         (1119, "pandas", 9, 0, 6),
         # miceprotein
-        # 1 column has only missing values with object dtype
-        (40966, "liac-arff", 1, 76, 0),
-        # with casting it will be transformed to either float or Int64
+        (40966, "liac-arff", 1, 77, 0),
         (40966, "pandas", 1, 77, 0),
         # titanic
         (40945, "liac-arff", 3, 6, 0),