fixed issue with binning bool cols and with constant datetime cols

Author: mplatzer

mostlyai/qa/accuracy.py

@@ -1046,7 +1046,7 @@ def _clip(col, bins):
 
         if col.nunique() == 1:
             # ensure 2 breaks for single-valued columns
-            val = col.iloc[0]
+            val = col.dropna().iloc[0]
             upper_limit = [val + np.timedelta64(1, "D")] if not pd.isna(val) else []
             breaks = [val] + upper_limit
         else:
@@ -1115,6 +1115,7 @@ def bin_non_categorical(
 
 
 def bin_categorical(col: pd.Series, bins: int | list[str]) -> tuple[pd.Categorical, list[str]]:
+    col = col.astype("string[pyarrow]")
     col = col.fillna(NA_BIN)
     col = col.replace("", EMPTY_BIN)
     # determine top values, if not provided

tests/unit/test_accuracy.py

@@ -37,6 +37,7 @@
     trim_labels,
     calculate_correlations,
     plot_store_correlation_matrices,
+    bin_categorical,
 )
 from mostlyai.qa.sampling import pull_data_for_accuracy, sample_two_consecutive_rows
 from mostlyai.qa.common import (
@@ -496,6 +497,14 @@ def test_num_col_nans_only(self):
         df_counts = df["nans"].value_counts().to_dict()
         assert df_counts["(n/a)"] == 10
 
+    def test_bin_categorical(self):
+        x = pd.Series(["a", "b"] * 50 + ["x"])
+        col, _ = bin_categorical(x, 5)
+        assert len(col) == 101
+        x = pd.Series([True, False] * 50 + [np.nan] * 100, dtype="object")
+        col, _ = bin_categorical(x, 5)
+        assert len(col) == 200
+
     def test_bin_numeric(self):
         # test several edge cases
         cases = [
@@ -534,6 +543,10 @@ def test_bin_datetime(self):
                 ),
                 ["⪰ 2023-01-30 13:00:00.333000"] * 20,
             ),  # two values
+            (
+                pd.Series([pd.NaT, "2024-11-20"], dtype="datetime64[ns]"),
+                ["(n/a)", "⪰ 2024-Nov-20"],
+            ),  # single value with leading N/A
         ]
 
         for col, expected in cases: