Fix mixed categorical sorting crashes (issue #273) (#336)

Author: oegedijk

RELEASE_NOTES.md

@@ -8,6 +8,8 @@
 - Improve compatibility with AutoGluon/custom wrappers by coercing pandas `DataFrame` outputs from `predict_proba`/`predict` to numpy arrays before indexing in classifier/regression helper paths.
 - Harden one-vs-all scorer handling so `make_one_vs_all_scorer` also accepts classifiers whose `predict_proba` returns a pandas `DataFrame`.
 - Fix ExplainerHub `add_dashboard_route` after first request by allowing dynamic dashboard registration/setup during route-triggered add, and add a regression test for issue #269.
+- Fix issue #273: avoid crashes when sorting categorical values containing mixed types and NaNs across explainer setup, PDP/category ordering, and categorical plot paths.
+- Add regression tests for mixed-type categorical sorting and mixed target-label sorting fallbacks in explainers and plots.
 
 ## Version 0.5.6:
 

TODO.md

@@ -6,10 +6,9 @@
 - Rules: link an issue when possible; include size S/M/L; mark blockers.
 
 **Now**
-- [M][Explainers][#273] categorical columns with NaNs: sorting and column preservation.
+- [M/L][Components][#262] add filters for random transaction selection in whatif tab.
 
 **Next**
-- [M/L][Components][#262] add filters for random transaction selection in whatif tab.
 - [S][Methods][#220] get_contrib_df accepts list/array input.
 - [M][Components][#176] FeatureInputComponent hide parameters.
 - [M][Explainers][#198] LightGBM string categorical handling across SHAP/plots.

explainerdashboard/explainer_methods.py

@@ -4,6 +4,7 @@
     "safe_isinstance",
     "unwrap_calibrated_classifier",
     "align_categorical_dtypes",
+    "sorted_categorical_values",
     "guess_shap",
     "mape_score",
     "parse_cats",
@@ -341,6 +342,32 @@ def align_categorical_dtypes(
     return aligned
 
 
+def sorted_categorical_values(values):
+    """Sort categorical values safely when types are mixed.
+
+    Keeps original values but orders deterministically:
+    booleans, then numbers, then other values by type/value string.
+    """
+
+    def _is_na(value):
+        try:
+            return bool(pd.isna(value))
+        except Exception:
+            return False
+
+    def _sort_key(value):
+        if isinstance(value, np.generic):
+            value = value.item()
+        if isinstance(value, bool):
+            return (0, int(value))
+        if isinstance(value, (int, float)) and not _is_na(value):
+            return (1, float(value))
+        return (2, type(value).__name__, str(value))
+
+    clean_values = [value for value in values if not _is_na(value)]
+    return sorted(clean_values, key=_sort_key)
+
+
 def guess_shap(model):
     """guesses which SHAP explainer to use for a particular model, based
     on str(type(model)). Returns 'tree' for tree based models such as
@@ -1171,7 +1198,9 @@ def _model_input(data):
     if grid_values is None:
         if isinstance(feature, str):
             if not is_numeric_dtype(X_sample[feature]):
-                grid_values = sorted(X_sample[feature].unique().tolist())
+                grid_values = sorted_categorical_values(
+                    X_sample[feature].unique().tolist()
+                )
             else:
                 grid_values = get_grid_points(
                     X_sample[feature],
@@ -1752,7 +1781,7 @@ def node_pred_proba(node):
                     return class_counts[pos_label] / total
 
                 # If pos_label not found, try to map it to available class keys
-                available_classes = sorted(class_counts.keys())
+                available_classes = sorted_categorical_values(class_counts.keys())
                 if len(available_classes) == 0:
                     return 0.0
 

explainerdashboard/explainer_plots.py

@@ -31,7 +31,11 @@
 import plotly.graph_objects as go
 from plotly.subplots import make_subplots
 
-from .explainer_methods import matching_cols, safe_isinstance
+from .explainer_methods import (
+    matching_cols,
+    safe_isinstance,
+    sorted_categorical_values,
+)
 
 
 def plotly_prediction_piechart(predictions_df, showlegend=True, size=250):
@@ -1040,7 +1044,7 @@ def plotly_shap_violin_plot(
     ), f"{X_col.name} is not categorical! Can only plot violin plots for categorical features!"
 
     if cats_order is None:
-        cats_order = sorted(X_col.unique().tolist())
+        cats_order = sorted_categorical_values(X_col.unique().tolist())
 
     n_cats = len(cats_order)
 
@@ -1098,7 +1102,11 @@ def plotly_shap_violin_plot(
 
     for i, cat in enumerate(cats_order):
         col = 1 + i * 2 if points or X_color_col is not None else 1 + i
-        if cat.startswith(X_col.name + "_"):
+        if (
+            isinstance(cat, str)
+            and isinstance(X_col.name, str)
+            and cat.startswith(X_col.name + "_")
+        ):
             cat_name = cat[len(X_col.name) + 1 :]
         else:
             cat_name = cat
@@ -2347,7 +2355,7 @@ def plotly_residuals_vs_col(
 
     if not is_numeric_dtype(col):
         if cats_order is None:
-            cats_order = sorted(col.unique().tolist())
+            cats_order = sorted_categorical_values(col.unique().tolist())
         n_cats = len(cats_order)
 
         if points:
@@ -2506,7 +2514,7 @@ def plotly_actual_vs_col(
 
     if not is_numeric_dtype(col):
         if cats_order is None:
-            cats_order = sorted(col.unique().tolist())
+            cats_order = sorted_categorical_values(col.unique().tolist())
         n_cats = len(cats_order)
 
         if points:
@@ -2658,7 +2666,7 @@ def plotly_preds_vs_col(
 
     if not is_numeric_dtype(col):
         if cats_order is None:
-            cats_order = sorted(col.unique().tolist())
+            cats_order = sorted_categorical_values(col.unique().tolist())
         n_cats = len(cats_order)
 
         if points:

explainerdashboard/explainers.py

@@ -275,7 +275,7 @@ def __init__(
             col for col in self.regular_cols if not is_numeric_dtype(self.X[col])
         ]
         self.categorical_dict = {
-            col: sorted(self.X[col].dropna().unique().tolist())
+            col: sorted_categorical_values(self.X[col].unique().tolist())
             for col in self.categorical_cols
         }
         # Include an explicit NaN option for categorical columns with missing values.
@@ -1010,7 +1010,7 @@ def ordered_cats(self, col, topx=None, sort="alphabet", pos_label=None):
                 values = self.categorical_dict[col]
             else:
                 values = [v for v in X[col].unique().tolist() if not pd.isna(v)]
-                values = sorted(values)
+                values = sorted_categorical_values(values)
             if topx is None:
                 return values
             else:
@@ -2695,7 +2695,9 @@ def __init__(
         if hasattr(self.model, "classes_"):
             class_values = list(self.model.classes_)
         elif not self.y_missing:
-            class_values = sorted(pd.Series(self.y).dropna().unique().tolist())
+            class_values = sorted_categorical_values(
+                pd.Series(self.y).dropna().unique().tolist()
+            )
 
         if not self.y_missing:
             if class_values is None:

tests/test_categorical_sorting.py

@@ -0,0 +1,78 @@
+import numpy as np
+import pandas as pd
+import pytest
+
+from explainerdashboard import ClassifierExplainer
+from explainerdashboard.explainer_plots import (
+    plotly_actual_vs_col,
+    plotly_preds_vs_col,
+    plotly_residuals_vs_col,
+    plotly_shap_violin_plot,
+)
+
+
+class NoClassesModel:
+    """Classifier-like model without classes_ attribute."""
+
+    def predict_proba(self, X):
+        n = len(X)
+        p = np.full(n, 0.6)
+        return np.c_[1 - p, p]
+
+    def predict(self, X):
+        return (self.predict_proba(X)[:, 1] > 0.5).astype(int)
+
+
+@pytest.mark.parametrize(
+    "plot_fn,args",
+    [
+        (
+            plotly_shap_violin_plot,
+            (
+                pd.Series(["a", 1, np.nan, "b"], name="mixed_cat", dtype=object),
+                np.array([0.1, -0.2, 0.0, 0.3]),
+            ),
+        ),
+        (
+            plotly_residuals_vs_col,
+            (
+                np.array([1.0, 2.0, 3.0, 4.0]),
+                np.array([1.1, 1.9, 2.8, 4.2]),
+                pd.Series(["a", 1, np.nan, "b"], name="mixed_cat", dtype=object),
+            ),
+        ),
+        (
+            plotly_actual_vs_col,
+            (
+                np.array([1.0, 2.0, 3.0, 4.0]),
+                np.array([1.1, 1.9, 2.8, 4.2]),
+                pd.Series(["a", 1, np.nan, "b"], name="mixed_cat", dtype=object),
+            ),
+        ),
+        (
+            plotly_preds_vs_col,
+            (
+                np.array([1.0, 2.0, 3.0, 4.0]),
+                np.array([1.1, 1.9, 2.8, 4.2]),
+                pd.Series(["a", 1, np.nan, "b"], name="mixed_cat", dtype=object),
+            ),
+        ),
+    ],
+)
+def test_plot_functions_do_not_crash_on_mixed_categorical_values(plot_fn, args):
+    try:
+        plot_fn(*args)
+    except TypeError as exc:
+        pytest.fail(f"Mixed categorical sorting should not crash: {exc}")
+
+
+def test_classifier_explainer_without_classes_handles_mixed_target_labels():
+    X = pd.DataFrame({"a": [1, 2, 3, 4], "b": [0, 1, 0, 1]})
+    y = pd.Series([0, "1", 0, "1"], dtype=object)
+
+    try:
+        explainer = ClassifierExplainer(NoClassesModel(), X, y, shap="kernel")
+    except TypeError as exc:
+        pytest.fail(f"Mixed target-label sorting should not crash: {exc}")
+
+    assert len(explainer.labels) == 2

tests/test_datasets.py

@@ -22,6 +22,22 @@ def predict_proba(self, X):
         return self._model.predict_proba(X)
 
 
+class MixedCategoricalModelWrapper:
+    def __init__(self, model) -> None:
+        self._model = model
+
+    def _preprocessor(self, X):
+        return X.drop(["MixedCat"], axis=1)
+
+    def predict(self, X):
+        X = self._preprocessor(X)
+        return self._model.predict(X)
+
+    def predict_proba(self, X):
+        X = self._preprocessor(X)
+        return self._model.predict_proba(X)
+
+
 @pytest.fixture(scope="module")
 def categorical_nan_explainer(classifier_data):
     X_train, y_train, X_test, y_test = classifier_data
@@ -73,3 +89,39 @@ def test_get_row_from_input_preserves_nan_unmerged(categorical_nan_explainer):
     row["Name"] = "NaN"
     input_row = explainer.get_row_from_input(row.values.tolist())
     assert pd.isna(input_row.loc[0, "Name"])
+
+
+def test_mixed_type_categorical_with_nan_does_not_crash_sorting(classifier_data):
+    X_train, y_train, X_test, y_test = classifier_data
+    X_train = X_train.copy()
+    X_test = X_test.copy()
+
+    X_train["MixedCat"] = pd.Series(
+        ["alpha"] * len(X_train), index=X_train.index, dtype=object
+    )
+    X_test["MixedCat"] = pd.Series(
+        ["alpha"] * len(X_test), index=X_test.index, dtype=object
+    )
+
+    # Simulate mixed-type categorical values from upstream preprocessing artifacts.
+    X_train.loc[X_train.index[:8], "MixedCat"] = 1
+    X_train.loc[X_train.index[8:16], "MixedCat"] = np.nan
+    X_test.loc[X_test.index[:4], "MixedCat"] = 0
+    X_test.loc[X_test.index[4:8], "MixedCat"] = np.nan
+
+    model = RandomForestClassifier(n_estimators=5, max_depth=2, random_state=42)
+    model.fit(X_train.drop(["MixedCat"], axis=1), y_train)
+    wrapper = MixedCategoricalModelWrapper(model)
+
+    try:
+        ClassifierExplainer(
+            wrapper,
+            X_test,
+            y_test,
+            cats=["Sex", "Deck", "Embarked"],
+            shap="kernel",
+        )
+    except TypeError as exc:
+        pytest.fail(
+            f"Categorical sorting should not crash on mixed values + NaN: {exc}"
+        )