Fix bug on force plots to show feature values for specific rows in tr…

…aining data (#3044)

* Pass actual row data from correct index to the shap.force_plot call. Fixes bug where only first row feature values are show in all row plots.

* Lint edit

* Add release notes for bug #3045

* Forgot pr in release note line

* Lint edit - blank line

* Lint adjustment

* Add tests to check force plot feature values match those of the rows to explain. (#3044)

Co-authored-by: Angela Lin <angela97lin@gmail.com>
Co-authored-by: Freddy Boulton <alfonsoboulton@gmail.com>

docs/source/release_notes.rst

@@ -26,6 +26,7 @@ Release Notes
         * Added in error message when fit and predict/predict_proba data types are different :pr:`3036`
         * Fixed bug where ensembling components could not get converted to JSON format :pr:`3049`
         * Fixed bug where components with tuned integer hyperparameters could not get converted to JSON format :pr:`3049`
+        * Fixed bug where force plots were not displaying correct feature values :pr:`3044`
         * Included confusion matrix at the pipeline threshold for ``find_confusion_matrix_per_threshold`` :pr:`3080`
         * Fixed bug where One Hot Encoder would error out if a non-categorical feature had a missing value :pr:`3083`
         * Fixed bug where features created from categorical columns by ``Delayed Feature Transformer`` would be inferred as categorical :pr:`3083`

evalml/model_understanding/force_plots.py

@@ -28,29 +28,22 @@ def graph_force_plot(pipeline, rows_to_explain, training_data, y, matplotlib=Fal
 
     def gen_force_plot(shap_values, training_data, expected_value, matplotlib):
         """Helper function to generate a single force plot."""
-        # Ensure there are as many features as shap values.
-        assert training_data.shape[1] == len(shap_values)
-
-        # TODO: Update this to make the force plot display multi-row array visualizer.
-        training_data_sample = training_data.iloc[0]
         shap_plot = shap.force_plot(
-            expected_value,
-            np.array(shap_values),
-            training_data_sample,
-            matplotlib=matplotlib,
+            expected_value, np.array(shap_values), training_data, matplotlib=matplotlib
         )
         return shap_plot
 
     if jupyter_check():
         initjs()
 
     shap_plots = force_plot(pipeline, rows_to_explain, training_data, y)
-    for row in shap_plots:
+    for ix, row in enumerate(shap_plots):
+        row_id = rows_to_explain[ix]
         for cls in row:
             cls_dict = row[cls]
             cls_dict["plot"] = gen_force_plot(
                 cls_dict["shap_values"],
-                training_data[cls_dict["feature_names"]],
+                training_data[cls_dict["feature_names"]].iloc[row_id],
                 cls_dict["expected_value"],
                 matplotlib=matplotlib,
             )

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_force_plots.py

@@ -17,6 +17,25 @@
 )
 
 
+def validate_plot_feature_values(results, X):
+    """Helper to validate feature values returned from force plots"""
+    for row, result in enumerate(results):
+        for result_label in result:
+            # Check feature values in generated force plot correspond to input rows
+            row_force_plot = result[result_label]["plot"]
+            assert "features" in row_force_plot.data
+            plot_features = row_force_plot.data["features"]
+            feature_vals = [plot_features[k]["value"] for k in plot_features]
+
+            # Features in results depend on effect size; filter feature names
+            effect_features_ix = plot_features.keys()
+            effect_features = [
+                row_force_plot.data["featureNames"][i] for i in effect_features_ix
+            ]
+
+            assert all(feature_vals == X[effect_features].iloc[row].values)
+
+
 def test_exceptions():
     with pytest.raises(
         TypeError,
@@ -113,6 +132,8 @@ def test_force_plot_binary(
                     shap.plots._force.AdditiveForceVisualizer,
                 )
 
+        validate_plot_feature_values(results, X)
+
 
 @pytest.mark.parametrize(
     "rows_to_explain, just_data", product([[0], [0, 1, 2, 3, 4]], [False, True])
@@ -164,6 +185,8 @@ def test_force_plot_multiclass(
                     shap.plots._force.AdditiveForceVisualizer,
                 )
 
+        validate_plot_feature_values(results, X)
+
 
 @pytest.mark.parametrize(
     "rows_to_explain, just_data", product([[0], [0, 1, 2, 3, 4]], [False, True])
@@ -211,6 +234,7 @@ def test_force_plot_regression(
             assert isinstance(
                 result["regression"]["plot"], shap.plots._force.AdditiveForceVisualizer
             )
+        validate_plot_feature_values(results, X)
 
 
 @pytest.mark.parametrize("pipeline_class,estimator", pipeline_test_cases)