Adding Feature Value column to SHAP table. (#1064)

* Adding Feature Value column to SHAP table.

* Editing release_notes for PR 1064.

* Moving updated prediction explanation tests from pipeline_tests to model_understanding_tests.

* Using string formatting to round and making some minor tweaks to tests.

* Adding comment explaining mocked return value of pipeline._transform

docs/source/release_notes.rst

@@ -10,6 +10,7 @@ Release Notes
         * Added guide on installing with conda :pr:`1041`
         * Standardized error when calling transform/predict before fit for pipelines :pr:`1048`
         * Added `percent_better_than_baseline` to Automl search rankings and full rankings table :pr:`1050`
+        * Added "Feature Value" column to prediction explanation reports. :pr:`1064`
     * Fixes
         * Updated TextFeaturizer component to no longer require an internet connection to run :pr:`1022`
         * Fixed non-deterministic element of TextFeaturizer transformations :pr:`1022`

evalml/model_understanding/prediction_explanations/_algorithms.py

@@ -47,14 +47,13 @@ def _compute_shap_values(pipeline, features, training_data=None):
     if estimator.model_family == ModelFamily.BASELINE:
         raise ValueError("You passed in a baseline pipeline. These are simple enough that SHAP values are not needed.")
 
-    pipeline_features = pipeline._transform(features)
-    feature_names = pipeline_features.columns
+    feature_names = features.columns
 
     # This is to make sure all dtypes are numeric - SHAP algorithms will complain otherwise.
     # Sklearn components do this under-the-hood so we're not changing the data the model was trained on.
     # Catboost can naturally handle string-encoded categorical features so we don't need to convert to numeric.
     if estimator.model_family != ModelFamily.CATBOOST:
-        pipeline_features = check_array(pipeline_features.values)
+        features = check_array(features.values)
 
     if estimator.model_family.is_tree_estimator():
         # Because of this issue: https://github.com/slundberg/shap/issues/1215
@@ -68,7 +67,7 @@ def _compute_shap_values(pipeline, features, training_data=None):
             explainer = shap.TreeExplainer(estimator._component_obj, feature_perturbation="tree_path_dependent")
         if ws:
             logger.debug(f"_compute_shap_values TreeExplainer: {ws[0].message}")
-        shap_values = explainer.shap_values(pipeline_features, check_additivity=False)
+        shap_values = explainer.shap_values(features, check_additivity=False)
         # shap only outputs values for positive class for Catboost binary estimators.
         # this modifies the output to match the output format of other binary estimators.
         # Ok to fill values of negative class with zeros since the negative class will get dropped
@@ -94,7 +93,7 @@ def _compute_shap_values(pipeline, features, training_data=None):
             decision_function = estimator._component_obj.predict_proba
         with warnings.catch_warnings(record=True) as ws:
             explainer = shap.KernelExplainer(decision_function, sampled_training_data_features, link_function)
-            shap_values = explainer.shap_values(pipeline_features)
+            shap_values = explainer.shap_values(features)
         if ws:
             logger.debug(f"_compute_shap_values KernelExplainer: {ws[0].message}")
 

evalml/model_understanding/prediction_explanations/_user_interface.py

@@ -10,7 +10,7 @@
 from evalml.problem_types import ProblemTypes
 
 
-def _make_rows(shap_values, normalized_values, top_k, include_shap_values=False):
+def _make_rows(shap_values, normalized_values, pipeline_features, top_k, include_shap_values=False):
     """Makes the rows (one row for each feature) for the SHAP table.
 
     Arguments:
@@ -35,15 +35,20 @@ def _make_rows(shap_values, normalized_values, top_k, include_shap_values=False)
     for value, feature_name in features_to_display:
         symbol = "+" if value >= 0 else "-"
         display_text = symbol * min(int(abs(value) // 0.2) + 1, 5)
-        row = [feature_name, display_text]
+        feature_value = pipeline_features[feature_name].iloc[0]
+        if pd.api.types.is_number(feature_value) and not pd.api.types.is_bool(feature_value):
+            feature_value = "{:.2f}".format(feature_value)
+        else:
+            feature_value = str(feature_value)
+        row = [feature_name, feature_value, display_text]
         if include_shap_values:
-            row.append(round(shap_values[feature_name][0], 2))
+            row.append("{:.2f}".format(shap_values[feature_name][0]))
         rows.append(row)
 
     return rows
 
 
-def _make_table(shap_values, normalized_values, top_k, include_shap_values=False):
+def _make_table(shap_values, normalized_values, pipeline_features, top_k, include_shap_values=False):
     """Make a table displaying the SHAP values for a prediction.
 
     Arguments:
@@ -56,20 +61,21 @@ def _make_table(shap_values, normalized_values, top_k, include_shap_values=False
     Returns:
         str
     """
-    dtypes = ["t", "t", "f"] if include_shap_values else ["t", "t"]
-    alignment = ["c", "c", "c"] if include_shap_values else ["c", "c"]
+    n_cols = 4 if include_shap_values else 3
+    dtypes = ["t"] * n_cols
+    alignment = ["c"] * n_cols
 
     table = Texttable()
     table.set_deco(Texttable.HEADER)
     table.set_cols_dtype(dtypes)
     table.set_cols_align(alignment)
 
-    header = ["Feature Name", "Contribution to Prediction"]
+    header = ["Feature Name", "Feature Value", "Contribution to Prediction"]
     if include_shap_values:
         header.append("SHAP Value")
 
     rows = [header]
-    rows += _make_rows(shap_values, normalized_values, top_k, include_shap_values)
+    rows += _make_rows(shap_values, normalized_values, pipeline_features, top_k, include_shap_values)
     table.add_rows(rows)
     return table.draw()
 
@@ -78,24 +84,24 @@ class _TableMaker(abc.ABC):
     """Makes a SHAP table for a regression, binary, or multiclass classification problem."""
 
     @abc.abstractmethod
-    def __call__(self, shap_values, normalized_values, top_k, include_shap_values=False):
+    def __call__(self, shap_values, normalized_values, pipeline_features, top_k, include_shap_values=False):
         """Creates a table given shap values."""
 
 
 class _SHAPRegressionTableMaker(_TableMaker):
     """Makes a SHAP table explaining a prediction for a regression problems."""
 
-    def __call__(self, shap_values, normalized_values, top_k, include_shap_values=False):
-        return _make_table(shap_values, normalized_values, top_k, include_shap_values)
+    def __call__(self, shap_values, normalized_values, pipeline_features, top_k, include_shap_values=False):
+        return _make_table(shap_values, normalized_values, pipeline_features, top_k, include_shap_values)
 
 
 class _SHAPBinaryTableMaker(_TableMaker):
     """Makes a SHAP table explaining a prediction for a binary classification problem."""
 
-    def __call__(self, shap_values, normalized_values, top_k, include_shap_values=False):
+    def __call__(self, shap_values, normalized_values, pipeline_features, top_k, include_shap_values=False):
         # The SHAP algorithm will return a two-element list for binary problems.
         # By convention, we display the explanation for the dominant class.
-        return _make_table(shap_values[1], normalized_values[1], top_k, include_shap_values)
+        return _make_table(shap_values[1], normalized_values[1], pipeline_features, top_k, include_shap_values)
 
 
 class _SHAPMultiClassTableMaker(_TableMaker):
@@ -104,11 +110,11 @@ class _SHAPMultiClassTableMaker(_TableMaker):
     def __init__(self, class_names):
         self.class_names = class_names
 
-    def __call__(self, shap_values, normalized_values, top_k, include_shap_values=False):
+    def __call__(self, shap_values, normalized_values, pipeline_features, top_k, include_shap_values=False):
         strings = []
         for class_name, class_values, normalized_class_values in zip(self.class_names, shap_values, normalized_values):
             strings.append(f"Class: {class_name}\n")
-            table = _make_table(class_values, normalized_class_values, top_k, include_shap_values)
+            table = _make_table(class_values, normalized_class_values, pipeline_features, top_k, include_shap_values)
             strings += table.splitlines()
             strings.append("\n")
         return "\n".join(strings)
@@ -132,8 +138,9 @@ def _make_single_prediction_shap_table(pipeline, input_features, top_k=3, traini
     """
     if not (isinstance(input_features, pd.DataFrame) and input_features.shape[0] == 1):
         raise ValueError("features must be stored in a dataframe of one row.")
+    pipeline_features = pipeline._transform(input_features)
 
-    shap_values = _compute_shap_values(pipeline, input_features, training_data)
+    shap_values = _compute_shap_values(pipeline, pipeline_features, training_data)
     normalized_shap_values = _normalize_shap_values(shap_values)
 
     if pipeline.problem_type == ProblemTypes.REGRESSION:
@@ -143,7 +150,7 @@ def _make_single_prediction_shap_table(pipeline, input_features, top_k=3, traini
     else:
         table_maker = _SHAPMultiClassTableMaker(pipeline._classes)
 
-    return table_maker(shap_values, normalized_shap_values, top_k, include_shap_values)
+    return table_maker(shap_values, normalized_shap_values, pipeline_features, top_k, include_shap_values)
 
 
 class _ReportSectionMaker:

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_algorithms.py

@@ -88,7 +88,7 @@ def calculate_shap_for_test(training_data, y, pipeline_class, n_points_to_explai
     pipeline = pipeline_class({}, random_state=0)
     points_to_explain = training_data[:n_points_to_explain]
     pipeline.fit(training_data, y)
-    return _compute_shap_values(pipeline, points_to_explain, training_data)
+    return _compute_shap_values(pipeline, pd.DataFrame(points_to_explain), training_data)
 
 
 interpretable_estimators = [e for e in _all_estimators_used_in_search() if e.model_family not in {ModelFamily.XGBOOST, ModelFamily.BASELINE}]

evalml/tests/model_understanding_tests/prediction_explanations_tests/test_explainers.py

@@ -31,42 +31,42 @@ def test_explain_prediction_value_error(test_features):
         explain_prediction(None, input_features=test_features, training_data=None)
 
 
-explain_prediction_answer = """Feature Name   Contribution to Prediction
-                                =========================================
-                                 d                    ++++
-                                 a                    +++
-                                 c                     --
-                                 b                    ----""".splitlines()
+explain_prediction_answer = """Feature Name Feature Value Contribution to Prediction
+                               =========================================================
+                                 d           40.00          ++++
+                                 a           10.00          +++
+                                 c           30.00          --
+                                 b           20.00          ----""".splitlines()
 
 
 explain_prediction_multiclass_answer = """Class: class_0
 
-        Feature Name    Contribution to Prediction
-        =========================================
-            a                           +
-            b                           +
-            c                           -
-            d                           -
+        Feature Name Feature Value Contribution to Prediction
+       =========================================================
+            a           10.00                +
+            b           20.00                +
+            c           30.00                -
+            d           40.00                -
 
 
         Class: class_1
 
-        Feature Name       Contribution to Prediction
-        =========================================
-            a                          +++
-            b                          ++
-            c                           -
-            d                          --
+        Feature Name Feature Value Contribution to Prediction
+       =========================================================
+            a           10.00               +++
+            b           20.00               ++
+            c           30.00               -
+            d           40.00               --
 
 
         Class: class_2
 
-        Feature Name    Contribution to Prediction
-        =========================================
-            a                      +
-            b                      +
-            c                     ---
-            d                     ---
+        Feature Name Feature Value Contribution to Prediction
+        =========================================================
+            a          10.00            +
+            b          20.00            +
+            c          30.00           ---
+            d          40.00           ---
             """.splitlines()
 
 
@@ -96,6 +96,8 @@ def test_explain_prediction(mock_normalize_shap_values,
     pipeline = MagicMock()
     pipeline.problem_type = problem_type
     pipeline._classes = ["class_0", "class_1", "class_2"]
+    # By the time we call transform, we are looking at only one row of the input data.
+    pipeline._transform.return_value = pd.DataFrame({"a": [10], "b": [20], "c": [30], "d": [40]})
     features = pd.DataFrame({"a": [1], "b": [2]})
     table = explain_prediction(pipeline, features, top_k=2).splitlines()