explainability ease added. (#113)

* feat: core metrics

* position parity added

* Delete holisticai/explainability/metrics/core/contrast_metrics.py

* position parity added with pre-commit

* deleted contrast metrics

* returns added to the description

* multiclass testing added

* multiclass testing added with pre-commit.

* vscode deleted

* position parity changed to get fi dataframes, not indexes

* bug fixed

* rank alignment metric added.

* changed

* fixed.

* explainability ease added.

* fixed.

* fixed.

---------

Co-authored-by: crismunoz <cristian.munoz@holisticai.com>

holisticai/explainability/metrics/core/all_metrics.py

@@ -7,6 +7,9 @@
     importance_order_constrast,
     importance_range_constrast,
 )
+from holisticai.explainability.metrics.global_importance._explainability_level import (
+    compute_explainability_ease_score,
+)
 
 
 def position_parity(
@@ -63,3 +66,33 @@ def rank_alignment(
             for i in conditional_feature_importance
         ]
     )
+
+
+def explainability_ease(partial_dependence_list: list[dict]):
+    """
+    Parameters
+    ----------
+    partial_dependence_list: list[dict]
+        a list of dictionaries containing partial dependencies for each feature.
+        For multiclass classification, partial dependencies are computed for each class separately.
+        For binary classification, partial dependence is calculated only for the positive class,
+        resulting in a single dictionary in the list. Similarly, for regression, there's only one dictionary.
+
+    Returns
+    -------
+    float
+        explainability ease value, average explainability ease value for multiclass setting
+    """
+    if len(partial_dependence_list) == 1:
+        return compute_explainability_ease_score(
+            partial_dependence=partial_dependence_list[0]
+        )[0]
+    else:
+        return np.mean(
+            [
+                compute_explainability_ease_score(
+                    partial_dependence=partial_dependence
+                )[0]
+                for partial_dependence in partial_dependence_list
+            ]
+        )

tests/explainability/test_all_metrics.py

@@ -7,9 +7,13 @@
 
 from holisticai.datasets import load_dataset
 from holisticai.explainability.metrics.core.all_metrics import (
+    explainability_ease,
     position_parity,
     rank_alignment,
 )
+from holisticai.explainability.metrics.global_importance._explainability_level import (
+    compute_partial_dependence,
+)
 from holisticai.explainability.metrics.utils import get_index_groups
 
 
@@ -197,3 +201,45 @@ def test_multiclass_classification_rank_alignment():
         for _, index in index_groups.items()
     ]
     assert rank_alignment(feat_importance, conditional_feature_importance) is not None
+
+
+def test_binary_classification_explainability_ease():
+    X_train, X_test, y_train, _, _ = binary_classification_process_dataset()
+    model = train_model_classification(X_train, y_train)
+    pred = model.predict(X_test)
+    feat_importance = get_feat_importance(
+        x=X_test, y=pred, model=model, samples_len=len(X_test)
+    )
+    partial_dependence = compute_partial_dependence(
+        model=model, feature_importance=feat_importance, x=X_test, target=1
+    )
+    assert explainability_ease(partial_dependence_list=[partial_dependence]) is not None
+
+
+def test_regression_explainability_ease():
+    X_train, X_test, y_train, _, _ = regression_process_dataset()
+    model = train_model_regression(X_train, y_train)
+    pred = model.predict(X_test)
+    feat_importance = get_feat_importance(
+        x=X_test, y=pred, model=model, samples_len=len(X_test)
+    )
+    partial_dependence = compute_partial_dependence(
+        model=model, feature_importance=feat_importance, x=X_test, target=None
+    )
+    assert explainability_ease(partial_dependence_list=[partial_dependence]) is not None
+
+
+def test_multiclass_classification_explainability_ease():
+    X_train, X_test, y_train, _ = multiclass_classification_process_dataset()
+    model = train_model_classification(X_train, y_train)
+    pred = model.predict(X_test)
+    feat_importance = get_feat_importance(
+        x=X_test, y=pred, model=model, samples_len=len(X_test)
+    )
+    partial_dependence = [
+        compute_partial_dependence(
+            model=model, feature_importance=feat_importance, x=X_test, target=i
+        )
+        for i in np.unique(pred)
+    ]
+    assert explainability_ease(partial_dependence_list=partial_dependence) is not None