193 add isolation forrest conversion to treeexplainer (#289)

* Add basic implementation Isolation Forest

* Update conversion isoforest

* Update both conversion versions

* Update test iso tree and own scoring

* Update validation method

* Comment out benchmark for import route issue

* started check

* Add scaling

* Use updated values in convert_isolation_tree_shap_isotree

* ran check

* Disable value_updating

* Clean up isoforest support

* Reactivate benchmarks

* Lint and format

* Add tests for isolation forest conversion

* adds check agains shap implementation

* deleted check_shap.py

---------

Co-authored-by: Maximilian <maximilian.muschalik@gmail.com>

shapiq/explainer/tree/conversion/sklearn.py

@@ -4,6 +4,7 @@
 from typing import Optional
 
 import numpy as np
+from sklearn.ensemble._iforest import _average_path_length
 
 from shapiq.utils import safe_isinstance
 from shapiq.utils.types import Model
@@ -74,3 +75,97 @@ def convert_sklearn_tree(
         empty_prediction=None,  # compute empty prediction later
         original_output_type=output_type,
     )
+
+
+def average_path_length(isolation_forest):
+    max_samples = isolation_forest._max_samples
+    average_path_length = _average_path_length(
+        [max_samples]
+    )  # NOTE: _average_path_length func is equivalent to equation 1 in Isolation Forest paper Lui2008
+    return average_path_length
+
+
+def convert_sklearn_isolation_forest(
+    tree_model: Model,
+) -> list[TreeModel]:
+    """Transforms a scikit-learn isolation forest to the format used by shapiq.
+
+    Args:
+        tree_model: The scikit-learn isolation forest model to convert.
+
+    Returns:
+        The converted isolation forest model.
+    """
+    scaling = 1.0 / len(tree_model.estimators_)
+
+    return [
+        # convert_isolation_tree_shap_isotree(tree, features, scaling=scaling)
+        convert_isolation_tree(tree, features, scaling=scaling)
+        for tree, features in zip(tree_model.estimators_, tree_model.estimators_features_)
+    ]
+
+
+def convert_isolation_tree(
+    tree_model: Model,
+    tree_features,
+    class_label: Optional[int] = None,
+    scaling: float = 1.0,
+    average_path_length: float = 1.0,  # TODO fix default value
+) -> TreeModel:
+    """Convert a scikit-learn decision tree to the format used by shapiq.
+
+    Args:
+        tree_model: The scikit-learn decision tree model to convert.
+        class_label: The class label of the model to explain. Only used for classification models.
+            Defaults to ``1``.
+        scaling: The scaling factor for the tree values.
+
+    Returns:
+        The converted decision tree model.
+    """
+    output_type = "raw"
+    tree_values = tree_model.tree_.value.copy()
+    tree_values = tree_values.flatten()
+    features_updated, values_updated = isotree_value_traversal(
+        tree_model.tree_, tree_features, normalize=False, scaling=1.0
+    )
+    values_updated = values_updated * scaling
+    values_updated = values_updated.flatten()
+
+    return TreeModel(
+        children_left=tree_model.tree_.children_left,
+        children_right=tree_model.tree_.children_right,
+        features=features_updated,
+        thresholds=tree_model.tree_.threshold,
+        values=values_updated,
+        node_sample_weight=tree_model.tree_.weighted_n_node_samples,
+        empty_prediction=None,  # compute empty prediction later
+        original_output_type=output_type,
+    )
+
+
+def isotree_value_traversal(
+    tree, tree_features, normalize=False, scaling=1.0, data=None, data_missing=None
+):
+    features = tree.feature.copy()
+    corrected_values = tree.value.copy()
+    if safe_isinstance(tree, "sklearn.tree._tree.Tree"):
+
+        def _recalculate_value(tree, i, level):
+            if tree.children_left[i] == -1 and tree.children_right[i] == -1:
+                value = level + _average_path_length(np.array([tree.n_node_samples[i]]))[0]
+                corrected_values[i, 0] = value
+                return value * tree.n_node_samples[i]
+            else:
+                value_left = _recalculate_value(tree, tree.children_left[i], level + 1)
+                value_right = _recalculate_value(tree, tree.children_right[i], level + 1)
+                corrected_values[i, 0] = (value_left + value_right) / tree.n_node_samples[i]
+                return value_left + value_right
+
+        _recalculate_value(tree, 0, 0)
+        if normalize:
+            corrected_values = (corrected_values.T / corrected_values.sum(1)).T
+        corrected_values = corrected_values * scaling
+        # re-number the features if each tree gets a different set of features
+        features = np.where(features >= 0, tree_features[features], features)
+    return features, corrected_values

shapiq/explainer/tree/explainer.py

@@ -41,7 +41,7 @@ class TreeExplainer(Explainer):
 
     def __init__(
         self,
-        model: Union[dict, TreeModel, Any],
+        model: Union[dict, TreeModel, list, Any],
         max_order: int = 2,
         min_order: int = 1,
         index: str = "k-SII",
@@ -61,6 +61,7 @@ def __init__(
         # validate and parse model
         validated_model = validate_tree_model(model, class_label=class_index)
         self._trees: list[TreeModel] = copy.deepcopy(validated_model)
+        # TODO trees are made instance of list here, but in validation they are also but then converted back into single element if list is length 1
         if not isinstance(self._trees, list):
             self._trees = [self._trees]
         self._n_trees = len(self._trees)

shapiq/explainer/tree/validation.py

@@ -6,7 +6,11 @@
 
 from .base import TreeModel
 from .conversion.lightgbm import convert_lightgbm_booster
-from .conversion.sklearn import convert_sklearn_forest, convert_sklearn_tree
+from .conversion.sklearn import (
+    convert_sklearn_forest,
+    convert_sklearn_isolation_forest,
+    convert_sklearn_tree,
+)
 from .conversion.xgboost import convert_xgboost_booster
 
 SUPPORTED_MODELS = {
@@ -20,6 +24,8 @@
     "sklearn.ensemble._forest.ExtraTreesClassifier",
     "sklearn.ensemble.RandomForestRegressor",
     "sklearn.ensemble._forest.RandomForestRegressor",
+    "sklearn.ensemble.IsolationForest",
+    "sklearn.ensemble._iforest.IsolationForest",
     "lightgbm.sklearn.LGBMRegressor",
     "lightgbm.sklearn.LGBMClassifier",
     "lightgbm.basic.Booster",
@@ -42,8 +48,11 @@ def validate_tree_model(
     # tree model (is already in the correct format)
     if type(model).__name__ == "TreeModel":
         tree_model = model
-    elif isinstance(model, list) and all([type(m).__name__ == "TreeModel" for m in model]):
-        tree_model = model
+    # direct return if list of tree models
+    elif type(model).__name__ == "list":
+        # check if all elements are TreeModel
+        if all([type(tree).__name__ == "TreeModel" for tree in model]):
+            tree_model = model
     # dict as model is parsed to TreeModel (the dict needs to have the correct format and names)
     elif type(model).__name__ == "dict":
         tree_model = TreeModel(**model)
@@ -66,6 +75,10 @@ def validate_tree_model(
         or safe_isinstance(model, "sklearn.ensemble._forest.ExtraTreesClassifier")
     ):
         tree_model = convert_sklearn_forest(model, class_label=class_label)
+    elif safe_isinstance(model, "sklearn.ensemble.IsolationForest") or safe_isinstance(
+        model, "sklearn.ensemble._iforest.IsolationForest"
+    ):
+        tree_model = convert_sklearn_isolation_forest(model)
     elif safe_isinstance(model, "lightgbm.sklearn.LGBMRegressor") or safe_isinstance(
         model, "lightgbm.sklearn.LGBMClassifier"
     ):

shapiq/interaction_values.py

@@ -655,7 +655,7 @@ def plot_force(
         Returns:
             The force plot as a matplotlib figure (if show is ``False``).
         """
-        from shapiq import force_plot
+        from .plot import force_plot
 
         return force_plot(
             self,

tests/conftest.py

@@ -9,7 +9,7 @@
 import pytest
 from PIL import Image
 from sklearn.datasets import make_classification, make_regression
-from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
+from sklearn.ensemble import IsolationForest, RandomForestClassifier, RandomForestRegressor
 from sklearn.linear_model import LinearRegression, LogisticRegression
 from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
 
@@ -142,6 +142,25 @@ def rf_clf_model() -> RandomForestClassifier:
     return model
 
 
+# Isolationforest model
+@pytest.fixture
+def if_clf_model() -> IsolationForest:
+    n_samples, n_outliers = 120, 40
+    rng = np.random.RandomState(0)
+    covariance = np.array([[0.5, -0.1], [0.7, 0.4]])
+    cluster_1 = 0.4 * rng.randn(n_samples, 2) @ covariance + np.array([2, 2])  # general
+    cluster_2 = 0.3 * rng.randn(n_samples, 2) + np.array([-2, -2])  # spherical
+    outliers = rng.uniform(low=-4, high=4, size=(n_outliers, 2))
+
+    X = np.concatenate([cluster_1, cluster_2, outliers])
+    y = np.concatenate([np.ones((2 * n_samples), dtype=int), -np.ones((n_outliers), dtype=int)])
+
+    # X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, random_state=42)
+    model = IsolationForest(random_state=42, n_estimators=3)
+    model.fit(X, y)
+    return model
+
+
 @pytest.fixture
 def xgb_reg_model():
     """Return a simple xgboost regression model."""

tests/tests_explainer/tests_tree_explainer/test_tree_explainer.py

@@ -363,3 +363,29 @@ def test_xgboost_shap_error(xgb_clf_model, background_clf_data):
 
     # now the values surprisingly are the same
     assert np.allclose(sv_shap, sv_shapiq_rounded_values, rtol=1e-5)
+
+
+def test_iso_forest_shap(if_clf_model):
+    """Tests the shapiq implementation of TreeSHAP vs. SHAP's implementation for Isolation Forest."""
+
+    x_explain = np.array([0.125, 0.05])
+
+    # the following code is used to get the shap values from the SHAP implementation
+    # import shap
+    # model_copy = copy.deepcopy(if_clf_model)
+    # explainer_shap = shap.TreeExplainer(model=model_copy)
+    # baseline_shap = float(explainer_shap.expected_value)
+    # sv_shap = explainer_shap.shap_values(x_explain)
+    # print(sv_shap)
+    # print(baseline_shap)
+    sv_shap = np.array([-2.34951688, -4.55545493])
+    baseline_shap = 12.238305148044713
+
+    # compute with shapiq
+    explainer_shapiq = TreeExplainer(model=if_clf_model, max_order=1, index="SV")
+    sv_shapiq = explainer_shapiq.explain(x=x_explain)
+    sv_shapiq_values = sv_shapiq.get_n_order_values(1)
+    baseline_shapiq = sv_shapiq.baseline_value
+
+    assert baseline_shap == pytest.approx(baseline_shapiq, rel=1e-6)
+    assert np.allclose(sv_shap, sv_shapiq_values, rtol=1e-5)

tests/tests_explainer/tests_tree_explainer/test_tree_explainer_conversion.py

@@ -5,7 +5,11 @@
 
 from shapiq.explainer.tree.base import TreeModel
 from shapiq.explainer.tree.conversion.edges import create_edge_tree
-from shapiq.explainer.tree.conversion.sklearn import convert_sklearn_forest, convert_sklearn_tree
+from shapiq.explainer.tree.conversion.sklearn import (
+    convert_sklearn_forest,
+    convert_sklearn_isolation_forest,
+    convert_sklearn_tree,
+)
 from shapiq.utils import safe_isinstance
 
 
@@ -123,3 +127,14 @@ def test_skleanr_rf_conversion(rf_clf_model, rf_reg_model):
     assert isinstance(tree_model, list)
     assert safe_isinstance(tree_model[0], tree_model_class_path_str)
     assert tree_model[0].empty_prediction is not None
+
+
+def test_sklearn_if_conversion(if_clf_model):
+    """Test the conversion of a scikit-learn isolation forest model."""
+    tree_model_class_path_str = ["shapiq.explainer.tree.base.TreeModel"]
+
+    # test the isolation forest model
+    tree_model = convert_sklearn_isolation_forest(if_clf_model)
+    assert isinstance(tree_model, list)
+    assert safe_isinstance(tree_model[0], tree_model_class_path_str)
+    assert tree_model[0].empty_prediction is not None

tests/tests_explainer/tests_tree_explainer/test_tree_explainer_validate.py

@@ -7,7 +7,7 @@
 from shapiq.explainer.tree.validation import validate_tree_model
 
 
-def test_validate_model(dt_clf_model, dt_reg_model, rf_reg_model, rf_clf_model):
+def test_validate_model(dt_clf_model, dt_reg_model, rf_reg_model, rf_clf_model, if_clf_model):
     """Test the validation of the model."""
     class_path_str = ["shapiq.explainer.tree.base.TreeModel"]
     # sklearn dt models are supported
@@ -20,6 +20,10 @@ def test_validate_model(dt_clf_model, dt_reg_model, rf_reg_model, rf_clf_model):
     for tree in tree_model:
         assert safe_isinstance(tree, class_path_str)
     tree_model = validate_tree_model(rf_reg_model)
+    for tree in tree_model:
+        assert safe_isinstance(tree, class_path_str)
+    # sklearn isolation forest is supported
+    tree_model = validate_tree_model(if_clf_model)
     for tree in tree_model:
         assert safe_isinstance(tree, class_path_str)