def _compute_sklearn_tree_importances(self,
estimator,
X=None,
weighted=True,
normalize=True):
"""Compute MDI importances following scikit-learn API on a tree-like model
Parameters
----------
X : np.ndarray, pd.DataFrame or None, optional
data used to compute MDI; when None use original dataset, by default None
weighted : bool, optional
if MDI should be computed based on weighted node samples, by default True
normalize : bool, optional
if MDI should be normalized after computing, by default True
Returns
-------
np.ndarray
MDI importances in same order than the features
"""
importances = np.zeros(self.n_features_)
nodes = get_sklearn_nodes_from(estimator, X=X, weighted=weighted)
for n in nodes:
if isinstance(n, Node):
left, right = nodes[n.left], nodes[n.right]
importances[
n.feature] += self._compute_impurity_importance_from(
n, left, right)
importances /= nodes[0].n_node_samples
if normalize:
importances = _normalize(importances, axis=None)
return importances
def _compute_sklearn_forest_importances(self,
X=None,
weighted=True,
normalize=True):
"""Compute MDI importances following scikit-learn API on a forest-like model
Parameters
----------
X : np.ndarray, pd.DataFrame or None, optional
data used to compute MDI; when None use original dataset, by default None
weighted : bool, optional
if MDI should be computed based on weighted node samples, by default True
normalize : bool, optional
if MDI should be normalized after computing, by default True
Returns
-------
np.ndarray
MDI importances in same order than the features
"""
trees = self.estimator.estimators_
importances = np.zeros(self.n_features_)
n_estimators = 0
for e in trees:
if isinstance(e, BaseDecisionTree):
n_estimators += 1
importances += self._compute_sklearn_tree_importances(
e, X=X, weighted=weighted, normalize=normalize)
else: # specific case of sklearn gradient boosting models
for e_ in e:
if e_.tree_.node_count > 1:
n_estimators += 1
importances += self._compute_sklearn_tree_importances(
e_, X=X, weighted=weighted, normalize=False)
importances /= n_estimators
if normalize:
importances = _normalize(importances, axis=None)
return importances
def test__normalize():
array = np.array([1, 2, 3, 4])
normalized_array = _normalize(array, axis=None)
expected_array = np.array([.1, .2, .3, .4])
np.testing.assert_allclose(normalized_array, expected_array)
def test__normalize_error():
array = np.array(['a', 'b', 'c', 'd'])
with pytest.raises(TypeError):
_normalize(array, axis=None)
def test__normalize_zero_sum():
array = np.array([0, 0, 0, 0])
normalized_array = _normalize(array, axis=None)
expected_array = np.array([.0, .0, .0, .0])
np.testing.assert_allclose(normalized_array, expected_array)