def test_predict(fitted_random_forest, mocker):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_random_forest)
spy = mocker.spy(mdi, 'interpret')
_ = mdi.predict(X=None, weighted=True, normalize=True)
spy.assert_called_once_with(X=None, weighted=True, normalize=True)
def test_interpret_values(fitted_decision_tree, fitted_random_forest,
fitted_gradient_boosting, fitted_ada_boost):
mdi = MeanDecreaseImpurity(use_precompute=False)
for e in (fitted_decision_tree, fitted_random_forest,
fitted_gradient_boosting, fitted_ada_boost):
mdi.fit(e)
importances = mdi.interpret(X=None, weighted=True, normalize=True)
np.testing.assert_allclose(e.feature_importances_, importances)
def test_interpret_precomputed(fitted_random_forest, mocker):
mdi = MeanDecreaseImpurity(use_precompute=True)
mdi.fit(fitted_random_forest)
spy = mocker.spy(mdi, '_compute_importances')
importances = mdi.interpret(X=None, weighted=True, normalize=True)
spy.assert_not_called()
assert np.array_equal(mdi.estimator.feature_importances_, importances)
def test_interpret_not_precomputed(fitted_random_forest, mocker):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_random_forest)
spy = mocker.spy(mdi, '_compute_importances')
importances = mdi.interpret(X=None, weighted=True, normalize=True)
spy.assert_called_once_with(X=None, weighted=True, normalize=True)
np.testing.assert_allclose(fitted_random_forest.feature_importances_,
importances)
def test_fit_ada_boost(fitted_ada_boost):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_ada_boost)
assert mdi.estimator == fitted_ada_boost
assert mdi.estimator is fitted_ada_boost
assert mdi._base == 'scikit-learn'
assert mdi._name == 'feature_importances_'
assert mdi._is_forest is True
assert mdi.n_features_ == fitted_ada_boost.estimators_[0].n_features_
assert mdi.n_features_ is fitted_ada_boost.estimators_[0].n_features_
def test_fit_tree(fitted_decision_tree):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_decision_tree)
assert mdi.estimator == fitted_decision_tree
assert mdi.estimator is fitted_decision_tree
assert mdi._base == 'scikit-learn'
assert mdi._name == 'feature_importances_'
assert mdi._is_forest is False
assert mdi.n_features_ == fitted_decision_tree.n_features_
assert mdi.n_features_ is fitted_decision_tree.n_features_
def test__compute_sklearn_forest_importances_for_gradient_boosting(
pseudo_gradient_boosting):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.estimator = pseudo_gradient_boosting
mdi.n_features_ = 4
importances1 = mdi._compute_sklearn_forest_importances(X=None,
weighted=False,
normalize=False)
importances2 = mdi._compute_sklearn_forest_importances(X=None,
weighted=False,
normalize=True)
importances3 = mdi._compute_sklearn_forest_importances(X=None,
weighted=True,
normalize=False)
importances4 = mdi._compute_sklearn_forest_importances(X=None,
weighted=True,
normalize=True)
expected_importances1 = np.array([0., .21, .21, 0.])
expected_importances2 = np.array([0., .5, .5, 0.])
expected_importances3 = np.array([0., .21, .21, 0.])
expected_importances4 = np.array([0., .5, .5, 0.])
np.testing.assert_allclose(importances1, expected_importances1)
np.testing.assert_allclose(importances2, expected_importances2)
np.testing.assert_allclose(importances3, expected_importances3)
np.testing.assert_allclose(importances4, expected_importances4)
def test__compute_importances_error(iris, fitted_random_forest):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_random_forest)
mdi._base = 'xgboost'
with pytest.raises(ValueError):
mdi._compute_importances(X=None, weighted=True, normalize=True)
def test_mdi_initialization_attributes():
mdi = MeanDecreaseImpurity(use_precompute=False)
assert mdi.use_precompute is False
assert mdi.estimator is None
assert mdi.n_features_ is None
assert mdi._base is None
assert mdi._name is None
assert mdi._is_forest is None
def test_mdi_initialization():
mdi = MeanDecreaseImpurity(use_precompute=False)
assert hasattr(mdi, 'use_precompute')
assert hasattr(mdi, 'estimator')
assert hasattr(mdi, 'n_features_')
assert hasattr(mdi, '_base')
assert hasattr(mdi, '_name')
assert hasattr(mdi, '_is_forest')
def test__compute_importances(fitted_random_forest, mocker):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_random_forest)
spy = mocker.spy(mdi, '_compute_sklearn_importances')
mdi._compute_importances(X=None, weighted=True, normalize=True)
spy.assert_called_once_with(X=None, weighted=True, normalize=True)
def test__compute_impurity_importance_from():
node = Node(index=1,
left=2,
right=3,
feature=0,
value=[.7, .3],
impurity=.42,
n_node_samples=100)
left = Leaf(index=2, value=[1., 0.], impurity=0., n_node_samples=70)
right = Leaf(index=3, value=[0., 1.], impurity=0., n_node_samples=30)
impurity_importance = MeanDecreaseImpurity._compute_impurity_importance_from(
node, left, right)
expected_impurity_importance = 42.
assert math.isclose(impurity_importance, expected_impurity_importance)
def test_fit_raise_error(iris, random_forest):
mdi = MeanDecreaseImpurity(use_precompute=False)
with pytest.raises(NotFittedError):
mdi.fit(random_forest)
svc = SVC(kernel='linear', C=1)
svc.fit(iris.data, iris.target)
with pytest.raises(TypeError):
mdi.fit(svc)
def test__compute_sklearn_importances_tree(fitted_decision_tree, mocker):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_decision_tree)
spy1 = mocker.spy(mdi, '_compute_sklearn_tree_importances')
spy2 = mocker.spy(mdi, '_compute_sklearn_forest_importances')
mdi._compute_sklearn_importances(X=None, normalize=True, weighted=True)
spy1.assert_called_once_with(fitted_decision_tree,
X=None,
weighted=True,
normalize=True)
spy2.assert_not_called()
def test__compute_sklearn_importances_forest(fitted_random_forest, mocker):
mdi = MeanDecreaseImpurity(use_precompute=False)
mdi.fit(fitted_random_forest)
spy1 = mocker.spy(mdi, '_compute_sklearn_tree_importances')
spy2 = mocker.spy(mdi, '_compute_sklearn_forest_importances')
mdi._compute_sklearn_importances(X=None, normalize=True, weighted=True)
last_estimator = mdi.estimator.estimators_[-1]
spy1.assert_called_with(last_estimator,
X=None,
weighted=True,
normalize=True)
spy2.assert_called_once_with(X=None, weighted=True, normalize=True)
