def explain_decision_tree(
clf,
vec=None,
top=_TOP,
target_names=None,
targets=None, # ignored
feature_names=None,
feature_re=None,
**export_graphviz_kwargs):
"""
Return an explanation of a decision tree classifier in the
following format (compatible with random forest explanations)::
Explanation(
estimator="<classifier repr>",
method="<interpretation method>",
description="<human readable description>",
decision_tree={...tree information},
feature_importances=[
FeatureWeight(feature_name, importance, std_deviation),
...
]
)
"""
feature_names = get_feature_names(clf, vec, feature_names=feature_names)
coef = clf.feature_importances_
tree_feature_names = feature_names
if feature_re is not None:
feature_names, flt_indices = feature_names.filtered_by_re(feature_re)
coef = coef[flt_indices]
indices = argsort_k_largest(coef, top)
names, values = feature_names[indices], coef[indices]
std = np.zeros_like(values)
export_graphviz_kwargs.setdefault("proportion", True)
tree_info = get_tree_info(clf,
feature_names=tree_feature_names,
class_names=target_names,
**export_graphviz_kwargs)
return Explanation(
feature_importances=[
FeatureWeight(*x) for x in zip(names, values, std)
],
decision_tree=tree_info,
description=DESCRIPTION_DECISION_TREE,
estimator=repr(clf),
method='decision tree',
)
def explain_rf_feature_importance(
clf,
vec=None,
top=_TOP,
target_names=None, # ignored
targets=None, # ignored
feature_names=None,
feature_re=None):
"""
Return an explanation of a tree-based ensemble classifier in the
following format::
Explanation(
estimator="<classifier repr>",
method="<interpretation method>",
description="<human readable description>",
feature_importances=[
FeatureWeight(feature_name, importance, std_deviation),
...
]
)
"""
feature_names = get_feature_names(clf, vec, feature_names=feature_names)
coef = clf.feature_importances_
trees = np.array(clf.estimators_).ravel()
coef_std = np.std([tree.feature_importances_ for tree in trees], axis=0)
if feature_re is not None:
feature_names, flt_indices = feature_names.filtered_by_re(feature_re)
coef = coef[flt_indices]
coef_std = coef_std[flt_indices]
indices = argsort_k_largest(coef, top)
names, values, std = feature_names[indices], coef[indices], coef_std[
indices]
return Explanation(
feature_importances=[
FeatureWeight(*x) for x in zip(names, values, std)
],
description=DESCRIPTION_RANDOM_FOREST,
estimator=repr(clf),
method='feature importances',
)
def get_feature_importances_filtered(coef,
feature_names,
flt_indices,
top,
coef_std=None):
# type: (...) -> FeatureImportances
if flt_indices is not None:
coef = coef[flt_indices]
if coef_std is not None:
coef_std = coef_std[flt_indices]
indices = argsort_k_largest(coef, top)
names, values = feature_names[indices], coef[indices]
std = None if coef_std is None else coef_std[indices]
return FeatureImportances.from_names_values(
names,
values,
std,
remaining=coef.shape[0] - len(indices),
)
def test_argsort_k_largest_empty():
x = np.array([0])
empty = np.array([])
assert _np_eq(x[argsort_k_largest(x, 0)], empty)
assert _np_eq(x[argsort_k_largest_positive(x, None)], empty)
def test_argsort_k_largest_None(x):
assert len(argsort_k_largest(x, None)) == len(x)
def test_argsort_k_largest_zero(x):
assert len(argsort_k_largest(x, 0)) == 0
def test_argsort_k_largest(x, k):
assume(len(x) >= k)
assume(len(set(x)) == len(x))
assume(not np.isnan(x).any())
assert (np.argsort(x)[-k:][::-1] == argsort_k_largest(x, k)).all()