def test_nontrivial_hierarchy_leaf_classification():
"""Test that a nontrivial hierarchy leaf classification behaves as expected.
We build the following class hierarchy along with data from the handwritten digits dataset:
<ROOT>
/ \
A B
/ \ / \ \
1 7 3 8 9
"""
class_hierarchy = {
ROOT: ["A", "B"],
"A": [1, 7],
"B": [3, 8, 9],
}
base_estimator = svm.SVC(gamma=0.001, kernel="rbf", probability=True)
clf = make_classifier(
base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
)
X, y = make_digits_dataset(targets=[1, 7, 3, 8, 9], )
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=RANDOM_STATE,
)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
assert_that(accuracy, is_(close_to(1., delta=0.02)))
def test_base_estimator_as_dict():
"""Test that specifying base_estimator as a dictionary mappings nodes to base estimators works."""
class_hierarchy = {
ROOT: ["A", "B"],
"A": [1, 7],
"B": [3, 8, 9],
}
clf = make_classifier(
base_estimator={
ROOT: KNeighborsClassifier(),
"B": svm.SVC(probability=True),
DEFAULT: MultinomialNB(),
},
class_hierarchy=class_hierarchy,
)
X, y = make_digits_dataset(
targets=[1, 7, 3, 8, 9],
as_str=False,
)
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=RANDOM_STATE,
)
clf.fit(X_train, y_train)
assert_that(
isinstance(clf.graph_.nodes[ROOT][CLASSIFIER], KNeighborsClassifier))
assert_that(isinstance(clf.graph_.nodes["B"][CLASSIFIER], svm.SVC))
assert_that(isinstance(clf.graph_.nodes["A"][CLASSIFIER], MultinomialNB))
def test_nmlnp_strategy_on_tree_with_dummy_classifier():
"""Test classification works on a tree graph when one of the nodes has out-degree 1 resulting in
creation of a "dummy" classifier at that node to trivially predict its child."""
# since NMLNP results in a mix of intermediate and lefa nodes,
# make sure they are all of same dtype (str)
class_hierarchy = {
ROOT: ["A", "B", "C"],
"A": ["1", "5", "6", "7"],
"B": ["2", "3", "8", "9"],
"C": ["4"],
}
base_estimator = svm.SVC(gamma=0.001, kernel="rbf", probability=True)
clf = make_classifier(
base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
prediction_depth="nmlnp",
stopping_criteria=0.9,
)
X, y = make_digits_dataset()
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=RANDOM_STATE,
)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
assert_that(list(y_pred), has_item("4"))
def test_nmlnp_strategy_with_float_stopping_criteria():
# since NMLNP results in a mix of intermediate and leaf nodes,
# make sure they are all of same dtype (str)
class_hierarchy = {
ROOT: ["A", "B"],
"A": ["1", "5", "6", "7"],
"B": ["2", "3", "4", "8", "9"],
}
base_estimator = svm.SVC(gamma=0.001, kernel="rbf", probability=True)
clf = make_classifier(
base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
prediction_depth="nmlnp",
stopping_criteria=0.9,
)
X, y = make_digits_dataset()
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=RANDOM_STATE,
)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
assert_that(list(y_pred), has_item("B"))
def classify_digits():
"""Test that a nontrivial hierarchy leaf classification behaves as expected.
We build the following class hierarchy along with data from the handwritten digits dataset:
<ROOT>
/ \
A B
/ \ / \ \
1 7 3 8 9
"""
class_hierarchy = {
ROOT: ["A", "B"],
"A": [1, 7],
"B": [3, 8, 9],
}
base_estimator = make_pipeline(
TruncatedSVD(n_components=24),
svm.SVC(
gamma=0.001,
kernel="rbf",
probability=True
),
)
clf = HierarchicalClassifier(
base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
)
X, y = make_digits_dataset(
targets=[1, 7, 3, 8, 9],
)
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=RANDOM_STATE,
)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
print("Classification Report:\n", classification_report(y_test, y_pred))
def test_nmlnp_strategy_on_dag_with_dummy_classifier():
"""Test classification works on a "deep" DAG when one of the nodes has out-degree 1 resulting in
creation of a "dummy" classifier at that node to triially predict its child.
This test case actually tests a few more subtle edge cases:
- String-based target labels with length > 1
- Multi-level degenerate sub-graphs, e.g some nodes having a sub-graph which is a path.
"""
# since NMLNP results in a mix of intermediate and lefa nodes,
# make sure they are all of same dtype (str)
class_hierarchy = {
ROOT: ["A", "B", "C"],
"A": ["1", "5", "6", "7"],
"B": ["2", "BC.1", "8", "9"],
"BC.1": ["3a"],
"C": ["BC.1"],
}
base_estimator = svm.SVC(gamma=0.001, kernel="rbf", probability=True)
clf = make_classifier(
base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
prediction_depth="nmlnp",
stopping_criteria=0.9,
)
X, y = make_digits_dataset()
y[where(y == "3")] = "3a"
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=RANDOM_STATE,
)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
assert_that(list(y_pred), has_item("3a"))