def test_roc_precalculated_t4():
"""test against precalculated data (with ties)
The plot should look like Fig. 2 in Mason & Graham (2002).
A naive implementation w/o regard for ties gives an incorrect
result in this situation.
"""
t4 = [(1, 100.0),
(1, 100.0),
(1, 100.0),
(1, 100.0),
(1, 80.0),
(0, 80.0),
(0, 80.0),
(1, 60.0),
(0, 40.0),
(0, 20.0),
(1, 0.0),
(0, 0.0),
(0, 0.0),
(0, 0.0),
(0, 0.0)]
rc = RocCurve.from_labels(*zip(*t4))
auc = rc.auc_score()
assert_almost_equal(auc, 0.839, 3)
def test_roc_simulated():
# Test Area under Receiver Operating Characteristic (ROC) curve
for _ in range(10):
y_true, probas_pred = simulate_predictions(1000, seed=random_seed())
rc = RocCurve.from_labels(y_true, probas_pred)
auc_expected1 = _auc(rc.fprs, rc.tprs)
auc_expected2 = auc_sklearn(y_true, probas_pred)
auc_actual = roc_auc_score(y_true, probas_pred)
assert_almost_equal(auc_expected1, auc_actual, 3)
assert_almost_equal(auc_expected2, auc_actual, 3)
def test_sample_perverse():
"""Perverese cases are 0.0 < AUL < 0.5
"""
clusters = [[1], [0, 0]]
aul1 = aul_score_from_labels(*clusters_to_labels(clusters))
aul2 = aul_score_from_clusters(clusters)
assert_almost_equal(aul1, 0.1111, 4)
assert_almost_equal(aul2, 0.1111, 4)
auc = RocCurve.from_clusters(clusters).auc_score()
assert_almost_equal(auc, 0.0, 4)
def test_sample_bad():
"""Bad clustering should score poorly
"""
clusters = [[1, 1, 0, 0], [0]]
aul1 = aul_score_from_labels(*clusters_to_labels(clusters))
aul2 = aul_score_from_clusters(clusters)
assert_almost_equal(aul1, 0.5, 4)
assert_almost_equal(aul2, 0.5, 4)
auc = RocCurve.from_clusters(clusters).auc_score()
assert_almost_equal(auc, 0.6667, 4)
def test_sample_neg_class1():
"""Similar to ``test_sample_perfect`` but have a negative of class 1
"""
clusters = [[1, 1, 1, 1, 1], [1], [0]]
aul1 = aul_score_from_labels(*clusters_to_labels(clusters))
aul2 = aul_score_from_clusters(clusters)
assert_almost_equal(aul1, 0.8690, 4)
assert_almost_equal(aul2, 0.8690, 4)
auc = RocCurve.from_clusters(clusters).auc_score()
assert_almost_equal(auc, 0.9167, 4)
def test_sample_cluster0_c0():
"""Similar to ``test_sample_perfect`` but have a cluster of class 0
"""
clusters = [[1, 1, 1, 1], [0, 0], [0]]
aul1 = aul_score_from_labels(*clusters_to_labels(clusters))
aul2 = aul_score_from_clusters(clusters)
assert_almost_equal(aul1, 0.6667, 4)
assert_almost_equal(aul2, 0.6667, 4)
auc = RocCurve.from_clusters(clusters).auc_score()
assert_almost_equal(auc, 1.0, 4)
def test_sample_cluster0_nh():
"""Same as in ``test_sample_perfect`` but cluster 0 not homogeneous
"""
clusters = [[1, 1, 1, 1, 0], [0], [0]]
aul1 = aul_score_from_labels(*clusters_to_labels(clusters))
aul2 = aul_score_from_clusters(clusters)
assert_almost_equal(aul1, 0.8, 4)
assert_almost_equal(aul2, 0.8, 4)
auc = RocCurve.from_clusters(clusters).auc_score()
assert_almost_equal(auc, 0.8333, 4)
def test_sample_perfect():
"""Perfect clustering
"""
clusters = [[1, 1, 1, 1, 1], [0], [0]]
aul1 = aul_score_from_labels(*clusters_to_labels(clusters))
aul2 = aul_score_from_clusters(clusters)
assert_almost_equal(aul1, 1.0, 4)
assert_almost_equal(aul2, 1.0, 4)
auc = RocCurve.from_clusters(clusters).auc_score()
assert_almost_equal(auc, 1.0, 4)
def add_ranking_metrics(args, clusters, pairs):
"""Add metrics based on ROC and Lift curves
"""
args_metrics = utils.METRICS
if set(utils.ROC_METRICS) & set(args_metrics):
from lsh_hdc.ranking import RocCurve
rc = RocCurve.from_clusters(clusters, is_class_pos=class_is_positive)
if 'roc_auc' in args_metrics:
pairs.append(('roc_auc', rc.auc_score()))
if 'roc_max_info' in args_metrics:
pairs.append(('roc_max_info', rc.max_informedness()))
if set(utils.LIFT_METRICS) & set(args_metrics):
from lsh_hdc.ranking import aul_score_from_clusters as aul_score
clusters_2xc = ([class_is_positive(point) for point in cluster]
for cluster in clusters)
if 'aul_score' in args_metrics:
pairs.append(('aul_score', aul_score(clusters_2xc)))
def test_roc_precalculated_t2():
"""test against precalculated data
The plot should look like Fig. 1 in Mason & Graham (2002)
"""
t2 = [(1, 98.4),
(1, 95.2),
(1, 94.4),
(0, 92.8),
(1, 83.2),
(1, 81.6),
(1, 58.4),
(0, 57.6),
(0, 28.0),
(0, 13.6),
(1, 3.2),
(0, 2.4),
(0, 1.6),
(0, 0.8),
(0, 0.0)]
rc = RocCurve.from_labels(*zip(*t2))
auc = rc.auc_score()
assert_almost_equal(auc, 0.875, 3)