def test_cluster_threshold(self):
"""Expected error for threshold to similarity should be reasonable"""
n_tests = 50
dim = 15
tot_err = 0
for test in range(n_tests):
# Get some sets and their similarities
sets = (randset(), randset())
jsim = jaccard_similarity(*sets)
# Find the threshold at which they cluster together
for threshold in range(1, 100, 5):
threshold = float(threshold) / 100
bandwidth = get_bandwidth(dim, threshold)
num_bands = int(dim / bandwidth)
cluster = Cluster(width=num_bands * bandwidth,
bandwidth=bandwidth)
cluster.add_item(sets[0])
cluster.add_item(sets[1])
if len(cluster.get_clusters()) == 2:
tot_err += abs(jsim - threshold)
break
avg_error = float(tot_err) / n_tests
self.assertLessEqual(avg_error, 0.30)
def test_signature_similarity(self):
"""The probability that two sets' signatures match at some index are
equal is equal to the Jaccard similarity between the two
"""
n_tests = 100
expected_error = 1.0 / 10 # Expected error is O(1/sqrt(dim))
mh = MinHashSignature(10 * 10)
err = 0.0
for _ in xrange(n_tests):
# Create random sets and their signatures
sets = (randset(), randset())
sigs = map(mh.get_signature, sets)
# Calculate true Jaccard similarity, and sim of signatures
jsim = jaccard_similarity(*sets)
ssim = sigsim(*sigs, dim=100)
# Accumulate error
err += abs(jsim - ssim)
# Over n_tests large, we should be within upper bound of expected error
avg_err = err / n_tests
self.assertGreaterEqual(
expected_error,
avg_err,
msg="Accuracy test failed. (avg error: %f)" % avg_err)
def test_jaccard_nan():
"""Returns NaN for empty set
"""
sim = jaccard_similarity([], [])
assert_true(np.isnan(sim))
