def test_birch_hierarchy():
X, y = make_blobs(random_state=40)
brc = Birch(n_clusters=None, branching_factor=5,
compute_sample_indices=True)
brc.fit(X)
# make sure that leave nodes contain all the samples
n_leaves = 1
sample_id = []
current_leaf = brc.dummy_leaf_.next_leaf_
while current_leaf:
subclusters = current_leaf.subclusters_
for sc in subclusters:
assert sc.n_samples_ == len(sc.samples_id_)
sample_id += sc.samples_id_
current_leaf = current_leaf.next_leaf_
n_leaves += 1
assert_array_equal(np.sort(sample_id), np.arange(X.shape[0]))
# Verify that the resulting hierarchical tree is deeper than 1 level
# (i.e. subclusters of the root node are nor tree leaves )
assert len(brc.root_.subclusters_) < n_leaves
# Make sure that subclusters of the root_ node contain all the samples
sample_id = []
for sc in brc.root_.subclusters_:
sample_id += sc.samples_id_
assert sc.n_samples_ == len(sc.samples_id_)
assert_array_equal(np.sort(sample_id), np.arange(X.shape[0]))
# Pick a sample at random and make sure that reported samples_id_
# matches with the subcluster the sample is closest to
document_id = 45
document_in_subcluster = []
distance_to_centroid = []
for sc in brc.root_.subclusters_:
centroid = X[sc.samples_id_, :].mean(axis=0)
distance_to_centroid.append(((X[[document_id]] - centroid)**2).sum())
document_in_subcluster.append(document_id in sc.samples_id_)
assert np.argmin(distance_to_centroid) == \
np.nonzero(document_in_subcluster)[0][0]
# Make sure that we can recompute labels from tree leaves
labels2 = np.zeros(X.shape[0], dtype=int)
cluster_id = 0
for current_leaf in brc._get_leaves():
subclusters = current_leaf.subclusters_
for sc in subclusters:
labels2[list(sc.samples_id_)] = cluster_id
cluster_id += 1
assert np.unique(brc.labels_).shape == np.unique(labels2).shape
# The two methods yield approximately equal labels
assert v_measure_score(brc.labels_, labels2) > 0.95
评论列表
文章目录
