def test_birch_example_reproducibility(example_id):
# check reproducibility of the Birch example
rng = np.random.RandomState(42)
X, y = make_blobs(n_samples=1000, n_features=10, random_state=rng)
cluster_model = Birch(threshold=0.9,
branching_factor=20,
compute_sample_indices=True)
cluster_model.fit(X)
# assert len(cluster_model.root_.subclusters_[1].child_.subclusters_) == 3
htree, n_subclusters = birch_hierarchy_wrapper(cluster_model)
assert htree.tree_size == n_subclusters
# same random seed as in the birch hierarchy example
assert htree.tree_size == 78
sc = htree.flatten()[example_id]
if example_id == 34:
# this is true in both cases, but example_id fails on circle ci
assert sc.current_depth == 1
assert len(sc.children) == 3
assert_array_equal([sc['cluster_id'] for sc in htree.flatten()],
np.arange(htree.tree_size))
def test_birch_clusterig_single_nodes():
basename = os.path.dirname(__file__)
X = np.load(
os.path.join(basename, '..', 'data', 'ds_lsi_birch', 'data.npy'))
branching_factor = 5
mod = Birch(n_clusters=None,
threshold=0.1,
branching_factor=branching_factor,
compute_labels=False,
compute_sample_indices=True)
mod.fit(X)
htree, n_subclusters = birch_hierarchy_wrapper(mod)
# let's compute cluster similarity
for row in htree.flatten():
inertia, S_sim = centroid_similarity(X, row['document_id_accumulated'])
row['document_similarity'] = S_sim
row['cluster_similarity'] = inertia
assert htree.tree_size == n_subclusters
doc_count = 0
for el in htree.flatten():
doc_count += len(el['document_id'])
el.current_depth
el.document_id_accumulated
assert doc_count == len(htree['document_id_accumulated'])
assert doc_count == X.shape[0]
assert htree.document_count == X.shape[0]
# make sure that we have no clusters with a single child
assert sum(len(el.children) == 1 for el in htree.flatten()) == 0
def test_birch_make_hierarchy(dataset, optimal_sampling):
if dataset == 'random':
np.random.seed(9999)
X = np.random.rand(1000, 100)
normalize(X)
branching_factor = 10
elif dataset == 'birch_hierarchical':
basename = os.path.dirname(__file__)
X = np.load(
os.path.join(basename, '..', 'data', 'ds_lsi_birch', 'data.npy'))
branching_factor = 2
mod = Birch(n_clusters=None,
threshold=0.1,
branching_factor=branching_factor,
compute_labels=False,
compute_sample_indices=True)
mod.fit(X)
htree, n_subclusters = birch_hierarchy_wrapper(mod)
# let's compute cluster similarity
for row in htree.flatten():
inertia, S_sim = centroid_similarity(X, row['document_id_accumulated'])
row['document_similarity'] = S_sim
row['cluster_similarity'] = inertia
assert htree.tree_size == n_subclusters
doc_count = 0
for el in htree.flatten():
doc_count += len(el['document_id'])
el.current_depth
el.document_id_accumulated
assert doc_count == len(htree['document_id_accumulated'])
assert doc_count == X.shape[0]
assert htree.document_count == X.shape[0]
if optimal_sampling:
s_samples_1 = compute_optimal_sampling(htree,
min_similarity=0.85,
min_coverage=0.9)
for row in s_samples_1:
assert len(row['document_similarity']) == 1
assert len(row['document_id_accumulated']) == 1
s_samples_2 = compute_optimal_sampling(htree,
min_similarity=0.85,
min_coverage=0.2)
s_samples_3 = compute_optimal_sampling(htree,
min_similarity=0.9,
min_coverage=0.9)
assert len(s_samples_1) > len(s_samples_2)
assert len(s_samples_1) < len(s_samples_3)
We start by computing BIRCH clustering on some random structured data,
"""
import numpy as np
from sklearn.datasets import make_blobs
from freediscovery.cluster import Birch, birch_hierarchy_wrapper
rng = np.random.RandomState(42)
X, y = make_blobs(n_samples=1000, n_features=10, random_state=rng)
cluster_model = Birch(threshold=0.9,
branching_factor=20,
compute_sample_indices=True)
cluster_model.fit(X)
###############################################################################
#
# Next we wrap each subcluster in the cluster hierarchy
# (``cluster_model.root_``) with the
# :class:`~freediscovery.cluster.BirchSubcluster` class
# that allows easier manipulation of the hierarchical tree.
htree, _ = birch_hierarchy_wrapper(cluster_model)
print('Total number of subclusters:', htree.tree_size)
###############################################################################
#
# Visualizing the hierarchy
# -------------------------
class birch(object):
def __init__(self,
threshold=0.5,
branching_factor=20,
n_clusters=None,
outlier_threshold=0.5):
self.threshold = threshold
self.branching_factor = branching_factor
self.n_clusters = n_clusters
self.outlier_threshold = outlier_threshold
self.Birch_clusterer = Birch(threshold=self.threshold,
branching_factor=self.branching_factor,
n_clusters=self.n_clusters,
compute_sample_indices=True)
# Fitting the model with train_X
def fit(self, data):
self.data = data
#self.data.drop(self.data.columns[len(self.data.columns)-1], axis=1, inplace=True)
self.Birch_clusterer.fit(self.data)
#Defines and builds the Cluster Feature Tree
def get_cluster_tree(self):
self.htree, n_clusters = birch_hierarchy_wrapper(self.Birch_clusterer)
clusters = {}
max_depth = 0
for i in range(n_clusters):
node = bcluster()
sub_cluster = self.htree.flatten()[i]
node.set_cluster_id(sub_cluster['cluster_id'])
depth = sub_cluster.current_depth
node.set_depth(depth)
if depth > max_depth:
max_depth = depth
if i not in clusters.keys():
clusters[i] = {}
if sub_cluster.current_depth == 0:
node.set_parent()
else:
node.set_parent(clusters[sub_cluster.parent['cluster_id']])
cluster_size = sub_cluster['cluster_size']
node.set_size(cluster_size)
data_points = sub_cluster['document_id_accumulated']
data_points_names = self.data.iloc[
data_points].index.values.tolist()
node.set_data_points(data_points_names)
centroid = self.data.iloc[
sub_cluster['document_id_accumulated'], :].mean(axis=0).values
node.set_centroid(centroid)
d1, d1_v = self.calculate_d1(centroid, data_points)
d2 = self.calculate_d2(centroid, data_points, d1_v)
node.add_d1(d1)
node.add_d2(d2)
node.calculate_threshold(self.outlier_threshold)
clusters[i] = node
self.cluster_tree = clusters
return self.cluster_tree, max_depth
#Calculate the d1 distance(point farthest away from centroid)
def calculate_d1(self, centroid, data_points):
d1 = 0
u = centroid
d1_v = None
for point in data_points:
v = point
distance = euclidean(u, v)
if distance > d1:
d1 = distance
d1_v = v
return d1, d1_v
#Calculate the d2 distance(point farthest away from d1 and its distance from centroid)
def calculate_d2(self, centroid, data_points, d1_v):
d2_d1 = 0
u = d1_v
d2_v = None
for point in data_points:
v = point
distance = euclidean(u, v)
if distance > d2_d1:
d2_d1 = distance
d2_v = v
d2 = euclidean(centroid, v)
return d2
# Display's the tree
def show_clutser_tree(self):
self.htree.display_tree()
# Prediction Function with height based prediction with outlier detection
def predict(self, test_X, depth):
predicted = []
for test_instance in test_X.iterrows():
test_sample = test_instance[1].values
min_distance = float('inf')
selected_cluster = None
for cluster_id in self.cluster_tree:
if self.cluster_tree[cluster_id].depth != depth:
continue
u = self.cluster_tree[cluster_id].centroid
v = np.asarray(test_sample, dtype='float64')
distance = euclidean(u, v)
if distance < min_distance:
min_distance = distance
selected_cluster = cluster_id
self.cluster_tree[selected_cluster].add_test_points(
test_instance[0])
# Outlier identifier
#if self.cluster_tree[selected_cluster].check_outlier(min_distance):
# self.cluster_tree[selected_cluster].add_outlier_points(test_instance[0])
#_predicted_label = self.cluster_tree[selected_cluster].classifier.predict([test_sample])
#self.cluster_tree[selected_cluster].add_predicted(_predicted_label)
predicted.append(selected_cluster)
return predicted
# Model certification creator
def certify_model(self, cluster_tree, test_y):
for cluster_id in cluster_tree:
if len(cluster_tree[cluster_id].test_points) == 0:
continue
cluster_tree[cluster_id].set_test_labels(
test_y[cluster_tree[cluster_id].test_points].values)
precision = metrics.precision_score(
cluster_tree[cluster_id].test_labels,
cluster_tree[cluster_id].predicted,
average='weighted')
recall = metrics.recall_score(cluster_tree[cluster_id].test_labels,
cluster_tree[cluster_id].predicted,
average='weighted')
f1_Score = metrics.f1_score(cluster_tree[cluster_id].test_labels,
cluster_tree[cluster_id].predicted,
average='weighted')
score = {
'precision': precision,
'recall': recall,
'f1_Score': f1_Score
}
cluster_tree[cluster_id].set_score(score)