def kmeans_plus_plus(self, num_clusters=100, exponent=2):
'''
Select k initial clusters by kmeans++ scheme, see https://en.wikipedia.org/wiki/K-means%2B%2B
'''
X = self.data
n, p = X.shape
clusters = np.zeros([num_clusters, p])
first_row = np.random.randint(0, n)
clusters[0] = X[first_row]
X = np.delete(X, first_row, 0)
print "Starting kmeans++ initialization..."
for i in range(1, num_clusters):
index = 0
rank = 0
for j in range(len(X)):
sample = X[j]
biggest_distance_to_cluster = 0
for k in range(i):
distance = stats_utils.euclidean_distance(clusters[k], sample)
if distance > biggest_distance_to_cluster:
biggest_distance_to_cluster = distance
weight = biggest_distance_to_cluster ** exponent
new_rank = pow(np.random.rand(), 1 / weight) # weighted reservoir sampling
if new_rank > rank:
rank = new_rank
index = j
clusters[i] = X[index]
X = np.delete(X, index, 0)
return clusters
def get_closest_cluster_index(self, sample, clusters):
smallest_distance = np.Inf
index = None
for k in range(len(clusters)):
distance = stats_utils.euclidean_distance(clusters[k], sample)
if distance < smallest_distance:
smallest_distance = distance
index = k
return index
