Skip to content
Python implementation of Density-Based Clustering Validation
Python Shell
Branch: master
Clone or download
christopherjenness Merge pull request #7 from christopherjenness/profiling
Fixes #6

53X speed improvement by using numpy instead of looping.
Latest commit b28e70a Aug 25, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
DBCV Remove unneeded code Aug 25, 2018
plots Add plots Mar 23, 2017
profiling Add updated profiling results Aug 25, 2018
tests Clean with flake8 Jul 4, 2017
.gitignore Add .coverage Nov 26, 2017
.travis.yml Use conda instead of pip Jul 4, 2017
LICENSE fix typo: DBVC -> DBCV Sep 23, 2017
requirements.txt Add dbcv testing Jul 4, 2017 Modified to exclude tests/plots and only use necessary packa… Sep 14, 2017


Python implementation of Density-Based Clustering Validation


Moulavi, Davoud, et al. "Density-based clustering validation." Proceedings of the 2014 SIAM International Conference on Data Mining. Society for Industrial and Applied Mathematics, 2014.


What is DBCV

How do you validate clustering assignmnets from unsupervised learning algorithms? A common method is the Silhoette Method, which provides an objective score between -1 and 1 on the quality of clustering. The silhouette value measures how well an object is classified in its own cluster instead of neighboring clusters. The silhouette (and most other popular methods) work very well on globular clusters, but can fail on non-glubular clusters such as:


Here, we implement DBCV which can validate clustering assignments on non-globular, arbitrarily shaped clusters (such as the example above). In essence, DBCV computes two values:

  • The density within a cluster
  • The density between clusters

High density within a cluster, and low density between clusters indicates good clustering assignments.


Here, I deliberately picked an example of clusters that density based clustering works well on.

from sklearn import datasets
import matplotlib.pyplot as plt
import seaborn as sns

noisy_moons = datasets.make_moons(n_samples=n_samples, noise=.05)
X = noisy_moons[0]
plt.scatter(X[:,0], X[:,1])


What happens when we try K-means clustering on these non-globular clusters?

from sklearn.cluster import KMeans

kmeans =  KMeans(n_clusters=2)
kmeans_labels = kmeans.fit_predict(X)
plt.scatter(X[:,0], X[:,1], c=kmeans_labels)


...Not so great. What about HDBSCAN, a density based clustering method?

import hdbscan

hdbscanner = hdbscan.HDBSCAN()
hdbscan_labels = hdbscanner.fit_predict(X)
plt.scatter(X[:,0], X[:,1], c=hdbscan_labels)


That's pretty good. To assess the quality of clustering, using Density-Based Clustering Validation, we call DBCV

from scipy.spatial.distance import euclidean

kmeans_score = DBCV(X, kmeans_labels, dist_function=euclidean)
hdbscan_score = DBCV(X, hdbscan_labels, dist_function=euclidean)
print(kmeans_score, hdbscan_score)

K means returns a DBCV score of -0.71, and HDBSCAN returns a score of 0.60.

You can’t perform that action at this time.