About This Website

This website is an illustration of ADCN algorithm. You can access the paper at ADCN: An Anisotropic Density-Based Clustering Algorithm for Discovering Spatial Point Patterns with Noise

.

NMI: Illustration of ADCN-KNN performance with different Eps & MinPts

This is a 3D visualization of ADCN-KNN performance with different Eps & MinPts. We use Normalized Mutual Information (NMI) index to quantify the "goodness" of the quality of clustering result. We use the spiral case as an example.


Rand: Illustration of ADCN-KNN performance with different Eps & MinPts

This is a 3D visualization of ADCN-KNN performance with different Eps & MinPts. We use Rand index to quantify the "goodness" of the quality of clustering result. We use the spiral case as an example.


Illustration of the reason why border points are exluded during the clustering process of ADCN-Eps and ADCN-KNN

The figure above shows the clustering result of ADCN-KNN (Eps = 24, MinPts = 3). Because ADCN uses ellipses instead of circles to get the Eps-ellipse-neighborhood of a point. When the algorithm is expanding one cluster along a zigzag linear feature like the figure shown above, the computed ellipse will be extremely flat like the “red ellipse” shown here. The Eps-ellipse-neighborhood of Point Pi is 4 points within the “red ellipse” including Pi itself. One interesting finding is that the only “black point”, we indicate as Pj, which is within the “red ellipse” should be a “Noise Point” according to both the clustering result and the result from human perception. Notice that Pj is a border point now with respect to Pi. The reason why ADCN-KNN can label Pj as a “Noise Point” is that Pj is only a border point but not a core point according to the current parameter combination (Eps = 24, MinPts = 3). However, if ADCN includes all the border points into the clusters during clustering process, Pj will be always labeled as a “Cluster Point” which contradicts with the result of human perception.

In short, if ADCN includes border points, during the clustering expanding process along a zigzag linear feature, many noise points on the both sides of this linear feature will be included in this cluster. In contrast, by excluding the border points and tuning Eps and MinPts, ADCN can better capture the irregular shapes of the clusters, especially the geographic features with linear or zigzag shapes.


Download Links of ADCN test platform and test data

Click here to download the TEST PLATFORM zip files:

Click here to download the TEST DATA zip files:


Contact

Gengchen Mai

Address:

4829 Ellison Hall

University of California, Santa Barbara

Santa Barbara, CA 93106, USA

Email:

gengchen_mai@geog.ucsb.edu

Website:

http://www.geog.ucsb.edu/~gengchen_mai/

Copyright © Gengchen Mai