RapidMiner

How can I validate a DBSCAN clustering using only internal criteria?

Wisdom logo Registration now open for RapidMiner Wisdom Americas | New Orleans | October 10-12, 2018   Learn More
Highlighted
Learner III agucaba123
Learner III

How can I validate a DBSCAN clustering using only internal criteria?

Hello, I'm trying to do a validation of different clustering models using ONLY internal criteria. With centroid-based clustering, like K-means and K-medoid, I used DB index and an extension that evaluates the silhouette index. My problem is that DB and silhouette indexs are not available for DBSCAN, and the others operators of RapidMiner Studio like density, or item distrubution make no sense to me in this case.

 

I saw this post, but I couldn't find an answer: https://community.rapidminer.com/t5/RapidMiner-Studio-Forum/Cluster-Performance-DBScan-and-agglomera...

By the way, I readed that in previous versions of RapidMiner existed an operator called "Cluster internal validation". https://community.rapidminer.com/t5/RapidMiner-Studio-Forum/Cannot-find-the-cluster-internal-validat...

Is this operator still available? 

 

2 REPLIES

Re: How can I validate a DBSCAN clustering using only internal criteria?

Hi @agucaba123,

 

I'm not aware of an operator called "Cluster internal validation".

However, you can eventually calculate the Silhouette Coefficient using a Python script.

If you are interested in, can you share your dataset and your process in order to see if it's possible.

 

Regards,

 

Lionel

Learner III agucaba123
Learner III

Re: How can I validate a DBSCAN clustering using only internal criteria?

Hi Lionel. I can't share the dataset but I tried to apply a Silhouette coeficient and the result was this:

 

DBSCAN.png

 

I looped the epsilon parameter between 0,1 and 2. The MinsPoints were defined as 5, 10 and 20. What does it means the Silhouette index in each case? Is it useful for validation in this clustering method? Because when the epsilon parameter rises, the segmentation is worse (the numbers under the value of epsilon are the sizes of the clusters)

 

Thanks for your time.