commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alina Ciobanu <>
Subject Re: [Math] Contributions to the clustering module (maybe GSoC)
Date Sat, 07 Feb 2015 20:53:09 GMT

I finally figured out my schedule for this summer and the conclusion is that I would be able
to dedicate about 20 hours per week for the GSoC project. As far as I understand, this is
about half of what is expected from a GSoC student, so unfortunately I think I should not
apply this year. I want to contribute to the Commons Math library nonetheless.

Best regards,
      From: Thomas Neidhart <>
 To: Commons Developers List <> 
 Sent: Tuesday, February 3, 2015 1:17 AM
 Subject: Re: [Math] Contributions to the clustering module (maybe GSoC)
On 02/02/2015 10:36 PM, Alina Ciobanu wrote:
> Hello Thomas,
> Thank you for the answer. I hope I will be able to clarify my schedule for the summer
in about a week from now and I will decide whether I should apply to GSoC this year or not.
I will let you know as soon as I can. Until then, I will shortly describe my first ideas below:
> 1. Spectral clustering [1] - It basically maps the data in a lower-dimensional space
(relying on the eigenvectors of the similarity matrix) and performs (k-means) clustering there.
This method can resolve a wide variety of problems, regardless of the form of the clusters.
It could be implemented efficiently using the Commons Math linear algebra module.
> 2. Mean shift algorithm [2] - I didn't grasp all the details of the algorithm yet, but
I find it very interesting. As far as I understand, it has been primarily used in pattern
recognition and computer vision. I discovered it while searching for an algorithm that does
not require the number of clusters as input parameter. I think it would be a good addition
to Commons Math besides DBSCAN, from this point of view.
> 3. Clustering evaluation methods3.1. The Silhouette Coefficient [3] - accounts for the
intra-cluster and inter-cluster distance to assign a score in [-1, 1] to a clustering.3.2.
External clustering evaluation [4] - when gold standard is available for the clustered data,
it can be used to asses the performance of a clustering algorithm.
> Suggestions are more than welcome. If you have requests from users for specific clustering
algorithms, please let me know.

You proposals sound good, as a pointer to already existing feature
requests you can take a look at:

 * Optics algorithm -
 * HAC algorithm -

Cluster evaluation would also be very interesting, I already wanted to
do something in this direction but could not find the time.

btw. by coincidence, we received a reminder about this years GSOC just
today, the deadline is 13-02-2015 to submit a project proposal with
project ideas.


To unsubscribe, e-mail:
For additional commands, e-mail:

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message