Hierarchical clustering: distance threshold

[vmirly]

So, the output of hierarchichal clustering results can be determined either by number of clusters, or by the a distance thereshold to cut the tree at that threshold. However, scikit learn only supports one way!

class sklearn.cluster.AgglomerativeClustering(n_clusters=2, ...

As a suggestion, is it possible to add the other option, to give distance_threshold as input argument and get as many clusters are created in the output! So, one of this argument should be given, either n_clusters or the threshold, not both!

[agramfort]

sure it is possible !

code is open and we always welcome good contributions...

[eyaler]

actually there are more options for the stopping/threshold criteria:
https://docs.scipy.org/doc/scipy-0.15.1/reference/generated/scipy.cluster.hierarchy.fcluster.html#scipy.cluster.hierarchy.fcluster
http://www.mathworks.com/help/stats/cluster.html

the inconsistency coefficient seems interesting and is the default in matlab and scipy...

[mjbommar]

You might also review some of the literate on graph modularity as a more robust measure of cluster quality. The igraph package (with python bindings) has a large number of these functions available for reference.

Some discussion here also re: clustering and "communitiy detection", agglomerative/divisive/spectrall approaches, etc.:
http://arxiv.org/abs/0906.0612
http://bommaritollc.com/2012/06/summary-community-detection-algorithms-igraph-0-6/

[VathsalaAchar]

Do you mind if I pick this up? I have implemented clustering in C++ years ago so I'll brush up and get back with questions.

[jnothman]

I think you'd be welcome to

…

On 11 May 2017 at 20:12, Vathsala Achar ***@***.***> wrote: Do you mind if I pick this up? I have implemented clustering in C++ years ago so I'll brush up and get back with questions. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3796 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz66oNne-dsdHaN8gyrMBFploImN5bks5r4t8ngaJpZM4CyRYi> .

[VathsalaAchar]

Thanks @jnothman
I saw that @vmichel has made the necessary changes for a distance threshold (and no PR?), but I was interested in the criterion that scipy.fcluster uses (mention by @eyaler) and was wondering if that would be useful along with the distance threshold. Or is it a feature of it's own?

[jnothman]

designing an api that is extensible to alternate stopping criteria may bea good idea, but it's often good to implement one at a time

…

On 12 May 2017 7:45 pm, "Vathsala Achar" ***@***.***> wrote: Thanks @jnothman <https://github.com/jnothman> I saw that @vmichel <https://github.com/vmichel> has made the necessary changes for a distance threshold (and no PR?), but I was interested in the criterion that scipy.fcluster uses (mention by @eyaler <https://github.com/eyaler>) and was wondering if that would be useful along with the distance threshold. Or is it a feature of it's own? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3796 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz668aVTFBtfzK_lPIIEOJUhCcNc0Vks5r5CpNgaJpZM4CyRYi> .