Re: [Databionic-ESOM-User] automated data clustering?

[<fa...@my...>]

Dick,

Prof. Ultsch has developed a method for automated clustering, see

http://www.uni-marburg.de/fb12/datenbionik/pdf/pubs/2005/ultsch05clustering
http://www.uni-marburg.de/fb12/datenbionik/pdf/pubs/2005/ultsch05ustarc

If you contact him directly, he might be willing to share the matlab =20
code with you. I know that a student is working on integrating the =20
method into the ESOM tools but I don't know if and when it would be =20
available.

For the comparison of clusterings or evaluation with know results I =20
recommend to look up measures to compare classifications, for example =20
precision/recall, sensitivity/specificity. you can average these =20
measures over multiple clustering runs. Please be more specific if =20
these general methods do not suffice.

best
fabian

Quoting Greg Dick <gd...@be...>:

> Dear ESOM users,
>
> I am using ESOM to cluster DNA sequences from environmental  =20
> microorganisms, based on genome wide signatures (tetranucleotide  =20
> frequency).  Overall I am very happy with the results and it has  =20
> proven to be an extremely valuable tool for our research group.   =20
> There are two areas that we are hoping to develop further and I am  =20
> curious if anyone has suggestions or comments:
>
> (1) Are there any automated methods for clustering data?  The  =20
> boundaries for our clusters range from obvious to questionable.   =20
> While this variable strength of clustering is useful information in  =20
> itself, we would like to develop an automated method for defining  =20
> clusters in order to avoid potential errors in where we draw the  =20
> lines (it is not always entirely clear how to do so).
>
> (2) Are there statistical tools that have been developed or applied  =20
> to ESOM to evaluate the robustness of clustering (ideally on a  =20
> per-cluster basis)?  We are interested in such an analysis, which  =20
> would either be based on the U-matrix distance structure and/or an  =20
> evaluation of the accuracy of the clustering (for much of our data  =20
> we know the true cluster affiliations).
>
> Any suggestions or references relevant to these areas would be  =20
> greatly appreciated.
>
> Greg Dick
> Postdoctoral Research
> University of California, Berkeley
>
>