Radford Neal wrote:
> There is a huge literature on determining the number of components
> AFTER seeing the data.  In my opinion, it is mostly misdirected.
> I'd recommend that you consider carefully whether your real
> objective is connected to determining the number of components in a
> Gaussian mixture.  I doubt that it is.

There is a very actual need to identify specific clusters/components.
Human mind functions on the basis of identified concepts and groups.
For example, there are many symptoms in medicine. The
clusters/components are diseases, the necessary simplifications
physicians use to be able to diagnose, decide and cure. A very
interesting problem is to examine the medical data and identify cases
where two different diseases are really a single one, or when a single
disease should be separated into several types. For example:

L Peelen, N Peek, K Zwinderman: Statistical methods to compare
different definitions of disease with an application to severe sepsis.
http://magix.fri.uni-lj.si/idamap/idamap2003/Peelen.pdf

We're not always modelling the truth, we're sometimes trying to help
people understand the world by simplifying it. Of course, there is no
need to constrain a computer to human cognitive limitations, and here
Radford's point stands.

-- 
mag. Aleks Jakulin
http://ai.fri.uni-lj.si/aleks/
Artificial Intelligence Laboratory,
Faculty of Computer and Information Science, University of Ljubljana.


.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to