Skip to content

clustering

2 messages · EdwardSeelefant@aol.com, Christian Hennig

#
Dear R-users,

I have a two - dimensional data set which needs to be clustered into
groups:
I'm searching for groups of points which show a positive
correlation (in a twodimensional plot of the data set), but I do not have
any knowledge about how many groups there might be.
Do you know of a clustering algorithm in R (or
in general) which can use a-priori information about the cluster's shape,
in particular that the points of a given cluster should lie on a line?
The "EMclust" algorithm does allow to chose a general shape, such as
"elliptical" for example. However, this method proved not to be
successful.

Thank you very much,
Yours sincererly

Philip
#
Dear Philip,

as you have already suggested, mclust would be the first choice. Did you
try to run EMclust with emModelNames="VEV"? If you send me the data, I can
take a look at it and see what went wrong.

There is also methodology for clusterwise linear regression (CLR), where the
points of a cluster lie on a line, but there is a distinction between
dependent and independent variable, which may or may not be what you
want. As far as I know, there is nothing for CLR on CRAN, but my R-package
fpc contains a function fixreg for linear regression fixed point clusters
(reference Hennig, Journal of Classification 19, 249-276 (2002)). It can be
obtained under
http://www.math.uni-hamburg.de/home/hennig/fixreg/fixreg.html

Furthermore I have a program for ML-estimation of linear regression
mixtures (reference deSarbo and Cron, Journal of Classification 5, 249-282 
(1988)), but undocumented and for private use. Ask me if you are
interested.

Best,
Christian
On Wed, 23 Apr 2003 EdwardSeelefant at aol.com wrote: