R help - Apr 2006 - passing known medoids to clara() in the cluster package

Greetings,

I have had good success using the clara() function to perform a simple cluster 
analysis on a large dataset (1 million+ records with 9 variables). 

Since the clara function is a wrapper to pam(), which will accept known medoid 
data - I am wondering if this too is possible with clara() ... The 
documentation does not suggest that this is possible. 

Essentially I am trying to implement a "supervised classification" of
numerous
geographic data layers. The "unsupervised" approach using clara()
works well,
but I feel the output classes would be more meaningful if I were able to let 
clara() know about the classes that I have in mind.

Is this at all feasible, or am I trying to accomplish something that is not 
possible?

Cheers,


-- 
Dylan Beaudette
Soils and Biogeochemistry Graduate Group
University of California at Davis
530.754.7341

>>>>> "DylanB" == Dylan Beaudette <dylan.beaudette
at gmail.com>
>>>>>     on Sun, 9 Apr 2006 19:28:44 -0700 writes:
    DylanB> Greetings, I have had good success using the clara()
    DylanB> function to perform a simple cluster analysis on a
    DylanB> large dataset (1 million+ records with 9 variables).

    DylanB> Since the clara function is a wrapper to pam(),
    DylanB> which will accept known medoid data - I am wondering
    DylanB> if this too is possible with clara() ... The
    DylanB> documentation does not suggest that this is
    DylanB> possible.

indeed, it doesn't --  because it's not yet possible.
I (as maintainer of "cluster") had added the ``known medoid''
option to pam() a while ago last June (for  cluster version 1.10.0),
and had left a note my TODO file to do the same for clara().

Unfortunately it's not true that clara() was a wrapper to pam()
as you state above.

Given your wish and clear "use case" situation, I'm more
motivated to approach this particular 'TODO' item!

Martin Maechler, ETH Zurich

    DylanB> Essentially I am trying to implement a "supervised
    DylanB> classification" of numerous geographic data
    DylanB> layers. The "unsupervised" approach using clara()
    DylanB> works well, but I feel the output classes would be
    DylanB> more meaningful if I were able to let clara() know
    DylanB> about the classes that I have in mind.

    DylanB> Is this at all feasible, or am I trying to
    DylanB> accomplish something that is not possible?

Hello Martin,

I am not sure that PAM works with pre-defined medoids. The avg.width is
always the same and when i check the medoids that were used for the
calculation they are the same as before, i.e. before defining my own
medoids.

Please advise
BAHKO

--
View this message in context:
http://r.789695.n4.nabble.com/passing-known-medoids-to-clara-in-the-cluster-package-tp797643p4632089.html
Sent from the R help mailing list archive at Nabble.com.