Re: another attempt

[http://dummy.us.eu.org/robert (Robert)]

 > From: Leonid <http://www.gmail.com/~l>
 > Date: Wed, 4 May 2005 19:14:32 -0400
 >
 > I do not know about multiple max/min. What's the rigorous meaning?

I just mean looking for local minima/maxima versus global minima/maxima.
That's usually the difficult problem.

 > OC1 is a good idea if I understand it correctly since it looks for
 > slanted hyperplanes to separate the clusters and that's what it really
 > is all about.

Yes, that's what it's all about.  Split, split, split.

 > The only issue is that the clusters are quite close "at
 > 0". Question: do you know if OC1 allows for some errors (usually in
 > any classification problem there is a sparse cloud of errors around
 > the clouds -- the vacuum is not absolute)?

It probably does.  Marc's software tried to adjust the numbers so the
error rate wouldn't get too large; it was usually something we had to keep
in mind when running it.

 > Also -- how does OC1 look
 > for the candidate planes -- I couldn't make this out from their
 > website; I'll try to poke around more...

I'm looking at Marc's thesis right now.  He used the goodness function
used in "Clustering Algorithms" by Hartigan 1975.  I have no idea what OC1
itself used.

 > How can CART work? Doesn't it require some learning? It is a
 > supervised system i.e. I need to keep telling them what right and
 > what's wrong...

I have never used CART.  I've only seen it compared with Bayes and it
compares favorably.  It also is on-par with OC1, supposedly.

 > It should be some min/max formula. Same concern about
 > neural networks.

Never used neural nets.

 > Am I mistaken?

I don't know.  I don't know about CART; never downloaded the software,
never investigated it thoroughly.  There's also C4.5, which you might look
at.

 > Thanks
 > 
 > L