DM: SOMs

[Warren Sarle]

I should have changed the subject line in my previous reply, which had
the extremely uninformative title of "Re: DM: My introduction".

I had asked Daniel X. Pape:
> > Why do you use SOMs? Why not ordinary VQ? What properties of your
> > textual and image collections do you expect the dimensions of the
> > SOM grid to reveal?

Dan replied:
> Well, what I specifically do in my research group is create SOMs in
> order to automatically categorize a data collection to allow the 
>user
> to _browse_ the collection. Once the SOM is created, I am using it 
>to
> create 2D and 3D interfaces to allow the user to graphically browse 
>the
> collections. Another way I am using them is to automatically 
>categorize
> a search result set for easy subsequent browsing - for example, if 
>you
> do a search on AltaVista you might get 2000 results... a SOM could
> categorize the results so the user could easily pick the one or two
> hundred most relevant results.

That's an interesting application that I had not thought of before.
I can see how a SOM could be quite useful for browsing. But I still
wonder if the dimensions of the SOM grid correspond to any 
interpretable
properties of the data. 

Do you have some way of automatically labelling the clusters, or do
you do it manually? Or do you have some other way of guiding the user
around the SOM?


> >From what I understand of vector quantization methods, there are 
>two
> reasons why I don't use them: One, for the most part, they are
> _supervised learning_ methods.

The terminology in this area can be quite confusing. "VQ" actually
refers to unsupervised methods. I just added an answer to the neural
net FAQ on this topic; see "How many kinds of Kohonen networks exist?"
in ftp://ftp.sas.com/pub/neural/FAQ.html .


> > I have tried to use the WEBSOM application at
> > http://websom.hut.fi/websom/comp.ai.neural-nets/html/root.html
> > to search for articles in comp.ai.neural-nets, and I found it 
>quite
> > useless. Dejanews works far better.
> 
> The results you got were different because the two tools you used
> (WEBSOM and DejaNews) were designed for two different things. It
> depends on how you were searching. If you were searching
> comp.ai.neural-nets for a specific term or author or article, then 
>of
> course DejaNews will be better - DejaNews uses very powerful and 
>very
> fast search methods - but at their heart, they are just simple 
>string
> matching methods.

Initially I was looking for articles on hardware--keywords such as
"hardware", "implementation", "VLSI"--and didn't find anything at all.
As an experiment, I then tried looking for genetic algorithms, and I
found a few articles scattered around the grid. In the course of 
further
aimless wandering, I discovered that various other topics seemed to be
widely scattered, including the topic of SOMs. Generally, the articles
were not clustered in a way conducive to either searching or browsing.

Of course, it would be much easier to search for topics if the SOM
were better labelled. In this particular example, there are large
regions of the grid with no labels at all. 

I hope your SOMs are more successful than WEBSOM!

-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not 
necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
* Do not send me unsolicited commercial, political, or religious 
email *