Re: [S] Combining Image and cluter-tree plots side by side

Jae K. Lee (
Thu, 21 May 1998 14:33:18 -0400

First, thank Patrick for answering me with an intersting idea.
I also feel to clarify what I meant by "side-by-side cluster-tree and
image plots."
Data mining of a huge database, in my opinion, is often done by the
combination of statistical and heuristic approaches. In our case we want
to visualize highly-multidimensional data into some cluster-oredered
image plot; the random or original ordered image plots, of course, don't
reveal any interesting patterns. Then, the question is now how we interactively
view the image plot with the entries ordered by a cluster algorithm.
We currently use the hierarchical cluster algorithm (hclust function).
We can make a cluster-tree with legible entry labels up to 2-300 drug
or gene components. Then, considering an image plot produced with this
cluster order, one naturally is interested in reading off entry lables
of some interesting areas of the image plot. If this is one time task,
we just print them out (still with quite a number of difficulties
to compare one by one) and
cross-compare the image and cluster-tree plots. But, when we need to do the
same task many more multiple times, the story is different, and it is
desirable to make those two plots matched together.
The idea Patrick suggested, making a pointwise box, is actually implemented
in the pc-based "Noesys" software and presumably many others.
I believe this would be one of the nice gadgets that Splus developers
need to consider in the next generation of the product.
Hope to get more of your response.


Jae Lee

On May 21, 6:27pm, Patrick Burns wrote:
> Subject: Re: [S] Combining Image and cluter-tree plots side by side
> Jae K. Lee wrote:
> > We are heavily using Splus, mostly on SGI Unix, due to memory and speed
> > problems, even though we have all new versions of pc-based Splus4.0 and
> > We found that using CLUSTER-TREE and IMAGE plots side-by-side, we can
> > many intersting information from such a large matrix.
> > However, there are several technical difficulties:
> First off, let me say that this is the sort of question that I think there
> be moreof on the list.
> >
> >
> > 1. Since we cann't read off labels of such a large number of entries,
> > and since we often find some intersting local areas of the image plot,
> > we hope to be able to zoom in or restrict to some ares of the image plot
> > from
> > our eye or some heuristic checking. Then, regenerate the restricted
> > and cluster-tree plots side-by-side (I know some other graphical
> > such as AVS and Noesys, provide such a capability, but not Splus).
> I doubt that there is a clean way to do this (though I stand to be
> have a vague plan for a kludgy way to do it. You plot your full image,
> then click on two spots that you want to be the upper left and lower right
> corners of the new image -- "locator" is used for this. Then you have the
> coordinates of the points, which you can use to subset the matrix and
> draw a new image. So I think that the function that used "locator" would
> only need to have the matrix as input.
> >
> >
> > 2. We want to match the sizes of image and cluster-tree plots, so that we
> > can direclty read off what are the entries column-by-column or
> > of the image and cluster-tree plots
> > (in this case we hope to have a way even for small matrices like 200 X
> >
> >
> I don't understand what it is that you are really doing here. I suggest that
> youresubmit a question to S-news with a more detailed explanation.
> Good luck,
> Pat Burns
>-- End of excerpt from Patrick Burns

Jae K. Lee, Ph.D.                         Tel:  (301) 496-9572
National Cancer Institute                 Email:
National Institutes of Health              
Bldg. 37, Room 5D-02
9000 Rockville Pike
Bethesda, MD 20892                        Fax:  (301) 402-0752
This message was distributed by  To unsubscribe
send e-mail to with the BODY of the
message:  unsubscribe s-news