comp.lang.idl-pvwave archive: archive

Home » Public Forums » archive » clustering

Show: Today's Messages :: Show Polls :: Message Navigator
E-mail to friend

Re: clustering [message #54959 is a reply to message #54850]

Mon, 23 July 2007 11:22

Vince Hradil
Messages: 574
Registered: December 1999

Senior Member

On Jul 23, 12:14 pm, little davey <dave-kel...@cox.net> wrote:
> Is it the case that you MUST use standardize() before you call
> CLUST_WTS()? The documentation does not say so, but I suspect from
> the code, and, I actually tried it with the initial poster's data, and
> got 4 clusters (he wanted 5, but this would appear to be a tough data
> set to cluster, as variables "2" and "3" are close to each other).
>
> As I posted an hour ago, part of the source code for CLUST_WTS() is:
>
> ;Normalized uniformly random cluster weights.
> ;;Weights = RANDOMU(SEED, N_Variables, N_Clusters) + Zero
> av1 = average(array[0,*])
> av2 = average(array[1,*])
> Weights = RANDOMU(SEED, N_Variables, N_Clusters) + Zero
> for k = 0L, N_Clusters-1 do Weights[*,k] = $
> (Weights[*,k] / TOTAL(Weights[*,k])) * Variable_Wts
>
> However, the variables AV1 and AV2 are NOT USED ANYWHERE IN THE CODE,
> so I suspect that the data is not "normalized" correctly in
> CLUST_WTS. The use of STANDARDIZE() may be necessary for CLUST_WTS to
> work.
>
> -- Dave K --

No, you just have to be careful about your data. Here's an example:

IDL> a=[[3,55],[4,54],[8,55],[9,56]]
IDL> plot, a[0,*], a[1,*], psym=4

"obviously", there are two clusters.

IDL> wts= clust_wts(a,n_clusters=2)
IDL> print, wts
1.99307 27.5069
6.42612 55.0837
IDL> oplot, wts[0,*], wts[1,*], psym=5
(one of the points doesn't even show up, oh, and fix the xrange, too)
IDL> plot, a[0,*], a[1,*], psym=4, xrange=[0,60], yrange=[0,60]
(uh-oh, maybe there is only one cluster)
IDL> oplot, wts[0,*], wts[1,*], psym=5
IDL> print, cluster(a,wts,n_clusters=2)
1
1
1
1
D'oh!

NOW:
IDL> sa = standardize(float(a))
IDL> plot, sa[0,*], sa[1,*], psym=4
IDL> swts= clust_wts(sa,n_clusters=2)
IDL> print, swts
-0.833454 -0.669169
0.864960 0.669170
IDL> oplot, swts[0,*], swts[1,*], psym=5
IDL> print, cluster(sa,swts,n_clusters=2)
0
0
1
1
Tada!

The reason the first (unstandardized) doesn't work is that the overall
variance is much larger than the variance within each coordinate, so
the clusters get "attracted" (if you will) to one of the coordinates
and disregards the other.

You have to "stay in touch" with your data - Black boxes are okay, as
long as you know what's going on inside.

Report message to a moderator

[Message index]

		clustering By: nivedita.raghunath on Tue, 17 July 2007 08:36
		Re: clustering By: Vince Hradil on Thu, 19 July 2007 10:30
		Re: clustering By: nivedita.raghunath on Thu, 19 July 2007 09:12
		Re: clustering By: Vince Hradil on Thu, 19 July 2007 13:46
		Re: clustering By: Conor on Wed, 18 July 2007 12:07
		Re: clustering By: Vince Hradil on Mon, 23 July 2007 11:22
		Re: clustering By: James Kuyper on Mon, 23 July 2007 11:16
		Re: clustering By: little davey on Mon, 23 July 2007 10:14
		Re: clustering By: little davey on Mon, 23 July 2007 09:08
		Re: clustering By: Conor on Mon, 23 July 2007 06:49
		Re: clustering By: nivedita.raghunath on Mon, 23 July 2007 06:45
		Re: Clustering By: Kai Mühlbauer on Tue, 01 November 2011 03:01
		Re: Clustering By: Kai Muehlbauer on Mon, 31 October 2011 08:40
		Re: Clustering By: Jeremy Bailin on Tue, 01 November 2011 12:16
		Re: Clustering By: kidpix on Mon, 31 October 2011 04:56
		Re: Clustering By: Kai Mühlbauer on Fri, 28 October 2011 10:07
		Re: clustering By: kisCA on Wed, 07 March 2012 13:37
		Re: clustering By: Russell[1] on Wed, 07 March 2012 09:32

Previous Topic:	Re: Move data to structure fields without copying
Next Topic:	Annotate Uncertainty [Upper & Lower] on a plot?

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

]

Current Time: Fri Oct 10 05:23:06 PDT 2025

Total time taken to generate the page: 0.08313 seconds