Re: K-means and SOM


Greg Heath wrote:
> Thiru wrote:
> > Hi,
> >
> > I would like to know, in what way clustering achieved by K-means is
> > different from that of SOMs.
>
> Impossible to answer. In general, there are several K-means objective
> functions and tens of versions of K-means algorithms. Some are
> deterministic and some are probabilistic (especially w.r.t. initial
> conditions). Each one will yield a different configuration.
>
> Sorry, I don't have MATLAB on this machine, so I can't get specific
> about the options available with the MATLAB version.
>
> If you are looking for a 2 or 3 dimensional visualization
> of the data, Kohonen recommends SOM with a gradually shrinking
> neighborhood. Otherwise he recommends K-means, which is
> equivalent to SOM without neighborhood shrinking.
>
> Personally, I have found faster visualizations by first obtaining
> clusters using K-means, then obtaining a 2 or 3 dimensional
> visualization using Sammon's algorithm on the cluster centers.
>
> You'll have to search for that algorithm, I don't have it anymore.

I just went to Google Groups and got 57 hits on

sammons algorithm

Hope this helps.

Greg
> If you are in a hurry, a linear approximation to the nonlinear results
> of SOM and Sammon can be obtained by just projecting the
> data on the dominant principal component plane. However, for
> complicated data, the results may be very different.
>
> Hope this helps.
>
> Greg

.

Relevant Pages

  • k-means clustering and data type
    ... clustering: Is it all right that if k has the same number like OTUs ... then k-means will work like UPGMA? ... Which data type needs k-means cluster algorithm? ...
    (sci.bio.misc)
  • Re: NP-hardness of k-means clustering
    ... > I checked the compendium first of all, ... This one is the closest to k-means. ... clustering algorithm as implied in the *references*, ... Eray Ozkural ...
    (comp.theory)
  • Re: NP-hardness of k-means clustering
    ... > I checked the compendium first of all, ... This one is the closest to k-means. ... clustering algorithm as implied in the *references*, ... Eray Ozkural ...
    (sci.math)
  • Re: normalized cross correlation and kmeans
    ... I have to implement k-means clustering on using the normalized cross ... correlation as a measure of distance of similarity. ... Better to use single-link clustering. ... How does one use normalized cross correlation in pdist function??? ...
    (comp.soft-sys.matlab)
  • normalized cross correlation and kmeans
    ... I have to implement k-means clustering on using the normalized cross ... I have 60 random points as data, such that they are a unit distance ... How does one use normalized cross correlation in pdist function??? ...
    (comp.soft-sys.matlab)