[Sigia-l] RE: Cluster analysis is dead - was Jakob wrong? (TryMDS)

[Boniface Lau]

> From: Peter Merholz
>  
> I just wrote about it on my site, w/r/t to cluster analysis and card
> sorting:
> http://www.peterme.com/archives/000402.html

In that article, you wrote:

PM> Concern has been raised on the list about whether cluster analysis
PM> is sufficient -- it produces a single hierarchical presentation of
PM> the concepts, an analysis that, frankly, attempts to meet that
PM> lowest common denominator. I think we can learn from the packaged
PM> foods industry that such an approach falls short.
[...]
PM> It's also yet another data point that we have to get away from
PM> hierarchical structures to more faceted ones 

According to:

http://www.mathpsyc.uni-bonn.de/doc/delbeke/delbeke.htm

MS> Cluster analysis models or ultrametric tree models, are equally
MS> applicable to proximity data including two-way (asymmetric) square
MS> and rectangular data as well as three-way two-mode data. The main
MS> difference with the MDS models is that most models for cluster
MS> analysis lead to a hierarchical structure. 
[...]
MS> A comparative study with both real and simulated data (see Everitt
MS> and Rabe-Hesketh, 1997) showed that data with an underlying
MS> hierarchical structure result in a better fit when using tree
MS> models while data with an underlying spatial structure result in a
MS> better fit when using multidimensional scaling models.  

To determine whether multidimensional scaling (MDS) or cluster
analysis is more appropriate, people should look into the underlying
information structure.


PM> -- allowing users to make their own "hierarchy" as they move
PM> through our information spaces.

Not sure what you meant by that. Could you give an example?


Boniface