[Sigia-l] Cluster analysis is dead - was Jakob wrong?

[Todd Warfel]

Trent,

Cluster analysis and card sorting is used to define "patterns." It's 
not definitive, nor is it the silver bullet. It's a great tool that 
offers insight into a suggested structure based on real user input.

We've been using Excel to do our cluster analysis in most cases. It's 
not as robust an analytical tool as EzSort, but EzSort simply doesn't 
provide what we need (e.g. run on a Mac).

In our experience, the ability to get down to three levels of 
granularity (e.g. parent, child, sub-child) is sufficient (to address 
item #1 below). So, keep that in mind when the "programmer" starts 
running the "what about this" speech. Don't make it any more 
complicated than it needs to be. There's little (no) need to program 
for edge cases here. Don't get sucked into that.

As to point #2, well, that's where the patterns come into play. The 
cluster analysis should reveal patterns of year vs. genre sorting. And 
then it's up to the IA to determine which method is best. It's entirely 
possible that this leads us to the decision that the default is by 
genre and an alternative method of search/browse is to use by year.

So, it's not that important that more than one pattern is revealed. 
What is important is that the software/method/tool used has the 
capability to reveal these patterns.

On Oct 7, 2004, at 9:50 PM, Trent Mankelow wrote:

> 1) When a similarity matrix is constructed, the process that creates 
> the
> matrix only captures information about cards that were grouped 
> together.
> If a person in the test decides they want to nest some cards below some
> others, the parent child information is lost in the matrix 
> construction.  That
> makes the technique blind to the test subjects card hierarchies.
>
> 2) The similarity matrix and resulting clusters represent an average.
> The card sorts may capture more than one fundamental 'mode', for 
> example
> the top level classifications could be year versus genre in a card sort
> about movies.  If the modes are fundamentally opposed, such as year vs
> genre, then the kind of average you would get out of cluster analysis
> with all the sorts included could be a non representative average, ie
> junk."

Cheers!

Todd R. Warfel
Partner, Design and Usability Specialist
MessageFirst | making products easier to use
--------------------------------------
Contact Info
voice: 	(607) 339-9640
email: 	twarfel at messagefirst.com
web: 	www.messagefirst.com
aim: 		twarfel at mac.com
--------------------------------------
In theory, theory and practice are the same.
In practice, they are not.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/enriched
Size: 2857 bytes
Desc: not available
Url : http://mail.asis.org/mailman/private/sigia-l/attachments/20041007/82b931dd/attachment.bin