I'm a newbie in cluster analysis so please excuse me if my question seems to be very basic. I'm using SPSS and Matlab for performing cluster analysis in a variety of datasets. Dendograms are great for visualising the results. However, they are not helpful for comparing the results produced by different clustering methods or algorithms. Is there a standard way to do this? Does anybody had use a Database for this purpose? I'm kind of proficient in Access-VBA, but I'm lost in what data I should store (resemblance matrix, cophenetic matrix, etc) and how to transform them in a comprehensive database matrix.
$\begingroup$
$\endgroup$
1
-
$\begingroup$ I have merged your two accounts, Diego. You'll need at least 50 reputation points to be able to leave comment everywhere on this site. $\endgroup$– chlOct 1, 2012 at 12:55
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
9
Wikipedia has a section dedicated to comparing ("evaluating") clustering results:
https://en.wikipedia.org/wiki/Cluster_analysis#Evaluation_of_clustering_results
However, these methods will essentially give you some number.
Cluster analysis is an explorative method, as such you'll have to investigate the discovered structure manually, whether it is more sensible than the structure you knew before, or not. One might say that there is no "good" or "bad" in cluster analysis, only "new" or "already known". But there is "new and useful" and "new but uninteresting" IMHO.
There is some work in the "multi view" clustering community on how to compare clusterings. If you want to generate a diverse clustering, you have to measure similarity.
See e.g. MultiClust 2011: http://dme.rwth-aachen.de/en/MultiClust2011#prog
answered Oct 1, 2012 at 8:53
-
$\begingroup$ Thank you Anony-Mousse for the links... they look very interesting and I'll read them through. Regarding my data: I don't know the "structure" of the data before hand. This is one of the aims I'm pursuing by using hierarchical clustering. However, it seems very unpractical to have to perform all this job manually. Can't believe there is no way to store the results in a simple database for further comparisons. I mean, similarity/dissimilarity matrices are just "calculations", and the linkage table represents the structure of the data. Are you completely sure that the only way of doing this is b $\endgroup$– DiegoOct 1, 2012 at 10:29
-
$\begingroup$ If you know the structure before, you don't need to run cluster analysis at all... what do you expect to find then, if you know the result already? $\endgroup$ Oct 1, 2012 at 13:19
-
$\begingroup$ Btw, Dendrograms only work for a very simple case of clustering methods, namely linkage clustering. $\endgroup$ Oct 1, 2012 at 13:20
-
$\begingroup$ Yeap Anony-Mousse, that's what I'm saying... I don't know the structure of the data. And yes, I'm using hierarchical agglomerative clustering. But that's not my question. So let me rephrase it: How can I use the numeric results of a cluster analysis in order to rebuild the dendogram results in a table(s) or matrix(ces) manner that can be suitable for being used in a database following the relational model, with the aim of further comparing the results of different methods, metrics, etc? $\endgroup$– DiegoOct 1, 2012 at 14:30
-
$\begingroup$ Well, only hierarchical clustering will actually give you a Dendrogram, so you will not be able to compare anything except hierarchical clustering. And well, I don't use SPSS, so I don't know if it can serialize and deserialize dendrograms. There is little use in storing it in anything else but a binary blob I guess, as you'll be analyzing full dendrograms, not parts. $\endgroup$ Oct 1, 2012 at 15:31
$\begingroup$
$\endgroup$
I made a mistake by calling the "measure" I want to get steps. I meant heights. It seems that Matlab can provide me what I'm looking for, by using the inconsistent function:
Now using the right "term" my google search returns results. It seems I don't have to put too much effort in getting the precise values of the heights according to the next quote:
Few investigators place much reliance on the precise values of the heights in a dendrogram, partly because of the difficulties referred (...); and patly because there is sometimes interest in comparing dendrograms obtained by analyzing the same data set by more than one method of analysis, when the heights in the different dendrograms may not be commesurate.
Thank you very much Anony-Mousse for your help
All the best,
Diego