Comments on: Clustergram: visualization and diagnostics for cluster analysis (R code)

By: Andreas Weidenhiller

Andreas Weidenhiller — Tue, 21 Dec 2010 11:12:59 +0000

Hi! I like this function and have two suggestions:

1. changing the noise generation command in the clustergram function to
noise <- unlist(tapply(line.width, clusters.vec, function(x){c=cumsum(x); c-mean(c)}))[order(seq_along(clusters.vec)[order(clusters.vec)])]
better arranges the lines around the center points

2. One use for color which I have found is this: If there is some grouping of the data which is known beforehand, one might want to mark the groups by these colors. (At least I did this – any one else interested?)

By: Ben Haller

Ben Haller — Fri, 17 Dec 2010 18:23:25 +0000

I’m liking using some alpha in the lines; I have a big dataset (300,000 points), and the alpha makes it easier to see how many lines go from one node to another.

By: Tal Galili

Tal Galili — Mon, 20 Sep 2010 15:13:53 +0000

Hello Deryl,

The function presented in the post offers kmeans with it’s default configurations.

From looking at

?kmeans

You could find that in the details section there is the answer:

The algorithm of Hartigan and Wong (1979) is used by default. Note that some authors use k-means to refer to a specific algorithm rather than the general method: most commonly the algorithm given by MacQueen (1967) but sometimes that given by Lloyd (1957) and Forgy (1965). The Hartigan–Wong algorithm generally does a better job than either of those, but trying several random starts is often recommended.

By: Tal Galili

Tal Galili — Mon, 20 Sep 2010 15:11:04 +0000

Hello Allan, Thank you for the offer. Personally, I’d prefer to allow a different solution of the function for different runs for (lazy) evaluation of the stability of the clustering solution.

BTW, you could also insure to always get the same result by using

some.number < - 666
set.seed(some.number)

Cheers,
Tal

By: Deryl

Deryl — Mon, 20 Sep 2010 15:05:22 +0000

Thanks for the email. I appreciate your response. Our team has addressed how to handle the missing values and we are looking over the resulting clustergrams. We are writing up a presentation for an upcoming conference, and so please entertain another query, if you don’t mind: Please let me know, what kind of kmeans algorithm does R implement here? For example, in Matthias Schonlau’s 2004 article (Computational Statistics, 19:95-111), he displays several clustergrams using SAS, Stata, and Splus implementations of random or deterministic kmeans algorithms (pp. 102-103). Which would be most similar to what is going on here with your R code? Thanks again for your insight.

By: Deryl

Deryl — Sun, 19 Sep 2010 06:38:19 +0000

Your presentation is very clear, even for someone with only basic knowledge of statistical programs and syntax. I have a very large data set (350,000+ observations) of national educational survey data that I need to cluster, and this implementation looks promising as a way to decide the number of clusters. It worked well for me using a small subset of the data for testing. But, how might I be able to work with missing data? When I include a larger sample data set with records with missing values, I get errors returned. Are there any solutions you could suggest? Any methods for filling in missing data that I should avoid? Thanks for any insight you can offer.

By: Silvana

Silvana — Thu, 17 Jun 2010 20:37:13 +0000

I appreciate the initiative you took to create this pattern in R.
I need to present a paper on this subject at university.
Do you know if the generation of clustergram is already implemented in R? Where can I find the mathematical calculations that are used in Clustergram?

Thanks,
Silvana

By: Tal Galili

Tal Galili — Thu, 17 Jun 2010 04:51:40 +0000

Cheers Ali

And for the others,
Martin refers to me complimenting his post here:
http://spiltmartini.com/2010/05/20/infographics/

By: Ali Martin

Ali Martin — Thu, 17 Jun 2010 00:07:49 +0000

Thanks for the comment on the infographic. I wish I could take credit for it. I don’t know much about clustergrams, but clustergram 6 reminds me of the brachial plexus (the plexus of nerves underneath the collarbone).

By: Tal Galili

Tal Galili — Wed, 16 Jun 2010 08:38:16 +0000

Hello Hadley,
Thank you for the code and suggestions.
I extended this post to link to the ggplot2 implementation you wrote.

With regards,
Tal