[R] model-based question-better readable version

eewwaaww at interia.pl eewwaaww at interia.pl
Thu Jul 5 15:54:43 CEST 2007

It is going to be easy question to you. I've  started to interest in model-based clustering.

Adrian E. Raftery "Recent Advances in Model-Based Clustering: Image Segmentation and Variable Selection" (www.stat.washington.edu/Raftery)showed that we can compare different classification methods using BIC
statistic. For diabetes dataset the best model is VVV model with 3 classes- for this model the BIC curve reaches the highest value and the error rate=12%
BIC curve for EII model~k-means is much under the VVV model curve and the error rate equals 18%, so k-means (EII) is worse then VVV, what is clear for me.

I would like to apply model-based to economic data set (GDP, life expectancy data of UE countries), because I am PhD  student of University of Economics in Poland.
Using this data (standardized) I get the best model EEV (2 classes), EII (k-means) curve is under EEV curve what suggests that k-means is worse then EEV, but class error for EII equals 0 and for EEV= 6% (and more for another variables), why?

Even applying  iris data we get lower class error for EII model (10%)  than for VEV (33%) for 2 classes,   in spite of that VEV model and others models curves are above EII model at the BIC plot.

For this data BIC chooses VEV for 2 clusters while the right number of classes, given in
column "Species"

My second question is: when model-based clustering (for which data sets, are there any special type of data) is better than k-means (kmeans), hierarchical clustering(hclust)?

I am looking forward to hearing from you.       

Best regards, 

Rozdajemy bilety na koncert
więcej na >> http://link.interia.pl/f1ae9

More information about the R-help mailing list