I have calaculated a SOM (with the kohonen package in R, 18x18 heaxagonal grid, 500 iterations, 92 variables, 1189 cases) and am currently trying to access it's usability. Calling a functions for topological error (nodedist) 1.354823. This as well as average quantization error of 10.34 indicate huge dissimilarities in my data. Increasing the grid lead to higher topological errors and using a 10x10 grid almost doubled the quantisation error. I have calculated one in WEKA with the same 18x18 setup and it outputs clusters and the means, SE of the variables. Additionally it includes a variable called value, which I believe to state the value of the variable in the cluster, maybe similar to the weight (?).

I am very new to this field as my data basically forced me into it (shiloette coeffiient for h.cluster did not 'find' any structure). Therefore I am now a bit unsure on what to do with the result, as most literature only deals with "this case worked". Can I still report it with these error measurements? Also, it gives me constellations of the variables within it's "clusters", can I still use them or should I rather treat that like a super-insignificant p value thing?

I am very grateful for help and have not included data examples, as my main question is how to deal with the above stated output properly.


PS: Although predictive value would have been a desirable outcome it was mainly computed in order to observe if different groups map together at the same spot in such a map. Which they do not (already know how to report that) but my main concern is, could I still use the clusters WEKA generated?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.