Dear all;
I used SOM and K-Means clustering methods in MATLAB to cluster my data. After the clustering, is there any function that I can use to test which of them has the higher performance since there is no prior knowledge about the data?
Thank you.
Hi Musa,
Unlike classification methods, comparing clustering methods is a critical issue that can be accomplished by using information of the class attribute. Precisely, converting clustering problem into a classification one, where the clusterer is a base method of the utillized classifier algorithm.
HTH.
Samer
.
part of the answer is in your question : "the higher performance" ... if you use those words, you must have some idea of what kind of "performance" you are interested in for your particular application ; for instance :
do you need a good quantization error ?
(k-means is better than som, for the same number of centers ; the difference is often not too high but stems from the additional topological constraint in som)
do you need a good visualization ?
(go for a som)
do you need to get a "holistic view" of your data ?
(go for a som)
do you need a fast algorithm ?
(go for k-means)
such questions (and the ranking between them in terms of importance) can only be answered from your research perspective
.
You may want to test first the performance on time consumption, for this use the matlab function timeit(). I believe thought, that you want to track the performance on the quality of the clusters, i.e. to quantify the error. For this, in case of K-means for example, one option is to compute the relative distance, of all cluster instances to the initial cluster centroids (sum-up the intra-cluster error), in other words, sum-up the squared distances from data-points to their cluster-center and divide each summand by the number of elements in the cluster, respectively. In case of the SOM, the cluster centroids are the prototype vectors. I hope this helps.
In unsupervised learning such as clustering you never can tell for sure
but you can get an idea of the correctness of tthe solution by computing several measures;
silhouet index
Gamma-W, Gamma-B, Gamma-WB (Huizinga)
AIC and BIC tables
the best test as validation is:
compute statistics on external variables, variables of history, present and foture/outcome/prognosis
test the info in the real world: ask people in practice if this makes sense
succes
Eddy
- Badges
- Science topic
- Similar topics
- Instrumentation Engineering
- Instruments