I want to do clustering using Gaussian Mixture modeling (GMM) on a set of data which is a 5-dimension vector of real values $(x_1,x_2,x_3,x_4,x_5)$. However the clustering result were pretty bad, despite the fact that the cluster seems very clear to human.

After verifying inital parameters and prior settings, I found a possible reason why I wouldn't get a good result. The gaussian mean learned by the GMM are the actually mean I want, however the variance does not represent the true variance of the cluster. It always gives a very small value.

After checking the distribution of the input, I think the reason why a small variance was givin by the Training was due to the fact that my data are pretty "discrete", there were a lot of huge peaks in it. Those input are indeed result of a linear transformation happended on a interger input.

Example of the pdf of part of the data on all the dimensions for all classes: Image

I was wondering if there is anyway for me to transform data look like that where there are so much discrete peaks, in order to use it in GMM in a better way? Or perhaps I should try other method?


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.