enter image description here

How to select number of clusters using GMM when the elbow test (AIC/BIC vs n_components) results in a graph like this?

  • 6 seems to be what you are looking for, if you are looking for a test there are a few, for ex. using clusGap statistic. – user2974951 Sep 25 at 8:51
up vote 4 down vote accepted

Welcome to CV!

This plot shows how the AIC and BIC change as a function of the number of clusters. While the AIC continuous to decrease with a larger number of clusters, you can see that the BIC stops decreasing after $k=6$ clusters. For this reason, you could choose $k = 6$.

Another way to choose the 'best' number of clusters is by considering the elbow(s) of the figure. The elbow of a function is a point after which the decrease becomes notably smaller. An elbow is a heuristic, so there is no exact way to determine which value best describes this point. For example, one could argue that the AIC & BIC both stop decreasing as much after $k = 5$ clusters, while someone else might argue that this is after $k = 6$ clusters. You could even argue that the biggest decrease has already happened after $k = 2$ clusters.

Lastly, you don't have to choose any number of clusters just because AIC/BIC/whatever suggested you do so. If you have some a priori reason to assume that there should be $k = 3$ clusters, then that might be a better choice.

In short: An elbow in this context is a heuristic guide to decide the number of clusters if you have no other reason to assume a certain number of clusters.

Your Answer

 
discard

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.