I know that there are several methods (Elbow method, silhouette analysis etc) which can help me to find the best number of clusters.

My questions are:

  • What are the best methods to use in Expectation Maximization (EM) Algorithm?

  • Can I use elbow method (as if it were for K-means algorithm) to get the "best" number of clusters, but to use in EM Algorithm?

Thanks in advance.

  • $\begingroup$ "Elbow" or SSwithin criterion is a raw, unnormalized source of more advanced indices such as CH (Calinski-Harabasz) or DB (Davies-Bouldin). Read a brief outline at the bottom in this answer $\endgroup$ – ttnphns Feb 28 '20 at 19:09

The "elbow" method never works well...

Instead, use any of the other internal measures such as Silhouette, Dunn, CH Index, BIC, AIC, C index, etc. that are much better and should also work for EM.

If you really want to stick with the worst method (Elbow), you could try to look for an elbow in the score function of EM too. Just as with k-means we'd expect the loglikelihood to improve as you increase the number of clusters. There could be an elbow there too, where the rate of improval slows down.


Your Answer

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

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