1
$\begingroup$

I perform clustering of time series into k=[2,N] number of clusters by using either DTW+kmedoids or DTW+single linkage+hierarchical clustering (HC), as advised in a previous post: Dynamic Time Warping Clustering

Regarding the evaluation of optimum number of clusters, I want to use expectation–maximization (EM) Gaussian Mixture Models (GMM) and determine the k that maximizes the log-likelihood for each approach.

My questions are:

  • What should be the input dataset in EM/GMM? The DTW similarity distance or the raw dataset?
  • Can the initializations in EM/GMM be the clusters centers of either the kmedoids or randomly selected seeds from the HC clusters?
  • Should I run k-fold cross-validation of the input dataset with the EM/GMM and return the average log-likelihood value?
$\endgroup$
  • $\begingroup$ In fact..based on this paper you can cluster time series with GMM/EM. I am working on how to implement with the well-known scikit-learn tools however. For an evaluation of hte optimal number of clusters, BIC/AIC are what you want to observe. These are also in the library for convenience. The link contains a great summary of how to use GMM/EM on certain data but not time series unfortunately. I'd like to hear of your updates as well. $\endgroup$ – HSL Mar 11 '18 at 15:29
1
$\begingroup$

Gaussian Mixture Modeling assumes ypur input data are coordinates in $R^d$ and contain Gaussian-shaped clusters. I don't think you can use this on time series. Also, GMM is a clustering approach on its own, and I don't see how you could use that to evaluate other clusterings.

$\endgroup$

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.