1
$\begingroup$

Is there a formulation of DTW that allows to deal with uncertainty or to utilize knowledge of the probabilistic structure of the signal? I would like to have a formulation similar to HMM but stay in DTW paradigm.

Why do I need that? I have several examples of a time series pattern that I want to find in the data. The amount of examples is not enough to build a complex model (e.g. HMM). On the other hand, I would like to leverage all the examples given at once and not use these examples one after another. I am wondering if there is DTW formulation that allows to account for that.

$\endgroup$
1
$\begingroup$

If you want to "leverage all the examples given at once", then you can optimally average all the patterns [a] into a single template and use that.

However, you should not do that if the patterns are polymorphic.

A good resource for DTW is [b]

[a] http://www.cs.ucr.edu/~eamonn/ICDM_2014_DTW_average.pdf [b] http://www.cs.unm.edu/~mueen/DTW.pdf

$\endgroup$
  • $\begingroup$ Yes, I would like "leverage all the examples given at once" but averaging might be too risky, since yes, template are similar but sometimes might have high variation. So, I am searching for a method that would provide me some uncertainty on every points when performring DTW. $\endgroup$ – fractile Jan 10 '18 at 9:06

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.