Comparing Euclidean distances with dynamic time warping (DTW): Will Euclidean distance perform better than DTW when clustering time series that all have the same length and sampling interval? Are there arguments in favor of DTW for this special case?
$\begingroup$
$\endgroup$
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
It depends on what you mean by performance - in terms of computing power DTW suffers from computational lethargy and Euclidean distance method dosnt. DTW is superior when it comes to classification and clustering.
Euclidean distance, which assumes the ith point in one sequence is aligned with the ith point in the other, will produce a pessimistic dissimilarity measure. The non-linear Dynamic Time Warped alignment allows a more intuitive distance measure to be calculated.
see Making Time-series Classification More Accurate Using Learned Constraints
I have used both DTW and Euclidean
Happy clustering !!!!
$\begingroup$
$\endgroup$
It totally depends on if you want your clustering to be invariant to dilation in time. For example if you have a speech signal we often want to be invariant to the speed at which people speak. If you're clustering music maybe you wouldn't want to use DTW.
$\begingroup$
$\endgroup$
You can check this great answer here: https://stats.stackexchange.com/a/22228/193114.
In general, it depends on the application you are working on. Euclidean distance is better than DTW when temporal alignment is not what you want. Euclidean distance is better than DTW when you wish to group time series that behave exactly the same at each time.
