Dynamic time warping 1: Motivation

Link to full playlist on DTW: • Dynamic time warping (...

[@mohamedrefaat197]

Beautiful! Thanks for this "series".

[@kamperh]

Very happy you enjoy it! :)

[@mahyar24]

As a Data Scientist, I'm a long-time youtube learner who watched hundreds of hours of DS content. THIS WAS ONE OF THE BESTS! It reminded me of @statquest. Thanks a lot!

[@kamperh]

Thanks a ton for the encouragement! :D

[@richsajdak]

Fantastic job! This is one of the best explanations of DTW I've seen

[@benjavides]

Great explanation and visualizations, thanks!

[@manishakowdiki8261]

nice explaination

[@julianray6802]

Best explanation I have come across!

[@shrur3527]

thank u so much

[@TheMangz1611]

could you share python code to do this?

[@codewithyouml8994]

How we could generate these types of simulations?

[@halbarads5406]

Hi, thanks for the video. I am working with a database where I have the total monthly customer spend over 5 years on a product. My goal is to try to see which customers tend to increase or decrease their consumption at the same time by similar magnitudes. For that I am making clusters with the variable time and with the variation in consumption with respect to the previous month. For this I used k-means using the Euclidean distance as a measure. Do you consider that this is okay or should I use dynamic time warping as a distance measure? Greetings and thanks.

[@kamperh]

When choosing a distance metric, the crucial question is what you would like to do. Different distance metrics makes different underlying assumptions. So in DTW, for instance, the metric is explicitly developed so that, if the same pattern occurs at different points within the signal, you will still get a low distance (high similarity). From what you are saying, I don't think you actually want this, since you have the phrase "consumption at the same time", although I'm not entirely sure what that means. With Euclidean distance, you will always be comparing the first time-point of one consumer to the first time-point of another; in some cases this might be exactly what you want to do, where in other cases this might be exactly the opposite of what you want to do. Long answer to basically say: there is never one distance metric that will give you best performance in all cases. DTW is really good at some things, but not everything. Hope that helps! :)

[@halbarads5406]

@@kamperh thanks for your answer, i'll be waiting for more content :D

[@SarahSanchez-b2w]

Allen Shirley Miller Elizabeth Thomas Linda

[@amvsekai]

Clear and nice explanation.. liked it ❤

[@1ssbrudra]

I think it is important that we stop calling it a "distance metric" because its mathematically inaccurate, its a measure at best. www.cs.unm.edu/~mueen/DTW.pdf

[@afraaalobaid8256]

That's very helpful! Thank you very much

[@kamperh]

Huge pleasure! :)

[@user-gs9ze8us4b]

flawless explanation👍🏻

[@nschweiz1]

Great video series! The algorithm video was the one that finally got me to "get" DTW!

[@kad3nce]

Very helpful. Thank you for sharing! A question I have… I am looking for an algorithm to compare a human drum performance captured over midi to a click track. Answering the first question you pose in this video would provide a kind of performance accuracy score (e.g. “how closely was the drummer able to re-create the signal generated by the click track?”). And answering the second question posed in your video, perhaps while also presenting data on how far offset the corresponding points are, could give a drummer insight as to whether they are behind or in front of the beat, and by how much, etc. DTW seems close to what I need, except that the algorithm seems to be dependent on the sampling rate matching between the two signals. In this case, the human drummer may add or drop notes altogether relative to the click track, and thus end up with a different number of samples that likely never precisely line up with the click track at any point in time. Could you recommend a way to process these two times series that would enable me to answer the two questions you posed in this video?

[@CarlosMagno-uz8de]

Such a great explanation. The best on the topic that I have come across!

[@xuxu4068]

This video helps me a lot, thank you

[@kamperh]

Very happy it helped!! :D

[@afwanherucahya]

Does DTW could do a forecast or predictions?

[@darh78]

What a great explanation of DTW!

[@diegocanales]

Exelent video!👏

[@stat511]

Thank you for sharing

[@AntonKeyter]

Herman die kwaliteit van hierdie produksie oor DTW is uistekend, wel gedaan!! Selfs die kode wat jy deel help my om DTW in C++ skryf. Hierdie playlist help my om natuurlike herhalings van patrone in the FX mark te identifiseer. Baie dankie, doen asb meer sulke produksies, jy't werkik 'n talent. - (Hoop jy kry hierdie boodskap 😀)

[@kamperh]

Dankie Anton! Dis so lekker om so boodskap in Afrikaans te kry! :D

[@Quizbeebuz]

who , you really look similar to Mark Zuckerberg in every single way ! great explanation btw

[@kamperh]

Don't know if that's a complement or an insult ;) But happy you enjoyed the explanation! :)

[@udomatthiasdrums5322]

love it!!

[@AB-bp9fi]

Very good.

[@jonathanz9889]

Sorry if this might be offensive but do you happen to be color blind? In the wikipedia GIF the blue curve you mention is actually purple.

[@alexmatt4012]

Another great tutorial starting with an unnecessary "so".