[ENH] wrap/implement Global Alignment DTW

[TonyBagnall]

Describe the feature or idea you want to propose

there is an interesting paper here, that describes an extension of DTW
https://arxiv.org/pdf/2002.03848.pdf
described here
https://rtavenar.github.io/hdr/parts/01/dtw/dtw_gi.html

Describe your proposed solution

It might already be in tslearn, in which case we can maybe wrap it, although might be better to port it in for numba reasons?

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[SalmanDeveloperz]

Hi Maintainers,

I am interested in contributing to this issue. I wanted to confirm if the Global Alignment DTW feature is already implemented in tslearn. If not, would you prefer it to be wrapped from an existing implementation or ported directly for performance optimization using numba?

I am happy to start working on this and would appreciate any guidance or additional context.
Thanks

[baraline]

Hi @SalmanDeveloperz,

By philosophy, we prefer to have our own implementation instead of warper when it's possible/reasonable. I think @chrisholder would be best suited to guide you on the details this issue and the state of the distance module.

Meanwhile, I'm sure you can start with some the implemented dtw inside the distance module and try to make the modification pour global alignment in a copy of it. That should get you started !

[tanishy7777]

By philosophy, we prefer to have our own implementation instead of warper when it's possible/reasonable. I think @chrisholder would be best suited to guide you on the details this issue and the state of the distance module.

Hi, Can I work on this? It seems like no one is working on it?

[tanishy7777]

@aeon-actions-bot assign @tanishy7777

[tanishy7777]

I wanted to check in regarding the implementation of DTW with Global Invariances (DTW-GI). I’ve gone through the paper (link) and also (link).

As suggested in #444 (comment) I will start implementing DTW-GI from scratch.

Since I’m new to Numba, I’d appreciate any guidance or recommended resources on the same. In the meantime, I'll first focus on getting a working implementation without Numba, then optimize it afterwards to use Numba.

@chrisholder @baraline @TonyBagnall

[baraline]

Sounds good, I would advice you to take a look at the other dtw implementation in aeon to have an idea of how to use numba when you come to that step.

[tanishy7777]

Also another small thing I found while going through the documentation for dtw

Shouldnt $i \in [1, s)$ instead of $i\in(1, m)$ ?

Sounds good, I would advice you to take a look at the other dtw implementation in aeon to have an idea of how to use numba when you come to that step.

[chrisholder]

Hi @tanishy7777 I am not 100% sure what you mean but this is what that comment is meant to explain:

• s is the length of the alignment path
• n is the length of the first time series
• m is the length of the second time series.
• A individual path value i.e. (5, 3) where the first value (5) is in the time series of length n and the second value (3) is in the time series of length m. The individual values (i.e. 5) cannot be > than the length of the time series (i.e. n or m). Hence 1 < i < m.

I maybe wrong and I think some of these docs need to be revisted as they don't read very well imo.

Regarding the global alignment DTW let me know if you need any help with anything best place to reach me is on the slack.