[MRG] add TimeSeriesScaleMeanMaxVariance()

[tonylee2016]

When I was using the package for K-shape, one issue found is that the original TimeSeriesScaleMeanVariance may amplify small changes (noises) in the signal. Because that it normalizes each signal per its own variance.

Here I am adding a new scaling class that preserves the amplitude relation across signals, this scaling method may be useful for some applications.

TimeSeriesScaleMeanMaxVariance is similar to the existing TimeSeriesScaleMeanVariance with one exception, the max std of all signals is used for normalization. As a result, the amplitude relation across signals is preserved.

[tonylee2016]

the method won't pass check_methods_subset_invariance, It is expected, I've excluded it from the test.

[GillesVandewiele]

Split up over multiple lines.

[GillesVandewiele]

And add space to right and left of ==

[GillesVandewiele]

Are these comments still relevant? Does it need to be stateful? I.e. should there be a fit() and transform() where in the fit() the self.std is stored and used in the transform() method?

[GillesVandewiele]

Hi,

Thank you for this PR! I think this could indeed be useful for some use cases, but I will let @rtavenar confirm this.

I left 2 very minor remarks. I checked out the check_methods_subset_invariance from sklearn and it indeed makes sense that the results change depending on the sub-batch you provide to the method (as it uses the max std of that batch to normalise). However, we must ensure that only the relevant tests are skipped, and that as many as possible tests are performed. Are we sure that adding _skip_test as a tag does not skip more tags than needed?

[GillesVandewiele]

One other thing that we could consider is, instead of making this part of the main codebase, to create an example/tutorial "defining a custom preprocessor" and then using this specific preprocessing technique. I can imagine that there are more alternatives to normalise out there, and they only have minimal differences. Including all of those into tslearn could quickly result in a bloated code-base.

[rtavenar]

Hi @tonylee2016
I think this is closely related to #280 and #285

The scalers we have implemented so far only rely on per-time series information. And we seem to have more and more requests like this one.

I would be OK to have such scalers in tslearn but I strongly believe that we should put effort in a clear documentation of the differences between scalers.

Also, maybe it would be worth it to have names that explicitly state that these scalers rely on global information, or have a subpackage dedicated to "global" scalers.

I am open to suggestions on these points and would like to clarify that before considering the add of these preprocessors in tslearn.

[tonylee2016]

Hi, @GillesVandewiele and @rtavenar

thanks for the comments and extra info. After reading #280 and #285, here are my thoughts on a more generic scaler framework

1. move the std and mu calculation to fit(). As [WIP] Fit method at preprocessing file #280 #285 is already doing it.
2. add an optional scaler Infos gathered from fit(): so that user know what data the scaler is based on
3. better document and an example for TimeSeriesScaleMeanMaxVariance() to show the different effects

I can spend some time on 3 in this PR. But 2 and 1 may be handled by #285?

@GillesVandewiele _skip_test will skip all estimator checks. However, I did not find a way to only skip check_methods_subset_invariance.

[codecov-commenter]

Codecov Report

Merging #333 (28ea436) into main (87495c1) will decrease coverage by 0.16%.
The diff coverage is 28.57%.

❗ Current head 28ea436 differs from pull request most recent head 2819398. Consider uploading reports for the commit 2819398 to get more accurate results

@@            Coverage Diff             @@
##             main     #333      +/-   ##
==========================================
- Coverage   94.70%   94.54%   -0.17%     
==========================================
  Files          59       59              
  Lines        4511     4525      +14     
==========================================
+ Hits         4272     4278       +6     
- Misses        239      247       +8     

Impacted Files	Coverage Δ
tslearn/preprocessing/__init__.py	100.00% <ø> (ø)
tslearn/preprocessing/preprocessing.py	89.10% <28.57%> (-9.75%)	⬇️
tslearn/tests/sklearn_patches.py	92.59% <0.00%> (+0.52%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 87495c1...2819398. Read the comment docs.