Disregard NaNs in TimeSeriesScalerMeanVariance

[eliwoods]

If a time series with any NaN values is passed to TimeSeriesScalerMeanVariance.fit_transform() the transformed time series returns as all NaN

E.g:

In [1]: from tslearn.preprocessing import TimeSeriesScalerMeanVariance                                                                

In [2]: from numpy import nan                                                                                                         

In [3]: TimeSeriesScalerMeanVariance().fit_transform([0, nan, 6])                                                                     
Out[3]: 
array([[[nan],
        [nan],
        [nan]]])

This could be fixed by using numpy.nanmean() and numpy.nanstd() in TimeSeriesScalerMeanVariance.transform instead of the current implementation. This also makes tslearn in line with sklearn in terms of NaN's and preprocessing: scikit-learn/scikit-learn#10404

[codecov-io]

Codecov Report

Merging #175 into dev will decrease coverage by 0.28%.
The diff coverage is 76.19%.

@@            Coverage Diff             @@
##              dev     #175      +/-   ##
==========================================
- Coverage   93.97%   93.69%   -0.29%     
==========================================
  Files          22       22              
  Lines        3039     2585     -454     
==========================================
- Hits         2856     2422     -434     
+ Misses        183      163      -20

Impacted Files	Coverage Δ
tslearn/__init__.py	75% <100%> (ø)	⬆️
tslearn/preprocessing.py	91.93% <100%> (ø)	⬆️
tslearn/tests/test_shapelets.py	100% <100%> (ø)	⬆️
tslearn/tests/test_estimators.py	89.28% <58.33%> (-5.59%)	⬇️
tslearn/tests/sklearn_patches.py	88.6% <0%> (-2.85%)	⬇️
tslearn/shapelets.py	91.2% <0%> (-2.58%)	⬇️
tslearn/tests/test_metrics.py	100% <0%> (ø)	⬆️
tslearn/tests/test_barycenters.py	100% <0%> (ø)	⬆️
tslearn/svm.py	98.43% <0%> (ø)	⬆️
... and 6 more

---

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 40cf0c9...2996607. Read the comment docs.

[rtavenar]

Hi @eliwoods

Thanks for the suggestion, this makes a lot of sense.

I would suggest the following improvements:

• Do the same for TimeSeriesScalerMinMax
• Add doctest examples to make the behaviour explicit for end-users
• Change the destination branch to dev such that your work can be included in the next tslearn version to come

Once all this is done, one last thing would be to document your contribution in CHANGELOG.md.

[pep8speaks]

Hello @eliwoods! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

• In the file tslearn/preprocessing.py:

Line 114:80: E501 line too long (87 > 79 characters)