xai init

[dumjax]

Explanability: Shap with Regression models #609

Summary

First draw of what should be explainability in darts. We use shap library to explain time series, in the case of RegressionModel.

[brunnedu]

Really cool addition to darts! 🚀 Added a few comments, may look at it in more detail later.

[hrzn]

It looks like a great start and I really like the result :) I also like the structure with separate explainer objects of different kinds living in the darts.explainability submodule.

A couple of points:

• I know it's because it's still a draft, you will have to pay attention to the arguments for each methods. I saw many cases that fail currently (when some argument is None, or not provided for instance).
• You should rethink a bit what goes in the super-class vs the implementation (I think the super-class should not contain much logic except what is very general and common to all present and future explainers).
• Writing unit tests will help you catch some of the issues.
• Also beware of the listing, you should follow the setup here
• Are you sure you need the class RegressionShapExplainers? I'm wondering if it brings something, as to me it looks like an extra layer and some extra code in a place where we don't really need the encapsulation, as its state is more or less the same as ShapExplainer. But I might be missing something.

[codecov-commenter]

Codecov Report

Base: 93.98% // Head: 93.82% // Decreases project coverage by -0.16% ⚠️

Coverage data is based on head (0dee794) compared to base (2988059).
Patch coverage: 90.90% of modified lines in pull request are covered.

❗ Current head 0dee794 differs from pull request most recent head 4915d32. Consider uploading reports for the commit 4915d32 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #909      +/-   ##
==========================================
- Coverage   93.98%   93.82%   -0.17%     
==========================================
  Files          73       78       +5     
  Lines        8200     8505     +305     
==========================================
+ Hits         7707     7980     +273     
- Misses        493      525      +32     

Impacted Files	Coverage Δ
darts/utils/data/encoders.py	98.88% <ø> (ø)
darts/explainability/explainability_result.py	80.00% <80.00%> (ø)
darts/explainability/shap_explainer.py	88.09% <88.09%> (ø)
darts/explainability/explainability.py	96.42% <96.42%> (ø)
darts/explainability/__init__.py	100.00% <100.00%> (ø)
darts/models/forecasting/regression_model.py	98.17% <100.00%> (-0.15%)	⬇️
darts/utils/data/tabularization.py	100.00% <100.00%> (ø)
darts/utils/timeseries_generation.py	96.15% <100.00%> (ø)
darts/timeseries.py	92.34% <0.00%> (-0.06%)	⬇️
...arts/models/forecasting/torch_forecasting_model.py	87.08% <0.00%> (-0.06%)	⬇️
... and 2 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[hrzn]

It looks great overall @dumjax , nice one!

I have some comments, but mostly unimportant things. The only somewhat important things IMO are

• Adding unit tests
• Fixing the case where no past or future covariates are given
• Improving the docstrings in a few places (I've made small suggestions, but I think we can improve the explanations and formatting in a few places)
• Factoring out the X/y computation as discussed

I'm really looking forward for us to release it 🚀

[hrzn]

I feel we're 95% there, but there are still a few important things to adapt and to polish before we can merge.
I've made lots of small comments, mainly about the documentation (pls double-check how it renders in HTML before merging), variable naming (I think target_names is confusing, we should use target_components) and also having the horizons start at 1 instead of 0 to be consistent with our forecasting models.

[dennisbader]

Looks good, thanks a lot for this!

[dennisbader]

I merged my PR with encoder updates yesterday :) If you apply the two suggestions (with the other suggestion), everything should be up to date!

Suggested change

	) = self.model.encoders.encode_train(
	target=background_series,
	past_covariate=background_past_covariates,
	future_covariate=background_future_covariates,
	)
	) = self.generate_fit_encodings(
	series=background_series,
	past_covariates=background_past_covariates,
	future_covariates=background_future_covariates,
	)