Probatus v3.0.0+ missing features & issues.

[ReinierKoops]

Issue Description
Probatus version 3.0.0+ removed some functionality that wasn't updated for multiple years, needed refactoring and was better done by other packages. When asking a sample of our users about this functionality, we noticed that as we suspected it wasn't used. Also with respect to maintainability, the fact it was out of sync with our other code and the potential to improve our core functionality in the future we chose to drop it.

If you make use of this package, but found that now some of its functionality that you previously used was removed in version 3.0.0, please report this here. Then we'll encourage you to use version 2.X.X in the meantime while we will look into reintroducing a modernized version of it. Also feel free to mention new ideas you might have.

[jeong2624]

I ran the function shap_elimination.fit_compute(X, y), but the error occurs on line 493 of feature_elimination.py.
ValueError: all the input array dimensions except for the concatenation axis must match exactly.

What should I do if I can't do it even though I fit X and y in data frames and series formats, respectively?

[ReinierKoops]

@jeong2624 could you please provide a code example and your probatus version?

[ReinierKoops]

• Update to latest Pandas, Numpy and SHAP is needed (2.0+)
• Try to simplify fit_params for ShapRFECV tree based models (lgbm, catboost and xcgboost should have updated ways of inserting parameters)
• Create a better toolset for unit tests (update and have sensible checks in place, have tests for many different kinds of setups: (regression, classification, ...), (single, multi, ...), (tree-based, linear, ....))
• Update the docs, notebooks, have more extensive examples for the code, perhaps add examples in the docstring as well
• Investigate if Scikit-learn pipelines are possible to be supported
• By default, one class (Label 0 or 1) was selected to be plotted. Now with multiclass support, it makes sense to give user control to also select the to be plotted class. This also means the shap_calc should return the whole shap_values such that when plotting the correct class can be selected.
• Create better plots / use the newer plots of SHAP