BUG: predict_parts with type="shape_wrapper" does not work with scikit pipelines

[nilslacroix]

I found a bug or missing feature with scikit pipelines in dalex.

When you try to create an explainer with predict_parts and type="shap_wrapper like this:

sh_mary = exp.predict_parts(mary, type='shap_wrapper', B = 10, label=mary.index[0])

It causes an error. This bug is reproducible with your titanic example.
I would be thankful for any workaround.

The error message is:

EDIT:

Furthermore if you take a look at the explainer output it seems that numpy.ndarrays do not work properly
but the error message literally states that it expects an array:

  -> predict function  : Accepts only pandas.DataFrame, numpy.ndarray causes problems.

TypeError                                 Traceback (most recent call last)
Input In [9], in <cell line: 2>()
      1 import shap
----> 2 sh_mary = exp.predict_parts(mary, type='shap_wrapper', B = 10, label=mary.index[0])

File ~\miniconda3\envs\Master_ML\lib\site-packages\dalex\_explainer\object.py:320, in Explainer.predict_parts(self, new_observation, type, order, interaction_preference, path, N, B, keep_distributions, label, processes, random_state, **kwargs)
    317 else:
    318     raise TypeError("Wrong type parameter.")
--> 320 _predict_parts.fit(self, new_observation, **kwargs)
    322 if label:
    323     _predict_parts.result['label'] = label

File ~\miniconda3\envs\Master_ML\lib\site-packages\dalex\wrappers\_shap\object.py:71, in ShapWrapper.fit(self, explainer, new_observation, shap_explainer_type, **kwargs)
     49 """Calculate the result of explanation
     50 
     51 Fit method makes calculations in place and changes the attributes.
   (...)
     67 None
     68 """
     69 from shap import TreeExplainer, DeepExplainer, GradientExplainer, LinearExplainer, KernelExplainer
---> 71 checks.check_compatibility(explainer)
     72 shap_explainer_type = checks.check_shap_explainer_type(shap_explainer_type, explainer.model)
     74 if self.type == 'predict_parts':

File ~\miniconda3\envs\Master_ML\lib\site-packages\dalex\wrappers\_shap\checks.py:48, in check_compatibility(explainer)
     46 if 'arrays_accepted' in explainer.model_info:
     47     if not explainer.model_info['arrays_accepted']:
---> 48         raise TypeError("'predict_function' not compatible with the 'shap' package")
     49 else:
     50     # data was None but now is available
     51     try:

TypeError: 'predict_function' not compatible with the 'shap' package

[hbaniecki]

So this is basically due to shap not working with scikit pipelines. If you run the following code in the titanic example:

from shap import TreeExplainer
TreeExplainer(exp.model, exp.data)

you obtain

File {PATH}\shap\explainers\_tree.py:979, in TreeEnsemble._
_init__(self, model, data, data_missing, model_output)
    977     self.base_offset = model.init_params[param_idx]
    978 else:
--> 979     raise Exception("Model type not yet supported by TreeExplainer: " + str(type(model)))
    981 # build a dense numpy version of all the tree objects
    982 if self.trees is not None and self.trees:

Exception: Model type not yet supported by TreeExplainer: <class 'sklearn.pipeline.Pipeline'>

I guess I could only improve the error message of TypeError: 'predict_function' not compatible with the 'shap' package to explain this behaviour.

[nilslacroix]

Damn it. I guess the only working solution would then be to call the preprocessor on X, use the estimator and transformed X in shap.explainer and then try to make an inverse transform of the computed shap values, if they scale right.

[hbaniecki]

TreeSHAP estimates explanations based on the structure of the model, which relies on variables and their distributions during train time. I have no clear intuition if inversing the transformation (in all unimaginable possible variations) would provide reasonable explanation results.