Add support to Imbalanced-learn pipelines in ClassifierExplainer

When I try to generate a `ClassifierExplainer``on a imblearn pipeline I get the following error:

TypeError: All intermediate steps should be transformers and implement fit and transform or be the string 'passthrough'
'SMOTETomek(random_state=42)' (type <class 'imblearn.combine._smote_tomek.SMOTETomek'>) doesn't

Full traceback:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-75-2a92cf12a19d> in <module>
      1 from explainerdashboard import ClassifierExplainer, InlineExplainer
----> 2 explainer = ClassifierExplainer(best_model, X_test, y_test)

/opt/conda/lib/python3.7/site-packages/explainerdashboard/explainers.py in __init__(self, model, X, y, permutation_metric, shap, X_background, model_output, cats, cats_notencoded, idxs, index_name, target, descriptions, n_jobs, permutation_cv, cv, na_fill, precision, labels, pos_label)
   1999                             cats, cats_notencoded, idxs, index_name, target,
   2000                             descriptions, n_jobs, permutation_cv, cv, na_fill,
-> 2001                             precision)
   2002 
   2003         assert hasattr(model, "predict_proba"), \

/opt/conda/lib/python3.7/site-packages/explainerdashboard/explainers.py in __init__(self, model, X, y, permutation_metric, shap, X_background, model_output, cats, cats_notencoded, idxs, index_name, target, descriptions, n_jobs, permutation_cv, cv, na_fill, precision)
    138             if shap != 'kernel':
    139                 pipeline_model = model.steps[-1][1]
--> 140                 pipeline_transformer = Pipeline(model.steps[:-1])
    141                 if hasattr(model, "predict") and hasattr(pipeline_transformer, "transform"):
    142                     X_transformed = pipeline_transformer.transform(X)

/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py in inner_f(*args, **kwargs)
     61             extra_args = len(args) - len(all_args)
     62             if extra_args <= 0:
---> 63                 return f(*args, **kwargs)
     64 
     65             # extra_args > 0

/opt/conda/lib/python3.7/site-packages/sklearn/pipeline.py in __init__(self, steps, memory, verbose)
    116         self.memory = memory
    117         self.verbose = verbose
--> 118         self._validate_steps()
    119 
    120     def get_params(self, deep=True):

/opt/conda/lib/python3.7/site-packages/sklearn/pipeline.py in _validate_steps(self)
    169                                 "transformers and implement fit and transform "
    170                                 "or be the string 'passthrough' "
--> 171                                 "'%s' (type %s) doesn't" % (t, type(t)))
    172 
    173         # We allow last estimator to be None as an identity transformation

TypeError: All intermediate steps should be transformers and implement fit and transform or be the string 'passthrough' 'SMOTETomek(random_state=42)' (type <class 'imblearn.combine._smote_tomek.SMOTETomek'>) doesn't

Versions System: python: 3.7.10 | packaged by conda-forge | (default, Feb 19 2021, 16:07:37) [GCC 9.3.0] executable: /opt/conda/bin/python machine: Linux-5.4.120±x86_64-with-debian-buster-sid

Python dependencies: pip: 21.1.2 setuptools: 49.6.0.post20210108 sklearn: 0.24.2 imblearn: 0.8.0 explainerdashboard: latest numpy: 1.19.5 scipy: 1.6.3 Cython: 0.29.23 pandas: 1.2.4 matplotlib: 3.4.2 joblib: 1.0.1 threadpoolctl: 2.1.0

[oegedijk]

I’m not very familiar with these types of pipelines, but seems you could do something like this to get the transformed input dataframe, and then extract the model:

X_transformed = pd.DataFrame(
     make_pipeline(*[step[1] for step in pipeline.steps[:-1]]).transform(X_test),
     columns=X_test.columns
 )
 model = steps[-1][1]

explainer = ClassifierExplainer(model, X_transformed, y_test)

actually maybe I could add some support for this…

[Abdelgha-4]

Ah I see, this is regrettable as I would love to exploit your works for imbalanced data models. Is there any work around for this in the time being ?