.. -- mode: rst --

|pypi_version|_ |pypi_downloads|_

.. |pypi_version| image:: https://img.shields.io/pypi/v/shap-selection.svg .. _pypi_version: https://pypi.python.org/pypi/shap-selection/

.. |pypi_downloads| image:: https://pepy.tech/badge/shap-selection/month .. _pypi_downloads: https://pepy.tech/project/shap-selection

===== SHAP-Selection: Selecting feature using SHAP values

Due to the increasing concerns about machine learning interpretability, we believe that interpretation could be added to pre-processing steps. Using this library, you will be able to select the most important features from a multidimensional dataset while explaining your decisions!

To use SHAP-Selection, you will need:

• SHAP <https://github.com/slundberg/shap>_

Instalation

.. code:: python

   pip install shap-selection

Citation

.. code:: bibtex

   @INPROCEEDINGS{MarcilioJr2020shapselection,  
     author={W. E. {Marcílio} and D. M. {Eler}}, 
     booktitle={2020 33rd SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI)},   
     title={From explanations to feature selection: assessing SHAP values as feature selection mechanism},   
     year={2020},  
     pages={340-347},  
     doi={10.1109/SIBGRAPI51738.2020.00053}
   }

Usage

To use SHAP-Selection, you must have a trained model. It works both for classification and regression purposes!

Load a dataset

.. code:: python

   iris_data = load_iris()

   X, y = iris_data.data, iris_data.target
   feature_names = np.array(iris_data.feature_names)

   X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)

Fit a model

.. code:: python

   model = cb.CatBoostClassifier(verbose=False)    
   model.fit(X_train, y_train)

Use SHAP-Selection

.. code:: python

   from shap_selection import feature_selection

   # please, use agnostic = True to use with any model...
   # agnostic = False will only work with tree-based models
   feature_order = feature_selection.shap_select(model, X_train, X_test, feature_names, agnostic=False)

Support

Please, if you have any questions feel free to contact me at [email protected]