trelawney 0.3.1

trelawney

Trelawney is a general interpretability package that aims at providing a common api to use most of the modern interpretability methods to shed light on sklearn compatible models (support for Keras and XGBoost are tested).

Trelawney will try to provide you with two kind of explanation when possible:

• global explanation of the model that highlights the most importance features the model uses to make its predictions globally

• local explanation of the model that will try to shed light on why a specific model made a specific prediction

The Trelawney package is build around:

• some model specific explainers that use the inner workings of some types of models to explain them:

  • LogRegExplainer that uses the weights of the your logistic regression to produce global and local explanations of your model

  • TreeExplainer that uses the path of your tree (single tree model only) to produce explanations of the model

• Some model agnostic explainers that should work with all models:

  • LimeExplainer that uses the Lime package to create local explanations only (the local nature of Lime prohibits it from generating global explanations of a model

  • ShapExplainer that uses the SHAP package to create local and global explanations of your model

  • SurrogateExplainer that creates a general surogate of your model (fitted on the output of your model) using an explainable model (DecisionTreeClassifier,`LogisticRegression` for now). The explainer will then use the internals of the surrogate model to explain your black box model as well as informing you on how well the surrogate model explains the black box one

Quick Tutorial (30s to Trelawney):

Here is an example of how to use a Trelawney explainer

>>> model = LogisticRegression().fit(X, y)
>>> # creating and fiting the explainer
>>> explainer = ShapExplainer()
>>> explainer.fit(model, X, y)
>>> # explaining observation
>>> explanation =  explainer.explain_local(X_expain)
[
    {'var_1': 0.1, 'var_2': -0.07, ...},
    ...
    {'var_1': 0.23, 'var_2': -0.15, ...} ,
]
>>> explanation =  explainer.graph_local_explanation(X_expain.iloc[:1, :])

>>> explanation =  explainer.feature_importance(X_expain)
{'var_1': 0.5, 'var_2': 0.2, ...} ,
>>> explanation =  explainer.graph_feature_importance(X_expain)

History

0.1.0 (2019-10-02)

• First release on PyPI.