Welcome to the Shapash documentation!#
Shapash is a Python library that makes machine learning models interpretable and understandable by everyone. Its plots and outputs use explicit labels for features and their values, so data scientists can understand their models and share the results with people who don’t work with data. Shapash is developed by data scientists at MAIF.
Company |
GitHub Account |
Website |
|---|---|---|
|
How Shapash works, in a few lines of code.
Install Shapash and its optional extras.
Notebooks for every feature.
Classes, methods and parameters.
What Shapash does#
Clear, labeled explanations. Global and local explainability plots display readable feature names and values instead of technical ones.
A web app to explore your model. Navigate from global to local explainability, filter subsets on explanatory or additional features, and compare correct and wrong predictions. See the live demo.
Summarized local explanations. Keep only the most influential features of each prediction, and export the summary to a pandas DataFrame:
Explainability quality metrics. Evaluate how far you can trust an explanation with three metrics: stability, consistency and compacity.
An audit report. Generate a standalone HTML report that records the data, the model and its explainability, as a basis for an audit document. See the example report.
From exploration to deployment. The SmartPredictor object brings predictions and summarized
explanations to production, through an API or in batch mode.
Text classifiers (preview). NlpExplainer explains text classifiers at the word level, with its
own web app, what-if analysis and label-noise detection.
Compatibility#
Regression, binary classification and multiclass classification.
Most scikit-learn, LightGBM, CatBoost and XGBoost models.
Contributions computed with SHAP or LIME, or provided directly, for example from another explainability method.
Preprocessing with category_encoders, a scikit-learn
ColumnTransformer, or a simple dictionary of feature names.
Shapash is released under the Apache 2.0 license.