This library provides a framework for building interactive dashboards to explain the inner workings of machine learning models, targeting data scientists and non-technical stakeholders. It simplifies the process of visualizing model performance, feature importances, and individual predictions, making complex AI models more transparent and interpretable.

How It Works

The library leverages Plotly Dash to create interactive web applications. It integrates with popular ML libraries (scikit-learn, XGBoost, LightGBM, CatBoost, Skorch) and calculates explainability metrics like SHAP values, permutation importances, and partial dependence plots. Its modular design allows for custom dashboard layouts and the aggregation of multiple dashboards into an "ExplainerHub."

Quick Start & Requirements

• Install via pip: pip install explainerdashboard or conda: conda install -c conda-forge explainerdashboard.
• Supports scikit-learn, xgboost, catboost, lightgbm, and skorch models.
• Official documentation: explainerdashboard.readthedocs.io
• Example notebook: launching dashboards

Highlighted Details

• Interactive exploration of model performance, feature importances, SHAP values, and "what-if" scenarios.
• Supports visualization of individual decision trees for tree-based models.
• Can export dashboards to static HTML for easy sharing and CI/CD integration.
• Modular design allows for custom layouts and integration into larger ExplainerHubs.

Maintenance & Community

• Developed by Oege Dijk.
• Documentation available at explainerdashboard.readthedocs.io.
• Example deployed dashboard: titanicexplainer.herokuapp.com

Licensing & Compatibility

• License: Not explicitly stated in the README.

Limitations & Caveats

Calculations for SHAP interaction values and permutation importances can be slow for large datasets; options exist to disable or approximate these. Memory usage can be substantial for large datasets, with suggestions for optimization like using float32 precision or external data storage.