Alibi is a Python library for explaining machine learning models, targeting ML engineers and researchers. It provides implementations of black-box and white-box, local and global explanation methods, enabling deeper model understanding and debugging.

How It Works

Alibi offers a scikit-learn-inspired API with initialize, fit, and explain steps. It supports a wide range of explanation techniques, including Anchors, Integrated Gradients, KernelSHAP, and Counterfactuals, catering to tabular, text, and image data. The library distinguishes between black-box (prediction function only) and white-box (model access required) methods, offering flexibility in model compatibility.

Quick Start & Requirements

• Install via pip: pip install alibi
• Optional dependencies: alibi[ray] for distributed computation, alibi[shap] for SHAP support.
• Supports TensorFlow/Keras models for specific methods.
• See Documentation for detailed examples.

Highlighted Details

• Comprehensive suite of explanation methods: ALE, Partial Dependence, Anchors, CEM, Counterfactuals, Integrated Gradients, SHAP, Trust Scores, and more.
• Supports tabular, text, and image data modalities.
• Includes methods for model confidence and prototype generation.
• Offers both local (instance-specific) and global (dataset-wide) explanations.

Maintenance & Community

• Developed by Seldon.io.
• See GitHub for contributors and activity.

Licensing & Compatibility

• Source-available license.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

Some methods require specific model types (e.g., TensorFlow/Keras) or additional dependencies like Ray for distributed explanations. The breadth of supported models varies per explanation technique.