Welcome to BEExAI documentation!

BEExAI is a Python library for benchmarking explainability methods on tabular data. It supports a wide range of explainability methods and evaluation metrics. It is designed to be easy to use and to allow fast obtention of benchmark results.

Major features include:

• Automatic preprocessing of tabular data

• Training of several models including scikit-learn and PyTorch Neural Network models.

• Computation of attributions for explainability methods from Captum

• Computation of evaluation metrics for explainability methods for robustness, faithfulness and complexity

Contents

Introduction

• Installation
  • Pre-requisites
  • Install from PyPI
  • Install from source
• Usage
  • Setup the configuration file
  • Notebooks
  • Load data and train a model
  • Compute explaination attributions and evaluation metrics
• Technical details
  • Supported models
  • Supported explaination methods
  • Evaluation metrics
• Metrics
  • Complexity
  • Faithfulness
  • Robustness

Examples

• Step by step
  • How to train a model on any dataset
  • Get explainability scores
  • Get explainability metrics
• Specific cases
  • End to end Neural Network classification
  • End to end XGBoost regression
• Benchmark
  • Download datasets from OpenML
  • Benchmark results

API Reference

• Add custom objects to the API
  • Add custom explainer
  • Add custom metric
  • Examples
• beexai
  • beexai package

GitHub repository https://github.com/SquareResearchCenter-AI/BEExAI