FairML
A Python toolbox for auditing machine learning models to detect and quantify bias in black-box predictions.
What is FairML?
FairML is a Python toolbox that audits machine learning models for bias by quantifying the relative significance of input features. It helps analysts determine if a black-box predictive model exhibits unintentional discrimination based on protected characteristics like race or gender. The toolbox is designed to assess fairness in models deployed in critical domains such as credit scoring, insurance, and employment.
Data scientists, machine learning engineers, and analysts who build or deploy predictive models in regulated or high-impact domains and need to audit for fairness and compliance. Researchers focused on algorithmic fairness and responsible AI will also find it valuable.
FairML provides a practical, model-agnostic toolkit that works with any black-box predictor, making fairness auditing accessible without requiring model interpretability. Its unique combination of model compression and multiple ranking algorithms offers a robust methodology to quantify and visualize discriminatory patterns.
Overview
FairML is an end-to-end Python toolbox designed to audit predictive models by quantifying their relative dependence on input features, enabling assessment of potential discrimination. It addresses the critical need to audit black-box models deployed in high-stakes domains like credit, insurance, and employment, where unintentional bias based on race, gender, or other protected characteristics can have serious societal impacts.
Key Features
- Model Auditing — Quantifies the relative significance of a model's inputs to assess fairness and discriminatory extent.
- Black-Box Compatibility — Works with any predictive model that provides a predict function, without requiring internal model details.
- Multiple Ranking Algorithms — Leverages four input ranking algorithms and model compression to measure predictive dependence.
- Bias Visualization — Generates feature dependence plots to visually communicate audit results.
- Interaction Analysis — Optional flag to check model dependence on feature interactions.
Philosophy
FairML is built on the principle that transparency and accountability are essential for responsible AI, providing tools to systematically audit models for fairness without requiring interpretability of the underlying algorithm.
Related Projects
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
