MIFair: Mutual Information Framework for Fair ML
MIFair serves as a comprehensive framework for evaluating and reducing bias in machine learning, grounded in mutual information principles. It tackles issues related to intersectionality, multiclass scenarios, and adaptability. The framework includes a metric template and an in-processing mitigation strategy, drawing from Prejudice Remover, which characterizes group fairness as the statistical independence between variables derived from predictions and sensitive attributes. It also creates links with fairness concepts such as independence and separation, while accommodating intersectionality, intricate subgroup dynamics, and multiclass classification through regularization-based training techniques.
Key facts
- MIFair is a unified framework for bias assessment and mitigation.
- It is based on mutual information.
- Addresses intersectionality, multiclass settings, and flexibility.
- Provides a flexible metric template.
- Uses an in-processing mitigation method inspired by Prejudice Remover.
- Defines group fairness as statistical independence between prediction-derived variables and sensitive attributes.
- Establishes equivalences with independence and separation fairness notions.
- Supports intersectionality, complex subgroup structures, and multiclass classification.
Entities
—