Bias-inducing geometries: An exactly solvable data model with fairness implications.

Area of Science:

• Statistical Physics
• Machine Learning
• Data Science

Background:

• Machine learning (ML) models can inherit and amplify human biases present in training data.
• Understanding the mechanisms of bias inheritance is crucial for developing fair AI systems.

Purpose of the Study:

• To investigate the role of data geometry in the emergence of ML bias.
• To analytically characterize bias inheritance and evaluate mitigation strategies.

Main Methods:

• Development of an exactly solvable high-dimensional model of data imbalance.
• Application of statistical physics tools to analyze ML model properties.
• Evaluation of loss-reweighing and coupled learning model strategies.

Main Results:

• Analytical predictions for fairness assessment observables derived from the model.
• Identification of incompatibilities between different fairness criteria.
• Demonstration that coupled learning models achieve superior fairness-accuracy trade-offs.

Conclusions:

• Data geometry significantly influences ML bias.
• Existing fairness criteria can be incompatible.
• Proposed coupled learning strategy offers an effective approach to mitigate bias in ML models.