Fairness and Accuracy Under Domain Generalization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Ethics

Background:

• Machine learning (ML) models face fairness concerns in high-stakes applications.
• Existing fairness methods assume identical training and deployment data distributions, which is often violated in practice.
• Previous research on domain generalization primarily focused on accuracy transfer, neglecting fairness.

Approach:

• This paper investigates the transfer of both fairness and accuracy under domain generalization, where test data may come from unseen domains.
• Theoretical bounds for unfairness and expected loss during deployment are developed.
• Sufficient conditions for perfect transfer of fairness and accuracy through invariant representation learning are derived.

Key Points:

• A novel learning algorithm is designed to ensure fair ML models maintain high fairness and accuracy across changing deployment environments.
• The algorithm is guided by theoretical findings on invariant representation learning.
• Experimental validation on real-world data confirms the algorithm's effectiveness.

Conclusions:

• The proposed approach enables robust and fair ML models that generalize well to new, unseen data domains.
• This work bridges the gap between ML fairness and domain generalization by considering both accuracy and fairness transfer.
• The developed algorithm offers a practical solution for deploying fair and accurate ML systems in dynamic environments.