Minimax Pareto Fairness: A Multi Objective Perspective.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Ensuring fairness in machine learning models is crucial, especially in sensitive applications.
• Existing fairness metrics often involve trade-offs and may require sensitive attribute information during prediction.
• Unbalanced classification problems exacerbate fairness challenges.

Purpose of the Study:

• To formulate group fairness as a multi-objective optimization problem.
• To propose a novel fairness criterion that achieves minimax risk and Pareto efficiency.
• To develop an optimization algorithm applicable to deep neural networks without requiring test-time sensitive attributes.

Main Methods:

• Formulating group fairness as a multi-objective optimization problem with separate objectives for each sensitive group's risk.
• Proposing a fairness criterion based on minimax risk and Pareto efficiency.
• Developing a compatible optimization algorithm for deep neural networks.

Main Results:

• The proposed method achieves minimax risk and Pareto efficiency across all sensitive groups.
• The framework effectively reduces worst-case classification errors in unbalanced datasets.
• Demonstrated favorable comparisons against existing approaches in real-world case studies.

Conclusions:

• The novel multi-objective optimization framework provides a robust approach to group fairness.
• The method's ability to avoid test-time sensitive attribute access broadens its applicability.
• The approach offers a promising direction for developing equitable and high-performing machine learning models.