# Minimax Pareto Fairness: A Multi Objective Perspective

**Authors:** Natalia Martinez, Martin Bertran, Guillermo Sapiro  
**arXiv:** [2011.01821](https://arxiv.org/abs/2011.01821)  
**Date:** 2020-11-03  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2011.01821v1/

## Key points

- Authors Natalia Martinez, Martin Bertran, and Guillermo Sapiro explore group fairness as a multi-objective optimization problem
- Proposed novel fairness criterion aims to minimize risk while being Pareto-efficient for all sensitive groups
- Approach avoids unnecessary harm and can produce the best zero-gap model based on policy requirements
- Simple optimization algorithm presented, compatible with deep neural networks and does not require access to sensitive attributes during testing
- Method demonstrated effectiveness in real case studies across various domains such as income prediction, ICU patient mortality prediction, skin lesion classification, and credit risk assessment
- Results show approach outperforms existing methods in terms of both fairness and accuracy

## Summary

In their research paper "Minimax Pareto Fairness: A Multi Objective Perspective," authors Natalia Martinez, Martin Bertran, and Guillermo Sapiro explore the formulation and formal characterization of group fairness as a multi-objective optimization problem. They propose a novel fairness criterion where a classifier aims to minimize risk while also being Pareto-efficient for all sensitive groups. This approach not only avoids unnecessary harm but also has the potential to produce the best zero-gap model based on policy requirements. To implement their framework, the authors present a simple optimization algorithm that is compatible with deep neural networks. Importantly, this method does not require access to sensitive attributes during testing, making it suitable for reducing worst-case classification errors in unbalanced scenarios. The effectiveness of this methodology is demonstrated through real case studies covering various domains such as income prediction, ICU patient mortality prediction, skin lesion classification, and credit risk assessment. The results highlight how their approach outperforms existing methods in terms of both fairness and accuracy. Overall, "Minimax Pareto Fairness: A Multi Objective Perspective" offers an innovative and comprehensive perspective on addressing group fairness concerns in machine learning models using multi-objective optimization techniques.