PFERM: A Fair Empirical Risk Minimization Approach with Prior Knowledge

Area of Science:

• Machine Learning
• Artificial Intelligence
• Biomedical Informatics

Background:

• Ensuring fairness in machine learning is critical to prevent biased predictions based on sensitive attributes.
• Strict fairness often leads to reduced accuracy, especially with prevalence disparities, limiting practical applications.
• Group prevalence differences, like higher Alzheimer's disease rates in women, necessitate tailored fairness approaches.

Purpose of the Study:

• To develop a machine learning framework that integrates prior knowledge of group prevalence ratios into fairness constraints.
• To address the trade-off between predictive accuracy and fairness in classification models.
• To create a more practical and equitable machine learning approach for sensitive applications.

Main Methods:

• Introduction of 'prior knowledge for fairness' by incorporating prevalence ratio information.
• Development of the Prior-knowledge-guided Fair ERM (PFERM) framework.
• Minimizing expected risk within a function class under a novel fairness constraint.

Main Results:

• The PFERM framework effectively balances accuracy and fairness.
• Empirical results demonstrate the preservation of fairness without sacrificing predictive accuracy.
• The approach proves effective even with significant group prevalence disparities.

Conclusions:

• Prior knowledge integration offers a flexible solution to the accuracy-fairness dilemma in machine learning.
• PFERM provides a practical method for building fairer and more accurate classifiers.
• Accounting for prevalence ratios is essential for equitable machine learning decision-making.