Search-based Automatic Repair for Fairness and Accuracy in Decision-making Software.

Area of Science:

• Computer Science
• Artificial Intelligence
• Software Engineering

Background:

• Machine learning (ML) decision-making software can exhibit fairness issues, unfairly treating individuals based on sensitive attributes like gender or race.
• Existing bias mitigation techniques often reduce model accuracy to achieve fairness, posing a challenge for responsible software development.

Purpose of the Study:

• To present a novel multi-objective search-based method for repairing fairness issues in ML software.
• To demonstrate an approach that simultaneously enhances both fairness and accuracy in binary classification models, without the typical accuracy trade-off.

Main Methods:

• Developed a novel search-based algorithm for bias mitigation in ML models.
• Applied the method to Logistic Regression and Decision Trees, widely used models in software fairness research.
• Compared the proposed approach against seven state-of-the-art bias mitigation techniques using three fairness metrics.

Main Results:

• The proposed method successfully increased both accuracy and fairness in 61% of the studied cases.
• In contrast, existing state-of-the-art methods consistently decreased accuracy while attempting to reduce bias.
• The novel approach enables improvements in fairness without compromising predictive performance.

Conclusions:

• This research offers the first multi-objective search-based bias mitigation approach for binary classification that does not trade accuracy for fairness.
• Software engineers can now improve the fairness of ML models without the concern of accuracy degradation.
• The method facilitates the creation of more responsible and equitable ML-powered decision-making systems.