Achieving Fairness Without Harm via Selective Demographic Experts.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Machine learning (ML) systems are increasingly used in healthcare, necessitating fairness and high predictive performance.
• Current bias mitigation methods often create a fairness-accuracy trade-off, negatively impacting certain demographic groups.
• This performance degradation is unacceptable in high-stakes medical applications like clinical diagnosis.

Purpose of the Study:

• To develop a fairness-without-harm approach for ML in human-centered domains.
• To mitigate bias in ML models without compromising predictive accuracy for any demographic group.
• To address the ethical and practical challenges of fairness-accuracy trade-offs in clinical settings.

Main Methods:

• Learning distinct data representations tailored for different demographic groups.
• Implementing a no-harm constrained selection mechanism.
• Utilizing demographic experts, comprising group-specific representations and personalized classifiers.

Main Results:

• The proposed approach successfully achieved fairness without harm across multiple datasets.
• Evaluated on medical datasets (eye disease, skin cancer, X-ray diagnosis) and face datasets.
• Demonstrated effectiveness in balancing fairness and accuracy for diverse demographic groups.

Conclusions:

• The fairness-without-harm approach is effective for ML in sensitive domains like healthcare.
• This method offers a viable solution to the ethical dilemma of fairness-accuracy trade-offs.
• Future work can explore broader applications of this technique in AI-driven decision-making.