Tuning Fairness by Balancing Target Labels.

Area of Science:

• Artificial Intelligence
• Machine Learning Ethics
• Algorithmic Fairness

Background:

• Machine learning models are increasingly deployed in critical decision-making processes.
• Ensuring fairness in these models is crucial for public trust and equitable outcomes.
• Bias in machine learning can lead to discriminatory outputs, disproportionately affecting certain demographic groups.

Purpose of the Study:

• To develop a unified framework for mitigating bias in machine learning outputs.
• To address fairness concerns in applications like loan approvals and job candidate selection.
• To provide a more intuitive and practical approach to controlling algorithmic fairness.

Main Methods:

• Introduced a latent target output within probabilistic models to control bias.
• Formulated fairness as a marginalization problem, avoiding complex constrained optimization.
• Unified several group fairness notions, including Demographic Parity and Equality of Opportunity.

Main Results:

• The proposed latent target output formulation effectively controls bias in machine learning outputs.
• The marginalization approach simplifies fairness implementation, allowing reuse of standard toolboxes.
• Directly varying fairness target rates offers intuitive control over the level of fairness achieved.

Conclusions:

• The latent target output method provides a unified and practical framework for achieving group fairness in machine learning.
• This approach offers a more interpretable and controllable alternative to existing methods relying on indirect parameters.
• Implementing this method can enhance the fairness of AI systems, fostering greater public confidence.