What are common fairness metrics used in AI systems, and how do they help in mitigating bias during model evaluation?

Define fairness metrics (e.g., demographic parity, equalized odds) and explain their roles in quantifying bias. Highlight how they identify disparities in model outcomes across groups, enabling targeted interventions. Emphasize their use in evaluating trade-offs between fairness and model performance during development.

Common fairness metrics include demographic parity (equal outcomes across groups), equalized odds (equal true/false positive rates), and accuracy parity (equal accuracy across groups). These metrics help identify bias by quantifying disparities in model predictions. For example, in hiring algorithms, demographic parity ensures equal acceptance rates for protected groups. However, enforcing fairness may reduce model accuracy, requiring trade-off analysis. Real-world applications include credit scoring, where equalized odds prevent unequal loan approval rates. Metrics also guide mitigation strategies, such as reweighting training data or adjusting decision thresholds. While these metrics are critical, they may conflict (e.g., demographic parity vs. accuracy parity), necessitating context-specific choices based on ethical priorities and domain requirements.