Demographic Parity Tails for Regression
New framework enforces demographic parity only in critical distribution tails, preserving accuracy elsewhere.
A team from LAMA and SAMM has introduced a novel fairness framework for regression models called 'Demographic Parity Tails for Regression.' The core innovation addresses a key limitation of standard demographic parity (DP), which enforces statistical independence between model predictions and sensitive attributes (like race or gender) across the entire distribution. The researchers argue that this blanket constraint can unnecessarily degrade a model's overall predictive performance, especially when fairness concerns are only critical in specific areas, such as the high or low ends of the prediction range.
Their methodology builds on optimal transport theory, a mathematical framework for comparing probability distributions. By focusing fairness enforcement solely on targeted 'tails' of the prediction distribution across different demographic groups, the approach allows for more nuanced and context-sensitive interventions. The team developed an interpretable algorithm that leverages the geometric structure of optimal transport to implement these partial constraints. They provide theoretical risk bounds and fairness guarantees for their method and validate its effectiveness through experiments, demonstrating a practical path to balancing accuracy and equity in real-world AI systems like loan approval or risk assessment models.
- Targets fairness constraints to distribution tails, not the entire prediction output, using optimal transport theory.
- Aims to reduce the accuracy trade-off typical of full demographic parity enforcement in regression models.
- Provides theoretical guarantees and experimental validation for a more practical fairness-accuracy balance.
Why It Matters
Enables more practical, accurate, and targeted fairness interventions in critical AI systems like finance and hiring.