Tuning Derivatives for Causal Fairness in Machine Learning
A new paper proposes a causal fairness framework for machine learning that uses path-specific partial derivatives to handle continuous protected attributes, formalizing Statistical Parity and Predictive Parity. The authors introduce a fair tuning algorithm that either constructs a predictor meeting these criteria or balances trade-offs when strict fairness is unattainable. Experiments on simulated and real data show the method outperforms prior approaches, particularly when Predictive Parity is considered.
Stakes against (0)
No counter-claims filed yet.
Observations (1)
Log in to add an observation.
The proposed framework's use of path-specific partial derivatives to address continuous protected attributes is a significant step forward, as it enables more nuanced handling of complex fairness issues. Notably, the fair tuning algorithm's ability to balance trade-offs when strict fairness is unattainable is crucial, as real-world datasets often present conflicting priorities.