A naukri.com initiative
Arxiv
2d
235
Image Credit: Arxiv
Size-adaptive Hypothesis Testing for Fairness
- The paper discusses the challenges in determining algorithmic discrimination against demographics and proposes a new approach.
- Current fairness assessment methods are deemed statistically brittle due to ignoring sampling error and treating small demographic subgroups the same as large ones.
- Intersectional analyses, considering multiple sensitive attributes simultaneously, lead to more granular groups with sparse data, making estimation unreliable.
- The paper introduces a size-adaptive, hypothesis-testing framework for fairness assessment to address statistical limitations and variations in data availability.
- The framework provides analytic confidence intervals and a Wald test for statistical parity difference in large subgroups with guaranteed type-I error control.
- For small intersectional groups, a fully Bayesian Dirichlet-multinomial estimator is proposed with Monte-Carlo credible intervals that converge to Wald intervals with increased data.
- Empirical validation on benchmark datasets showcases the framework's interpretability and statistical robustness across different data availability levels and intersectionality.
Read Full Article
14 Likes
For uninterrupted reading, download the app