Machine learning systems are increasingly used in sensitive domains such as healthcare, finance, and law. A central concern is whether these systems treat different population groups fairly. Traditionally, fairness notions have focused on single protected attributes (e.g., gender or ethnicity), requiring models to perform equally well across the groups defined by that attribute.
However, in real-world populations, protected attributes overlap: gender, ethnicity, religion, and socioeconomic status interact to form subgroups such as Black women or Asian men from lower-income backgrounds. These overlapping groups can face unique disadvantages that are not captured when fairness is enforced only at the single-attribute level. This challenge is sometimes called intersectional fairness.
In the medical domain, for example, it is not sufficient to guarantee that predictions are equally accurate for men and women as a whole — it is equally important that subgroups such as Black men or Asian women receive predictions of comparable quality. Ignoring these intersections risks fairness gerrymandering, where a system appears fair overall but fails specific subgroups.
This thesis aims to:
- Survey and analyze existing notions of group fairness for overlapping groups, such as subgroup fairness, differential fairness, and multicalibration.
- Implement and evaluate efficient measures for auditing intersectional fairness in predictive models.
- Experimentally assess fairness in real-world datasets, with a focus on medical prediction tasks, to understand how overlapping group fairness plays out in practice.
The project will contribute both a conceptual understanding of intersectional fairness notions and a practical evaluation framework for measuring and mitigating disparities across overlapping groups in applied machine learning.
If you want to learn more or get involved, please contact Anne-Marie George or Pooya Zakeri.