Reproduce “On the Fairness of Causal Algorithmic Recourse” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Julius von Kügelgen, Amir-Hossein Karimi, Umang Bhatt, Isabel Valera, Adrian Weller, Bernhard Schölkopf
A causal framework distinguishes fairness in algorithmic predictions from fairness in the effort required to overturn an unfavorable decision. The paper introduces group- and individual-level measures of causal recourse, proves that common fairness guarantees do not ensure fair recourse, and shows on synthetic data and Adult that causal intervention costs can reveal disparities hidden by geometric distance metrics.
Key insights
- Predictive fairness, group-level recourse fairness, and individual-level recourse fairness are distinct: satisfying one does not generally guarantee the others.
- Euclidean or boundary distance may misrepresent actionable burden when features are causally linked; causal cost can be either higher or lower than the corresponding geometric estimate.
- Using only non-descendants can guarantee individual-level causal recourse fairness under stated assumptions, but may reduce accuracy; adding correctly inferred exogenous variables can recover baseline-level performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.