Reproduce “SPARC: Scalable Path-Specific Counterfactual Fairness via Causal Conditional Independence” 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
Bowei Tian, Yexiao He, Ziyao Wang, Meng Liu, Yongkai Wu, Ang Li
The SPARC framework introduces a scalable approach to Path-Specific Counterfactual Fairness, reformulating it into a causal conditional independence constraint to address limitations in high-dimensional data. This method achieved a superior balance between predictive performance and fairness, demonstrating up to a 5x reduction in Equalized Odds on datasets like TCGA-LUAD, while preserving task-relevant information in domains such as medical imaging.
Key insights
- Path-Specific Counterfactual Fairness can be scaled to high-dimensional data by reducing the intractable problem of counterfactual probability estimation to a tractable causal conditional independence constraint (I(S; X | Y) = 0).
- This causal conditional independence condition is theoretically proven to be sufficient for eliminating the direct causal effect of a sensitive attribute and implies observable fairness metrics like Equalized Odds.
- The framework effectively disentangles sensitive attribute information from legitimate, task-relevant features in high-dimensional inputs, ensuring predictions are based on appropriate diagnostic signals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.