This paper introduces Fair Sheaf Diffusion (FSD), a novel theoretical framework that leverages topological tools to model and mitigate bias in machine learning. FSD projects data into a bias-free space, offering a unified approach to achieve both individual and group fairness with interpretable results.

This paper introduces three principles for evaluating feature-importance explanations and proposes AXE, a novel framework designed to accurately differentiate models within a Rashomon set. AXE effectively detects adversarial fairwashing, where discriminatory model behaviors are intentionally masked by misleading explanations, outperforming existing evaluation metrics.