This session investigates the theoretical foundations of fair regression constrained by demographic parity within the unawareness framework—where explicit disparate treatment is disallowed. Expanding on existing models that permit such treatment, the talk addresses how to characterize optimal fair regression functions when minimizing quadratic loss, demonstrating how the solution maps to a barycenter problem with optimal transport costs.

Furthermore, the session examines the interplay between optimal fair cost-sensitive classification and fair regression. It shows that nestedness among classifier decision sets is both necessary and sufficient for equivalence between the two, allowing optimal classifiers to be thresholded directly from the fair regression function (and vice versa).

Paper: https://jmlr.org/beta/papers/v27/24-1338.html

Register here!