Fair Classifiers via Transferable Representations

This talk explores group fairness in text classification, focusing on achieving fair treatment between sensitive groups like gender. Charlotte Laclau presents an approach that extends the Wasserstein Independence measure for developing unbiased neural text classifiers. Learn how adversarial training can induce Wasserstein independence between representations learned for target labels and sensitive attributes. Discover how domain adaptation techniques can eliminate the need for sensitive attribute data during training. The 48-minute presentation provides both theoretical foundations and empirical evidence supporting this methodology for creating fair classifiers through transferable representations. This research presentation from Télécom Paris was delivered at the Institut des Hautes Etudes Scientifiques (IHES) and is available on CARMIN.tv, a French video platform specializing in mathematics and interdisciplinary scientific content.