Abstract
We introduce the Differential Adjusted Parity (DAP) loss to produce unbiased informative representations. DAP utilises a differentiable variant of the adjusted parity metric to create a unified objective function that combines downstream task classification accuracy and its inconsistency across sensitive feature domains. A key element in this approach is the use of Soft Balanced Accuracy, which makes the metric differentiable while remaining suitable for imbalanced problems. In contrast to previous non-adversarial approaches, DAP does not suffer a degener-acy where the metric is satisfied by performing equally poorly across all sensitive domains. On Adult and COMPAS, DAP outperforms several adversarial models on downstream task accuracy and fairness. The largest gains reach 22.5%, 44.1% and 40.1% on demographic parity, equalized odds and sensitive feature accuracy, respectively, when compared to the best performing adversarial approach for each metric.