Eliminating Latent Discrimination: Train Then Mask

Nowadays, many sensitive decision-making tasks rely on automated statistical and machine learning algorithms. Examples include targeted advertising, credit scores and loans, college admissions, prediction of domestic violence, and even investment strategies for venture capital groups. There has been a growing concern about errors, unfairness, and transparency of such mechanisms from governments, civil organizations and research societies [2, 33, 40]. That is, whether or not we can prevent discrimination against protected groups and attributes (e.g., race, gender, etc). Clearly, training a machine learning algorithm with the standard aim of loss function minimization (i.e., high accuracy, low prediction error, etc) may result in predictive behaviors that are unfair towards certain groups or individuals [18, 29, 42]. In many real-world applications, we are not allowed to use some sensitive features. For example, EU anti-discrimination law prohibits the use of protected attributes (directly or indirectly) for several decision-making tasks [13]. A naive approach towards fairness is to discard sensitive attributes from training data. However, if other (seemingly) nonsensitive variables are correlated with the protected ones, the learning algorithm may use them to proxy for protected features in order to achieve a lower loss.