Reviews: Avoiding Discrimination through Causal Reasoning

This paper formulates fairness for protected attributes as a causal inference problem. It is a well-written paper that makes an important connection that discrimination, in itself, is a causal concept that can benefit from using causal inference explicitly. The major technical contribution of the paper is to show a method using causal graphical models for ensuring fairness in a predictive model. In line with social science research on discrimination, they steer away from thinking about a gender or race counterfactual which can be an unwieldy or even absurd counterfactual to construct---what would happen if someone's race is changed---and instead focus on using proxy variables and controlling causal pathways from proxy variables to an outcome. This is an important distinction, and allows us to declare proxy variables such as home location or name and use them for minimizing discrimination.