If fairness is a counterfactual concept then lots of things can be biased
This week we look at fairness and its evil twin bias. Our starting point will be, roughly, that fairness as an idea is essentially counterfactual. Specifically, I or my algorithm treat you fairly in an allocation decision with respect to one of your attributes, for example your gender, if I would have allocated you the same thing had your gender been different.
While not the only way to think about fairness, this has some intuitive appeal. However, many other things about you would have been caused to be different had your gender been different, even assuming the counterfactual ‘you with another gender’ is conceivable. Things get interestingly difficult rather quickly and much legal and moral argument depends implicitly on answers to the ensuing difficulties, for example the concrete question about whether fairness requires that no allocation model depends on your gender. We will examine circumstances where this simple heuristic may fail as well as situations where competing definitions of fairness are in principle incompatible. We will try to use our causal inference tools to make some sense of the issue.
Kusner et al. (2017) ‘Counterfactual Fairness’ Advances in Neural Information Processing Systems.
Russell et al. (2017) ‘When worlds collide: Integrating different counterfactual assumptions in fairness’ Advances in Neural Information Processing Systems.
Pleiss et al. (2017) ‘On Fairness and Calibration’. Arxiv