Subjective fairness: Fairness is in the eye of the beholder

Fairness is a desirable property of decision rules applied to a population of individuals. For example, college admissions should be decided on variables describing merit, but may also need to take into account the fact that certain communities are inherently disadvantaged. At the same time, individuals should not feel that another individual in a similar situation obtained an unfair advantage. All this must be taken into account while still caring about optimizing for a decision maker’s utility function. In particular, for a given distribution over a population, we wish to derive a decision rule that takes into account a merit variable, but also ensures fairness for members of disadvantaged groups. The problem becomes even more challenging when we take into account potential uncertainties in decision making models, which can even make strict notions of fairness impossible to satisfy. As an example, consider the problem of fair prediction with disparate impact as defined by Chouldechova [2016]. Informally, their formulation defines a statistic a such that true category y (also called outcome or true label) is conditionally independent of a sensitive variable z given the statistic and the model parameters θ, i.e. y ⊥ z | a, θ. When we face uncertainties in our modeling assumptions, the natural thing is to impose that the conditional independence holds if we marginalize the parameters out, i.e. y ⊥ z | a. As we argue later in the paper, such a condition is impossible to satisfy, even if it holds for every possible parameter value, and we must incorporate subjectivity when model parameters are uncertain. We instead develop a natural, and widely applicable framework for fairness that relies on the available information. We develop algorithms for achieving a few different notions of fairness within the subjective framework, and in particualr recently proposed concepts of fairness that are grounded in concepts of similarity and conditional independence. We argue that a suitable notion of similarity in the Bayesian setting is distributional similarity conditioned on the observations. For the latter, as independence is difficult to achieve uniformly in the Bayesian setting, we suggest a relaxation, for which we provide a small experimental demonstration.