A note on the impossibility of “fairness” – Thomas Miconi
“Various measures can be used to determine bias or unfairness in a predictor. Previous work has already established that some of these measures are incompatible with each other. Here we show that, when groups differ in prevalence of the predicted event, several intuitive measures of fairness (ratio of positive predictions to event occurrence, probability of positive prediction given actual occurrence or non-occurrence, and probability of occurrence given positive or negative prediction) are all mutually exclusive: if one of them is equal among groups, the other two must differ. The only exceptions are for perfect, or trivial (always-positive or always-negative) predictors. As a consequence, any non-perfect, non-trivial predictor can always be portrayed as biased or unfair under a certain perspective. The result applies to all predictors, algorithmic or human. We conclude with possible ways to handle this effect when assessing and designing prediction methods.”

Subjects: stat.AP cs.AI stat.ML
http://arxiv.org/abs/1707.01195
http://arxiv.org/pdf/1707.01195.pdf