Abstract

In this chapter, we primarily consider the task of binary classification, where each individual i has some observable attributes Xi and a binary label Yi ∈ {0, 1}; where appropriate, they will also have a protected attribute Ai . For individual fairness in Section 17.2, the fair algorithm only observes Xis, and must assign an outcome Yˆ i to each individual; for group and multi-group fairness in Sections 17.3 and 17.4, the algorithm takes in a training set of n observations containing (Xi, Yi) pairs (and Ai in Section 17.3), and must produce a binary classifier for use on future observations that maps Xi to a predicted outcome Yˆ i. In all settings, the mapping from Xi to Yˆ i must respect the fairness constraint.