The Effects of Changing the Causal Model of a Counterfactually Fair Predictor

Abstract

The body of research on fairness in machine learning has been growing in recent years. One de nition of fairness that has been proposed is that of counterfactual fairness, in which one tries to model and take into ac- count how variables in uence each other in a biased world. This is done via a causal model, which needs to be assumed. This thesis investigates the e ects on the outcomes of a counterfactually fair predictor, when its underlying causal model is changed.