Training machine learning models for fair decisions faces two key challenges: The \emph{fairness-accuracy trade-off} results from enforcing fairness which weakens its predictive performance in contrast to an unconstrained model. The incompatibility of different fairness metrics poses another trade-off -- also known as the \emph{impossibility theorem}. Recent work identifies the bias within the observed data as a possible root cause and shows that fairness and predictive performance are in fact in accord when predictive performance is measured on unbiased data. We offer a causal explanation for these findings using the framework of the FiND (fictitious and normatively desired) world, a "fair" world, where protected attributes have no causal effects on the target variable. We show theoretically that (i) classical fairness metrics deemed to be incompatible are naturally satisfied in the FiND world, while (ii) fairness aligns with high predictive performance. We extend our analysis by suggesting how one can benefit from these theoretical insights in practice, using causal pre-processing methods that approximate the FiND world. Additionally, we propose a method for evaluating the approximation of the FiND world via pre-processing in practical use cases where we do not have access to the FiND world. In simulations and empirical studies, we demonstrate that these pre-processing methods are successful in approximating the FiND world and resolve both trade-offs. Our results provide actionable solutions for practitioners to achieve fairness and high predictive performance simultaneously.

A decision can be defined as fair if equal individuals are treated equally and unequals unequally. Adopting this definition, the task of designing machine learning models that mitigate unfairness in automated decision-making systems must include causal thinking when introducing protected attributes. Following a recent proposal, we define individuals as being normatively equal if they are equal in a fictitious, normatively desired (FiND) world, where the protected attribute has no (direct or indirect) causal effect on the target. We propose rank-preserving interventional distributions to define an estimand of this FiND world and a warping method for estimation. Evaluation criteria for both the method and resulting model are presented and validated through simulations and empirical data. With this, we show that our warping approach effectively identifies the most discriminated individuals and mitigates unfairness.

However, a fundamental question remains: What is fairness? This question is not so easy to answer and is often skipped; instead of asking "what is fairness", the questions of "how to measure fairness of ML models" and "how to make ML models fair" are pursued. This paper does not intend to criticize individual approaches that address those latter questions and often propose important solutions. Rather, the aim is to make explicit the premises that underlie the various understandings of fairness and the approaches to solving fairness problems. In doing so, a largely concordant understanding can be elaborated that is based on a rich foundation in the history of philosophy. Subsequently, we show that the conception of fairness depends on multilayered normative evaluations; any discussion of fairML is reliant on adopting those normative stipulations. The basis for fair decisions is always the question of the equality of the people treated with respect to the subject matter concerned. With this decision basis, a decision rule is to be established, which in turn can be adapted to the concrete needs as a result of normative stipulations. Based on this basic concept of fairness, we turn to the questions of to what extent ML models can induce unfair treatments in automated decision-making (ADM), and of how to implement these normative stipulations in training an ML model and in using its predictions in ADM.