"Patriarchy Hurts Men Too." Does Your Model Agree? A Discussion on Fairness Assumptions
Marco Favier, Toon Calders
TL;DR
This work interrogates the fairness pipeline by exposing implicit monotonic assumptions in how bias is introduced into data. It formalizes a framework with $p_f(\boldsymbol{y}|x)$, $p_u(\boldsymbol{y}|x)$ and a biasing function $\beta$ so that $p_u(\boldsymbol{y}|x)=\beta(p_f(\boldsymbol{y}|x),x)$, and defines decisions $\boldsymbol{d}:X\to\Delta^n$ under a Pareto order to study fairness. Using this setup and $\text{Weller's}$ theorem, the paper connects fair optimization to the biasing structure, deriving conditions like 'Affirmative Action' and 'Double Standard' and showing how many conventional fairness assumptions imply monotonicity. It also discusses within-group reranking and shows that if such reranking occurs, several core assumptions fail, underscoring the need to explicitly state bias models. Overall, the work clarifies when pursuing fairness constraints is meaningful and when observed fairness improvements may be tautological under monotone bias.
Abstract
The pipeline of a fair ML practitioner is generally divided into three phases: 1) Selecting a fairness measure. 2) Choosing a model that minimizes this measure. 3) Maximizing the model's performance on the data. In the context of group fairness, this approach often obscures implicit assumptions about how bias is introduced into the data. For instance, in binary classification, it is often assumed that the best model, with equal fairness, is the one with better performance. However, this belief already imposes specific properties on the process that introduced bias. More precisely, we are already assuming that the biasing process is a monotonic function of the fair scores, dependent solely on the sensitive attribute. We formally prove this claim regarding several implicit fairness assumptions. This leads, in our view, to two possible conclusions: either the behavior of the biasing process is more complex than mere monotonicity, which means we need to identify and reject our implicit assumptions in order to develop models capable of tackling more complex situations; or the bias introduced in the data behaves predictably, implying that many of the developed models are superfluous.