[Paper Review] What Is Fairness? On the Role of Protected Attributes and Fictitious Worlds

Ludwig Bothmann, Kristina Peters+1 more|arXiv (Cornell University)|May 19, 2022

Ethics and Social Impacts of AI4 citations

TL;DR

This paper formalizes fairness in machine learning by grounding it in philosophical principles and causal reasoning, proposing a fictitious, normatively desired (FiND) world where protected attributes (PAs) have no causal effects. It argues that predictive performance is essential for fairness, not a trade-off, and introduces a warped-world framework to approximate the FiND world, enabling causal fairness evaluation without PAs while ensuring fair but unequal treatment of unequal individuals.

ABSTRACT

A growing body of literature in fairness-aware machine learning (fairML) aims to mitigate machine learning (ML)-related unfairness in automated decision-making (ADM) by defining metrics that measure fairness of an ML model and by proposing methods to ensure that trained ML models achieve low scores on these metrics. However, the underlying concept of fairness, i.e., the question of what fairness is, is rarely discussed, leaving a significant gap between centuries of philosophical discussion and the recent adoption of the concept in the ML community. In this work, we try to bridge this gap by formalizing a consistent concept of fairness and by translating the philosophical considerations into a formal framework for the training and evaluation of ML models in ADM systems. We argue that fairness problems can arise even without the presence of protected attributes (PAs), and point out that fairness and predictive performance are not irreconcilable opposites, but that the latter is necessary to achieve the former. Furthermore, we argue why and how causal considerations are necessary when assessing fairness in the presence of PAs by proposing a fictitious, normatively desired (FiND) world in which PAs have no causal effects. In practice, this FiND world must be approximated by a warped world in which the causal effects of the PAs are removed from the real-world data. Finally, we achieve greater linguistic clarity in the discussion of fairML. We outline algorithms for practical applications and present illustrative experiments on COMPAS data.

Motivation & Objective

• To bridge the gap between centuries of philosophical fairness discourse and modern fairness-aware machine learning (fairML).
• To clarify the normative foundations of fairness by distinguishing fairML's role from socio-political and legal responsibilities.
• To demonstrate that fairness issues can emerge even without protected attributes, challenging the perceived trade-off between fairness and predictive performance.
• To formalize a fictitious, normatively desired (FiND) world where PAs have no causal effects, and propose a warped-world approximation for practical fairness evaluation.
• To enable fair but unequal treatment (vertical equity) through normative stipulations and causal modeling, avoiding the pitfalls of one-size-fits-all fairness metrics.

Proposed method

• Formalizes fairness using a philosophical framework rooted in equality of treatment and normative stipulations.
• Introduces the concept of a fictitious, normatively desired (FiND) world where protected attributes (PAs) have no causal effects on outcomes.
• Proposes a warped world as a practical approximation of the FiND world, achieved by removing the causal effects of PAs from real-world data.
• Employs causal inference techniques to estimate and remove the causal impact of PAs, enabling training and evaluation of ML models in the warped world.
• Defines fairness criteria based on counterfactual reasoning and individual well-calibration, ensuring models are fair even without PAs.
• Outlines a three-question normative framework for non-ML experts to define fairness goals, focusing on societal values rather than technical metrics alone.

Experimental results

Research questions

• RQ1 How can philosophical conceptions of fairness be consistently formalized and applied to machine learning in automated decision-making (ADM)?
• RQ2 To what extent can fairness problems arise in ML models even in the absence of protected attributes (PAs)?
• RQ3 How can causal reasoning be systematically integrated into fairness evaluation and model training to ensure normatively sound outcomes?
• RQ4 What is the role of predictive performance in achieving fairness, and why is it not a trade-off but a prerequisite?
• RQ5 How can a fair but unequal treatment (vertical equity) be achieved through formalized normative stipulations and causal modeling?

Key findings

• Fairness problems can emerge in ML models even without protected attributes, primarily due to poor individual well-calibration, not because of the presence of PAs.
• Predictive performance is not in conflict with fairness but is a necessary condition for achieving fairness, challenging the common trade-off narrative.
• The proposed fictitious, normatively desired (FiND) world framework enables a causal understanding of fairness by imagining a world where PAs have no causal effects.
• The warped world, used to approximate the FiND world, allows practical training and evaluation of ML models under causal fairness constraints.
• The framework supports vertical equity by enabling fair but unequal treatment of unequal individuals through normative stipulations, not just equal treatment.
• The approach reduces ambiguity in fairness discussions by clearly separating normative decisions (e.g., defining fairness goals) from technical ML implementation.