What is Fairness? On Protected Attributes and Fictitious Worlds

Ludwig Bothmann, Kristina Peters, Bernd BischlEWAF 2025 · OpenReview

TL;DR

This paper bridges the gap between philosophical discussions of fairness and its application in machine learning (ML). It proposes a formal framework for fairness in automated decision-making (ADM) systems, emphasizing that fairness problems can arise even without protected attributes (PAs) and that predictive performance is crucial for achieving fairness. The authors introduce the concept of a "fictitious, normatively desired (FiND) world" where PAs have no causal effects, which serves as a target for building fairer ML models.

What Problem Does This Paper Address?

The field of fairness-aware machine learning (fairML) has exploded, with numerous metrics and methods proposed to mitigate unfairness in ML models used for automated decision-making (ADM). However, the authors argue that a fundamental question remains largely unaddressed: "What is fairness?" This lack of a clear, foundational definition leaves a significant gap between centuries of philosophical discourse on fairness and the ML community's recent adoption of the concept. The paper aims to bridge this gap by formalizing a consistent concept of fairness derived from philosophical considerations and translating it into a practical framework for designing and evaluating ML models in ADM systems. They also highlight that fairness issues can emerge even when protected attributes (PAs) are not explicitly used, and that predictive performance is not an antagonist to fairness but rather a prerequisite.

Methods

The core of the paper's methodology lies in formalizing a philosophical concept of fairness and translating it into a computational framework. Key elements include:

  • Formalizing Fairness: The authors build upon Aristotle's concept of justice, distinguishing between "equality" and "equity" and defining fairness based on normative stipulations about how individuals should be treated. They introduce definitions for arithmetic, geometric, and modified fair treatment, emphasizing that fairness is always a result of a normative choice.
  • The FiND World: A central concept is the "fictitious, normatively desired (FiND) world." This is a hypothetical world where protected attributes (PAs) have no causal effects on the target variable. This world serves as a benchmark for fairness, allowing for the removal of direct and indirect causal influences of PAs.
  • Warping: Since the FiND world is inaccessible, the authors propose approximating it through a "warped world." This involves transforming real-world data to simulate the absence of PA causal effects. The paper outlines the need for warping methods to create this data for training and evaluating ML models.
  • Causal Considerations: The framework heavily relies on causal reasoning, particularly in understanding how PAs influence outcomes and how to disentangle these influences in the FiND world. Directed Acyclic Graphs (DAGs) are used to visualize these causal relationships.

Datasets

The paper mentions using the COMPAS dataset for illustrative experiments. This dataset is commonly used in fairness research due to its relevance to criminal justice and the presence of protected attributes like ethnicity. The authors use it to demonstrate their framework's application in a real-world scenario.

Key Insights and Ideas

  • Fairness is Normative, Not Just Technical: The paper strongly emphasizes that fairness is not solely a technical problem solvable by algorithms. It requires explicit normative decisions about what constitutes fair treatment and how to define task-specific merit.
  • Predictive Performance is Essential for Fairness: Contrary to a common perception of a tradeoff, the authors argue that high predictive performance is a necessary condition for achieving fairness. An ML model that poorly predicts outcomes cannot lead to fair decisions, even if it attempts to satisfy fairness metrics.
  • Fairness Problems Exist Without Protected Attributes: The paper demonstrates that even in the absence of PAs, issues like information coarsening (grouping individuals based on features) and estimation errors in ML models can lead to unfairness.
  • The FiND World as a Fairness Target: The concept of a FiND world, where PAs have no causal influence, provides a clear theoretical target for building fairer systems. It allows for a principled way to remove unwanted causal pathways.
  • Warping as a Practical Approximation: The introduction of the "warped world" offers a pragmatic approach to approximating the ideal FiND world, making the theoretical framework applicable to real-world data.

Key Results

The paper focuses on theoretical contributions and a framework for understanding and addressing fairness, rather than presenting quantitative performance metrics in the traditional sense.

  • Formal Framework: The primary result is the development of a formal framework for understanding fairness in ADM systems, grounded in philosophical concepts and incorporating causal reasoning.
  • Distinction between Real and FiND Worlds: The paper clearly delineates the differences between the real world (with PA influences) and the FiND world (without them), providing a conceptual basis for fairness interventions.
  • Illustrative Experiments: The authors present illustrative experiments on the COMPAS dataset to demonstrate the application of their framework, though detailed quantitative results are not the main focus. These experiments aim to highlight the practical implications of their theoretical concepts.

Limitations and Future Work

The authors acknowledge several limitations and point towards future research directions:

  • Defining the FiND World is Normative and Controversial: The specific definition of the FiND world, including which paths from PAs are considered unfair and should be removed, is a highly normative and potentially contentious issue.
  • Warping Methods and Validity Measurement: Developing concrete and robust warping methods to approximate the FiND world and establishing rigorous techniques to measure the validity of these warps are left as future work. This includes addressing uncertainties related to real-world causal DAGs.
  • PA-Focus is Future Work: While the paper focuses on PA-neutrality, the authors suggest that a corresponding framework for PA-focus (e.g., affirmative action) would be a valuable future research direction, fitting within the same philosophical concept of normative, task-specific equality.
  • Applied Use Cases: The paper prioritizes theoretical clarity over fully applied use cases, with illustrative examples provided in the appendix. Future work could focus on developing more comprehensive workflows for applied ADM use cases.
  • Uncertainty in Causal Discovery: The process of deciding on a DAG for the real world, which is crucial for warping, involves uncertainty that needs to be accounted for in future research.

Summary updated

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.
BibTeX
@inproceedings{bothmann2025what,
  author = {Ludwig Bothmann and Kristina Peters and Bernd Bischl},
  title = {{What is Fairness? On Protected Attributes and Fictitious Worlds}},
  booktitle = {EWAF},
  year = {2025},
  url = {https://openreview.net/forum?id=QjAAnLzEPy},
}
Figure