[Paper Review] Fairness in Machine Learning: Lessons from Political Philosophy

Reuben Binns|arXiv (Cornell University)|Dec 10, 2017

Ethics and Social Impacts of AISocial Sciences32 references390 citations

TL;DR

The paper surveys philosophical theories of discrimination and egalitarianism to clarify what 'fairness' means in machine learning and to guide which fairness metrics and approaches are appropriate in different contexts.

ABSTRACT

What does it mean for a machine learning model to be `fair', in terms which can be operationalised? Should fairness consist of ensuring everyone has an equal probability of obtaining some benefit, or should we aim instead to minimise the harms to the least advantaged? Can the relevant ideal be determined by reference to some alternative state of affairs in which a particular social pattern of discrimination does not exist? Various definitions proposed in recent literature make different assumptions about what terms like discrimination and fairness mean and how they can be defined in mathematical terms. Questions of discrimination, egalitarianism and justice are of significant interest to moral and political philosophers, who have expended significant efforts in formalising and defending these central concepts. It is therefore unsurprising that attempts to formalise `fairness' in machine learning contain echoes of these old philosophical debates. This paper draws on existing work in moral and political philosophy in order to elucidate emerging debates about fair machine learning.

Motivation & Objective

• Clarify how philosophical concepts of discrimination map onto machine learning fairness concerns.
• Explore how different egalitarian frameworks inform the selection and prioritization of fairness metrics in ML.
• Highlight the contextual and historical factors that should influence fairness interventions in ML systems.

Proposed method

• Survey and synthesize key philosophical theories of discrimination (mental-state vs. non-mental-state accounts).
• Review egalitarian debates (currency of egalitarianism, spheres of justice, luck vs. desert, deontic justice).
• Connect philosophical analyses to common ML fairness measures (demographic parity, accuracy equity, equality of opportunity, disparate mistreatment).
• Discuss how historical, sociological, and contextual factors should inform feature selection and fairness interventions.

Experimental results

Research questions

• RQ1 What counts as discrimination in algorithmic decision-making, and when is it morally problematic?
• RQ2 How do different egalitarian theories translate into criteria for fair ML (e.g., equality of outcome vs. equality of opportunity, luck egalitarian considerations)?
• RQ3 In what ways should historical and sociological contexts influence fairness assessments and mitigation strategies in ML?
• RQ4 What is the relevance of representational harms versus distributive harms in ML fairness?

Key findings

• Discrimination notions from philosophy (mental-state and generalisation-based accounts) have limited direct applicability to algorithms, suggesting alternative fairness foundations.
• Egalitarian theories offer a richer framework (currency of egalitarianism, spheres of justice, luck vs. desert, deontic justice) to justify and prioritize fairness interventions in ML.
• Fairness in ML often involves trade-offs between incompatible metrics (e.g., accuracy equity vs. equalized false positive rates), which can be better understood through deontic and historical considerations.
• Representational harms (e.g., biased cultural representations) require different fairness objectives than distributive harms, posing unique challenges for ML systems.
• Contextual factors such as the sphere of justice and historical injustices should guide which fairness metrics and mitigations are appropriate in a given setting.