Source-linked AI summary
On the Apparent Conflict Between Individual and Group Fairness
Reuben Binns
TL;DR
The paper addresses the apparent conflict between individual and group fairness measures. It uses egalitarian accounts and the concept of individual justice to argue that the measures need not embody conflicting principles, while both fall short of individual justice.
Problem
Individual and group fairness are often treated as important but conflicting measures, motivating attempts to minimise trade-offs between them.
Method
The paper compares fairness concepts from fair machine learning with egalitarianism, consistency, and individual justice from political and legal philosophy.
Results
Individual and group fairness measures do not necessarily reflect different normative principles, but both fail to satisfy individual justice.
Takeaways & Limitations
Fairness measures should be selected through context-specific questions about an algorithmic system's purpose and the sources of unfairness.
Takeaways & Limitations
Individual fairness requires a task-relevant similarity mapping that is difficult to obtain and does not readily generalize to unseen individuals.
Abstract
from arXiv · showhide
A distinction has been drawn in fair machine learning research between `group' and `individual' fairness measures. Many technical research papers assume that both are important, but conflicting, and propose ways to minimise the trade-offs between these measures. This paper argues that this apparent conflict is based on a misconception. It draws on theoretical discussions from within the fair machine learning research, and from political and legal philosophy, to argue that individual and group fairness are not fundamentally in conflict. First, it outlines accounts of egalitarian fairness which encompass plausible motivations for both group and individual fairness, thereby suggesting that there need be no conflict in principle. Second, it considers the concept of individual justice, from legal philosophy and jurisprudence which seems similar but actually contradicts the notion of individual fairness as proposed in the fair machine learning literature. The conclusion is that the apparent conflict between individual and group fairness is more of an artifact of the blunt application of fairness measures, rather than a matter of conflicting principles. In practice, this conflict may be resolved by a nuanced consideration of the sources of `unfairness' in a particular deployment context, and the carefully justified application of measures to mitigate it.
1 INTRODUCTION
The paper argues that the apparent conflict between individual and group fairness reflects underspecified assumptions and context rather than fundamentally opposed principles. It develops this argument through Fair-ML theory and philosophical analysis, with implications for how practitioners justify measures.
- Fair-ML research distinguishes group fairness, which targets statistical parity across protected groups, from individual fairness, which targets similar outcomes for similar people.The distinction has motivated proposals to reduce trade-offs between the measures.
- The paper critically assesses whether these measures are genuinely in conflict and finds that their apparent opposition need not exist at the level of principle.It argues that both can reflect the same moral and political concerns.
- For researchers, the paper identifies a lack of sustained theoretical discussion about the principles supposedly underlying the conflict between individual and group measures.Existing work has largely focused on technical ways to reconcile the measures.
- For practitioners, selecting a fairness measure requires sound justification because Fair-ML governance measures will be applied across public- and private-sector decision systems.The paper connects this need for justification to careful consideration of the relationship between the two fairness approaches.
- The paper further argues that individual fairness does not achieve truly individualized justice because it still generalizes across people who share relevant features.On this view, individual fairness becomes a special case of group fairness.
- In concrete deployments, apparent conflicts should be examined through the moral and empirical assumptions about the decision procedure and the fairness concerns in that context.The paper presents these assumptions as the source of apparent conflict when both measures seem to disagree.
2 ‘INDIVIDUAL’ AND ‘GROUP’ FAIRNESS IN FAIR-ML
Fair-ML literature operationalizes group fairness through statistical parity and individual fairness through similarity or merit, but both approaches face important limitations and apparent trade-offs. Subsequent proposals attempt to reconcile them, without resolving the underlying conceptual debate.
- 2.1.1 Group fairness measures.: Group fairness measures assess statistical parity between people with different protected attributes, including parity in outcomes or error rates.Statistical parity can require equal proportions of protected groups in outcome classes, while other measures compare false-positive or false-negative rates.
- 2.1.1 Group fairness measures.: Group fairness may miss discrimination affecting intersectional or legally unrecognized groups, while testing every subgroup can scale poorly and encourage overfitting.Intersectional combinations such as black women may experience unfairness even when gender and race are assessed separately.
- 2.1.2 Individual fairness measures.: Individual fairness gives similar predictions or decisions to people who are similar with respect to task-relevant features, usually as defined by a distance metric.The metric is intended to represent similarity for the task or decision context.
- 2.1.2 Individual fairness measures.: A central practical limitation of individual fairness is that decision-makers may not know how to obtain a task-relevant similarity mapping for all individuals or generalize it to unseen people.This difficulty is acknowledged as a major shortcoming of the initial formulation.
- 2.1.2 Individual fairness measures.: Later approaches relax the need for complete hand-specified mappings by learning distance functions through clustering or judgments from ethical experts.Other proposals instead define fairness through individual merit, linking predictions to the true probability of the relevant label.
- 2.1.3 Reconciliation between individual and group measures.: Despite reconciliation proposals, individual and group fairness appear to trade off when protected groups differ in feature and label distributions.Individual fairness can object to unequal outcomes for similar people, whereas group fairness can object to disparities produced by individual-level treatment.
3 IS THERE A GROUP / INDIVIDUAL DISTINCTION IN PRINCIPLE?
The paper argues that individual and group fairness need not represent conflicting principles: both can reflect shared moral and political concerns when their assumptions and context are made explicit.
- Philosophical mapping suggests that individual and group fairness are different ways of reflecting the same underlying moral and political concerns.
- The apparent conflict may arise because one underlying normative concept can be operationalised in different ways.
- The paper’s account is necessarily partial, drawing on a narrow literature and selecting two norms plausibly underlying the Fair-ML distinction.
- 3.1 Consistency: Consistency motivates individual fairness through the idea that similar cases should be treated alike and identical cases should receive the same judgement.
- 3.1 Consistency: Consistency alone does not specify which features determine likeness, leaving both individual and group fairness potentially inconsistent depending on what is included or excluded.
- 3.2 Egalitarianism: Egalitarian theories provide a framework for evaluating decisions in allocations such as jobs, welfare, loans, education, and bail, including which features should affect outcomes.
- 3.2 Egalitarianism: Protected characteristics such as gender and race are typically treated by egalitarian theories as attributes that do not justify unequal distributions.
- 3.2 Egalitarianism: Individual fairness can incorporate egalitarian concerns by defining task-relevant similarity through norms and causal assumptions about responsibility and unchosen circumstances.
4 INDIVIDUAL FAIRNESS VS INDIVIDUAL JUSTICE
The paper distinguishes individual justice from individual fairness, arguing that algorithmic individual-fairness models still generalise from prior cases and therefore do not achieve truly individualised assessment.
- Individual justice: Individual justice requires assessing people on their own qualities and circumstances rather than generalisations about their group membership.Legal philosophy calls this individualised or particularised justice, also known as Einzelfallgerechtigkeit.
- Individual justice: Individualised justice can involve additional information and reasoning about mitigating circumstances that an algorithm does not consider.The paper presents human involvement as potentially serving this aim in automated decision-making.
- Individual fairness: Individual fairness remains feature-based: people sharing the same selected features receive the same outcome through generalisation from similar prior cases.This makes individually-fair models individually unjust according to the individual-justice account.
- Implication: Individual fairness cannot be truly individualised while it treats decisions as yes-or-no outputs over a delimited range of input values.The paper characterises it as a fine-grained form of group generalisation defined by locations in task-relevant metric space.
- Conclusion: Thus, individual fairness does not preserve the essence of individual justice, despite their apparent similarity.Both involve generalisation, although individual fairness uses finer-grained categories.
5 DISSOLVING THE CONFLICT IN PRACTICE
In practice, the paper argues that apparent conflicts between group and individual fairness usually reflect different assumptions about the sources of disparities, not a choice between fundamentally opposed fairness principles.
- Framing the conflict: The practical conflict is attributed to unstated or conflicting assumptions rather than to selecting individual versus group fairness measures.Once assumptions are resolved, either measure may reflect the resulting fairness concerns.
- College admissions: Admissions concerns can involve both group disparities in acceptance rates and unequal outcomes for similarly scored applicants.These concerns are commonly treated with group and individual metrics respectively.
- College admissions: Fairness objectives should include more than predicted academic success, including avoiding disadvantages arising from circumstances outside applicants’ control.Decision-makers must agree which factors are beyond control and how they influence observed application features.
- Financial lending: Individual and group approaches can share the same assumptions that gender disparities in loan defaults reflect unjust structures.Both may seek predictions that ignore structural effects, even though men may object to losing opportunities.
- Financial lending: Equal calibration can preserve those structural assumptions while addressing calibration disparities rather than eliminating outcome disparities.Under these assumptions, equal repayment records can support similar predictions without requiring outcome parity.
- Worldviews: Raw similarity and statistical parity conflict when disparities are attributed respectively to personal choices and unjust structures.Changing the variants can reverse which individual and group measures conflict.
- Worldviews: Agreeing which differences reflect choice, luck, or structural injustice is empirically and politically difficult.This difficulty extends to deciding whether decision-makers are responsible for reinforcing those differences.
- Worldviews: Fairness measures become appropriate only after the relevant worldview about structural bias and feature distortion is specified.Under a group-skew assumption, adjusted individual metrics may be justified; under WYSIWYG, raw similarity or equal calibration may be appropriate.
6 CONCLUSION
Individual and group fairness measures need not embody different normative principles, but both fail to satisfy individual justice. Their practical application requires examining algorithmic purpose, data assumptions, and the kinds of unfairness involved.
- Individual and group fairness measures can both reflect consistency and egalitarianism, despite their apparent conflict.
- Both fairness families fail to satisfy the principle of individual justice, despite their surface-level similarity to it.
- Applying either fairness family requires asking what the algorithm should select, how qualifications are generated and recorded, and how structural inequalities shape features.
- The apparent conflict reflects differing empirical and normative assumptions associated with measures, not an exclusive ethical basis for either family.
- Examining those assumptions can redirect attention from model adjustments toward changing unjust underlying structures.