DataTerminology.comNEW
Ensuring equitable treatment or outcomes across individuals or groups in data-driven decision systems, according to a chosen fairness criterion.
DESCRIPTION
Fairness in machine learning formalizes notions of equitable treatment or outcomes. Multiple definitions exist, such as demographic parity, equalized odds, and individual fairness, which can conflict. Choosing the right definition depends on context, harms, and legal constraints. Fairness requires measurement, stakeholder input, and often trade-offs with utility or privacy. It is an ongoing process, monitored throughout model operation.
ALTERNATIVE DEFINITIONS
Wikipedia
In machine learning, algorithmic fairness is the study of how to define, detect, and mitigate unwanted bias in algorithms.
Frames fairness as a field studying definitions and mitigation of bias.
Google Developers
Fairness in machine learning means ensuring that your models’ predictions do not discriminate against individuals or groups.
Focuses on non-discrimination in model predictions.
RELATED TERMS
Bias & Fairness Equalized Odds Demographic Parity Individual Fairness Calibration Disparate Impact Causal Fairness Explainability Human Oversight Responsible AI
COMPETING TERMS
Predictive Accuracy
RELATIONS
KNOWLEDGE AREAS
AI Data Science AI Governance Mathematics
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer AI Product Manager Data Governance Manager Chief Data Officer
EXTERNAL SOURCES
- Wikipedia - Fairness (machine learning)
- Barocas, Hardt, Narayanan - Comprehensive fairness textbook
- ArXiv - Equalized odds fairness paper
- NIST - Measurement of AI characteristics
TOP QUESTIONS
Why is Fairness important in AI systems?
Unfair systems can harm individuals and erode trust. Fairness reduces disparate impacts and supports legal compliance and social responsibility.
Category: Relevancy
How do you choose a Fairness definition?
Consider context, harms, legal constraints, and stakeholder values; often, multiple candidate metrics are evaluated for trade-offs.
Category: HowTo
Can Fairness and accuracy conflict?
Yes, constraints may reduce some accuracy measures; however, better data and modeling can improve both fairness and performance.
Category: Opinion
What are popular Fairness metrics?
Demographic parity, equalized odds, equal opportunity, calibration within groups, and individual fairness are common metrics.
Category: Overview
How do you validate Fairness post-deployment?
Monitor disaggregated metrics, audit decisions, gather user feedback, and retrain when drift or harms are detected.
Category: Validation
What is the biggest misunderstanding about Fairness?
That one metric fits all; fairness is context-dependent and often involves irreconcilable trade-offs between definitions.
Category: Misunderstanding
What techniques improve Fairness?
Pre-processing reweighting, in-processing constraints, post-processing thresholding, and causal approaches help mitigate disparities.
Category: Techniques
How does Fairness relate to regulation?
Anti-discrimination laws and sector rules influence acceptable fairness criteria and documentation requirements.
Category: Regulation
Who should be involved in Fairness decisions?
Data scientists, domain experts, legal, ethics, and impacted stakeholders should collaborate on definitions and acceptance criteria.
Category: Collaboration