DataTerminology.comNEW
A fairness notion where predictions remain unchanged under counterfactual changes to protected attributes.
DESCRIPTION
Counterfactual fairness uses causal models to require that an individual’s prediction would be the same in a counterfactual world where their protected attribute differed. It formalizes fairness at the individual level rather than groups. This approach demands explicit causal assumptions and high-quality data, and can be challenging to validate. It is powerful for reasoning about interventions and unfair pathways. Adoption is growing where causal expertise is available.
ALTERNATIVE DEFINITIONS
Kusner, Loftus, Russell, Silva (Counterfactual Fairness)
A predictor is counterfactually fair if its prediction for any individual is the same in the actual world and in a counterfactual world where the individual belonged to a different demographic group.
Original definition from the seminal paper introducing counterfactual fairness.
Fairness and Machine Learning (Barocas, Hardt, Narayanan)
Counterfactual fairness requires that the prediction for an individual be invariant under interventions on protected attributes in a causal model.
Emphasizes interventions within structural causal models.
NIST AI Risk Management Framework
Causal and counterfactual methods can support fairness by evaluating how changes to protected attributes would affect outcomes in well-specified models.
Places counterfactual fairness within a broader risk and governance context.
RELATED TERMS
Causal Inference Structural Causal Model Individual Fairness Do-Calculus Protected Attribute Causal Graph Interventions Fairness Metrics Explainability Sensitivity Analysis
COMPETING TERMS
RELATIONS
- Broader: Fairness
KNOWLEDGE AREAS
AI Data Science AI Governance Mathematics
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer Data Governance Manager
EXTERNAL SOURCES
- Wikipedia - Fairness (machine learning)
- arXiv - Original paper on concept
- Fairness and Machine Learning - Textbook coverage of causal fairness
- NIST - Risk framing for causal fairness
TOP QUESTIONS
Why use Counterfactual Fairness instead of group metrics?
It addresses individual-level fairness and causal pathways, avoiding some paradoxes of group metrics. It’s valuable when causal knowledge is available and justified.
Category: Guidance
What data do I need for Counterfactual Fairness?
High-quality data on relevant covariates and domain knowledge to build structural causal models are essential, along with assumptions you can justify and test.
Category: Requirements
What are common pitfalls with Counterfactual Fairness?
Unjustified causal assumptions, omitted variables, and misuse of proxies can yield false assurances of fairness.
Category: Mistakes
How do you evaluate Counterfactual Fairness empirically?
Estimate structural models, simulate interventions on protected attributes, and compare predicted outcomes for individuals.
Category: How-To
What is the main criticism of Counterfactual Fairness?
It can be impractical due to modeling complexity and untestable assumptions, especially in high-dimensional settings.
Category: Opinion
Is Counterfactual Fairness compatible with Equalized Odds?
Not necessarily; different fairness notions can conflict. Multi-objective optimization and policy guidance are needed.
Category: Comparison
Where to start learning Counterfactual Fairness?
Read the Kusner et al. paper, then study causal inference basics with Pearl’s SCMs and fairnessbook chapters.
Category: Getting Started
How does Counterfactual Fairness relate to explainability?
Both can use causal models; counterfactuals explain “what-if” changes while also supporting fairness testing.
Category: Relation
What is a risk of applying Counterfactual Fairness incorrectly?
False confidence from misspecified causal graphs can hide harmful biases.
Category: Risks