Wiki Home
Group fairness
Definition
Group fairness requires a model to behave equitably across groups defined by a sensitive attribute such as race, sex, or age. The requirement is stated as a statistical constraint on the model’s predictions conditioned on the sensitive attribute and the true label. Demographic parity asks that the prediction be independent of the sensitive attribute; equalized odds asks for equal true- and false-positive rates across groups; equality of opportunity asks for equal true-positive rates (Hardt et al., 2016). The constraint is enforced by pre-processing the training data to remove bias, in-processing through a penalty or adversarial term in the training objective, or post-processing the model’s outputs to satisfy the chosen metric.
The notion is group-level and statistical. It bounds disparities between subgroup statistics and makes no promise about any individual. Different metrics encode different goals and cannot in general be satisfied at once.
Papers that use this concept
- SoK: Unintended Interactions among Machine Learning Defenses and Risks — treats group fairness as a defense and surveys how enforcing it shifts a model’s susceptibility to privacy and security risks.
- PAL*M: Property Attestation for Large Generative Models — fairness is an example attestable property: its attribute-distribution measurement underlies fairness auditing, and it cites zero-knowledge certificates of fairness as the cryptographic route to the same goal.
Variants and traps
- Demographic parity, equalized odds, and equality of opportunity are distinct constraints; satisfying one generally precludes the others.
- The guarantee is over subgroup statistics, so an individual can still be treated in a way the group-level metric does not register.
See also
References
- Hardt, M., et al. “Equality of Opportunity in Supervised Learning.” Advances in Neural Information Processing Systems (NeurIPS), 2016.