Skip to main content
Back

Shapley Meets Rawls Unifies AI Fairness Audits and Feature Attribution

Researchers show the Shapley value can both quantify and attribute group unfairness, cutting inference cost versus bootstrap on Census Income data.

Paper, Full Access
  • Paper reframes standard group fairness criteria as differences of group-Shapley values, unifying measurement and explanation.
  • A two-stage nested Shapley attributes the group gap to individual features with axiomatic guarantees.
  • Framework extends to five Efficient-Symmetric-Linear values, aggregated by majority voting for robust attributions.
  • Plug-in variance estimator enables inference without bootstrap, cutting runtime versus 1,000-replication baselines.
  • On UCI Census Income, Age, Hours and Marital status drive gender unfairness.
  • Experiments ran on the Toubkal supercomputer using a single 56-core CPU node, no GPU.

Fairness audits measure gaps between groups, while SHAP-style tools assign model outputs to features. The arXiv preprint Shapley Meets Rawls connects those tasks by expressing group-fairness gaps as differences between Shapley allocations. The resulting pipeline measures a gap, tests its statistical significance, and decomposes it into feature-level contributions.

One allocation rule, three outputs

The Shapley value comes from cooperative game theory, where it divides a coalition's total payoff among participating players. In model explanations, the players are features and the payoff is a prediction. The paper instead constructs a game over demographic groups and treats an outcome metric as the quantity to allocate.

For fairness criteria defined as differences between group metrics, the authors show an equivalent representation as a difference between group-Shapley values. Demographic parity, for example, compares positive prediction rates across groups. Equal opportunity compares true-positive rates, while equalized odds examines both true-positive and false-positive rates. Each relevant metric can become the payoff in the group-level game.

The decomposition relies on three properties of Shapley-style allocations:

  • Efficiency: allocations sum to the total value being distributed.
  • Symmetry: players with identical effects receive identical allocations.
  • Linearity: allocations for combined games equal the sum of their separate allocations.

Those properties let the fairness gap inherit a structured decomposition, so the feature contributions add back to the measured group difference.

How the proposed audit is assembled
Stage Calculation Output
Measure Difference between group-Shapley values