Equality of Opportunity in Supervised Learning

Inside the paper

The problem

Predictions used for decisions such as lending can treat protected groups differently, while simply withholding a protected attribute does not prevent other features from encoding it.

The idea

The authors define fairness criteria in terms of prediction errors conditional on the actual outcome, and show how an existing predictor can be adjusted to satisfy them.

How it works

• For equalized odds, compare positive prediction rates across protected groups separately for people with each actual outcome.
• Equal opportunity relaxes this requirement to the group with the favorable actual outcome, matching true positive rates across groups.
• Use group-specific, possibly randomized post-processing of an existing predictor to satisfy the selected constraint while minimizing a chosen prediction loss.
• For scores, examine group receiver operating characteristic curves to find feasible thresholds or mixtures of thresholds that meet the constraint.

What it showed

The paper gives theoretical results for these adjustments and illustrates tradeoffs using a credit-score case study. It also shows that statistics of predictions, outcomes and group membership alone cannot distinguish every underlying source of unfairness.

Read it for

Read Section 2 for the definitions and Figure 1 with Section 4 for the adjustment. Section 7 shows the credit example; the proofs and Section 5’s optimality bounds can wait until the setup is familiar.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.