Mitigating Group Disparity

## The problem being solved A disparity here means the model's behaviour differs between groups defined by a protected attribute such as sex, race, age or disability status - one group sees a higher false positive rate, a lower selection rate, or a worse recall than another. Suppose the audit is done and the gap is real and material. The question is what you can actually turn. Every mitigation touches the pipeline in exactly one of three places, and the taxonomy is worth memorising because it maps one-to-one onto what access you have and what you are allowed to do. ## Pre-processing: change the data You modify the training set before any model sees it, so the disparity is smaller in the signal the learner is given. - **Reweighing** attaches a weight to each training row so that, in the weighted data, group membership and the label are statistically independent. Rows in under-represented (group, favourable outcome) cells get weights above one; over-represented cells get weights below one. - **Relabelling** or *massaging* flips the labels of a small number of borderline rows near the decision boundary to balance the outcome rates. It is more invasive - you are editing ground truth - and needs a strong argument that the original labels were themselves biased. - **Representation repair** transforms the features so the protected attribute becomes hard to recover from them. The defining property: the protected attribute is consumed at training time and never again. The shipped scorer is group-blind. That is a large practical advantage. ## In-processing: change the fit You modify the objective the learner optimises. - **A penalty term.** Add a differentiable surrogate of the disparity to the loss: `loss = task_loss + lambda * disparity`. Turning up `lambda` buys a smaller gap and gives back accuracy. On a job-ad delivery model, an equalized-odds penalty that closed a 9-point false positive rate gap cost about 1.5 AUC points - a real and typical order of magnitude. - **A hard constraint.** Solve the same task subject to the disparity staying under a bound, rather than trading it off softly. - **Adversarial debiasing.** Attach a second head - an adversary - that tries to predict the protected attribute from the main model's representation or output. The adversary is trained to succeed; the main model is trained to do its own job *and* to make the adversary fail. On a customer-service escalation model an adversary might try to read the speaker's dialect off the representation. At convergence the representation carries little group information. The catch is that min-max training is unstable, and the guarantee only covers what that particular adversary is capable of detecting. Again the protected attribute is only needed while training. ## Post-processing: change the decision The fitted model is untouched; you change how its scores become decisions. - **Per-group thresholds.** Accept above 0.52 for one group and above 0.61 for another. Mechanically the most direct lever on a measured gap, because it acts on exactly the quantity being measured. - **Reject-option rules.** Define a band of scores near the boundary where the model is least confident, and inside that band favour the disadvantaged group or route the case to a human reviewer. The defining property is the mirror image of the other two: these rules read the protected attribute *at decision time*, which many deployments cannot do and many regulators treat very differently from using it during training. ## Choosing between them Ask, in order: (1) Do I control training, or only a vendor's scores? If only scores, post-processing is your only option. (2) Will the protected attribute be present, accurate and permitted at decision time? If not, post-processing is out. (3) Do I believe the labels? If the labels themselves encode the historical bias, a mitigation applied downstream of them is polishing a corrupted target, and data work comes first. ## The accuracy cost A constraint shrinks the feasible set, so the best training accuracy achievable under a fairness constraint can only match or fall below the unconstrained optimum - a fairness fix that appears free in-sample usually means the constraint was not binding. Out-of-sample, a constrained model occasionally scores *better*; read that as evidence the disparity came from fitting noise in a thin slice, and go check, rather than as a free lunch. Whichever family you pick, the fix is not done until the gap is re-measured on held-out data, the residual gap is written down, and monitoring is in place - mitigations decay as base rates and populations drift.