Mitigation at Pre-, In- and Post-Processing

Once a disparity is diagnosed and a criterion chosen, the intervention can act on the data, on the training objective, or on the outputs. The three differ in how much they cost, how much control they give, and, decisively, whether they are legally available.

Pre-processing

Modify the training data so a model fitted on it is less disparate: reweight examples so groups contribute equally, resample, or learn a representation from which the protected attribute cannot be predicted while task-relevant information is retained.

It is model-agnostic, so it survives a change of algorithm, and it is auditable in the sense that the transformation can be inspected. It gives weak guarantees, since nothing forces the downstream model to be fair, and adversarial representation learning in particular tends to remove less of the attribute than its objective suggests.

In-processing

Change what is optimised: add a fairness constraint or penalty to the objective, or train adversarially so that a discriminator cannot predict the attribute from the representation. Reductions approaches convert a fairness-constrained problem into a sequence of cost-sensitive classification problems, which is attractive because it works with any base learner.

This gives the most direct control, since the criterion is in the objective, and it typically achieves the best accuracy at a given fairness level because the tradeoff is optimised rather than imposed afterwards. The costs are that it requires access to training, ties the fix to the model, and adds a hyperparameter whose setting is the fairness-accuracy tradeoff made numerical.

Post-processing

Adjust the outputs: group-specific thresholds, or the randomised derivation that achieves equalised odds from any score. Applicable to a model you cannot retrain, including a third-party one, which is often the deciding practical advantage. It gives exact satisfaction of the chosen criterion, and it can require discarding information, since equalising rates from an unequal score sometimes means randomising.

The legal constraint

Post-processing with group-specific thresholds explicitly uses the protected attribute in the decision. In several jurisdictions and several domains that is disparate treatment and is unlawful regardless of its intent or effect, which removes the technically cleanest intervention from the available set.

This is the tension that makes fairness work legally awkward: measuring disparity requires the attribute, and the most direct corrections use it in a way the law may forbid. The interventions that survive are usually pre-processing and in-processing methods that use the attribute during development but not at inference, and confirming that distinction with counsel before building is cheaper than discovering it after.

When it breaks

The tradeoff is real and should be quantified. Constraining a model costs accuracy, and the honest presentation is a curve showing accuracy against disparity across the constraint strength, with the chosen operating point marked and justified. A single number hides the choice that was made.

Mitigations are brittle to distribution shift. A model constrained to satisfy a criterion on the training distribution can violate it on a shifted one, and the violation is silent. Fairness metrics belong in production monitoring, not only in pre-deployment evaluation.

Enforcing parity can harm the group it protects. Lowering a threshold for a group to equalise selection rates admits more people who will then fail, and where failure is costly, a loan default, a course they cannot complete, the intervention has transferred harm rather than removed it. Whether that is an improvement depends on facts outside the model.

The pipeline is not the system. A perfectly constrained model feeding a human who overrides it in a patterned way produces a disparate outcome from a fair model. Evaluating the model rather than the decision is the most common scoping error in fairness work.