Bake it into training
In processing methods modify the learning algorithm itself so fairness is enforced while the model trains. Rather than cleaning data first or adjusting outputs later, the objective itself rewards both accuracy and fairness.
How it is done
- Penalty terms: add a regularizer to the loss that grows when a fairness metric is violated, trading accuracy against fairness by a tunable weight.
- Hard constraints: solve a constrained optimization that bounds the unfairness while maximizing accuracy.
- Adversarial training: a second network tries to predict the protected attribute from the model, and the main model is trained to defeat it, removing that information.
Tradeoffs
In processing can target a specific fairness metric precisely, because it optimizes for it directly. The cost is complexity: training is harder, less portable across model types, and tied to the chosen metric.
Key idea
In processing methods add fairness penalties, constraints, or adversaries to the training objective so the model learns to be fair directly, giving precise control at the price of more complex training.