InProcessing

Inprocessing is a category of algorithmic fairness techniques that incorporate bias mitigation directly into the training of a predictive model. Unlike pre-processing approaches that modify the data before learning or post-processing methods that adjust predictions after training, inprocessing embeds fairness constraints within the optimization objective or the model architecture itself. The goal is to reduce the model's reliance on protected attributes such as race, gender, or age while maintaining predictive accuracy.

Common inprocessing approaches include adversarial debiasing, constrained optimization, and regularization methods. Adversarial debiasing trains a secondary model to predict the protected attribute from the primary model’s internal representations; the primary model is updated to both perform its task well and hinder the adversary from succeeding. Constrained optimization and regularization add fairness constraints directly to the loss function, for example enforcing demographic parity or equalized odds through Lagrangian multipliers. Fair representation learning aims to remove information about sensitive attributes from latent representations so that these attributes have little to no influence on predictions.

Practical considerations include the requirement for sensitive attribute data during training, the dependence of results on the chosen fairness metric, and potential trade-offs with accuracy. Inprocessing methods may offer stronger or more direct fairness guarantees for the trained model under specific assumptions, but they do not guarantee fairness in all deployment contexts or under distribution shifts. They are widely discussed in machine learning fairness literature and applied in domains such as lending, hiring, and risk assessment.