Strategies to Mitigate Bias

A bias finding is the start of an engineering task, not a verdict. Mitigations fall into three families, ordered by where they act in the pipeline.

Before training (pre-processing)

• Fix representation — collect or re-weight data so under-represented groups have adequate samples.
• Audit features — remove or transform proxies for protected attributes (postcode is the classic proxy for ethnicity or income).
• Question the label — if the historical outcome you're predicting was itself biased (past hiring decisions, past arrests), no amount of modelling fixes it. Consider a cleaner target.

During training (in-processing)

• Add fairness constraints or penalties to the loss function.
• Train per-group calibration.

After training (post-processing)

• Threshold adjustment — per-group decision thresholds chosen to equalise TPR/FPR. Often the fastest fix, but document the reasoning carefully.
• Human review lanes — route borderline scores to human decision-makers (this also strengthens your Art. 14 human-oversight story).

Whatever you do — document it

Under Art. 10, examining and mitigating bias is an explicit data-governance duty for high-risk systems. Record: the metric, the finding, the mitigation chosen, the alternatives rejected, and the re-measured result. Then re-run the bias report on fresh data at a fixed cadence — mitigation drifts as data drifts, which is what the monitoring module is for.