BiasClean v3.0 offers an optional, experimental layer of protection: model-level fairness constraints using Support Vector Machines (SVM).
Where BiasClean’s core mitigation works on the dataset — rebalancing and correcting disparities in the data itself — SVM Fairness Enforcement works at the model level, adding fairness-aware constraints directly into the prediction process. This provides fairness-aware predictions alongside BiasClean’s mitigated data, giving an additional layer of protection beyond data-level correction alone.
How it works: Support-vector-machine rebalancing is applied to correct disparities identified during the audit stage, while working to preserve the dataset’s overall predictive accuracy.
Important — this is experimental. Enabling SVM Fairness Enforcement may introduce accuracy trade-offs. It is offered as an optional toggle at upload time, not a default behaviour, so you can choose whether the added fairness constraint is worth the potential impact on predictive performance for your specific use case.