Projects Open source Since 2023
EquiPy
A Python package for post-processing fairness: it takes a trained model and returns a calibrated, fairer one, for a single sensitive attribute or several at once.
What it does
Much fairness tooling asks you to retrain. EquiPy works the other way round: it sits after your model and transports its score distribution towards a Wasserstein barycenter, so that the distributions seen by each protected group coincide, either exactly or partially, along a tunable fairness path.
The multi-attribute case is the interesting one. Correcting for gender and then for age is not the same as correcting for both, and naive sequential correction undoes itself. EquiPy implements the sequential mechanism from our AAAI 2024 paper, which composes corrections in a way that keeps earlier guarantees intact.
Why it matters in insurance
A pricing model cannot usually be retrained on demand: it is validated, documented and sometimes filed with a regulator. Post-processing lets you audit and correct an existing model, and quantify what the correction costs in premium accuracy. In our experience that last number is the one a risk committee asks for first.
Related work
- A sequentially fair mechanism for multiple sensitive attributes, AAAI 2024
- Fairness in multi-task learning via Wasserstein barycenters, ECML-PKDD 2023
- Mitigating discrimination in insurance with Wasserstein barycenters, BIAS 2023