Recently, I started writing a series of posts exploring bias in AI and different ways to mitigate it in a workflow in greater detail. In my last two blogs, I covered reweighing as a mitigation technique within the pre-processing stage of modeling, and adversarial debiasing during the in-processing (model training) stage of the machine learning workflow.
The third stage in the machine learning (ML) pipeline where we can intervene to reduce bias is called post-processing. Post-processing algorithms are mitigation steps that can be applied to the model predictions. On Fairness and Calibration [1], Equality of Opportunity in Supervised Learning [2] and Decision Theory for Discrimination-aware Classification [3] are among the different post-processing bias mitigation techniques proposed from academic literature.

#artificial-intelligence #machine-learning #data-science