Option 3: Post-Processing
If we decide to mitigate bias at the post-processing stage, we will have to take the predictions from the model and perform a separate process to reduce bias. We will need as inputs: existing model predictions, labels and group membership; with this we can create a new set of less biased predictions, for example using model optimization. Notice that in order to do this you will need access to protected attributes at the point of inference, which is not always possible.
aif360 proposes the following post-processing mitigation techniques:
Equalized Odds Postprocessing, which alters output labels to optimize equalized odds. It uses linear programming in order to find probabilities with which to modify the labels (Pleiss 2017, Hardt 2016).
Calibrated Equalized Odds Postprocessing, which optimizes over calibrated model outputs to find probabilities with which to alter output labels with an equalized odds objective(Pleiss 2017, Hardt 2016).
Reject Option Classification, which gives favorable outcomes to unprivileged groups and unfavorable outcomes to privileged groups in a confidence band around the decision boundary with the highest uncertainty (Kamiran 2012).
An example of how to do this in a binary classification problem in recruitment can be found on our notebook here, or downloading the following file:
Last updated