Subgroup Fairness, Instantaneous Fairness, and Distributional Repair

[JoeJohnson01]

Subgroup Fairness, Instantaneous Fairness, and Distributional Repair

The Subgroup Fairness and Instantaneous Fairness algorithms focus on minimizing disparities between different groups over time and across subgroups. These algorithms expand on a conventional fairness model to address scenarios where fairness might vary over different time periods or among subgroups with multiple sensitive attributes.

The Distributional Repair algorithm focuses on conditional independence. The idea is to reduce the correlation between sensitive and non-sensitive features in a dataset through use of an OT plan, increasing fairness in the dataset ahead of training.

Related Paper

For additional information, and results of model evaluation, please refer to the paper linked below, written by myself as part of my Masters Dissertation at Imperial College London

Practical_Aspects_of_Fairness_in_AI (1).pdf

References:

Abigail Langbridge, Anthony Quinn, and Robert Shorten
Quan Zhou, Jakub Marecek, and Robert N. Shorten

Closes #549

[rahulnair23]

HI @JoeJohnson01 - thanks. I've started the initial review. For now, could you have look at the following:

1. The DCO has been incorrectly signed. Could you sign off the commit correctly?
2. The LDS methods involve several dependencies. Could you include relevant packages in the setup, so that the methods are pip installable? Currently, throws missing import errors.
3. There appears to be code relating to two solvers. Do you need one or both? Your links suggest that SDPA is preferred.
4. For distributional repair - do you have any evaluations of downstream tasks? It would be useful if the demonstrator notebook has a more concrete use case where data repair improves fairness outcomes. Currently, it ends just by discussing repair quality. How would you like end-users to invoke this preprocessing step?