Practical Fairness

Chapter 5. Fairness In-Processing

This chapter focuses on fairness interventions that can be used during the process of training a model. While most guidance will encourage a fairness intervention as early in the data pipeline as possible, sometimes the earliest stages of fairness interventions, related to data pre-processing, won’t be available. Some potential cases where data pre-processing won’t be available could include the following:

• Proprietary or licensed data stores that may come with rules about how they can be stored or modified.

• Organizational guidelines that express disapproval of modifying baseline data sets. We can imagine some people having philosophical objections even if they are otherwise in favor of fairness interventions.

• Data sets that are so large that pre-processing interventions are not computationally feasible given your resources.

• A desire for the simplest pipeline and overall fewer numbers of steps from raw data to results.

If these or any similar use cases describe your constraints, you should consider in-processing as an optimal point in the data modeling pipeline for a fairness intervention. In-processing refers to any fairness intervention that modifies the training process for a machine learning model. If we want to think more broadly than machine learning training pipelines, an in-processing intervention would be an intervention in which the decision-making process itself is modified by changing the feedback that is used to assess whether ...