Other Methods and Summary

In this section, we have discussed representatives of each mitigation method. However, the entire landscape is broader than these approaches! There are numerous algorithms available in different packages. Let’s briefly discuss some of them:

Preprocessing

• Learning Fair Representations (LFR): This method aims to encode original features into a latent space that retains essential information while simultaneously obfuscating information about protected attributes. Such a representation can be used with any classification or regression model.

• Optimizer preprocessing: Another method for altering the dataset representation to minimize the influence of sensitive features.

• Correlation removal: Correlations between protected attributes and valuable features can lead to unfair results. Removing such correlations makes it less likely for the algorithm to rely on them.

In-processing

• GerryFair: An algorithm for learning models (linear regression, SVM, or decision trees) that are fair with respect to rich subgroups. A rich subgroup is a linear combination of sensitive features. As a result, the model can easily handle multiple protected attributes.

• Prejudice removal: A classification algorithm that uses a fairness metric as a regularization term.

Post-processing

• Equalized odds post-processing: A dedicated method for satisfying the equalized odds constraint.

• Rejection option: This method modifies predictions within a model confidence band by assigning favorable predictions for underprivileged groups and unfavorable ones for privileged groups.

Many methods listed here are implemented in another popular fairness toolkit, aif360.

Choosing the algorithm

Unfortunately, there is no simple rule for choosing an algorithm. However, we can narrow down possible options depending on a specific metric to optimize and the problem type. The following table shows method compatibility and can help with the selection process.