3. Fairness II

How can a threshold be determined if the distributions of 2 groups are very different?

• 2 group-specific thresholds.

  m is a metric of interest, e.g., FPR, acceptance rate

What are 3 approaches to ensure fair classification?

• Pre-processing: Change training data.

• In-processing: Change model training procedure.

• Post-processing: Change model decisions.

→ In-processing is most flexible.

What is a problem in training models under fairness constraints?

Non-convex for many well-known classifiers (logistc, SVM).

→ Hard to compute efficiently.

How can Hyperparameter Optimization be used for fairness?

• Hyperparameters have massive impact on performance.

• Train model under different hyperparameters and observe trade-off between accuracy and unfairness.