Chapter 7 – Decreasing Bias and Achieving Fairness

1. No. There might be proxies in our models for sensitive attributes, but not in our models.
2. Salary and income (in some countries), occupation, a history of a felony charge.
3. Not necessarily. Satisfying fairness according to demographic parity wouldn’t necessarily result in a model being fair according to equalized odds.
4. Demographic parity is a group fairness definition to ensure that a model’s predictions are not dependent on a given sensitive attribute, such as ethnicity or sex.

Equalized odds is satisfied when a given prediction is independent of the group of a given sensitive attribute and the real output.

5. Not necessarily. For example, there could be feature proxies for 'sex' among top contributors to model predictions.
6. We can use explainability techniques to identify potential biases in our models and then plan to improve them toward fairness. For example, we can identify...