Explainability, Fairness, and Bias
What this domain tests
Explainability, fairness, and bias questions test whether you can choose the right explanation for the audience and interpret fairness metrics without overclaiming.
Core rules
- Global explanations describe model behavior overall; local explanations justify an individual decision. Regulated adverse-action contexts often require reason codes.
- Inherently interpretable models simplify explanation; post-hoc methods such as SHAP or LIME approximate behavior and can be unstable.
- Fairness definitions include demographic parity, equalized odds, and predictive parity. They can conflict; no model satisfies all simultaneously in general.
- Bias can enter through sampling, labels, proxies, feature engineering, and deployment feedback.
- Mitigation can be pre-processing, in-processing, or post-processing; each has trade-offs.
Worked example
A hiring model selects 30% of male applicants and 21% of female applicants. The selection-rate ratio is 0.70, below the 0.80 four-fifths rule, indicating disparate impact. Removing gender does not fix it because job history and referral variables may proxy gender. The team should test proxies, consider less discriminatory alternatives, and document the chosen fairness criterion and trade-offs.
Traps
Reporting one fairness metric as definitive; assuming removal of a protected attribute removes bias; treating SHAP values as causal; and giving explanations that are unstable or incomprehensible to the decision user.