Practice — Responsible ML: Interpretability, Fairness & Privacy (6 questions)

Your credit-risk committee has seen the following explanation for one declined applicant from a gradient-boosted model. The model outputs log-odds of repayment; the base value (average over a 100-row background of applicants) is 0.99.

(a) What does "additivity" mean here, and what is the model's predicted repayment probability for this applicant? (b) Why is TreeSHAP able to give these values exactly and quickly, when Shapley values normally need all coalitions of features, and which two variants of TreeSHAP exist? (c) A product manager says: "Income has a positive SHAP value, so if we raise applicants' stated income their approval chance goes up and we should advise it." Answer. (d) What would you change if the committee asks why the reasons would differ if you used approved applicants as the background?

Two teams ask you which fairness metric to use.

1. Hiring screen. A model ranks resumes and the top 30 percent go to a recruiter. The applicant pool has different group shares per role.
2. Emergency triage. A model flags patients who need urgent attention among those arriving at an emergency department; a missed flag can be fatal, and a false flag costs a nurse's time.

For each: pick a primary criterion among demographic parity, equal opportunity, equalized odds, predictive parity and calibration within groups; name the check you would add; say what you would do about the label (what the model is trained to predict); and state what you give up. Then explain why you cannot simply demand all of them.