Databricks ML Associate Practice Question 27: Interpretable Model Choice
Databricks Certified Machine Learning Associate · Test 3 · question 7 of 10
A lender must give every declined applicant a written reason citing the specific factors that counted against that individual application. A regulator requires the reason to follow from the model's own decision rule rather than from an approximation produced afterwards. Predictive performance across the candidate models is close enough that it is not the deciding factor.
Which model should the team choose?
Show answer and explanation
Answer:
Logistic regression is intrinsically interpretable. The score for an applicant is a weighted sum of their feature values, so each feature's contribution to *that* decision is the coefficient times the value, a signed, per-applicant quantity that is the decision rule itself rather than a reconstruction of it. That is what an adverse-action reason requires, and it is why regulated credit decisioning leans on linear models even where an ensemble scores marginally better.
Option A is the arrangement the regulator has ruled out. SHAP and LIME produce approximations of a black-box model's behaviour after the fact, and they are explicitly not the model's own decision rule.
Option C confuses global with local explanation. feature_importances_ describes which features mattered across the whole forest; it says nothing about why one applicant was declined, and it is unsigned, so it cannot say whether a feature helped or hurt.
Option D has the same limitation as C with more complexity. Early stopping controls overfitting and has no bearing on explaining an individual prediction.
Reference: Linear models (scikit-learn)