AIF-C01 Practice Question 69: Interpretable Model Tradeoff

The Machine Learning Lens names both halves of this situation as anti-patterns: *prioritizing model performance metrics over interpretability when business or regulations requires explainability*, and *using complex models when simpler, more interpretable alternatives would meet business requirements*. Its guidance states that *complex models like deep neural networks may deliver higher accuracy but are often harder to interpret*, while *simpler models like decision trees or linear regression provide more straightforward explanations but might sacrifice some performance*, and advises that if high explainability is required, consider inherently interpretable models like decision trees, rule-based systems, or linear models. AWS even uses credit approval as its worked example. Two points of AUC is precisely the "might sacrifice some performance" trade the regulation compels here (option B).

- A is wrong because aggregate metrics describe behaviour across a whole population. They cannot tell an individual applicant why *their* application was declined, which is what the regulator is asking for. - C is wrong on two counts: a reviewer who cannot see why the model decided is no better placed to give the reason, and the safeguard would reach only applicants who complain rather than every decline. - D is wrong because picking whichever model gives the nicer answer is not a transparency measure. The reason given would not describe how the decision is actually made, and two models can disagree for reasons nobody can articulate.

Reference: MLPERF06-BP02 Evaluate model explainability