AIF-C01 Practice Question 28: Feature Attribution Explainability
AWS Certified AI Practitioner (AIF-C01) · Test 3 · question 8 of 10
A lender's regulator asks the bank to show, for one specific rejected loan application, which input features pushed that decision toward a rejection and by how much. The model is a gradient-boosted tree deployed on Amazon SageMaker AI, and the bank already reports its overall accuracy. Which approach produces the evidence the regulator asked for?
Show answer and explanation
Answer:
The regulator is asking an explainability question about one prediction. SageMaker Clarify uses a model-agnostic feature attribution approach built on SHAP (Shapley values), which assigns each feature an importance value for a particular prediction, so you can answer "why did the model reject this applicant" and show the contribution of each input. Clarify can do this after training or per-instance during inference, which is exactly the per-application evidence required (option D).
- A is wrong because Model Monitor watches for data and quality drift over time. Knowing that today's traffic resembles the training distribution says nothing about why one applicant was rejected. - B is wrong because aggregate metrics describe how often the model is right across a whole dataset. The bank already reports accuracy, and the regulator asked for something a global metric cannot give: the reasoning behind a single decision. - C is wrong because Guardrails filters the inputs and outputs of generative AI models on Amazon Bedrock. It does not apply to a tabular gradient-boosted tree, and blocking a topic is a control, not an explanation of a decision already made.
Reference: Model Explainability with Amazon SageMaker Clarify