MLA-C01 Practice Question 55: Linear Learner Target Recall
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 6 · question 5 of 10
An insurer trains a binary classifier with the SageMaker AI built-in Linear Learner algorithm to flag potentially fraudulent claims for investigation. A regulator requires the model to catch at least 90% of fraudulent claims in validation. Investigators are overloaded, so among the models that meet that requirement, the insurer wants the one that sends them the fewest legitimate claims.
How should the engineer configure the training job?
Show answer and explanation
Answer:
Linear Learner trains several models in parallel and uses binary_classifier_model_selection_criteria to pick the final model on the validation data. With precision_at_target_recall, recall is held at target_recall while precision is maximized. Catching 90% of fraud is a recall floor, and sending investigators the fewest legitimate claims means the fewest false positives, which is the highest precision. That is exactly this criterion (option A).
- B is wrong because it swaps the roles: it holds *precision* at 0.9 and maximizes recall, which does not guarantee the regulator's 90% recall floor. - C is wrong because F-beta with beta 2 favors recall in a single blended score, but it sets no hard floor. The chosen model could fall short of 90% recall, or give up precision without need. - D is wrong because accuracy (the default) is dominated by the many legitimate claims in an imbalanced dataset, and adjusting a threshold by hand afterward bypasses the built-in selection meant for this trade-off.
Reference: Linear learner hyperparameters