AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 2
Question 10
A retailer has historical labeled transactions marked returned/not returned, with numeric and categorical predictors. The business needs a probability for each new transaction. Which problem formulation is appropriate?
- Unsupervised k-means clustering of unlabeled purchase histories.
- Supervised binary classification with a calibrated probability output.
- Continuous-value regression to predict shipping weight.
- Sequence forecasting of total weekly sales without transaction labels.
Explanation
Correct Answer
B:Binary classification maps labeled examples to the two outcomes and can produce a per-transaction probability for the positive outcome.
Incorrect Answers
A:Clustering discovers groups but does not directly learn the supplied returned/not-returned target or its calibrated probability.
C:Shipping-weight regression predicts a different continuous target, not the returned class specified in the requirement.
D:Aggregated demand forecasting answers a time-series question and does not classify individual labeled transactions.
Question 11
A manufacturing team must predict the remaining service hours of a machine from temperature, age and vibration. Historical records contain numeric hours until failure, not categorical labels. Which target type is appropriate?
- Binary classification of whether the motor type is A or B.
- Unsupervised segmentation with no target field.
- Regression against continuous remaining operating hours.
- Multiclass classification of the machine factory identifier.
Explanation
Correct Answer
C:Regression models a continuous numeric target such as remaining service hours and supports numeric error evaluation.
Incorrect Answers
A:Classifying a motor type would solve an equipment-label problem rather than predicting remaining lifetime.
B:Unsupervised segmentation does not use the known numeric lifetime target for learning.
D:Classifying a factory ID does not estimate how many hours remain before the machine requires service.
Question 12
A support organization has millions of unlabeled chat ticket embeddings and wants to discover recurring themes for a routing taxonomy before human labels exist. Which initial modeling approach best matches the available data?
- Train a supervised binary classifier requiring known urgent/not-urgent labels.
- Apply clustering to the embedding vectors and review the resulting groups with subject experts.
- Predict each ticket creation timestamp using continuous-value regression.
- Fit a supervised multiclass model using agent IDs as if they were verified topic labels.
Explanation
Correct Answer
B:Clustering can explore natural similarity groups in unlabeled embeddings; experts can then validate whether groups reflect useful themes.
Incorrect Answers
A:Without defined urgency targets, a supervised binary classifier lacks the ground-truth labels it must learn.
C:Predicting timestamps does not discover semantic categories for routing the chat tickets.
D:Agent IDs capture assignment history, not an independently validated set of issue-topic labels.
Question 13
A 10-hour model-training job uses SageMaker Managed Spot Training and occasionally loses capacity. Restarting from epoch zero consumes too much time. What mechanism best preserves useful progress between interruptions?
- Increase the number of random training seeds.
- Persist model checkpoints to the configured S3 location and restore from the latest saved checkpoint.
- Disable all CloudWatch logs to shorten startup.
- Replace validation data with additional training examples.
Explanation
Correct Answer
B:SageMaker managed training can synchronize checkpoint files to S3 and restore them on restart after a Spot interruption.
Incorrect Answers
A:Seed selection controls randomness but does not preserve learned parameter state when an instance is interrupted.
C:Removing logs does not retain model weights, optimizer state or the current epoch across interrupted instances.
D:Moving holdout examples into training creates evaluation leakage but does not implement interruption recovery.
Question 14
A SageMaker automatic model tuning job compares XGBoost classifiers for a fraud detector. The selected business metric is validation ROC-AUC, where larger values are better. How must the tuning objective be configured?
- Set the objective metric to validation AUC and the optimization direction to Maximize.
- Set the metric to AUC but use Minimize so the smallest area is selected.
- Use training-job duration as the sole metric because it always correlates with ROC-AUC.
- Use validation RMSE with no comparison to the fraud-classification objective.
Explanation
Correct Answer
A:SageMaker HPO selects candidates by an explicitly specified metric and direction; higher validation AUC is desirable here.
Incorrect Answers
B:Minimizing AUC would intentionally choose worse ranking discrimination under the stated objective.
C:Short training duration can be useful for cost but does not establish classification discrimination and is not the specified metric.
D:Regression RMSE measures numeric errors and does not replace the agreed binary-classification AUC objective.
Question 15
A hyperparameter search wastes GPU hours on training trials whose validation loss has clearly stopped improving relative to stronger trials. The algorithm reports the selected metric after each epoch. Which change best limits this waste?
- Turn off the validation metric so poor trials run without comparisons.
- Increase the maximum epochs for all trials without a stopping criterion.
- Enable supported automatic early stopping in the SageMaker tuning job.
- Select model variants solely by their training-set accuracy.
Explanation
Correct Answer
C:Automatic early stopping can terminate underperforming training jobs based on their reported objective metrics across epochs.
Incorrect Answers
A:Disabling metrics prevents the tuning service from assessing which trials are underperforming.
B:Allowing more epochs for every weak candidate typically increases compute consumption without adding selection discipline.
D:Training accuracy can be optimistic and does not provide the validation evidence needed to stop poor trials.
Question 16
A hospital triage screening model flags rare emergencies. Missing an actual emergency is far more costly than generating a false alert. Which validation metric should receive strongest emphasis, while also reviewing precision?
- Overall accuracy, because the majority class can dominate without issue.
- Mean absolute percentage error on categorical labels.
- Recall for the emergency (positive) class.
- Training dataset row count, regardless of predictions.
Explanation
Correct Answer
C:Positive-class recall measures the fraction of actual emergencies identified, matching the cost of false negatives.
Incorrect Answers
A:Overall accuracy can be very high by predicting the majority non-emergency class and still missing most emergencies.
B:Percentage error is a regression-style calculation and is not defined as the relevant classification miss metric.
D:Row count is a data-coverage measure and cannot show whether the model catches the positive clinical events.
Question 17
A team compares two numeric demand forecasts. Both are evaluated on the same units and test periods; the manager wants to penalize occasional large prediction errors strongly. Which metric fits?
- Micro-averaged multiclass F1.
- Root mean squared error on the held-out numeric forecast values.
- Categorical confusion-matrix accuracy.
- The number of examples used by the training DataLoader.
Explanation
Correct Answer
B:RMSE squares residuals before averaging and taking a root, making relatively large numeric errors contribute more strongly.
Incorrect Answers
A:F1 summarizes precision and recall for classification; it is not a squared-error measure for numeric forecasts.
C:Classification accuracy requires discrete labels rather than continuous numeric forecast values.
D:Training example count does not directly quantify held-out prediction error or the severity of large misses.
Question 18
An established SageMaker Clarify customer needs to explain why a specific approved credit model produced a particular score. The requirement is per-prediction feature attribution rather than aggregate endpoint traffic metrics. What technique is directly relevant?
- Use SHAP-based feature attribution for the individual prediction and document the chosen baseline.
- Increase the endpoint instance size and treat lower latency as a feature explanation.
- Create an S3 inventory report and use object counts as evidence of a decision factor.
- Change the model package approval status without analyzing the input record.
Explanation
Correct Answer
A:SHAP-based attributions estimate each input feature’s contribution to the particular model output relative to a defined baseline.
Incorrect Answers
B:Serving capacity can affect latency but cannot explain the contribution of input fields to the model score.
C:S3 inventory summarizes objects and metadata; it does not compute feature-level prediction attribution.
D:Model approval is a deployment governance decision, not an explanation of one scored transaction.
Question 19
A quality-assurance team submits 200 MB images to an ML model. Processing one image takes 12 minutes, and callers can poll for results instead of waiting on an HTTP connection. Which SageMaker hosting option fits these constraints?
- Real-time Inference with a regular 60-second invocation response limit.
- Serverless Inference with a 4 MB payload limit.
- Asynchronous Inference with queued requests and S3-backed results.
- A multi-variant real-time endpoint that still requires a synchronous prediction response.
Explanation
Correct Answer
C:Asynchronous inference queues large requests, supports processing up to one hour, and returns results without holding a synchronous connection open.
Incorrect Answers
A:Normal real-time responses must fit the synchronous invocation duration; the described 12-minute operation exceeds the stated regular limit.
B:Serverless inference is designed for shorter, smaller requests and the 200 MB payload exceeds its documented maximum.
D:Adding variants changes how live traffic is routed but does not turn ordinary real-time inference into a long-running queued service.
Question 20
A retailer needs predictions on an S3 archive of 40 million records once per month. There is no interactive application, and keeping a model-serving endpoint active between jobs is unnecessary. Which deployment mode is most suitable?
- A permanently provisioned real-time SageMaker endpoint.
- SageMaker Batch Transform on the available input objects.
- An asynchronous endpoint that queues every row independently all month.
- A serverless endpoint invoked manually once per record in a tight single-thread loop.
Explanation
Correct Answer
B:Batch Transform is designed for large offline datasets when no persistent serving endpoint is required.
Incorrect Answers
A:A persistent endpoint would incur unnecessary serving overhead when predictions are only generated as an offline job.
C:An async endpoint is useful for queued inference requests, not a natural replacement for a single bulk offline transform job.
D:Per-record serverless calls add invocation complexity and are not the straightforward bulk S3 inference mechanism.
