AWS Certified Machine Learning - Specialty Practice Exam 2025: Latest Questions

A data science team stores raw CSV training data in Amazon S3. They need to run SQL queries to quickly understand schema and basic aggregates without provisioning servers. Which approach is MOST appropriate?

A company wants to build a near-real-time feature store pipeline. Events are ingested continuously, transformed, and written to Amazon S3 and Amazon Redshift for training and analytics. Which AWS service is BEST suited to perform continuous data transformations with minimal operational overhead?

A machine learning engineer is training an XGBoost model on tabular data and wants to reduce the impact of features with large numeric ranges. Which preprocessing step is MOST appropriate?

A team deploys a model to a SageMaker endpoint and wants to detect data drift by comparing recent inference feature distributions with the training baseline. Which SageMaker capability is designed for this purpose?

A retail company has a highly imbalanced fraud dataset (0.2% fraud). They train a binary classifier and see 99.8% accuracy in validation, but business stakeholders report many fraud cases are still missed. Which evaluation metric is MOST appropriate to focus on?

A team uses Amazon SageMaker to train a model on data stored in S3. Training jobs are slow because the container waits for all data to download before starting. The dataset is several terabytes and is updated frequently. Which approach will MOST likely improve throughput and start time?

A company trains models in SageMaker and must ensure reproducibility and traceability for audits. They need to track datasets, code, hyperparameters, and resulting model artifacts across experiments. Which combination is MOST appropriate?

A team is building a churn prediction model. During EDA they discover that the feature "customer_last_contact_date" is recorded after a customer cancels service, and thus would not be available at prediction time. What is the MOST likely issue and best action?

A company needs a multi-account MLOps setup. Data scientists in a development account should train and register models, but only a centralized production account can deploy models to production endpoints. Approvals are required before deployment, and deployments must be automated. Which architecture BEST meets these requirements?

A team trains a neural network for product recommendations using SageMaker. Training loss decreases steadily, but validation loss starts increasing after several epochs. They suspect overfitting and want an automated way to stop training at the optimal point without manual monitoring. What should they do?

A data science team stores raw training data in Amazon S3. They need to run the same feature engineering steps reliably for both training and inference, including one-hot encoding and scaling, and want the transformations to be versioned and reproducible. Which approach is MOST appropriate?

A model is deployed to a SageMaker endpoint. The team observes that real-time predictions have started to drift from expected values, but no application errors are logged. They want an automated way to detect changes in the statistical properties of live inference data compared to training data. What should they use?

A dataset includes user IDs, event timestamps, and a binary label indicating churn. The data scientist suspects label leakage because a feature derived from events after the churn decision may be present. Which exploratory analysis is MOST appropriate to confirm leakage risk?

A company trains a binary classifier where only 0.5% of records are positive. The business cares most about catching positives while keeping false alarms manageable. Which evaluation approach is MOST appropriate during model selection?

A team trains models in SageMaker and wants to track datasets, code, hyperparameters, and model artifacts for governance and reproducibility across experiments. Which solution BEST supports this requirement with minimal custom work?

An ML engineer is training an XGBoost model with one-hot encoded categorical features, resulting in a very high-dimensional sparse matrix. Training is slow and memory intensive. Which approach is MOST appropriate to improve training efficiency while keeping model quality similar?

A retail company wants to forecast daily demand for thousands of SKUs. They have multiple related time series with holiday effects and would like a managed approach that can incorporate item-level metadata (e.g., category) and known future events. Which service and feature set is MOST appropriate?

A team deploys a TensorFlow model to a SageMaker real-time endpoint. They see intermittent 5xx errors and high latency during traffic spikes. The model itself is stable, and the payload size is small. Which action is the BEST first step to improve availability and latency under variable load?

A model is trained to predict loan default. The validation AUC is unusually high. Investigation reveals that a feature 'recovery_amount' is included, which is only known after a borrower defaults. The model is already in production. What is the MOST appropriate remediation plan to reduce business risk while correcting the issue?

A company needs near-real-time fraud detection. Events are written to a Kinesis Data Stream and must be enriched with recent user aggregates stored in an online feature store. The model must be invoked with low latency and the feature definitions must be consistent with training. Which architecture BEST meets these requirements?

A retail company stores raw clickstream logs in Amazon S3. A data scientist must create a feature table aggregated by user and day and keep it continuously updated for model training in Amazon SageMaker. The solution should minimize operational overhead and support incremental processing. Which approach is best?

A machine learning team built a binary classifier on an imbalanced dataset where the positive class occurs in 0.5% of records. Accuracy is high, but the model is missing many positives. Which evaluation approach is most appropriate to assess performance and select a threshold?

A team wants to compare several algorithms and hyperparameter ranges for a text classification problem in Amazon SageMaker. They also want to avoid overfitting during tuning. Which configuration best meets these requirements?

A SageMaker real-time endpoint suddenly starts returning 5XX errors after a new model is deployed. CloudWatch shows frequent container restarts and a spike in memory utilization. The model artifact size increased significantly compared to the previous version. What is the most likely cause and best immediate fix?

A healthcare company is building a training pipeline on Amazon SageMaker using data stored in Amazon S3. The S3 objects are encrypted with SSE-KMS. The training job fails with an AccessDenied error when attempting to read the training data. The IAM role for the training job has s3:GetObject permission on the bucket. What additional configuration is required to resolve the issue securely?