50 AWS Certified Machine Learning - Specialty Practice Questions: Question Bank 2025

A team stores raw clickstream logs as JSON in Amazon S3. They want to run ad-hoc SQL queries to quickly validate schema changes without provisioning any servers. Which AWS solution is most appropriate?

A data scientist is exploring a labeled dataset in Amazon SageMaker Studio and suspects one class is severely underrepresented. What is the quickest way to quantify label distribution before modeling?

A company wants a managed service to automatically build, train, and tune classification models from tabular data without requiring the team to select algorithms manually. Which AWS service best meets this requirement?

A model endpoint in Amazon SageMaker suddenly starts returning increased latency. The team wants to identify whether the slowdown is due to CPU saturation or memory pressure on the endpoint instances. Which AWS capability should they use?

A dataset contains customer records with missing values in multiple numeric columns and a few categorical columns with rare categories. The team plans to train an XGBoost model in SageMaker. What preprocessing approach is most appropriate to improve model quality and robustness?

A team is building a binary classifier and observes strong performance during training but significantly worse performance on a holdout set. They suspect data leakage caused by feature engineering. Which action is most effective to reduce leakage risk?

A company has a highly imbalanced fraud dataset where fraud is 0.2% of transactions. They care most about catching fraud while controlling false positives. Which evaluation approach is most appropriate?

A team wants to retrain a model whenever new labeled data arrives in S3. They also need approval and traceability for each production deployment. Which architecture best satisfies these requirements?

A company is training a deep learning model on a large dataset stored in S3 using SageMaker. Training time is high because data loading becomes a bottleneck, and the team wants to maximize GPU utilization. Which approach is MOST effective?

After deploying a model, a team notices a steady drop in prediction quality over several weeks. They suspect feature drift due to changing user behavior. They want automated detection and alerting, and the ability to inspect which features shifted. Which solution is best?

A data scientist needs to quickly calculate summary statistics and visualize distributions for a 50-GB dataset stored in Amazon S3. The goal is minimal setup and the ability to run SQL-like analysis. Which approach is MOST appropriate?

A team is building a real-time inference API on Amazon SageMaker endpoints. They need to capture a sample of request/response payloads for later inspection and model debugging, with minimal code changes to the service. Which feature should they enable?

A company wants to build a baseline binary classifier quickly with little feature engineering. The data is tabular and stored in Amazon S3. The team wants the service to automatically perform algorithm selection and hyperparameter tuning. Which option should they choose?

A data engineering team receives daily JSON logs in Amazon S3. Schema fields can appear or disappear over time. Analysts need to query the latest data without frequent manual schema updates. What is the BEST approach?

A team is training a classification model where positive examples are only 1% of the dataset. Initial training shows high accuracy but very poor recall on the positive class. Which change is MOST likely to improve the model’s ability to detect the minority class?

A company uses Amazon SageMaker to train models on a schedule. They need end-to-end reproducibility for audits: the exact code, hyperparameters, input data version, and resulting model artifacts for each run must be tracked. Which combination BEST meets this requirement?

A team is building a time series forecasting model for product demand. They create random train/test splits and observe unrealistically strong test performance. In production, forecast accuracy is much worse. What is the MOST likely cause and the BEST fix?

A computer vision team uses SageMaker built-in image classification. Training loss decreases steadily, but validation accuracy plateaus and then degrades. Which action is MOST appropriate to address the issue?

A regulated enterprise must deploy an ML inference endpoint in a private network with no public internet access. The endpoint must still pull the model artifacts securely and log metrics. Which architecture BEST satisfies these constraints?

A company trains a model in multiple Regions due to data residency requirements. They need consistent feature definitions and offline training datasets across Regions, while also serving low-latency online features for real-time inference in each Region. Which solution BEST addresses these needs with the LEAST duplication of logic?

A data science team is exploring a large dataset in Amazon S3 using Amazon SageMaker Studio. They want to quickly profile columns (missing values, distributions, correlations) and generate a shareable report without building custom code. Which SageMaker feature is best suited for this?

A team needs to load daily CSV files from Amazon S3 into Amazon Redshift for downstream analytics. They want a managed approach that can automatically infer and evolve schema when new columns appear, and they prefer not to manage servers. Which solution best meets these requirements?

A model has been trained in SageMaker, and a product team wants to run inference on millions of records stored in Amazon S3 once per day. Low latency is not required, but the solution must be cost-effective and managed. What is the best inference approach?

A company is building a churn prediction model where only 2% of customers churn. Initial training produces high accuracy but poor recall for churners. Which approach is MOST appropriate to improve the model’s ability to identify churners?

A team built a binary classifier in SageMaker. They observe that predicted probabilities are poorly calibrated: among predictions around 0.8, only ~60% are positive. They need well-calibrated probabilities for decisioning thresholds. Which technique is MOST appropriate?

A company stores raw clickstream JSON in Amazon S3. Analysts want to run ad-hoc SQL to explore nested attributes without predefining a schema and without loading data into a database. Which solution is best?

A team wants to operationalize feature generation so training and real-time inference use the exact same feature definitions. Features must be available with low-latency access for online predictions and also for offline backfills. Which SageMaker capability best addresses this requirement?

An object detection model is trained in SageMaker using images stored in S3. Training metrics are excellent, but production performance is poor. Investigation shows that images were randomly split into train/validation after upload; many near-duplicate images from the same video appear in both splits. What is the most likely issue and the best corrective action?

A regulated company must encrypt all training data and model artifacts and ensure that SageMaker training jobs cannot write output to any S3 bucket except a dedicated, encrypted bucket. They also want to prevent the job from having internet access. Which combination of controls best satisfies these requirements?

A team trains a gradient boosted model for demand forecasting. They included a feature "avg_sales_last_7_days" computed using the full dataset, including days after the prediction timestamp, due to an incorrect aggregation query. The model performs exceptionally well offline but fails in production. What is the BEST way to prevent this class of issue going forward?

A data science team wants to explore a large dataset stored in Amazon S3 using SQL without managing any infrastructure. They need to quickly compute summary statistics and filter rows for analysis. Which AWS service should they use?

A company stores raw clickstream events (JSON) in Amazon S3. The ML team needs to convert them into partitioned Parquet files and register the schema so analysts can query the curated dataset. Which approach is the MOST appropriate?

A binary classifier deployed for fraud detection shows high overall accuracy, but the fraud class is rare and many fraud cases are missed. Which metric is MOST appropriate to prioritize during evaluation?

A team needs to deploy a trained model to a REST endpoint and automatically scale the number of instances based on incoming request volume. Which SageMaker capability should they use?

An ML engineer notices that a model trained with SageMaker XGBoost performs much better on the training data than on the validation data. Which action is MOST likely to reduce overfitting?

A retailer wants to build a recommendation system using implicit feedback (clicks and purchases) and wants a managed solution that can train and host the recommender with minimal custom code. Which service/feature is the BEST fit?

A data scientist is preparing features and wants to prevent data leakage when creating time-based aggregates (for example, average spend in the last 7 days) for a model that predicts future customer churn. Which approach is BEST?

A team wants to operationalize an end-to-end ML workflow that includes data preprocessing, training, evaluation, model approval, and deployment. They also need to track lineage and artifacts for auditing. Which solution best satisfies these requirements with managed AWS capabilities?

A healthcare company must train a model on sensitive data stored in Amazon S3. The security team requires that the training job cannot access the public internet, and data must not leave the VPC. Which configuration meets these requirements?

A team is building a near-real-time feature pipeline. Events arrive on Amazon Kinesis Data Streams. They need to compute rolling aggregates (for example, 5-minute counts per user) and make the results available for low-latency online inference. Which architecture is MOST appropriate?

A data science team wants to quickly visualize distributions, correlations, and missing values for a tabular dataset stored in Amazon S3 before deciding on feature engineering steps. The team prefers a managed, interactive environment with minimal setup. Which approach is MOST appropriate?

A company uses Amazon SageMaker to train models and wants to track and compare multiple experiments (hyperparameters, metrics, and artifacts) across iterations. Which SageMaker capability is designed for this purpose?

A retailer has a severe class imbalance problem: only 0.3% of transactions are fraudulent. The team is evaluating a binary classifier and wants a metric that reflects performance on the minority class and is not dominated by true negatives. Which metric is MOST appropriate?

A company is building near-real-time features for a recommendation model. User events arrive continuously and must be available for both analytics and model training. The solution must support durable ingestion, replay, and near-real-time processing into an S3 data lake. Which architecture best meets these requirements?

A team is training a model with Amazon SageMaker and uses multiple sources: features in Amazon S3 and labels in Amazon Redshift. They need a repeatable way to join, transform, and export a training dataset, and they want to minimize custom code while keeping the pipeline serverless. What is the BEST approach?

A model performs well in offline validation but degrades after deployment. The data science team suspects feature distribution changes between training and live inference. Which SageMaker capability provides built-in monitoring to detect data drift and model quality issues over time?

A team is using linear models for a high-dimensional dataset with many correlated features. They want to reduce overfitting and automatically drive some feature weights to exactly zero to perform feature selection. Which regularization technique should they choose?

A team built an NLP classifier using subword tokenization. During inference, they observe unexpectedly high latency and memory usage on the endpoint, even at low traffic. Investigation shows that the model container repeatedly downloads tokenization assets from Amazon S3 on every invocation. What is the BEST fix?

A company needs to train a deep learning model on sensitive medical images. The security team requires that data remain encrypted at rest and that training instances do not have direct internet access. The team must still be able to pull training data from S3 and write model artifacts back to S3. Which solution meets these requirements with the LEAST operational overhead?

A team is developing a demand forecasting model where under-forecasting is far more costly than over-forecasting. They want to optimize the model to penalize underestimates more heavily while still training on a scalable managed service. Which approach is MOST appropriate?