AWS Certified Machine Learning - Specialty Intermediate Practice Exam: Medium Difficulty 2025

A data scientist is building a real-time fraud detection system that processes streaming transaction data. The system needs to handle 10,000 transactions per second with sub-second latency requirements. Feature engineering includes calculating rolling averages over the past 30 minutes. Which AWS architecture should be implemented?

A machine learning engineer notices that a training dataset stored in Amazon S3 contains customer reviews with significant class imbalance: 95% positive reviews and 5% negative reviews. The binary classification model consistently predicts positive class with 95% accuracy but fails to identify negative reviews. What is the MOST effective approach to address this issue?

A company is using Amazon SageMaker to train a deep learning model for image classification. Training jobs are taking 12 hours on ml.p3.2xlarge instances. The data scientist wants to reduce training time while maintaining model accuracy. The training dataset consists of 500,000 images stored in S3. Which combination of strategies would be MOST effective? (Choose TWO)

An ML team needs to perform feature engineering on a 2TB dataset containing structured customer data with 200 columns. They need to handle missing values, encode categorical variables, normalize numerical features, and create interaction features. The processed data will be used to train multiple models. Which AWS service combination provides the MOST scalable and cost-effective solution?

A data scientist discovers that their regression model has high training accuracy (R² = 0.95) but poor validation accuracy (R² = 0.65). The model uses 50 features with polynomial transformations up to degree 4. What is the BEST approach to address this issue?

A company deployed a SageMaker endpoint for a recommendation model that serves 1,000 requests per day during business hours (8 AM - 6 PM) but receives minimal traffic outside these hours. Each inference takes 200ms and the model size is 2GB. The current deployment uses a single ml.m5.xlarge instance running 24/7. How can the ML engineer optimize costs while maintaining performance?

An ML team is analyzing a dataset with 100 features to predict customer churn. Initial exploratory analysis shows that many features are highly correlated (correlation coefficients > 0.85). The team wants to reduce dimensionality while retaining the most important information for their logistic regression model. Which approach should they use?

A financial services company is deploying a credit risk assessment model that must provide explanations for each prediction to satisfy regulatory requirements. The model processes loan applications containing 30 features including income, credit history, and employment details. Which SageMaker solution provides the MOST comprehensive explainability?

A data scientist is building a neural network for time series forecasting of product demand. The training loss decreases steadily, but validation loss decreases initially then starts increasing after epoch 15. The learning curves show a growing gap between training and validation loss. What combination of techniques would BEST address this issue?

An ML team has trained a sentiment analysis model using SageMaker and needs to deploy it to handle variable traffic that ranges from 100 to 10,000 requests per minute. They require A/B testing capability to gradually shift 10% of traffic to a new model version while monitoring performance. The average inference time is 50ms. What deployment strategy should they implement?