AWS Machine Learning Specialty (MLS-C01) Complete Practice Exam

A) Partition by file size B) Hive-style partitioning on columns frequently used in filters (e.g., date, region) C) Random partitioning D) Apply compression only, no partitioning

Answer: B

Explanation: For Athena performance, use Hive-style partitioning: s3://bucket/data/year=2024/month=03/day=01/. Partitioning on columns commonly found in WHERE clauses reduces data scanned, cutting cost and query time. Combine with columnar formats like Parquet or ORC for even better performance.

A) On-demand Worker Type B) Increase to G.1X Worker Type C) Enable Job Bookmarks to avoid reprocessing D) Always use maximum DPUs

Answer: C

Explanation: Enabling AWS Glue Job Bookmarks tracks previously processed data so only new data is processed on each run. This prevents duplicate processing and reduces cost. Flex execution (flexible execution class) also helps with cost optimization.

A) No difference; they are the same service B) Streams supports custom consumers and reprocessing; Firehose delivers automatically to S3/Redshift/etc. C) Firehose has lower latency D) Streams can only store data in S3

Answer: B

Explanation: Kinesis Data Streams allows multiple independent consumers, parallel processing via shards, and reprocessing within the retention window. Kinesis Data Firehose is a fully managed, near-real-time service that delivers data to S3, Redshift, OpenSearch, or Splunk without code.

A) Faster model training B) Consistent feature reuse, prevention of training-serving skew, and point-in-time feature retrieval C) Cost reduction D) Automatic hyperparameter tuning

Answer: B

Explanation: SageMaker Feature Store manages ML features in an online store (low-latency lookup) and an offline store (large-scale batch training). Multiple teams can reuse features, training-serving skew is prevented, and time-based point-in-time queries avoid data leakage.

A) S3 bucket policies B) Access permissions at the database, table, column, and row levels C) EC2 instance access D) IAM policies only

Answer: B

Explanation: Lake Formation supports fine-grained access control down to the Glue Data Catalog database, table, and column level. Row filters further restrict which rows a specific user can see. This is more granular than IAM policies alone.

A) S3 + Athena B) Kinesis Data Streams + Kinesis Data Analytics (Apache Flink) C) DynamoDB + Lambda D) RDS + EC2

Answer: B

Explanation: Kinesis Data Streams ingests real-time data; Kinesis Data Analytics for Apache Flink performs windowed aggregations, filtering, and transformations in real time. Results can be written back to Kinesis Streams or S3.

A) AWS Glue Crawler B) AWS Glue Studio C) AWS Data Pipeline D) AWS Batch

Answer: A

Explanation: AWS Glue Crawler scans data sources such as S3, RDS, and DynamoDB, automatically infers schemas, and creates table metadata in the Glue Data Catalog. It can run on a schedule or on demand.

A) Set the S3 bucket to public B) Use Lake Formation Data Sharing or add the external account's ARN to the S3 bucket policy C) Email the data D) Use cross-region replication only

Answer: B

Explanation: Use Lake Formation cross-account sharing via AWS RAM (Resource Access Manager), or add the external account's IAM ARN to the S3 bucket policy. Lake Formation cross-account sharing preserves column/row-level access control while sharing data.

A) Use only SageMaker Training Jobs B) Define transformations in SageMaker Data Wrangler, then automate with SageMaker Pipelines C) Run pandas scripts manually D) Use AWS Batch

Answer: B

Explanation: SageMaker Data Wrangler provides a GUI with 300+ built-in transformations for feature engineering. Defined transformations can be exported as SageMaker Processing Jobs and integrated into SageMaker Pipelines to automate the end-to-end ML workflow.

A) A separate EC2 instance is required B) Connect a Lambda function as a Data Transformation C) Directly invoke a Glue ETL job D) Transformation is not possible

Answer: B

Explanation: Kinesis Data Firehose supports Lambda functions as data transformation targets. Each record passes through the Lambda function (for filtering, format conversion, enrichment, etc.) before being delivered to the destination such as S3.

A) Copy the entire dataset to local EBS first B) Use Pipe mode or FastFile mode instead of File mode C) Move data to EFS before training D) Decompress data before use

Answer: B

Explanation: SageMaker training data input modes: File mode copies all data before training starts. Pipe mode streams data from S3. FastFile mode (recommended) mounts S3 directly, allowing training to start immediately without copying all data first, saving time on large datasets.

A) Machine learning model training B) Visually design and run ETL pipelines without code C) Database management D) Real-time streaming processing only

Answer: B

Explanation: AWS Glue Studio provides a drag-and-drop interface for visually building ETL pipelines. It generates and monitors Spark-based ETL jobs without code, and allows direct editing of the generated Python or Scala code.

A) Amazon Macie B) AWS Shield C) Amazon GuardDuty D) AWS WAF

Answer: A

Explanation: Amazon Macie uses ML to automatically detect sensitive data (credit card numbers, SSNs, PII, etc.) in S3 buckets. Findings integrate with Security Hub, and Glue ETL can be used for masking.

A) Model serving B) Running ML workflow tasks such as preprocessing, postprocessing, model evaluation, and feature engineering C) Hyperparameter tuning D) A/B testing

Answer: B

Explanation: SageMaker Processing Jobs are fully managed and handle data preprocessing, feature engineering, model evaluation, and post-deployment analysis. They support built-in containers (Scikit-learn, Spark, XGBoost) or custom containers.

A) RDS → Lambda → S3 B) Sources → Kinesis Data Streams → Kinesis Data Firehose → S3 C) EC2 → EBS → S3 D) DynamoDB → S3 direct copy

Answer: B

Explanation: Multiple data sources send real-time data to Kinesis Data Streams. Kinesis Data Firehose applies buffering, compression, and encryption before storing to S3. Firehose also supports dynamic partitioning based on partition keys.

A) Performance improvement B) To handle custom data formats or delimiters that the default classifiers cannot recognize C) Cost reduction D) Security enhancement

Answer: B

Explanation: Default Glue classifiers support common formats like CSV, JSON, Parquet, and ORC. Custom classifiers (Grok, XML, JSON, CSV) are needed for fixed-width formats, non-standard delimiters, or non-standard data formats to ensure accurate schema detection.

A) SMOTE (Synthetic Minority Over-sampling Technique) B) Undersampling the majority class C) Adjusting class weights D) Collecting no additional data and using accuracy only

Answer: D

Explanation: Accuracy is misleading with class imbalance. Correct approaches: SMOTE (generate synthetic minority samples), undersampling the majority class, adjusting class weights, and collecting more data. Use metrics like AUC-ROC, F1-score, and Precision-Recall curves instead.

A) Download raw data and analyze locally B) Use Data Wrangler's Data Quality and Insights Report C) Manually write SQL queries D) Analyze on a separate EC2 instance

Answer: B

Explanation: SageMaker Data Wrangler's Data Quality and Insights Report automatically generates per-column statistics (mean, median, std, distribution), missing values, outliers, class imbalance, and data type information — enabling fast EDA.

A) No problem B) Curse of dimensionality from high-cardinality variables — use Target Encoding or embeddings C) Apply it to numeric variables too D) Label Encoding is always better

Answer: B

Explanation: Applying One-Hot Encoding to variables with hundreds or thousands of categories produces very sparse, high-dimensional vectors. Target Encoding (replacing categories with the target mean), embeddings, or Frequency Encoding are more effective alternatives.

A) Confusion Matrix B) SHAP (SHapley Additive exPlanations) C) ROC Curve D) Learning Curve

Answer: B

Explanation: SageMaker Clarify uses SHAP (based on Shapley game theory) to explain each feature's contribution to predictions. It provides both global explanations (feature importance over the entire dataset) and local explanations (per-prediction feature contributions).

A) Mean/median/mode imputation B) KNN imputation C) Drop rows/columns (when the proportion is small) D) Feed missing values directly into the model

Answer: D

Explanation: Most ML algorithms cannot handle missing values. Correct approaches: statistical imputation (mean, median, mode), KNN imputation, prediction-based imputation, or dropping rows/columns when the proportion is small. XGBoost handles missing values internally.

A) Min-Max Normalization B) StandardScaler (Z-score normalization) C) Robust Scaler (uses median and IQR) D) No scaling at all (except for tree-based models)

Answer: D

Explanation: Distance-based algorithms (KNN, SVM, neural networks) are sensitive to scale. Min-Max maps to [0,1]; StandardScaler produces mean=0, std=1; RobustScaler is robust to outliers. Tree-based models (Random Forest, XGBoost) are relatively scale-invariant, but other algorithms require scaling.

A) RMSE B) Class Imbalance (CI), Difference in Positive Proportions in Labels (DPL) C) AUC-ROC D) F1-score

Answer: B

Explanation: SageMaker Clarify provides pre-training bias metrics (DPL: difference in label proportion across groups; CI: class imbalance, etc.) and post-training model bias metrics (DPPL: difference in predicted positive proportions). These evaluate fairness across demographic groups.

A) Normalization B) Differencing and seasonal decomposition C) Applying PCA D) One-hot encoding

Answer: B

Explanation: To achieve stationarity in time-series data, apply differencing (first-order, seasonal). Use statsmodels' seasonal_decompose or STL decomposition to separate trend, seasonal, and residual components. This preprocessing is critical for ARIMA, SARIMA, and DeepAR.

A) Convert only to Unix timestamp B) Decompose into year, month, day, day-of-week, hour, week number, and other features C) Remove date features D) Use the raw string directly

Answer: B

Explanation: Extract year, month, day, day-of-week, time-of-day, quarter, week number, and whether it's a holiday from date/time fields. Cyclical features like day-of-week or month should be sin/cos-encoded to preserve continuity. This captures seasonality and weekly patterns.

A) Spearman correlation B) Pearson correlation C) Kendall's tau D) Chi-square statistic

Answer: B

Explanation: Pearson correlation measures the strength of the linear relationship between two continuous variables on a scale of -1 to 1. Spearman is rank-based (monotonic non-linear), Kendall measures rank concordance. Spearman is more commonly used in ML preprocessing due to its robustness to outliers.

A) Separate training required B) Use the Quick Model feature to automatically compute feature importance C) Manual correlation analysis D) Not possible

Answer: B

Explanation: Data Wrangler's Quick Model feature quickly trains a Random Forest on the selected target variable and visualizes each feature's importance. This is useful for feature selection and data quality assessment.

A) RMSE, MAE, R2 B) Precision, Recall, F1-score, AUC-ROC C) BLEU, Perplexity D) MAPE, SMAPE

Answer: B

Explanation: Classification metrics: Precision (true positives among predicted positives), Recall (true positives found among all actual positives), F1-score (harmonic mean of Precision and Recall), AUC-ROC (threshold-independent performance measure). For imbalanced data, F1-score and AUC-PR (area under the Precision-Recall curve) are especially important.

A) Linear Regression B) t-SNE, PCA, UMAP C) Random Forest D) KNN

Answer: B

Explanation: PCA is a linear dimensionality reduction that maximizes variance. t-SNE preserves local structure in high-dimensional data for cluster visualization. UMAP is faster than t-SNE and also preserves global structure. SageMaker includes PCA as a built-in algorithm.

A) RMSE (Root Mean Squared Error) B) MAE (Mean Absolute Error) C) R2 (coefficient of determination) D) MAPE

Answer: B

Explanation: RMSE squares errors, making it sensitive to outliers. MAE takes the mean of absolute errors, reducing outlier influence. Use RMSE when large errors deserve extra penalty. R2 measures how much data variance the model explains.

A) Tokenizing B) Removing stopwords C) Image augmentation D) TF-IDF vectorization

Answer: C

Explanation: Text preprocessing includes tokenizing, lowercasing, removing stopwords, stemming/lemmatization, and vectorization (TF-IDF, Word2Vec, etc.). Image augmentation is a computer vision technique, not a text preprocessing step.

A) Cross-validation B) Calculate VIF (Variance Inflation Factor) or use a correlation heatmap C) Adjust the learning rate D) Change batch size

Answer: B

Explanation: Multicollinearity arises from strong correlations between independent variables. A VIF greater than 10 indicates multicollinearity. A correlation heatmap provides a visual assessment. Solutions include Ridge Regression or PCA to address the issue.

A) Linear Learner B) Random Cut Forest (RCF) and IP Insights C) BlazingText D) DeepAR

Answer: B

Explanation: SageMaker Random Cut Forest (RCF) is an unsupervised anomaly detection algorithm that computes an anomaly score for each data point. IP Insights learns normal usage patterns for IP addresses and detects unusual access.

A) Precision increases and Recall decreases B) Recall increases and Precision decreases C) AUC-ROC increases D) No change

Answer: B

Explanation: Lowering the threshold predicts more positives, increasing Recall (catching more true positives) but decreasing Precision (more false positives). Use a lower threshold when minimizing false negatives is critical (e.g., fraud detection). AUC-ROC is a threshold-independent metric.

A) Both training and validation errors are high B) Training error is low but validation error is high C) Both training and validation errors are low D) Only validation error is low

Answer: B

Explanation: Overfitting occurs when a model fits the training data well but generalizes poorly to new data (validation/test). A large gap between training and validation errors signals overfitting. Solutions: more training data, regularization (L1/L2), dropout, early stopping, ensemble methods.

A) Complex ETL requiring custom code B) Visual data cleaning and transformation without code C) Real-time streaming data processing D) Model training

Answer: B

Explanation: AWS Glue DataBrew is a visual data preparation tool with 250+ pre-built transformations for cleaning and standardizing data without code. It is ideal for non-technical users to explore and prepare data.

A) Applying scaling only B) Creating polynomial features or feature crosses C) One-hot encoding D) Applying PCA

Answer: B

Explanation: Polynomial features (x1^2, x1*x2, etc.) and feature crosses capture non-linear interactions between variables. Google's Wide and Deep Learning explicitly creates cross features. However, dimensionality grows rapidly, so care is needed.

A) Faster training B) Uses all data for both training and validation, providing a more stable performance estimate C) Automated data augmentation D) Automatic hyperparameter tuning

Answer: B

Explanation: K-Fold cross-validation splits data into K folds and iterates K times, using each fold as the validation set once. All data participates in both training and validation, reducing high-variance estimation caused by a small validation set and providing a more reliable performance estimate.

A) Use all data with no weighting B) Stratified sampling to maintain class proportions C) Random sampling only D) Remove the minority class

Answer: B

Explanation: Stratified sampling preserves the original class proportions in each train/validation/test split. In sklearn use train_test_split(stratify=y). This ensures each set has a balanced class distribution, which is crucial for evaluating model performance on imbalanced data.

A) Difference in prediction accuracy B) Difference in the rate at which the model predicts positive labels across groups defined by a protected attribute C) Difference in number of model parameters D) Difference between training and validation errors

Answer: B

Explanation: DPPL measures the difference in the proportion of positive predictions across groups defined by a protected attribute (e.g., gender, age). A value close to 0 indicates a fairer model. A negative DPPL indicates bias against one group.

A) Increase num_round B) alpha (L1), lambda (L2) regularization, subsample, colsample_bytree C) Increase eta D) Unlimited max_depth

Answer: B

Explanation: XGBoost regularization: alpha (L1 regularization) and lambda (L2 regularization) penalize model complexity. subsample (row sampling ratio) and colsample_bytree (column sampling ratio) increase diversity. Limiting max_depth, increasing min_child_weight, and using early_stopping_rounds also prevent overfitting.

A) Always slower than random search B) Efficiently explores promising hyperparameter combinations based on previous experiment results C) No difference D) More expensive

Answer: B

Explanation: Bayesian optimization intelligently selects the next hyperparameter combination to try based on prior experiment results. It finds good hyperparameters with fewer experiments than random or grid search, saving time and cost. SageMaker supports Bayesian, Random, and Hyperband strategies.

A) Code is always required B) AutoML that automatically explores algorithms, feature preprocessing, and hyperparameters C) Supports only deep learning models D) Model explainability is not available

Answer: B

Explanation: SageMaker Autopilot is a fully managed AutoML service. It analyzes data, tries multiple algorithms (XGBoost, Linear, MLP, etc.) and feature transformations, and selects the best model. It provides generated notebook code for transparency and includes an Explainability Report.

A) BlazingText B) DeepAR C) KNN D) PCA

Answer: B

Explanation: SageMaker DeepAR is an autoregressive recurrent neural network (ARNN) optimized for time-series forecasting. Training jointly on multiple related time series produces better forecasts than individual series. It also outputs probabilistic predictions (forecast distributions).

A) Randomly initialize the entire model and train from scratch B) Freeze lower layers and train only the upper layers on the new dataset C) Apply the same strategy regardless of dataset size D) Transfer learning cannot be applied to images

Answer: B

Explanation: Common Transfer Learning approach: 1) Use a model pre-trained on ImageNet (VGG, ResNet, EfficientNet, etc.), 2) Freeze the lower feature-extraction layers, 3) Fine-tune only the upper task-specific layers (classification head) on the new dataset. Freeze more layers when data is scarce.

A) Single-node, multi-core B) Data Parallelism and Model Parallelism C) Serial training, parallel training D) GPU training, CPU training

Answer: B

Explanation: Data Parallelism: copies the same model to multiple GPUs, each processes a different batch, then gradients are synchronized. Uses SageMaker Data Parallel Library. Model Parallelism: splits the model itself across multiple GPUs — used when a model doesn't fit on a single GPU. Uses SageMaker Model Parallel Library.

A) Image classification B) Word2Vec embedding training and text classification (supervised/unsupervised) C) Time-series forecasting D) Recommendation systems

Answer: B

Explanation: SageMaker BlazingText supports two modes: 1) Word2Vec mode: generates word embeddings via unsupervised learning (Skip-gram, CBOW), and 2) Text Classification mode: supervised classification of text. Similar to fastText, it processes large text datasets quickly.

A) Binary classification B) Text generation tasks like machine translation and text summarization C) Regression D) Time-series forecasting

Answer: B

Explanation: BLEU (Bilingual Evaluation Understudy) evaluates machine translation quality by measuring n-gram overlap between generated text and reference translations. It is used to evaluate the SageMaker Seq2Seq algorithm (machine translation, text summarization).

A) To reduce cost B) To use libraries or algorithms not supported by SageMaker's built-in algorithms or frameworks C) For faster training D) For automatic distributed training

Answer: B

Explanation: BYOC (Bring Your Own Container) is used when you need an ML framework, custom algorithm, or specific library version not supported by SageMaker. Push a Docker image to ECR and use it in SageMaker. The image must conform to the SageMaker training/inference contract.

A) The model is more uncertain B) The model predicts text better (lower is better) C) The model is smaller D) Training is faster

Answer: B

Explanation: Perplexity measures how well a language model predicts test data. Mathematically it is the exponent of the cross-entropy loss on test data. Lower perplexity means the model predicts the next word with more confidence — indicating a better language model.

A) Always more expensive B) Up to 90% cost savings but interruptions possible; checkpointing to S3 is required C) Faster training speed D) No additional configuration needed

Answer: B

Explanation: SageMaker Managed Spot Training can save up to 90% compared to On-Demand. However, Spot instances can be interrupted at any time, so checkpoints must be saved to S3 to support resumption after interruption. Set MaxWaitTimeInSeconds to control maximum wait time.

A) CSV files B) RecordIO or augmented manifest format C) JPEG only D) TFRecord only

Answer: B

Explanation: SageMaker Object Detection (SSD/Faster RCNN-based) supports RecordIO-protobuf format and augmented manifest (JSON Lines with S3 image URLs + labels). Labels generated by SageMaker Ground Truth can be used directly.

A) Use complex deep learning models B) Transfer Learning, Data Augmentation, Few-shot Learning C) Train a model without data D) Remove regularization

Answer: B

Explanation: With limited data: 1) Transfer Learning — fine-tune a pre-trained model from a related domain, 2) Data Augmentation — expand the dataset by transforming existing data, 3) Few-shot Learning — learn new classes from a few examples, 4) Use pre-trained models from SageMaker JumpStart.

A) Always use K=3 B) Elbow Method or Silhouette Score C) Use the dataset size as K D) Choose randomly

Answer: B

Explanation: Elbow Method: plot WCSS (within-cluster sum of squares) against K and pick the "elbow" point where the curve flattens. Silhouette Score: measures intra-cluster cohesion and inter-cluster separation to evaluate K. SageMaker K-Means can automatically try multiple K values.

A) Bounding box coordinates B) A class label for each pixel C) A single image label D) A text description

Answer: B

Explanation: Semantic Segmentation assigns each pixel of an image to a specific class (e.g., person, car, road). The output is a mask the same size as the input image, where each pixel value is a class ID. Widely used in autonomous driving and medical imaging.

A) No difference B) Boosting sequentially corrects prior model errors; Bagging averages results of independent models C) Bagging is slower D) Boosting is only for regression

Answer: B

Explanation: Bagging (Bootstrap Aggregating): trains independent models in parallel and averages results; reduces variance (Random Forest). Boosting: trains models sequentially, focusing more on samples misclassified by the prior model; reduces bias (XGBoost, AdaBoost, LightGBM). Boosting generally outperforms single models but has a higher overfitting risk.

A) Binary classification B) Topic modeling to discover hidden themes in text documents C) Image classification D) Numeric prediction

Answer: B

Explanation: LDA is an unsupervised algorithm that discovers latent topics in a document collection. Each document is represented as a mixture of topics, and each topic is a distribution over words. Applications include news categorization and content recommendation.

A) Handles large-scale datasets B) Fast code development and debugging (run SageMaker code locally) C) Automatic distributed training D) Production deployment

Answer: B

Explanation: SageMaker Local Mode runs training/inference code locally (or in SageMaker Studio) without starting actual SageMaker instances. Useful for fast iterative development and debugging. Requires Docker. Not appropriate for large-scale data or distributed training.

A) No difference B) Real-time returns responses immediately; async queues large/slow requests and stores results in S3 C) Async is faster D) Only real-time supports auto-scaling

Answer: B

Explanation: Real-time endpoints: low-latency requests requiring immediate responses. Async inference: queues requests stored in S3 and stores results in S3 — suitable for large payloads (up to 1 GB) or long-running processing. Batch Transform: processes large batches in a one-off job.

A) Network latency only B) Data drift, model quality degradation, bias drift, and feature attribution changes C) Cost overruns only D) Code bugs only

Answer: B

Explanation: SageMaker Model Monitor has four monitor types: 1) Data Quality Monitor — detects statistical changes in input data, 2) Model Quality Monitor — detects degradation in prediction performance, 3) Bias Drift Monitor — detects changes in model bias, 4) Feature Attribution Drift Monitor — detects changes in feature importance.

A) To run a single model training job B) To automate and reproduce ML workflows (data processing → training → evaluation → deployment) C) Cost reduction only D) Data storage

Answer: B

Explanation: SageMaker Pipelines defines ML workflows as a DAG (Directed Acyclic Graph) and automates them. Pipeline steps (Processing, Training, Evaluation, RegisterModel, Deploy, etc.) can be chained with conditional execution, parameter passing, and experiment tracking. CI/CD integration enables MLOps.

A) Create two separate endpoints B) Use Production Variants to specify traffic weights and serve multiple models simultaneously C) Route with a Lambda function D) Distribute via CloudFront

Answer: B

Explanation: SageMaker Endpoint Production Variants allows serving multiple model versions from a single endpoint. Assign traffic weights to each variant (e.g., 90% current model, 10% new model), compare performance metrics, then gradually shift traffic.

A) Accelerate model training B) Compile and optimize models for specific hardware (edge devices, cloud instances) C) Data preprocessing D) Hyperparameter tuning

Answer: B

Explanation: SageMaker Neo compiles trained ML models for specific hardware targets (ARM, x86, NVIDIA GPU, Intel chips, etc.) using Apache TVM to reduce model size and increase inference speed. Particularly useful for IoT/edge device deployments.

A) Server errors; resolve by restarting B) Data drift (change in input data distribution); detect with Model Monitor, then retrain C) Network issues D) Code bugs

Answer: B

Explanation: The main cause of model degradation is data drift — when the distribution of production input data diverges from training data. SageMaker Model Monitor continuously monitors for drift, sends automatic alerts when drift is detected, and can trigger a retraining pipeline.

A) Continuously high-traffic real-time services B) Intermittent traffic; no infrastructure cost when idle C) Large-scale batch processing D) Sub-millisecond latency requirements

Answer: B

Explanation: SageMaker Serverless Inference charges only for actual usage with zero infrastructure cost when idle. It is well-suited for intermittent or unpredictable traffic patterns. There is a cold-start latency, so it is not appropriate for strict latency requirements.