MLS-C01 Mock Exam 77 — 25 Questions

Question 1

A fraud detection team trains an XGBoost binary classifier in Amazon SageMaker using the built-in algorithm. The label distribution is 99.5% legitimate and 0.5% fraudulent. The team wants to maximize the model's ability to rank fraudulent transactions highly without manually resampling the data. Which SageMaker built-in XGBoost hyperparameter should they tune to achieve this?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

scale_pos_weight. Class imbalance is addressed at the loss level by weighting the positive class, and the SageMaker built-in XGBoost algorithm provides scale_pos_weight for exactly that purpose. Setting it near the negative-to-positive ratio makes errors on fraudulent transactions costlier during boosting, improving the model's ability to rank rare events highly without altering the dataset.

Answer

num_round

Answer

subsample

Answer

max_depth

Question 2

A machine learning engineer is using Amazon SageMaker to train a model with the built-in Linear Learner algorithm. The dataset has a large number of features, and the engineer suspects that many are irrelevant. The engineer wants to automatically perform feature selection during training. Which hyperparameter should the engineer adjust to achieve this?

Accepted Answer

Set the l1 regularization hyperparameter to a positive value.. L1 regularization in Linear Learner adds a penalty equal to the absolute value of weights, which promotes sparsity and effectively performs feature selection by zeroing out irrelevant coefficients. L2 regularization only shrinks weights, while use_bias and target_precision do not influence feature selection. Thus, l1 is the correct hyperparameter.

Answer

Set the use_bias hyperparameter to false.

Answer

Set the target_precision hyperparameter to a high value.

Answer

Set the l2 regularization hyperparameter to a positive value.

Question 3

A data scientist is performing EDA on a large dataset stored in Amazon S3 as Parquet files, partitioned by date. The dataset contains 500 million rows and 100 columns, and the scientist needs to quickly compute summary statistics and correlations between numerical features without loading the entire dataset into memory. The scientist wants to minimize both cost and time. Which AWS service should the scientist use to run SQL queries directly on the S3 data and generate the required statistics?

Accepted Answer

Amazon Athena. Amazon Athena is the correct choice because it enables serverless SQL queries directly against data in Amazon S3, eliminating the need to load or copy data. It supports Parquet, which is columnar and efficient for analytical queries, and can compute summary statistics and correlations using SQL. This minimizes cost and time, aligning with the requirement to quickly explore a large dataset without managing infrastructure.

Answer

AWS Glue

Answer

Amazon Redshift Spectrum

Answer

Amazon EMR

Question 4

A data scientist is using Amazon SageMaker to train a model. The training job must write model artifacts to an Amazon S3 bucket. The data scientist wants to ensure that the training job can access the S3 bucket securely without embedding long-term AWS credentials in the training container. Which mechanism should the data scientist use to grant the training job access to the S3 bucket?

Accepted Answer

Attach an IAM role to the SageMaker training job with a policy that allows access to the S3 bucket.. SageMaker training jobs assume an IAM role specified when the job is created. This role provides temporary AWS credentials to the training container, enabling secure access to AWS services like S3. Attaching an IAM role with a least-privilege policy is the standard, secure method. Storing credentials in a Dockerfile or Secrets Manager is insecure or overly complex. Public S3 buckets are unacceptable. Therefore, using an IAM role is the correct approach.

Answer

Configure the S3 bucket to allow public read and write access so the training job can access it without credentials.

Answer

Use AWS Secrets Manager to store AWS credentials and retrieve them in the training script at runtime.

Answer

Store AWS access keys as environment variables in the training container's Dockerfile.

Question 5

A data scientist has trained a linear regression model using Amazon SageMaker and wants to deploy it for real-time predictions. The model expects a CSV input with a single record per request. The data scientist needs to create an endpoint that can be invoked with low latency. Which SageMaker deployment option should they use?

Accepted Answer

Create a real-time endpoint using the SageMaker hosting services.. Real-time endpoints provide persistent, low-latency inference for individual requests, making them ideal for the described use case. Batch transform is for offline processing, Serverless Inference may have cold starts, and Lambda is not a SageMaker deployment option. Therefore, creating a real-time endpoint is the correct choice.

Answer

Use SageMaker Serverless Inference to deploy the model.

Answer

Create a batch transform job to process the data.

Answer

Deploy the model to an AWS Lambda function using the SageMaker SDK.

Question 6

A data scientist is exploring a dataset in Amazon SageMaker Studio and wants to detect anomalies in a univariate time series of sensor readings. The data is approximately normally distributed, and the scientist needs a simple statistical method to flag outliers that are more than three standard deviations from the mean. Which method should the scientist use?

Accepted Answer

Z-score. The Z-score is the correct method because it quantifies the distance of a data point from the mean in units of standard deviation. For normally distributed data, a threshold of three standard deviations is a common rule for outlier detection. This method is straightforward to implement and interpret, and it directly satisfies the requirement to flag points more than three standard deviations from the mean.

Answer

DBSCAN clustering

Answer

Isolation Forest

Answer

Interquartile range (IQR) method

Question 7

A data scientist is using Amazon SageMaker to train a model on a dataset with a highly imbalanced binary target (99% negative, 1% positive). The data scientist wants to improve the model's ability to detect the positive class without collecting more data. Which two of the following approaches should be used? (Choose two.)

Accepted Answer

Use the 'scale_pos_weight' hyperparameter in the XGBoost algorithm.. For imbalanced binary classification, using 'scale_pos_weight' in XGBoost adjusts the loss to focus on the minority class, and applying SMOTE generates synthetic minority samples to balance the training data. Both methods directly address the imbalance without discarding data. Increasing epochs or using accuracy does not help, and undersampling discards valuable negative examples.

Answer

Increase the number of epochs in the training job.

Answer

Use accuracy as the evaluation metric for model selection.

Answer

Remove some negative class samples to balance the dataset.

Question 8

A machine learning engineer is using Amazon SageMaker to train a deep learning model for natural language processing. The model is trained on a large corpus and achieves high accuracy on the training set but poor performance on the validation set. The engineer suspects overfitting and wants to apply regularization techniques. Which two of the following SageMaker features or techniques can help mitigate overfitting? (Choose two.)

Accepted Answer

Use dropout layers in the neural network architecture.. Early stopping and dropout are both effective regularization techniques for deep learning. Early stopping prevents the model from overfitting by halting training when validation performance degrades. Dropout introduces randomness during training, forcing the network to learn redundant representations and reducing overfitting. Together, they address the high training accuracy and low validation accuracy observed.

Answer

Increase the learning rate.

Answer

Use a larger batch size.

Answer

Increase the number of epochs for training.

Question 9

A data scientist is using Amazon SageMaker to run a hyperparameter tuning job for a deep learning model. The tuning job is configured to maximize the validation accuracy. After several training jobs, the data scientist notices that the tuning job is not exploring a wide range of hyperparameter values and is converging to a suboptimal configuration. Which action should the data scientist take to improve the tuning job's exploration?

Accepted Answer

Change the hyperparameter tuning strategy to Random.. The tuning job is likely using the default Bayesian strategy, which can exploit promising hyperparameters too early and miss better configurations. Switching to a Random strategy ensures a more thorough exploration of the hyperparameter space by sampling values randomly, which can prevent premature convergence and lead to better overall performance. Other options do not directly address the exploration issue.

Answer

Use a smaller learning rate for the model.

Answer

Decrease the value of the MaxNumberOfTrainingJobs parameter.

Answer

Increase the value of the MaxNumberOfTrainingJobs parameter.

Question 10

A data scientist is using Amazon SageMaker to train a deep learning model for natural language processing. The training job runs on multiple GPU instances and uses the SageMaker distributed data parallel library. The data scientist observes that the training throughput decreases significantly as the number of GPUs increases. What is the most likely cause of this performance degradation?

Accepted Answer

The all-reduce communication overhead is increasing faster than the computational savings from adding GPUs.. The correct answer is that all-reduce communication overhead is increasing faster than computational savings. In distributed data parallel training, each GPU computes gradients on its local batch, and these gradients must be averaged across all GPUs using all-reduce. As the number of GPUs grows, the all-reduce operation involves more participants and more data exchange, which can saturate network bandwidth, especially across nodes. This communication cost can dominate, reducing overall throughput. Other factors like dataset size, batch size, or learning rate do not directly cause throughput to drop as GPUs increase.

Answer

The learning rate is not scaled appropriately for the larger effective batch size.

Answer

The training dataset is too small for the number of GPUs, causing each GPU to process very few batches.

Answer

The batch size per GPU is too large, causing memory swapping on each GPU.

Question 11

A data scientist is using Amazon SageMaker to train a linear learner model for a regression task. The dataset has 50 features, and the target variable is continuous. After training, the model's predictions on the validation set are consistently off by a large margin, and the loss is high. The data scientist suspects that the features have very different scales, and some are highly correlated. Which preprocessing step should be applied to improve the model's performance?

Accepted Answer

Standardize the features by subtracting the mean and dividing by the standard deviation, and consider removing highly correlated features.. Linear models assume that features are on comparable scales and that they are not highly correlated. Standardization addresses the scale disparity, allowing the optimizer to converge more efficiently. Removing highly correlated features reduces multicollinearity, which can cause unstable coefficient estimates and poor predictive performance. Together, these steps are essential for preparing data for a linear learner. Other options either misinterpret the data type or fail to address the root cause.

Answer

Apply one-hot encoding to all numerical features to convert them into categorical features.

Answer

Increase the number of epochs and use a smaller learning rate to allow the model to converge.

Answer

Use a non-linear activation function in the linear learner model to capture complex relationships.

Question 12

A machine learning engineer has deployed a real-time SageMaker endpoint backed by three production variants. The engineer wants to shift 10% of live inference traffic from the current model to a newly trained model that is already registered and hosted on the same endpoint, while keeping the remaining 90% on the current model. The change must take effect immediately without redeploying the endpoint. Which action should the engineer take?

Accepted Answer

Call UpdateEndpointWeightsAndCapacities with the new variant's DesiredWeight set to 0.1 and the current variant's DesiredWeight set to 0.9.. A multi-variant endpoint already hosts both models, so traffic can be rebalanced in place. UpdateEndpointWeightsAndCapacities changes each variant's DesiredWeight immediately, which is exactly the mechanism for gradual canary rollouts and A/B splits. Because the weights are fractions summing to 1, setting the new variant to 0.1 and the incumbent to 0.9 delivers the required 10/90 live split without recreating the endpoint or interrupting existing callers.

Answer

Enable automatic scaling on the new variant and rely on Application Auto Scaling to distribute incoming traffic between variants.

Answer

Call CreateEndpointConfig with new ProductionVariants and then CreateEndpoint, letting SageMaker replace the existing endpoint.

Answer

Modify the endpoint's data capture configuration so requests are mirrored to both variants and compare results offline.

Question 13

A data scientist is using Amazon SageMaker to train a model and needs to monitor the training job's resource utilization and debug performance issues. The training job is running on multiple instances and uses a custom training script. The data scientist wants to collect metrics such as CPU utilization, GPU utilization, and memory usage, and also wants to capture tensor-level information for debugging. Which two actions should the data scientist take to meet these requirements? (Choose two.)

Accepted Answer

Use SageMaker Debugger to capture tensors and analyze them for debugging.. CloudWatch metrics provide resource utilization monitoring for SageMaker training jobs out of the box. SageMaker Debugger captures tensors and allows debugging of training issues. Together, they meet the requirements. X-Ray is for tracing, lifecycle configurations are for notebook instances, and SageMaker Experiments is for experiment tracking, not tensor debugging.

Answer

Enable AWS X-Ray tracing for the training job to trace tensor operations.

Answer

Use Amazon SageMaker Experiments to track tensor values across training runs.

Answer

Attach a lifecycle configuration script to the training instances to log GPU utilization to Amazon S3.

Question 14

A company has a production SageMaker real-time endpoint serving a fraud detection model. The model was trained on transaction data from several months ago, and the fraud patterns have shifted. The team wants to detect when the live request data diverges from the training data distribution so they can trigger retraining, without building custom monitoring infrastructure. Which solution should they implement?

Accepted Answer

Enable SageMaker Model Monitor with a data quality baseline created from the training dataset, and configure a monitoring schedule with CloudWatch alerts.. SageMaker Model Monitor is the managed capability for this need: it generates a baseline of statistics and constraints from the training data, then a monitoring schedule analyzes captured endpoint data and reports violations to CloudWatch. Because the analysis and alerting are built in, the team gets distribution drift detection and an alarm path to retraining without writing custom monitoring code.

Answer

Attach a custom CloudWatch Logs subscription filter to the endpoint's container logs and write a Lambda function that computes KL divergence for each feature.

Answer

Create an A/B test between the current model and a retrained model, and use the traffic split results to infer data drift.

Answer

Enable SageMaker Debugger on the endpoint and configure rules to detect distribution changes in incoming requests.

Question 15

A retail company stores daily sales records as newline-delimited JSON in an Amazon S3 bucket under the prefix s3://retail-data/raw/sales/. A data engineer must register this data in the AWS Glue Data Catalog so that Amazon Athena can query it, and the schema will evolve over time as the vendor adds fields. The engineer wants Glue to automatically infer the schema and detect the JSON structure without writing a custom classifier. Which approach should the engineer use?

Accepted Answer

Create an AWS Glue crawler targeting s3://retail-data/raw/sales/, let it use the built-in JSON classifier, and run the crawler on a schedule so schema changes are picked up.. An AWS Glue crawler with the built-in JSON classifier reads newline-delimited JSON in S3 and automatically populates the Data Catalog with an inferred schema, which Athena then queries. Scheduling the crawler keeps the schema current as new fields appear, satisfying both the inference and schema-evolution requirements without custom code.

Answer

Use Amazon Kinesis Data Firehose to deliver the JSON records into the Data Catalog and let Firehose infer the schema for Athena.

Answer

Define the table manually in the AWS Glue Data Catalog using the CREATE TABLE statement in Athena, specifying every column and its data type.

Answer

Create an AWS Glue ETL job that reads the JSON with a DynamicFrame and writes it to a new S3 location as Parquet, then query the Parquet output with Athena.

Question 16

A data scientist is performing exploratory data analysis on a dataset with 50 features. The scientist wants to identify pairs of numerical features that are highly linearly related to avoid multicollinearity in a subsequent linear model. The scientist also wants to visualize the relationships between multiple features simultaneously. Which two techniques should the scientist use? (Choose two.)

Accepted Answer

Create a scatterplot matrix (pairplot). Computing the Pearson correlation coefficient matrix and creating a scatterplot matrix are the two techniques that directly address the need to identify highly linearly related feature pairs and visualize multiple feature relationships. The correlation matrix quantifies linear relationships, while the scatterplot matrix provides a visual overview, together enabling effective detection of multicollinearity and exploration of feature interactions.

Answer

Calculate variance inflation factor (VIF) for each feature

Answer

Perform principal component analysis (PCA)

Answer

Use t-SNE for dimensionality reduction

Question 17

A data scientist is using Amazon SageMaker to train a model on a dataset that contains categorical features with high cardinality (e.g., user IDs). The data scientist wants to use the built-in Factorization Machines algorithm. Which data format should the training data be in to handle these categorical features effectively?

Accepted Answer

RecordIO-Protobuf format with sparse encoding.. The correct answer is RecordIO-Protobuf format with sparse encoding. Factorization Machines are designed to handle sparse data, especially high-cardinality categorical features, by learning latent vectors. The RecordIO-Protobuf format efficiently stores sparse data by only recording non-zero entries, reducing memory and compute. CSV with one-hot encoding would be dense and inefficient, while JSON Lines and Parquet are not natively supported by the algorithm. Using the correct format ensures the model can leverage the sparsity for effective training.

Answer

CSV format with one-hot encoding for all categorical features.

Answer

JSON Lines format with nested categorical fields.

Answer

Parquet format with dictionary encoding for categorical columns.

Question 18

A machine learning engineer is using Amazon SageMaker to train a model on a large dataset stored in Amazon S3. The dataset is in CSV format and is 500 GB. The engineer wants to reduce training time by using SageMaker Pipe mode instead of File mode. However, the training script currently reads the entire CSV file into memory and performs preprocessing. What should the engineer do to adapt the training script for Pipe mode?

Accepted Answer

Modify the script to read the data as a streaming dataset using the SageMaker PipeModeDataset and apply preprocessing on the fly.. Pipe mode streams data from S3 to the training container, eliminating the need to download the entire dataset. The training script must use the PipeModeDataset to read the stream and perform preprocessing per batch. This adaptation allows efficient training on large CSV datasets without loading everything into memory.

Answer

Increase the instance count and use distributed training with File mode to handle the large dataset.

Answer

Convert the CSV files to TFRecord format and use the SageMaker Pipe mode with the TensorFlow dataset API.

Answer

Use SageMaker Augmented Manifest files to stream the CSV data and preprocess using the SageMaker TensorFlow estimator.

Question 19

A machine learning engineer needs to prepare a training dataset from a large collection of JSON files stored in Amazon S3. The files are compressed with Gzip and contain nested structures. The engineer wants to use AWS Glue to transform the data into Parquet format, flatten the nested fields, and store the result in a partitioned S3 location for efficient querying by Amazon Athena. The transformation must handle schema evolution and nested data without manual coding. Which AWS Glue feature should the engineer use to accomplish this?

Accepted Answer

AWS Glue DynamicFrame with relationalize and apply_mapping transforms. DynamicFrame transforms in AWS Glue are specifically built to handle nested and evolving schemas without manual coding. The relationalize transform flattens nested structures, and apply_mapping allows field-level transformations. This combination efficiently converts nested JSON into a flattened Parquet dataset suitable for Athena, meeting all stated requirements.

Answer

AWS Glue DataBrew with a recipe that uses the UNNEST function

Answer

AWS Glue crawler with a custom classifier for JSON and then use Athena to UNNEST

Answer

AWS Glue Studio with a custom Spark script using the explode() function

Question 20

A data analyst has a 12 GB Parquet dataset in Amazon S3 and needs to run interactive exploratory queries from an Amazon SageMaker Studio notebook. The analyst wants to query the data with SQL-like syntax, avoid loading the entire dataset into memory, and minimize cost without provisioning a persistent cluster. Which approach should the analyst use?

Accepted Answer

Register the data in the AWS Glue Data Catalog and query it with Amazon Athena directly from the notebook.. The requirements point to a serverless, SQL-capable engine that reads Parquet in place and scans only needed columns. Athena, with metadata in the AWS Glue Data Catalog, satisfies all of them: no cluster, pay-per-scan pricing, SQL syntax, and columnar pruning on Parquet. The other approaches either load everything into memory or require provisioned compute.

Answer

Load the full Parquet dataset into a pandas DataFrame in the notebook and run groupby operations.

Answer

Provision an Amazon EMR cluster with Spark, then connect the Studio notebook to the cluster's Livy endpoint.

Answer

Copy the dataset into an Amazon Redshift cluster and query it with the Redshift Data API.