AWS Certified Machine Learning Specialty MLS-C01 (MLS-C01) — Questions 451–525

Options A and C are correct. Adding L2 regularization penalizes large weights, reducing variance. Reducing model complexity (e.g., using a simpler model) also reduces overfitting.

Option B is wrong because adding more features increases complexity. Option D is wrong because increasing training epochs may lead to more overfitting. Option E is wrong because decreasing training data increases variance.

XGBoost supports scale_pos_weight to handle class imbalance, directly optimizing for F1. Linear Learner with balanced class weights can also help but typically optimizes log loss. K-Means is unsupervised.

PCA is for dimensionality reduction.

Increasing the number of shards increases parallelism, and using enhanced fan-out allows each consumer to have its own read throughput. Option B is wrong because decreasing retention period does not improve consumer performance. Option D is wrong because using a Lambda function may not help if the bottleneck is shard throughput.

Option E is wrong because using KCL (Kinesis Client Library) is already standard; not a direct fix.

Precision is low, recall is high. To improve precision, the data scientist can adjust the decision threshold or use a different algorithm. Option A (increasing regularization) may help but is not direct.

Option D (using class weights) addresses imbalance. Option C (collecting more data) may not help. Option E (changing to random forest) could improve precision.

Option B (removing features) might hurt.

Switching to Bayesian search (e.g., 'Bayesian' strategy) is more efficient because it uses past results to choose the next hyperparameters, potentially finding better values in fewer jobs. Increasing the number of jobs would increase cost. Random search might get lucky but is less efficient.

Changing the objective metric or scaling features would not directly improve the tuning process.

Option A is correct. The list-objects command returns up to 1000 objects per call. If there are more objects, pagination is needed.

Option B is incorrect because the query syntax is correct. Option C is incorrect because the prefix is fine. Option D is incorrect because the unit is bytes, so 1000000 is 1 MB.

Amazon S3 provides scalable object storage with versioning. Option A is wrong because EBS is block storage for single EC2 instances. Option C is wrong because EFS is file storage for shared access.

Option D is wrong because DynamoDB is a NoSQL database.

The table is not partitioned (empty PartitionKeys), and the SerDe expects comma-delimited files. If some files use a different delimiter (e.g., tab), those rows may not be parsed correctly, leading to fewer rows counted. Option A is wrong because the table schema matches the data.

Option B is wrong because Athena can handle many small files, but count should still include all rows. Option D is wrong because Athena does not skip files based on size unless explicitly filtered.

Option A is correct because side-by-side boxplots show distribution and outliers across categories. Option C is correct because overlapping histograms with transparency show distribution shapes. Option B is wrong because scatter plots require two numerical variables.

Option D is wrong because heatmaps are for correlation or contingency tables. Option E is wrong because bar charts are for categorical data, not numerical distributions.

Options B and D are correct. Using SageMaker Pipe Mode streams data from S3, reducing download time. Using multiple GPUs (distributed training) parallelizes computation.

Option A (batch size decrease) may slow training. Option C (larger images) increases computation. Option E (more epochs) increases training time.

Option B is correct because analyzing the missing pattern (e.g., MCAR, MAR, MNAR) guides imputation strategy. Option A is wrong because mode imputation is a method, not a diagnostic step. Option C is wrong because label encoding is for modeling, not imputation decision.

Option D is wrong because dropping rows may lose data; need to understand missingness first.

Options B and D are correct. Option A is wrong because scatter plot is for two variables. Option C is wrong because heatmap shows correlation.

Option E is wrong because bar chart is for categorical data.

Using SageMaker's distributed data parallelism (e.g., with SageMaker distributed training libraries) across multiple GPUs can significantly reduce training time by splitting the mini-batches across GPUs. Increasing instance type to a single larger GPU (e.g., p3.16xlarge) helps but is less effective than multi-GPU distribution. Hyperparameter tuning doesn't directly reduce training time.

Spot instances may interrupt.

Option D is correct because increasing the number of shards increases the capacity of the stream, allowing more parallel consumers and reducing backlog. Option A is wrong because reducing retention period does not affect processing speed. Option B is wrong because increasing retention period may increase backlog.

Option C is wrong because decreasing shards reduces capacity, worsening the issue.

Option A is correct because reducing max_depth prevents trees from growing too deep, which can cause numeric overflow. Option B is wrong because increasing eta helps convergence but not overflow. Option C is wrong because increasing num_round may worsen overflow.

Option D is wrong because subsample doesn't affect depth.

Option C is correct because mutual information can capture any dependency. Option A is wrong because Pearson correlation only captures linear. Option B is wrong because Spearman captures monotonic but not all non-linear.

Option D is wrong because ANOVA is for categorical vs continuous.

Sort keys help the query optimizer scan less data. Distribution keys reduce data shuffling. ANALYZE commands update statistics for the optimizer.

Option B (increasing node count) is costly and not always necessary. Option D (deleting old data) may help but is not a direct performance tuning technique.

Option C is correct because log transformation is commonly used to reduce right skewness. Option A is wrong because standardization does not change distribution shape. Option B is wrong because min-max scaling does not change shape.

Option D is wrong because one-hot encoding is for categorical variables, not continuous.

Option A is correct because early stopping terminates poor performing jobs early, saving time. Option B is wrong because reducing max runtime may not allow convergence. Option C is wrong because increasing max jobs would increase time.

Option D is wrong because random search is faster but may miss optimal hyperparameters; however, early stopping directly reduces time on bad trials.

Option C is correct because Isolation Forest is an ensemble method that isolates anomalies effectively in high-dimensional spaces without assuming distribution. Option A is wrong because Z-score assumes normal distribution. Option B is wrong because IQR is univariate.

Option D is wrong because DBSCAN is for clustering, not specifically for outlier detection.

Options B and E are correct. B: Visualizing the distribution helps identify shape and outliers. E: Reporting missing value counts is a standard EDA step.

A is wrong because imputation should be done after analysis. C is wrong because removing outliers without analysis is premature. D is wrong because normalization is not a first step.

Option C is correct because the policy allows s3:PutObject on 'my-bucket/*', but the training job may need to write to a different bucket or path. Option A is wrong because the actions are allowed. Option B is wrong because s3:PutObject is allowed.

Option D is wrong because the bucket is specified. Option E is wrong because the training job role is separate from the notebook role.

SageMaker provides pre-built deep learning containers (DLCs) for PyTorch. Using a SageMaker framework estimator with a pre-built PyTorch container is the easiest and most efficient. The framework estimator automatically handles the container and dependencies.

Creating a custom container is more work. Using the generic container requires manually installing dependencies. Using a notebook instance is for interactive development, not one-time training.

Option B is correct because the ETag can be used to verify file integrity. Option A is wrong because ContentLength shows size in bytes, ContentLength 1048576 is 1 MB. Option C is wrong because ETag is not a version ID; S3 versioning uses VersionId.

Option D is wrong because preprocessed metadata indicates the file has been processed.

A learning rate that is too high can cause the loss to plateau or diverge. Reducing the learning rate often helps. Increasing batch size may stabilize training but not directly address plateau.

Switching optimizer or adding regularization may help but are less direct.

Option C is correct because Data Wrangler allows specifying custom missing value symbols during import. Option A is wrong because dropping rows is premature. Option B is wrong because replacing after import is less efficient.

Option D is wrong because imputation is needed after treating missing values.

Option A is correct because using Pipe input mode streams data directly from S3 to the training algorithm without downloading, reducing startup time. Option B is wrong because FSx for Lustre is not needed for simple streaming. Option C is wrong because increasing instance storage does not address the data transfer issue.

Option D is wrong because using EBS optimized instances does not change the data loading mechanism.

Option C is correct because switching to Bayesian search will explore new regions based on previous results, potentially finding better hyperparameters. Option A is wrong because increasing the number of parallel jobs does not change the search strategy. Option B is wrong because decreasing the number of training jobs reduces exploration.

Option D is wrong because early stopping does not change the search strategy.

Option B is correct because converting CSV to optimized formats like Parquet or RecordIO reduces I/O overhead and improves throughput. Option A is wrong because increasing batch size may help but does not address I/O. Option C is wrong because changing to a GPU instance is not allowed per stem.

Option D is wrong because reducing epochs reduces accuracy, not training time effectively.

Option C is correct because the policy allows s3:GetObject only on the 'my-training-data' bucket, but Athena writes results to a different S3 location (specified in the query configuration). Without permission to read from that output bucket, the notebook cannot load the results. Option A is wrong because Athena actions are allowed.

Option B is wrong because listing the training data bucket is allowed. Option D is wrong because reading individual objects from the training data bucket is allowed.

Distributed data parallelism (Option B) splits the data across multiple GPUs, reducing per-worker memory load and speeding up training. Option A (batch normalization) does not reduce training time. Option C (Spot Instances) introduces interruptions and may increase total time.

Option D (SageMaker Debugger) is for monitoring, not performance.

Option D is correct because SMOTE generates synthetic samples for the minority class. Option A is wrong because it discards majority class data. Option B is wrong because scaling does not balance classes.

Option C is wrong because encoding is for categorical features.

AWS Glue Crawlers automatically scan data sources, infer schemas, and populate the AWS Glue Data Catalog. Option B is wrong because Glue ETL jobs are for transforming data. Option C is wrong because Glue DataBrew is for visual data preparation.

Option D is wrong because Glue Studio is a visual ETL development tool.

Option A is correct because using SPICE in-memory engine speeds up queries by caching data. Option B is wrong because sampling loses accuracy. Option C is wrong because reducing bin count is a trade-off.

Option D is wrong because Athena is query engine, but QuickSight already uses it; SPICE is better.

Option C is correct because object dtype typically indicates string or mixed types; converting to numeric allows mathematical operations. Option A is wrong because object dtype does not necessarily indicate missing values. Option B is wrong because object dtype is not automatically numeric.

Option D is wrong because normalization is done after conversion.

Increasing early stopping rounds and adding regularization (like gamma or lambda) helps reduce overfitting. Lowering learning rate with more rounds can also help. Option A (decreasing rounds) might underfit.

Option C (increasing max_depth) worsens overfitting. Option D (increasing num_round) with no regularization may overfit more.

Option B and D are correct. Converting to Parquet reduces data scanned due to columnar storage and compression. Partitioning by customer_id (which is frequently filtered) improves partition pruning.

Option A is wrong because increasing workers is not applicable to Athena (serverless). Option C is wrong because converting to JSON increases size. Option E is wrong because using S3 Select may not integrate with Athena directly.

Option C is correct because using Pipe input mode streams data directly from S3 to the training algorithm without writing to disk, reducing I/O latency. Option A is wrong because increasing instance size may help but incurs higher cost and may not address data loading bottleneck. Option B is wrong because SageMaker automatically manages data channels.

Option D is wrong because enabling data compression on the S3 objects does not reduce training time significantly and adds CPU overhead for decompression.

Option C is correct because SageMaker Debugger can profile the training job and identify bottlenecks. Option A may add overhead. Option B is not detailed.

Option D is for inference.

Option A is correct because ANOVA F-statistic tests if means across groups are significantly different. Option B is wrong because chi-square is for categorical features. Option C is wrong because mutual information is for feature selection but doesn't directly test separability.

Option D is wrong because variance ratio is not standard.

SageMaker Inference Recommender provides load testing and recommendations for instance type and endpoint configuration. It can help identify optimal settings to reduce latency spikes. Multi-model endpoints are for hosting multiple models, not directly for latency spikes.

Elastic Inference is for accelerating deep learning inference, not general latency. Automatic scaling adjusts capacity but not per-request latency.

Option B is correct because early stopping halts training when validation loss stops improving, preventing overfitting. Option A is wrong because adding more layers may worsen overfitting. Option C is wrong because increasing learning rate can cause divergence.

Option D is wrong because decreasing batch size may increase noise but not directly address overfitting.

Option A is correct: configuring auto scaling based on CPU utilization adds instances during peak and removes them during off-peak, meeting latency and cost goals. Option B (Batch Transform) is for offline inference. Option C (larger instance) is less cost-effective.

Option D (scale down) would worsen latency.

Option C is correct because dropping one of the highly correlated features reduces redundancy and mitigates multicollinearity. Option A is wrong because scaling does not address collinearity. Option B is wrong because PCA creates orthogonal components but reduces interpretability; dropping a feature is more straightforward.

Option D is wrong because adding polynomial features increases correlation.

Options B and D are correct. B: Visualizing missing patterns (e.g., with a missingno matrix) is a good EDA practice. D: Understanding the mechanism (MCAR, MAR, MNAR) is critical for choosing imputation method.

A is wrong because listwise deletion can introduce bias if data not MCAR. C is wrong because mean imputation reduces variance. E is wrong because missing values should be handled before model training, not after.

Option C is correct because a histogram or density plot reveals skewness visually. Option A is wrong because count plot is for categorical data. Option B is wrong because scatter plot shows relationship between two variables.

Option D is wrong because heatmap shows correlation, not skewness.

Option B is correct because Pipe mode streams data directly from S3, reducing disk I/O. Option A (File mode) writes to disk first, causing I/O wait. Option C (ShardedByS3Key) is for distributed training but does not reduce I/O.

Option D (RecordIO) reduces file size but still uses File mode. Option E (SageMaker batch transform) is for inference, not training.

Option D is correct because AUC measures the model's ability to distinguish between classes irrespective of threshold, and is suitable for imbalanced datasets. Option A is wrong because accuracy is misleading when classes are imbalanced. Option B is wrong because precision only considers positive predictions.

Option C is wrong because recall only considers actual positives.

Option C is correct because the IQR method identifies outliers as points outside 1.5*IQR from the quartiles, which is robust to non-normal distributions. Option A is incorrect because Z-score assumes normality and is sensitive to extreme outliers. Option B is incorrect because Mahalanobis distance is for multivariate outliers.

Option D is incorrect because PCA is for dimensionality reduction, not outlier detection.

Using SageMaker's production variants with auto-scaling can help handle traffic spikes without over-provisioning, thus managing latency and cost. Switching to a larger instance would increase cost. Reducing the number of instances would increase latency.

Enabling data capture adds overhead and increases latency.

Options A and B are correct. A is correct because imputation with mean/median fills missing values. B is correct because removing rows with missing values is a valid approach.

C is wrong because one-hot encoding is for categorical data, not missing values. D is wrong because PCA is for dimensionality reduction, not missing value handling. E is wrong because standardization is for scaling, not missing values.

Options A, B, and D are correct. A: Create a production variant with the new model. B: Set up a production variant for the existing model.

D: Configure the endpoint to route traffic between variants. Option C (CloudWatch) is monitoring, not part of setup. Option E (Batch Transform) is for batch, not real-time.

Option B is correct because checking the partition metadata using SHOW PARTITIONS or querying the information_schema helps verify that the partition structure is correct and that queries are using partition pruning. Option A is incorrect because compressing data reduces storage but does not directly affect partition pruning. Option C is incorrect because converting to Parquet improves columnar scanning but does not address partition misuse.

Option D is incorrect because increasing workers does not fix incorrect partition usage.

Options A and D are correct. Option A: The IAM role must have s3:GetObject permission for the bucket. Option D: The input data channel must specify the S3 URI with the correct prefix.

Option B is wrong because File mode does not require RecordIO. Option C is wrong because Pipe mode is not used. Option E is wrong because File mode does not require data in a single file.

Option A is correct because manually checking files confirms data presence. Option B is correct because MSCK REPAIR TABLE adds partitions not yet registered. Option C is correct because incorrect data types can cause filters to exclude rows.

Option D is wrong because the query does not automatically update partitions. Option E is wrong because converting to Parquet is not a troubleshooting step.

The 'ProvisionedThroughputExceededException' error in Amazon Kinesis Data Streams indicates that the data ingestion rate exceeds the write capacity of the stream's shards. Each shard supports up to 1 MB/s or 1,000 records/s for writes. If the clickstream data volume surpasses this limit, the Lambda function, which reads from the stream, will encounter this exception.

Increasing the number of shards scales the write capacity to match the data volume.

Option A is correct because deploying across multiple Availability Zones provides zone redundancy. Option C is correct because using multiple instances in the endpoint configuration ensures that if one instance fails, others handle traffic. Option D is correct because enabling auto-scaling can replace failed instances.

Option B is wrong because a single instance in one AZ is not resilient. Option E is wrong because a single AZ does not protect against AZ failure.