Data Preparation for Machine Learning — AWS Certified Machine Learning Engineer MLA-C01 practice questions

For continuous delivery of stream records to S3 with buffering, use which service?

1. Kinesis Data Streams only
2. Managed MSK cluster
3. Kinesis Data Firehose
4. SQS with poller Lambda

Answer: C — Kinesis Data Firehose

Firehose is the managed buffered delivery layer; raw Streams have no built-in sink; MSK adds ops; SQS is a queue.

SageMaker Data Wrangler is BEST used for which purpose?

1. Visual feature engineering flow
2. Distributed training clusters
3. Drift monitoring on endpoints
4. Endpoint autoscaling policy

Answer: A — Visual feature engineering flow

Data Wrangler is UI-driven data prep; autoscaling, training clusters, and drift are separate SageMaker features.

For splitting a classification dataset while preserving class ratios, use which?

1. Only random uniform split
2. Only time-based split
3. Stratified train-test split
4. Only leave-one-out split

Answer: C — Stratified train-test split

Stratified split preserves class proportions; random can imbalance; time-based is for series; LOO is for small data.

To share catalog metadata across accounts without copying data, use which service?

1. Lake Formation cross-account
2. DMS cross-account replication
3. Public S3 bucket exposure
4. S3 copy jobs cross-account

Answer: A — Lake Formation cross-account

Lake Formation grants cross-account catalog access; copy jobs duplicate; public buckets are unsafe; DMS is a DB tool.

Which is BEST for orchestrating a multi-step data pipeline across AWS services?

1. AWS Step Functions workflow
2. Only S3 event notifications
3. Only CloudWatch schedule rules
4. Only manual shell scripts

Answer: A — AWS Step Functions workflow

Step Functions orchestrates multi-step workflows; S3 events trigger single steps; CloudWatch schedules cron; scripts are ad hoc.

Ground Truth is BEST used for which stage of the ML lifecycle?

1. Labeled dataset creation
2. Only batch inference runs
3. Only endpoint autoscaling
4. Model training runtime only

Answer: A — Labeled dataset creation

Ground Truth is a managed labeling service; the other stages are separate SageMaker features.

For time-series validation, MOST fit split strategy is which?

1. Stratified random splits
2. Random k-fold shuffling
3. Time-ordered rolling folds
4. Bootstrap resample splits

Answer: C — Time-ordered rolling folds

Time-ordered rolling folds respect temporal order; random and bootstrap leak future info into training.

For imbalanced classification, SMOTE PRIMARILY works by which mechanism?

1. Synthesizing minority samples
2. Deleting majority-class rows
3. Random over-sample minority
4. Adding purely random noise

Answer: A — Synthesizing minority samples

SMOTE synthesizes new minority samples by interpolating neighbors; random duplicates and noise are different.

Detecting multicollinearity in linear features is BEST done by which measure?

1. Only test-set accuracy score
2. Only training loss reading
3. Variance Inflation Factor score
4. Only train-time GPU count

Answer: C — Variance Inflation Factor score

VIF quantifies multicollinearity; the others are metrics or infra unrelated to feature relationships.

Missing categorical values are MOST fit imputed by which strategy?

1. Zero-fill for all missing
2. Random category assignment
3. Mode or new UNKNOWN token
4. Deletion of whole column

Answer: C — Mode or new UNKNOWN token

Mode or UNKNOWN preserves distribution; random adds noise; zero-fill is not meaningful; deletion loses signal.

Standardization vs Normalization?

Standardize: mean 0, std 1 (Z-score). Normalize: scale to [0,1] (min-max). Standardize for assumes-normal models; normalize for distance-based.

Optimal file format for ML training?

Apache Parquet (columnar, compressed) or RecordIO (SageMaker pipe mode).

Amazon EMR?

Managed Hadoop/Spark cluster service. For large-scale data processing/feature engineering.

Best AWS service for batch data ingestion to S3?

AWS DataSync, AWS Glue, AWS Snow Family, AWS Transfer Family, Amazon Kinesis Firehose (for streaming).