AWS MLS-C01: Exploratory Data Analysis Practice Questions

One-hot encoding a 3,200-cardinality feature produces a very sparse 3,200-dimensional binary vector per record. The SageMaker Factorization Machines algorithm is specifically designed to work efficiently with sparse input (it accepts data in sparse RecordIO-protobuf format), making one-hot encoding a legitimate and common choice. Ordinal encoding assigns arbitrary integers (e.g., 1–3200) with no ordinal meaning, implying false numeric relationships between unrelated categories — this can mislead distance-based or linear models and is not recommended for nominal high-cardinality features. Target encoding reduces dimensionality to a single numeric column but is prone to target leakage and overfitting when applied naively without proper cross-validation folds, making 'always the safest choice' incorrect. Hashing (feature hashing) maps categories to a fixed-size hash space and produces sparse output compatible with Factorization Machines — the claim that it requires dense input is incorrect.

B) Correct: Amazon Athena is a fully serverless, interactive query service that runs SQL directly on data stored in S3 (including Parquet format) using the AWS Glue Data Catalog. It requires no infrastructure provisioning, making it ideal for ad-hoc EDA aggregations at scale. A) Incorrect: Redshift Spectrum can query S3 data but requires an active Redshift cluster to be provisioned and managed, which contradicts the serverless requirement. C) Incorrect: AWS Glue ETL jobs are designed for batch transformation pipelines, not interactive EDA; they also have higher startup overhead and are not query-oriented. D) Incorrect: SageMaker Processing can run Pandas scripts but requires launching a managed compute instance and loading data into memory, which is less scalable and efficient than Athena for SQL-based aggregations on large Parquet datasets.

Amazon Athena is the correct choice because it is a serverless, interactive query service that allows you to run standard SQL queries directly against data stored in Amazon S3 without any infrastructure provisioning. It is purpose-built for ad-hoc EDA at scale. — Amazon Redshift Spectrum requires a provisioned Redshift cluster as the query coordinator, so it is not fully serverless and adds unnecessary overhead for simple EDA. — AWS Glue ETL jobs use PySpark and are optimized for batch data transformation pipelines, not interactive SQL-based exploration. — Amazon EMR requires cluster provisioning and management, which contradicts the 'no server provisioning' requirement.

Median imputation is the correct choice for a heavily right-skewed distribution. The median is a robust statistic unaffected by extreme values or skew, making it a better central tendency measure than the mean in this scenario. — Mean imputation is problematic for skewed distributions because the mean is pulled toward the tail, causing the imputed values to be unrepresentative of the typical value. — Mode imputation is appropriate for categorical variables or discrete numeric features with a clear dominant value, not for continuous skewed numeric data. — Forward-fill is a time-series technique that uses the previous row's value and is inappropriate for cross-sectional tabular data with no meaningful row ordering.

A) Correct: The Z-score method is a standard statistical technique for detecting outliers in approximately normal distributions; values with |z| > 3 are typically flagged as outliers. B) Correct: The IQR method is a robust, distribution-free technique for outlier detection that does not assume normality; it is especially effective for skewed distributions. C) Correct: Isolation Forest is a well-established unsupervised machine learning algorithm for anomaly/outlier detection that works by randomly isolating observations; it is available in scikit-learn and can be run in a SageMaker Studio notebook. D) Incorrect: Min-Max Scaling rescales values to a fixed range but does not detect or remove outliers — extreme values will still dominate the scaled range and distort the scaling for other observations. E) Incorrect: Dropping an entire feature because it contains outliers is an extreme and inappropriate action; the correct approach is to handle the outlier observations (cap, transform, or remove them), not the feature itself.