Data Preparation for Machine Learning practice questions

Use Amazon Athena to convert the data to JSON format and store it in S3.

Why it fails: Athena is a query service, not a data transformation service.

Use AWS Glue to convert the data to Apache Hive format.

Why it fails: Hive format is not a standard file format; Parquet is columnar and efficient.

Use Apache Spark DataFrame to write the data as CSV with Snappy compression.

Why it fails: CSV is not columnar and does not optimize query performance as well as Parquet.

Apply data augmentation to the majority class by adding noise.

Why it fails: Data augmentation is not standard for tabular data and may introduce noise.

Use a weighted loss function during training to penalize misclassifications of the minority class.

Why it fails: Weighted loss is a training technique, not a data preparation step.

Apply random under-sampling to reduce the majority class to match the minority class size.

Why it fails: Under-sampling discards data and may lose important patterns.

Use partition pruning on the source data to reduce the amount of data processed.

Why it fails: Partition pruning reduces data scanned but does not directly improve resource consumption efficiency.

Switch the output format from Parquet to CSV to reduce processing overhead.

Why it fails: CSV is not columnar and may increase I/O and storage costs.

Use a larger instance type for the Glue job by increasing the number of DPUs.

Why it fails: More DPUs increase parallelism but also cost; may not be cost-effective.

Apply hash encoding to map categories to a fixed number of buckets.

Why it fails: Hash encoding may cause collisions and lose information.

Apply one-hot encoding to all categorical features.

Why it fails: One-hot encoding high-cardinality features creates too many columns, increasing dimensionality.

Apply label encoding to assign integer values to each category.

Why it fails: Label encoding implies an order that may not exist, leading to misleading patterns.

Apply feature selection to reduce the number of features.

Why it fails: Feature selection is optional and not a mandatory step for all models.

Remove outliers from the dataset.

Why it fails: Outlier removal is not always required; it depends on the data and model.

Include the raw PII in the training dataset and rely on the model to not memorize it.

Why it fails: Retaining raw PII in training data breaches privacy regulation and cannot be remedied by hoping the model avoids memorisation. Data Wrangler's PII detection transform exists precisely to mask, redact or drop such fields before training; that transform is the compliant path this option bypasses.

Use AWS Glue to copy the data to a separate bucket without any transformations.

Why it fails: Copying PII verbatim to another bucket preserves the sensitive data, so it fails the privacy requirement entirely. AWS Glue is tempting because it performs ETL at scale, but it would be correct only when transformations such as masking or tokenisation are actually applied during the copy.

Random undersampling of the majority class

Why it fails: Removes potentially valuable data.

Random oversampling of the minority class

Why it fails: Duplicates existing samples, risk of overfitting.

Apply class weights during model training

Why it fails: Affects loss function, not data preparation.

Label encoding

Why it fails: Assumes ordinal relationship, not suitable for nominal categories.

One-hot encoding

Why it fails: Creates too many features, leading to high dimensionality.

Frequency encoding

Why it fails: Replaces with count/frequency, may lose information.

Set k=1 to minimize bias

Why it fails: k=1 is prone to overfitting and noise.

Use all 100,000 rows to find neighbors for each missing value

Why it fails: Computationally expensive; consider sampling or approximate methods.

Use only the feature with missing values to find neighbors

Why it fails: Does not use information from other features.

Write a custom Python script using pandas dt.day_name()

Why it fails: Unnecessarily complex when built-in transform exists.

Use one-hot encoding on the timestamp

Why it fails: One-hot encoding is for categorical features, not extraction.

Use the 'extract' transform with format '%A'

Why it fails: Not a supported transform in Data Wrangler.

Apply log transformation after adding a constant

Why it fails: Log transform is for positive data; adding constant is arbitrary.

Remove all rows with negative sales values

Why it fails: Loses data on returns, which may be informative.

Impute negative values with the mean

Why it fails: Incorrectly treats negative values as missing; mean may be positive.

SMOTE to balance classes

Why it fails: Addresses class imbalance, not general overfitting.

One-hot encoding of all categorical features

Why it fails: Increases dimensionality, potentially worsening overfitting.

Log transformation of skewed features

Why it fails: Addresses skewness, not overfitting.

Use AWS Glue DataBrew to fill missing values with the median of total_purchases.

Why it fails: Median imputation, like mean imputation, does not preserve the distribution and can introduce bias.

Drop all records with missing total_purchases values.

Why it fails: Dropping 15% of records reduces sample size and may bias the model if missingness is not random.

Replace missing total_purchases with the mean of the non-missing values.

Why it fails: Mean imputation reduces variance and can distort relationships between variables.

Use AWS Glue to convert all timestamps to UTC, apply a mapping function to correct merchant category misspellings to a standard list, and drop records with missing fraud_label.

Why it fails: Dropping 5% of records may lose important fraud cases and introduce bias.

Use AWS Glue to convert timestamps to UTC, use a fuzzy matching algorithm to standardize merchant categories, and replace missing fraud_label with the mean value (0.05).

Why it fails: Mean imputation on a binary variable produces non-integer values, which are invalid for classification.

Use AWS Glue to convert timestamps to UTC, correct merchant categories by mapping known misspellings to correct names, and drop records with missing fraud_label.

Why it fails: Only mapping known misspellings may miss variants, and dropping missing labels causes data loss.

In AWS Glue DataBrew: 1) Impute age with the mean. 2) Apply Standard Scaler to all numeric columns. 3) Use Random Oversampling to balance the target column.

Why it fails: Imputing with mean does not remove invalid values like negative ages, and Standard Scaler may not be appropriate for all numeric columns.

In AWS Glue DataBrew: 1) Replace age with median. 2) Convert diagnosis_code to uppercase. 3) Apply SMOTE to balance the target column.

Why it fails: Replacing with median does not remove invalid values, and SMOTE may create synthetic samples that are not realistic for categorical data.

In AWS Glue DataBrew: 1) Remove rows where age is outside 0-120. 2) Drop diagnosis_code column. 3) Use Random Undersampling to balance the target column.

Why it fails: Dropping diagnosis_code loses potentially important information, and undersampling reduces sample size.

Use AWS Glue to group records by session_id and aggregate event_types into a list per session. Then apply a mapping function to standardize event_type names.

Why it fails: Grouping by session loses the sequential order of events, which is critical for sequence-based models.

Use AWS Glue to drop duplicate records based on all columns. Then drop the event_type column and use only numeric features for training.

Why it fails: Dropping event_type removes the core interaction type, which is essential for recommendation.

Use AWS Glue to impute event_type with the mode for records with inconsistent values. Then drop duplicate records based on user_id, product_id, and timestamp.

Why it fails: Mode imputation on event_type is not appropriate because it would assign a potentially incorrect event type to many records.

Add the libraries to the `requirements.txt` file in the same S3 bucket as the script

Why it fails: requirements.txt alone does not trigger installation; you must explicitly run pip.

Use Amazon EFS to store the libraries and mount them to the processing container

Why it fails: EFS is an unnecessary complexity for simple library dependencies.

Use the `pip install` command within the processing script at runtime

Why it fails: Runtime pip install increases job duration and may fail inconsistently.

Use a bucketed join with the same number of buckets

Why it fails: Without co-location, shuffling may still occur.

Broadcast join the larger dataset

Why it fails: Broadcast join is for small datasets, not both large.

Use a repartition on the join key before join

Why it fails: Repartition reduces but does not eliminate shuffling.

One-hot encoding

Why it fails: One-hot encoding creates 500 sparse binary columns, directly violating the requirement to avoid increasing dimensionality. It is the correct choice for low-cardinality nominal features, where the added columns remain manageable and no target leakage is introduced.

Frequency encoding

Why it fails: Frequency encoding maps each category to its occurrence count, discarding the target relationship entirely, so it cannot capture how neighbourhoods relate to price. It is tempting because it keeps dimensionality to one column and handles high cardinality, but it suits unsupervised or tree-based feature engineering, not target-aware regression encoding.

Label encoding

Why it fails: Label encoding assigns each of the 500 neighbourhoods an arbitrary integer, implying a false ordinal relationship that regression models will treat as numeric distance, so it cannot capture the target relationship. It is tempting because label encoding suits tree-based models, where arbitrary integer splits work fine, or genuinely ordinal features such as education levels.

Target encoding of high-cardinality categorical features

Why it fails: Target encoding can cause data leakage and reduces interpretability.

PCA dimensionality reduction

Why it fails: PCA transforms features into uninterpretable components.

One-hot encoding of categorical features

Why it fails: One-hot encoding adds many columns, complicating interpretability.