Normalization and ML Models

• Splits data into related tables so each fact is stored once. Removes redundancy and update anomalies, and keys keep the data consistent.
• Cost: more joins at read time.
• For ML and analytics, the training table is often denormalized on purpose into one wide row per entity, because reads matter more than write integrity.

2. Feature normalization (scaling)

• Min-max scaling squeezes values into 0 to 1. Z-score centers on the mean with unit variance.
• Needed for distance-based and gradient-based models: k-means, KNN, SVM, linear models, neural networks.
• Not needed for tree models (random forest, XGBoost).
• Compute the mean and standard deviation on the training split only, then reuse them on validation and test data. Otherwise you leak information.

-- Z-score in SQL (compute the statistics on training rows only in real pipelines)
SELECT id,
       (amount - AVG(amount) OVER ()) / NULLIF(STDDEV_SAMP(amount) OVER (), 0) AS amount_z
FROM transactions;

Interview line: normalize the database for integrity, denormalize the feature table for speed, and scale features inside the ML pipeline.