Apply Leakage-Safe Imputation and Ordinal Encoding

You are given training and test tables that include numeric columns, an ordered categorical column service_tier with business ordering bronze < silver < gold, and a binary target available only in training. Some feature values are missing, and the test set may contain a tier that never appeared in training. Describe how to create processed_train and processed_test using imputation and ordinal encoding while preventing any information from the test set from leaking into the pipeline. Cover when to call fit_transform versus transform, how to keep feature alignment intact, and how to deal with unknown categories.

Constraints & Assumptions

• The target exists only in the training table and must never be passed into a feature transformer.
• Numeric and categorical columns may be imputed with different strategies.
• The two processed outputs must share the same feature columns in the same order.
• Unknown or missing tier values must not be silently mapped to a higher business tier.

Clarifying Questions to Ask

• Is the tier ordering defined by the business, or should it be learned from the observed labels?
• Should missing tier values become their own level, or be imputed to a known level?
• Which estimator will use the transformed features, and does it support sentinel values?
• Will preprocessing be assessed inside cross-validation, or only on a single train/test split?

Hint — Fit once on training data: All learned statistics, such as medians and category mappings, come from the training fit and are applied unchanged to validation and test data.

What a Strong Answer Covers

• Separate numeric and categorical pipelines that are fit only on training features.
• Explicit ordinal categories plus a deliberate encoding for unknown values.
• Correct use of fit_transform on training data and transform on all other data.
• Stable column names, ordering, dtypes, and row alignment.
• Cross-validation implications and serialization of the fitted pipeline.

Follow-up Questions

• What kind of leakage happens if the imputer is fit before the cross-validation folds are created?
• When would one-hot encoding be safer than ordinal encoding?
• How would you track a rising rate of unseen service tiers after deployment?

Overview: Describe leakage-safe imputation and ordinal encoding for training and test data. Cover fit versus transform, business-defined category ordering, unseen levels, cross-validation, and stable output schemas.