You receive training and test tables with numeric columns, an ordered categorical column `service_tier` whose business order is `bronze < silver < gold`, and a binary training target. Some feature values are missing, and the test set may contain a tier not observed in training. Explain how to produce `processed_train` and `processed_test` using imputation and ordinal encoding without leaking information from the test set. Address when to use `fit_transform` versus `transform`, how to preserve feature alignment, and how to handle unknown categories. ### Constraints & Assumptions - The target is available only in the training table and must never enter a feature transformer. - Numeric and categorical columns may use different imputation strategies. - Both processed outputs must contain the same feature columns in the same order. - Unknown or missing tiers must not be silently promoted to a higher business tier. ### Clarifying Questions to Ask - Is the tier order fixed by the business or inferred from observed labels? - Should missing tier values form a separate level or be imputed to a known level? - Which estimator will consume the transformed features, and can it handle sentinel values? - Is preprocessing evaluated inside cross-validation or only on one train/test split? ```hint Fit once on training data Every learned statistic, including medians and category mappings, belongs to the training fit and is reused unchanged on validation and test data. ``` ### What a Strong Answer Covers - Separate numeric and categorical pipelines fitted only on training features. - Explicit ordinal categories and a deliberate unknown-value encoding. - Correct use of `fit_transform` on training data and `transform` elsewhere. - Stable column names, ordering, dtypes, and row alignment. - Cross-validation implications and serialization of the fitted pipeline. ### Follow-up Questions - What leakage occurs if the imputer is fit before cross-validation folds are created? - When would one-hot encoding be safer than ordinal encoding? - How would you monitor a rising rate of unseen service tiers after deployment?