In scikit-learn, how does ColumnTransformer apply transforms per column, and what happens to unlisted columns?

## The problem it solves Real tabular data is heterogeneous: numeric columns want scaling, categorical columns want encoding, a text column wants vectorising, and some columns want nothing at all. A single transformer cannot express that, and doing it by hand in pandas before the split is the most common source of leakage in a project. `ColumnTransformer` is the container that applies different transformers to different column subsets as one fitted object, so the whole thing can sit inside a pipeline and be fitted once on training data. ## How it is constructed The first argument is a list of `(name, transformer, columns)` triples. - `name` is a string used for parameter addressing and output-name prefixes; it must be unique. - `transformer` is any transformer object, or the string `'drop'` to discard those columns explicitly, or `'passthrough'` to keep them untouched. - `columns` may be a list of column names, a list of integer positions, a slice, a boolean mask, a single string, or a callable. `make_column_selector(dtype_include=...)` is the callable scikit-learn ships for selecting by dtype, which is how you say 'all numeric columns' without enumerating them. `make_column_transformer(...)` is the shorthand constructor that generates the names for you from the transformer class names. ## The two behaviours that catch people **Unlisted columns vanish.** `remainder='drop'` is the default. Add a column to your data, forget to add it to the transformer list, and it is simply gone from the model's input — no error, no warning, just a quieter model. `remainder='passthrough'` keeps them, appended after all the transformed blocks. You can also pass an estimator (for example `SimpleImputer()`) as `remainder` to apply one transform to everything not otherwise claimed. **Output column order changes.** The blocks are concatenated in the order the transformers are declared, each block internally ordered as that transformer emits. So a DataFrame with `age, city, income` transformed by a list that puts the `OneHotEncoder` on `city` first produces the city indicators before the scaled numeric columns. If any downstream code indexes the output by position — plotting coefficients, slicing a feature block — it must derive positions from `get_feature_names_out()` rather than from the input frame. ## Names, output types and inspection `get_feature_names_out()` returns the full output names. `verbose_feature_names_out=True` (the default) prefixes each with its transformer name and a double underscore, giving `num__age`, `cat__city_Berlin`; set it to `False` for bare names, which raises an error if that produces duplicates. `set_output(transform='pandas')` makes the container and its children return DataFrames, which is the single biggest quality-of-life improvement for debugging a preprocessing stack; `'polars'` is also supported. The output container type is decided by `sparse_threshold` (default `0.3`): if the fraction of sparse columns in the concatenated result exceeds it, you get a sparse matrix, otherwise everything is densified. That is why adding one wide `OneHotEncoder` block can flip your whole output from a NumPy array to a CSR matrix, and code downstream that calls array-only methods then breaks. After fitting, `named_transformers_` gives dictionary access to the fitted children (`ct.named_transformers_['cat'].categories_`), and `transformers_` is the resolved list including the remainder entry. ## Fitting semantics Each transformer sees only its own columns, and they are independent — `n_jobs` can fit them in parallel. Fitting happens on whatever data is passed to `ColumnTransformer.fit()`, which is why the container belongs inside a pipeline: then each cross-validation fold fits the encoders and scalers on that fold's training part only, and a category or a mean from the validation part cannot reach the model. ## How to answer Describe the triple, state plainly that unlisted columns are dropped by default, and mention that output order follows the transformer list. If you add `make_column_selector` for dtype-based selection and `get_feature_names_out()`/`set_output` for keeping the result interpretable, you sound like someone who has maintained one of these rather than copied it from a tutorial.