FeatureEngineer workflows¶

About¶

Feature preparation becomes part of the model pipeline. Define derived columns once, fit the raw-input and output-schema contract, and let prediction replay the same definitions before the estimator receives its features.

Why use it?¶

Use it when training and serving must share ordered, deterministic feature logic. It is especially useful for dependent features: calculate an intermediate column in one specification, consume it in the next, and drop it before the model boundary.

Examples¶

Input¶

Code¶

import polars as pl

from polars_learn import FeatureEngineer, FeatureSpec

raw = pl.DataFrame({"price": [10.0, 20.0], "quantity": [2, 3]})
engineer = FeatureEngineer(
    [
        FeatureSpec(pl.col("price") * pl.col("quantity"), alias="notional"),
        FeatureSpec(
            pl.col("notional").log1p(), alias="log_notional", drop=["quantity"]
        ),
    ]
)
prepared = engineer.fit_transform(raw)
assert prepared.columns == ["price", "notional", "log_notional"]
assert prepared["notional"].to_list() == [20.0, 60.0]
assert prepared["log_notional"].equals(pl.Series("log_notional", [20.0, 60.0]).log1p())
assert raw.columns == ["price", "quantity"]

Output¶

The assertions in the example verify the output against the implementation. Python examples are executed during documentation validation.

Practical notes¶

Put the engineer before the predictive estimator and pass raw data to both pipeline.fit and pipeline.predict. It records schemas rather than learned feature statistics; use fitted statistical preprocessing for training-derived means, scales, or vocabularies. Fit resolves schemas without evaluating rows; eager fit_transform executes feature expressions once during transform. See the complete prediction pipeline.