Feature engineering toolkit

What this project does

This small teaching project demonstrates how to build a simple, reviewable feature-engineering toolkit for a mixed-type tabular classification task. It emphasises practical steps: defining which raw fields are available at prediction time, composing imputers and encoders into a ColumnTransformer + Pipeline, preventing leakage from future-only fields, and comparing the engineered pipeline to a raw-feature baseline. The transformation is packaged with a version string and basic input checks so it can be reviewed for deployment.

Data

The dataset is synthetic and created inside the code. It contains numeric features, categorical features, and a datetime field. A deliberately injected column called “future_flag” simulates a future-only signal (available during training but not at prediction) so the project can demonstrate leakage checks. The synthetic target is binary.

How to build it

1. Create the mixed-type DataFrame in project.py using numpy and pandas, documenting which columns are allowed at prediction time.
2. Compose preprocessing: numeric imputer + standard scaler; categorical imputer + one-hot encoder (handle_unknown=’ignore’); custom datetime features (day-of-week, recentness).
3. Fit a classifier pipeline (RandomForestClassifier) and a simple baseline pipeline using only basic numeric features; evaluate on a held-out test split with fixed random_state.
4. Package the transform into a small wrapper class with a version string and an input-check method; include small example fixtures demonstrating expected inputs.

How to run

Run the module as a script. It will print a JSON object with evaluation metrics for the engineered pipeline and the baseline.

python project.py

How to test

Run pytest in the project directory. Tests verify the returned metric keys, value ranges, a conservative performance floor on the held-out accuracy, and determinism for a fixed seed.

Results

Measured by running python project.py (seed 42) on 2026-10-23; the tests in test_project.py assert the floor.

Metric	Value
version	v2
accuracy	0.8000
roc_auc	0.8588
accuracy_baseline	0.7080
n_train	750
n_test	250

Also returned (see python project.py output): example_input_rows (3 items).

Limitations

This is a small educational example using synthetic data. The synthetic patterns are simple and chosen so that the models learn reliably; real datasets will have more complex distributional shifts, require calibration, and need careful monitoring. The toolkit demonstrates common transformations but is not a full production feature-store library; it lacks persistent serialization, schema migration tooling and runtime monitoring.

Next steps

• Add persistence (safe on-disk serialization) and schema migration checks before deployment.
• Extend the datetime transformer to support time-zone aware logic and rolling-window aggregates.
• Add more extensive unit tests covering a broader variety of missing-column and unseen-category scenarios.

---

Full case study, build steps and evaluation checks: Feature engineering toolkit on networkershome.com

Built by Networkers Home · Learn the skills in the AI Engineer Course · Part of 100 AI projects with code