Build the pipeline

Fit preprocessing on training data, every time.

12 minutes. Mission: make mixed tabular data reproducible.

One object owns every learned step

from sklearn.compose import make_column_transformer
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import OneHotEncoder
from sklearn.pipeline import make_pipeline
from sklearn.linear_model import LogisticRegression

prep = make_column_transformer(
    (SimpleImputer(strategy="median"), numeric_columns),
    (OneHotEncoder(handle_unknown="ignore"), categorical_columns),
)
model = make_pipeline(prep, LogisticRegression(max_iter=1000))

A pipeline fits transforms only on each training partition, then applies them unchanged to held-out rows. This prevents preprocessing leakage and keeps training and prediction steps together.^[1]

Why the details matter

Median imputation supplies missing numeric values. One-hot encoding turns categories into indicator columns. handle_unknown="ignore" allows a category unseen during training without crashing prediction.

Check retrieval

Win: you can keep preprocessing and model fitting inside one cross-validatable object.

Primary source

Study scikit-learn’s mixed-types ColumnTransformer example.

PreviousNext: compare honestly