Introduction

In scikit-learn, transformer parameters inside a Pipeline are tuned the same way estimator parameters are tuned: by naming them with the pipeline step name, two underscores, and the parameter name. Once you understand that naming rule, grid-searching over preprocessing options becomes straightforward.

The main benefit of doing this inside a pipeline is correctness. During cross-validation, each fold fits the transformer only on the training split, which prevents data leakage from the validation split.

Use Step Names In param_grid

Suppose you have a pipeline with a scaler and a classifier:

1from sklearn.pipeline import Pipeline
2from sklearn.preprocessing import StandardScaler
3from sklearn.svm import SVC
4
5pipe = Pipeline([
6    ("scaler", StandardScaler()),
7    ("svc", SVC()),
8])

To tune transform arguments, prefix them with the step name:

1param_grid = {
2    "scaler__with_mean": [True, False],
3    "scaler__with_std": [True, False],
4    "svc__C": [0.1, 1, 10],
5}

Then pass the pipeline and the grid to GridSearchCV:

1from sklearn.model_selection import GridSearchCV
2
3grid = GridSearchCV(pipe, param_grid=param_grid, cv=5)
4grid.fit(X_train, y_train)
5
6print(grid.best_params_)

That is the standard pattern. Transformer parameters and estimator parameters are treated uniformly.

Tuning Different Transformer Choices

You can also grid-search over different transformer objects entirely by replacing the step value itself:

1from sklearn.decomposition import PCA
2from sklearn.preprocessing import MinMaxScaler
3
4param_grid = {
5    "scaler": [StandardScaler(), MinMaxScaler()],
6    "svc__C": [0.1, 1, 10],
7}

This is useful when the preprocessing strategy is part of the model design rather than just a fixed preparation step.

Why Pipelines Matter For Search

Without a pipeline, it is easy to fit a transformer on the full dataset before cross-validation, which leaks information. A pipeline prevents that mistake because each cross-validation split fits its own transformer instance internally.

That is why hyperparameter tuning over transform arguments belongs inside the pipeline definition instead of in ad hoc preprocessing code around it.

Searching Over Transformer Choice

Grid search is not limited to numeric settings such as with_mean=True or C=1.0. You can also search over which transformer is used in a given step, or even skip a step entirely by replacing it with passthrough. That makes the pipeline itself part of model selection instead of treating preprocessing as fixed.

Pipeline Search Is Still Normal Model Selection

Even though transformer arguments live earlier in the pipeline, they are still part of the predictive model. Scaling choices, feature selection settings, and dimensionality reduction parameters all change the hypothesis you are evaluating, so it is correct to include them in the same cross-validated search as the estimator hyperparameters.

passthrough Is Searchable Too

A preprocessing step can also be disabled through the parameter grid when that is part of the experiment design.

It also keeps preprocessing choices visible in the final best-parameter report.

That is useful when preprocessing is part of the experiment rather than a fixed assumption.

Common Pitfalls

• Forgetting the step__parameter naming convention in param_grid.
• Preprocessing outside the pipeline and introducing data leakage.
• Using the wrong step name after renaming a pipeline stage.
• Trying to tune parameters on an object that is not part of the pipeline.
• Searching too many preprocessing combinations without considering runtime cost.

Summary

• Tune transformer parameters in a pipeline with step__parameter names.
• 'GridSearchCV treats transformer and estimator parameters the same way.'
• Pipelines protect cross-validation from preprocessing leakage.
• You can search over parameter values or even over different transformer objects.
• Keep preprocessing inside the pipeline when it is part of model selection.