Scikit-learn Pipelines and Preprocessing Cheat Sheet

Updated 2026-05-15

🧠Study flashcards on this topic55 cards · spaced repetition→

Scikit-learn pipelines are workflow tools that chain preprocessing transformers and estimators into a single composable object. Located in sklearn.pipeline and sklearn.compose, they ensure reproducible data transformations, prevent data leakage during cross-validation, and streamline hyperparameter tuning. Pipelines enforce that each transformation step learned from training data (scaling means, encoding categories) is applied identically to validation and test folds. Key mental model: think of pipelines as assembly lines where each station (transformer) modifies the data in a consistent, repeatable way — transformers never see test data during fit, only during transform.

What This Cheat Sheet Covers

This topic spans 13 focused tables and 58 indexed concepts, 55 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core Pipeline ClassesTable 2: Numerical Scaling TransformersTable 3: Categorical Encoding TransformersTable 4: Missing Data ImputationTable 5: Feature Selection within PipelinesTable 6: Advanced Feature EngineeringTable 7: Custom TransformersTable 8: Hyperparameter Tuning with PipelinesTable 9: Cross-Validation with PipelinesTable 10: Pipeline Introspection and UtilitiesTable 11: ColumnTransformer Advanced FeaturesTable 12: Memory Caching with JoblibTable 13: Common Pipeline Patterns

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Pipeline Classes

These are the building blocks you compose everything else from. Pipeline chains steps in sequence, ColumnTransformer routes different columns to different transformers, and FeatureUnion runs transformers in parallel and stitches their outputs together — with make_* shortcuts that auto-name steps so you can skip the boilerplate tuples. Master these five and the rest of the library snaps into place around them.

Class	Example	Description
Pipeline	`Pipeline([('scaler', StandardScaler()), ('clf', LogisticRegression())])`	• Chains transformers sequentially with an optional final estimator • calls `fit_transform` on each step except last
make_pipeline	`make_pipeline(StandardScaler(), LogisticRegression())`	Convenience constructor that auto-generates step names (`'standardscaler'`, `'logisticregression'`) instead of requiring tuples
ColumnTransformer	`ColumnTransformer([('num', StandardScaler(), ['age', 'income']), ('cat', OneHotEncoder(), ['city'])])`	• Applies different transformers to different column subsets • concatenates results horizontally into single feature matrix
make_column_transformer	`make_column_transformer((StandardScaler(), ['age', 'income']), (OneHotEncoder(), ['city']))`	Convenience constructor for `ColumnTransformer` that omits need to name each transformer tuple
FeatureUnion	`FeatureUnion([('pca', PCA(n_components=2)), ('poly', PolynomialFeatures(degree=2))])`	• Applies multiple transformers in parallel to the same input data • concatenates outputs horizontally
FunctionTransformer	`FunctionTransformer(np.log1p, inverse_func=np.expm1)`	Wraps arbitrary Python function as a transformer with optional `inverse_transform` support for reversible operations

Table 2: Numerical Scaling Transformers

Many algorithms — anything distance- or gradient-based — behave badly when features live on wildly different scales, so putting them on common ground is usually step one. The choice between these comes down to your data: StandardScaler assumes a roughly Gaussian spread, RobustScaler shrugs off outliers using the median and IQR, MinMaxScaler squeezes into a fixed range, and the power and quantile transformers reshape the distribution itself toward normality.

Transformer	Example	Description
StandardScaler	`StandardScaler().fit_transform(X)`	• Centers features to zero mean and scales to unit variance • sensitive to outliers • assumes Gaussian-like distribution

The rest of this section is part of the full cheat sheet

Table 3: Categorical Encoding Transformers

Models speak numbers, so text categories have to be translated first — and the right translation depends on the structure. One-hot encoding gives each category its own column and suits unordered values; ordinal encoding maps to integers when there's a natural ranking or you're feeding a tree; target encoding folds in the mean label to tame high-cardinality columns. The Label* variants are the dedicated tools for transforming the y target rather than the features.

Transformer	Example	Description
OneHotEncoder	`OneHotEncoder(sparse_output=False, handle_unknown='ignore').fit_transform(X)`	• Converts categories to binary columns (one-hot vectors) • each category becomes a separate feature • `handle_unknown='ignore'` prevents errors on unseen values

The rest of this section is part of the full cheat sheet

Table 4: Missing Data Imputation

Real datasets have holes, and most estimators refuse to train until you fill them. SimpleImputer is the fast baseline that plugs in a mean or most-frequent value, while IterativeImputer and KNNImputer use the relationships between columns or nearby samples for smarter, multivariate fills. When the very fact that a value was missing carries signal, the indicator options preserve that missingness as its own feature.

Transformer	Example	Description
SimpleImputer	`SimpleImputer(strategy='mean').fit_transform(X)`	• Fills missing values with mean, median, most_frequent, or constant • univariate (per-column) imputation • default for quick baseline

The rest of this section is part of the full cheat sheet

Table 1: Core Pipeline Classes

Class	Example	Description
Pipeline	`Pipeline([('scaler', StandardScaler()), ('clf', LogisticRegression())])`	• Chains transformers sequentially with an optional final estimator • calls `fit_transform` on each step except last
make_pipeline	`make_pipeline(StandardScaler(), LogisticRegression())`	Convenience constructor that auto-generates step names (`'standardscaler'`, `'logisticregression'`) instead of requiring tuples
ColumnTransformer	`ColumnTransformer([('num', StandardScaler(), ['age', 'income']), ('cat', OneHotEncoder(), ['city'])])`	• Applies different transformers to different column subsets • concatenates results horizontally into single feature matrix
make_column_transformer	`make_column_transformer((StandardScaler(), ['age', 'income']), (OneHotEncoder(), ['city']))`	Convenience constructor for `ColumnTransformer` that omits need to name each transformer tuple
FeatureUnion	`FeatureUnion([('pca', PCA(n_components=2)), ('poly', PolynomialFeatures(degree=2))])`	• Applies multiple transformers in parallel to the same input data • concatenates outputs horizontally
FunctionTransformer	`FunctionTransformer(np.log1p, inverse_func=np.expm1)`	Wraps arbitrary Python function as a transformer with optional `inverse_transform` support for reversible operations

Table 2: Numerical Scaling Transformers

Transformer	Example	Description
StandardScaler	`StandardScaler().fit_transform(X)`	• Centers features to zero mean and scales to unit variance • sensitive to outliers • assumes Gaussian-like distribution

Table 3: Categorical Encoding Transformers

Transformer	Example	Description
OneHotEncoder	`OneHotEncoder(sparse_output=False, handle_unknown='ignore').fit_transform(X)`	• Converts categories to binary columns (one-hot vectors) • each category becomes a separate feature • `handle_unknown='ignore'` prevents errors on unseen values

Table 4: Missing Data Imputation

Transformer	Example	Description
SimpleImputer	`SimpleImputer(strategy='mean').fit_transform(X)`	• Fills missing values with mean, median, most_frequent, or constant • univariate (per-column) imputation • default for quick baseline