PyCaret
GitHub
Preprocessing

Scale and transform

Normalize numeric features and apply a power transform.

Two flags on the Experiment constructor control numeric scaling and distribution-shape correction. Both are off by default — turn them on when your data calls for them.

normalize=True — StandardScaler#

Z-scores every numeric column on the training set, then applies the same transform to the test set. Equivalent to:

from sklearn.preprocessing import StandardScaler
StandardScaler().fit(X_train).transform(X_test)

In PyCaret:

from pycaret.classification import ClassificationExperiment

exp = ClassificationExperiment(
    target="Purchase",
    normalize=True,        # adds StandardScaler to the numeric pipeline
).fit(data)

When to turn it on:

  • Distance-based models (KNN, SVM with RBF kernel, neural nets, K-Means) need scaled inputs to behave.
  • Linear models with regularization (Ridge, Lasso) penalize each feature's coefficient equally — unscaled features make the regularization unfair.

When to leave it off:

  • Tree-based models (RandomForest, GBM, LightGBM, XGBoost) are scale-invariant.

transformation=True — PowerTransformer (Yeo-Johnson)#

Applies a Yeo-Johnson transform to make the marginal distribution of each numeric column more Gaussian-like. Unlike Box-Cox, Yeo-Johnson handles negative values, so it's safe to use without pre-checking your data.

exp = ClassificationExperiment(
    target="Purchase",
    transformation=True,   # adds PowerTransformer
).fit(data)

When to turn it on:

  • Skewed numeric features and a model that assumes Gaussianish inputs (linear, MLP, naive Bayes).
  • Features with heavy tails dragging the mean / variance.

Combining both#

transformation is applied before normalize in the numeric branch of the column transformer. The order is fixed:

Imputer → (PowerTransformer if transformation) → (StandardScaler if normalize)

So with both on:

exp = ClassificationExperiment(
    target="Purchase",
    transformation=True,
    normalize=True,
).fit(data)

…each numeric column ends up roughly Gaussian with mean 0 and unit variance.

Inspecting the pipeline#

The fitted preprocessing pipeline is on exp._fit_state:

print(exp._fit_state["preprocess_pipeline"])
# Pipeline(steps=[('preprocess',
#   ColumnTransformer(transformers=[
#     ('numerical_pipeline',
#       Pipeline([('imputer', SimpleImputer()),
#                 ('transformer', PowerTransformer(...)),
#                 ('scaler', StandardScaler())]),
#       ['Age', 'Income', ...]),
#     ('categorical_pipeline',
#       Pipeline([('imputer', SimpleImputer(strategy='most_frequent')),
#                 ('encoder', OrdinalEncoder(...))]),
#       ['Gender', 'Region', ...]),
#   ]))])

Persisting the experiment via save_model carries the preprocessor along with the trained model — predict-time data flows through the same chain you trained on.

What's removed in 4.0#

The 3.x flags normalize_method=, transformation_method=, pca=, pca_method=, pca_components= are gone. The defaults (zscore + yeo-johnson) are wired in, and PCA was deemed niche enough to live in the user's preprocessing code rather than as a PyCaret knob. If you need PCA, fit it on the train set and feed the projected frame to Experiment.fit() directly.