When to use?

Scaling is important for distance-based algorithms like KNN and SVM.

When to use which?

General rules:

StandardScaler: the default choice; linear models, SVM, PCA, KNN.
MinMaxScaler: when a bounded range is needed (neural network inputs, pixels) and there are no outliers.
RobustScaler: when there are clear outliers.
None: not needed for tree-based models.

If unsure, compare a few with cross-validation; the choice is a hyperparameter too.

The leakage rule

The scaler is fit only on the training data. Fitting on all data leaks the test statistics into training and artificially inflates the score:

# WRONG
sc = StandardScaler().fit(X)              # includes the test
# RIGHT
sc = StandardScaler().fit(X_train)
X_train_s = sc.transform(X_train)
X_test_s = sc.transform(X_test)

In test data the min/max may exceed the training ones; seeing values outside 0-1 after MinMax is normal.

Exercise Out-of-range test value

Fit a MinMaxScaler on X_train only. Transform X_test and print it; see that the value outside the training range comes out above 1.

Show example solution

import numpy as np
from sklearn.preprocessing import MinMaxScaler

X_train = np.array([[10.0], [20.0], [30.0]])
X_test = np.array([[25.0], [40.0]])

mm = MinMaxScaler().fit(X_train)
print(mm.transform(X_test))

Test yourself

If income data includes a few billionaires, which scaler fits?
- RobustScaler
- MinMaxScaler
- No scaling needed
- StandardScaler always
Show answer

RobustScaler — The median and IQR aren't affected by outliers.

Learn the scaler on the training data only.

sc = StandardScaler().fit(____)

Show answer

sc = StandardScaler().fit(X_train)

Match the situations with the choices.

Default

StandardScaler
Pixels 0-255

MinMaxScaler
Outliers

RobustScaler
Random forest

No scaling needed

Next lesson: Pipeline

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies