Scikit-Learn · Class

MaxAbsScaler

MaxAbsScaler scales each feature by its largest absolute training value. It preserves zero-centered structure and sparsity because it does not subtract feature means.

How it works

MaxAbsScaler learns the largest absolute value in each feature during fit and uses those values to scale later data. It does not subtract feature means, so zero values remain zero; this also lets it preserve sparsity for supported sparse inputs.

Syntax

class sklearn.preprocessing.MaxAbsScaler(*, copy=True, clip=False)

Parameters

Name	Type	Description
`copy`	`bool, default=True`	Whether to copy the input before transforming; default is True.
`clip`	`bool, default=False`	Whether to clip transformed values to the fitted range; default is False.

Returns

Transformed data with the same shape as the input; the class page documents transform as returning an array or sparse matrix.

Key points

Each feature receives its own scale based on the largest absolute training value.
After fitting and transforming the training set, every feature's maximum magnitude is 1.
The transformation keeps the original zero locations because no mean is subtracted.
CSR and CSC sparse matrices are supported inputs.
fit learns the values needed by subsequent transform calls.
The estimator performs feature-wise, rather than global, normalization.

Common mistakes

It does not make outliers less influential; an extreme value still determines the feature's scale.
Do not use it when mean-centering is required; its design deliberately leaves feature offsets unchanged.
Sparse support is specific to CSR and CSC matrix formats documented on the class page.

Example: Scale centered data

from sklearn.preprocessing import MaxAbsScaler

X = [[1., -1., 2.], [2., 0., 0.], [0., 1., -1.]]
scaler = MaxAbsScaler().fit(X)
X_scaled = scaler.transform(X)
print(X_scaled)

Fits per-feature maximum absolute values, then transforms the data by dividing each feature by its fitted maximum.

When to use it

Zero-centered numeric data

Features are already centered around zero and need comparable magnitudes without mean subtraction.

Sparse feature matrices

A dataset is stored as CSR or CSC and scaling must retain sparse structure.

Train/test preprocessing

A preprocessing pipeline needs reusable feature scales learned from training data.

Interview questions

QHow does MaxAbsScaler transform features?

It divides each feature by that feature's maximum absolute training value, making the largest magnitude equal to 1. It does not center the features.

Follow-up: Why can it preserve sparse structure? Because it scales existing values without subtracting a mean, so stored zero entries remain zero.

QHow should it be used in a train/test workflow?

Call fit or fit_transform on training data, then call transform on new data. The learned per-feature scales are reused for later inputs.

Follow-up: What leakage risk must be avoided? Fit only on the training partition; fitting before the split lets test information influence the learned scales.

QWhen is MaxAbsScaler a suitable choice?

It supports sparse CSR and CSC matrices and is intended for data already centered at zero or sparse data. Unlike a centering transformer, it does not destroy sparsity.

Follow-up: When might another scaler be preferable? Use one with centering or robust statistics when zero-centered scaling is inappropriate or outliers need stronger treatment.

sklearn.preprocessing.maxabs_scale
sklearn.preprocessing.MinMaxScaler
sklearn.preprocessing.StandardScaler
sklearn.preprocessing.RobustScaler

Source: MaxAbsScaler — scikit-learn 1.9.1 documentation. Checked 2026-10-15.