MaxAbsScaler

Scales each feature by its maximum absolute value, mapping all features to the range $[-1, 1]$. Unlike StandardScaler and MinMaxScaler, MaxAbsScaler does not shift the data (no centering), so it preserves sparsity.

Formula

$$Z_{ij} = \frac{X_{ij}}{\max_i |X_{ij}|}$$

If $\max_i |X_{ij}| = 0$ for a feature (all zeros), the feature is left unchanged.

When to Use

• Sparse data: The only scaler (besides Normalizer) that preserves zero entries, making it ideal for sparse matrices.
• Centered data: When the data is already centered and only scaling is needed.
• Preserves sign: The sign of each entry is unchanged, unlike MinMaxScaler which shifts values.

Mirrors sklearn.preprocessing.MaxAbsScaler.

Constructor

Skigen::MaxAbsScaler<Scalar> scaler(bool clip = false);

Methods

Method	Description
fit(X)	Compute per-feature max absolute value
transform(X)	Scale to [-1, 1]
inverse_transform(Z)	Recover original scale
transform_inplace(X)	Scale in-place

Fitted Attributes

Accessor	Type	Description
max_abs()	RowVectorType	Per-feature max absolute value
scale()	RowVectorType	Per-feature scale factor

Example

#include <Skigen/Preprocessing>

Skigen::MaxAbsScaler scaler;
auto X_scaled = scaler.fit_transform(X);