MinMaxScaler

Linearly scales each feature to a specified range $[\text{min},\, \text{max}]$ (default $[0, 1]$), based on the per-feature minimum and maximum observed during fitting.

Formula

$$Z_{ij} = \frac{X_{ij} - X_j^{\min}}{X_j^{\max} - X_j^{\min}} \cdot (\text{max} - \text{min}) + \text{min}$$

If $X_j^{\max} = X_j^{\min}$ for a feature (zero range), the feature is set to $\text{min}$.

When to Use

• Bounded outputs: When the algorithm requires inputs in a fixed range (e.g., neural networks with sigmoid activations).
• Preserves zeros: Unlike StandardScaler, MinMaxScaler to $[0, 1]$ maps zero to a non-negative value, but does not center the data.
• Sensitive to outliers: A single extreme value can compress the rest of the data into a narrow range. For robustness, consider RobustScaler.

Mirrors sklearn.preprocessing.MinMaxScaler.

Constructor

Skigen::MinMaxScaler<Scalar> scaler({min, max}, bool clip = false);

Parameter	Default	Description
feature_range	{0, 1}	Target range
clip	false	Clip to range after transform

Methods

Method	Description
fit(X)	Compute per-feature min and max
transform(X)	Scale to target range
fit_transform(X)	Fit and transform in one call
inverse_transform(Z)	Recover original scale
transform_inplace(X)	Scale in-place
inverse_transform_inplace(Z)	Recover in-place

Fitted Attributes

Accessor	Type	Description
data_min()	RowVectorType	Per-feature minimum
data_max()	RowVectorType	Per-feature maximum
scale()	RowVectorType	Per-feature scale factor

Example

#include <Skigen/Preprocessing>

Skigen::MinMaxScaler scaler({0, 1});
scaler.fit(X);
auto X_scaled = scaler.transform(X);
// All values now in [0, 1]