Development build. This page describes main, not a released package. The latest published Lodestar.Preprocessing is 0.1.0 — read its documentation.

Home › Preprocessing › Feature scaling

Fits a scaler on a row-major sample matrix.

public static MaxAbsScaler Fit(ReadOnlySpan<double> samples, int featureCount, MaxAbsScalerOptions options = null)
public static MaxAbsScaler Fit(CsrMatrix samples, MaxAbsScalerOptions options = null)

Parameters — samples is the sample matrix, row-major: featureCount values per row. featureCount is how many values each row carries. options chooses whether Transform clips to [−1, 1]; null does not.

Returns — a fitted MaxAbsScaler.

Exceptions — ArgumentNullException when the sparse overload is given no matrix. ArgumentOutOfRangeException when featureCount is not positive. ArgumentException when samples holds no row, a partial one, or a non-finite value.

Example — a feature that is all zeros divides by 1 rather than by 0.

using Lodestar.Preprocessing;

double[] zeros = [0.0, 0.0, 0.0];

MaxAbsScaler scaler = MaxAbsScaler.Fit(zeros, featureCount: 1);

double maximum = scaler.MaximumAbsolute[0];  // => 0
double divisor = scaler.Scale[0];            // => 1

Remarks — the floor is max < 10·eps rather than max == 0, the rule MinMaxScaler.Fit states with its measurement: a feature whose largest absolute value is a few quadrillionths is constant to within what a division could resolve, and dividing by it would amplify noise rather than reveal signal.

MaximumAbsolute reports what was seen and Scale what is divided by, so the two disagree exactly on the features the floor caught — the pair above is the smallest case of that.

The sparse overload takes a CsrMatrix, and this is the scaler that suits one. It never subtracts, so a zero stays a zero, and Transform's sparse overload returns the shape that went in. A column with no stored value at all has a maximum absolute of 0 and a scale of 1, the floor reached a different way.

Applies to — net10.0, netstandard2.0.

See also — MaxAbsScaler, MaxAbsScaler.Transform.