Preprocessing minmaxscaler fit

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

MinMaxScaler.Fit

Fits a scaler on a row-major sample matrix.

public static MinMaxScaler Fit(ReadOnlySpan<double> samples, int featureCount, MinMaxScalerOptions options = null)

Parameters — samples is the sample matrix, row-major: featureCount values per row. featureCount is how many values each row carries. options chooses the range and whether Transform clips to it; null is [0, 1] without clipping.

Returns — a fitted MinMaxScaler.

Exceptions — ArgumentOutOfRangeException when featureCount is not positive, or the range is not two finite bounds with the low one below the high one. ArgumentException when samples holds no row, a partial one, or a non-finite value.

Example — a range of 1.11e-15, which is not zero and is still treated as constant.

using Lodestar.Preprocessing;

// Three values a quadrillionth apart: the range is 1.11e-15, below 10 * eps.
double[] nearlyConstant = [1.0, 1.0 + 1e-15, 1.0];

MinMaxScaler scaler = MinMaxScaler.Fit(nearlyConstant, featureCount: 1);

double range = scaler.DataRange[0];  // => 1.1102230246251565E-15
double scale = scaler.Scale[0];      // => 1

Remarks — the near-constant test is range < 10·eps, not range == 0. That is sklearn.preprocessing._data._handle_zeros_in_scale with no constant mask, which is how the reference calls it for this scaler, for MaxAbsScaler and for RobustScaler — StandardScaler is the one that passes a mask, and its rule is the two-pass variance bound instead. Move the example one decade out, to 1.0 + 4e-15, and the same shape is above the threshold and scales by about 2.5e14.

DataRange reports what was seen and Scale what is divided by, so the two disagree exactly on a feature the floor caught.

Non-finite input is refused, where the reference skips a NaN. Refusing rather than propagating because RobustScaler sorts, and a NaN in a sorted column comes back as a percentile nobody asked for; the three scalers answer alike.

Applies to — net10.0, netstandard2.0.

See also — MinMaxScaler, MinMaxScaler.Transform.