Preprocessing 0.2.0 robustscaler

Lodestar.Preprocessing 0.2.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Home › Feature scaling

RobustScaler

Centres each feature on its median and scales it by an interpercentile range.

public sealed class RobustScaler

Properties — FeatureCount and SampleCount are the shape it was fitted on. Centre is each feature's median and Scale the range it divides by, each nullable and null exactly when its own step is off — a simpler mapping than StandardScaler's, where turning centring off still fits a mean.

Example — one column with an outlier three decades out.

using Lodestar.Preprocessing;

double[] samples = [1.0, 2.0, 3.0, 4.0, 5000.0];

RobustScaler scaler = RobustScaler.Fit(samples, featureCount: 1);

double centre = scaler.Centre![0];  // => 3
double scale = scaler.Scale![0];    // => 2

double outlier = scaler.Transform(samples)[4];  // => 2498.5

Remarks — this is the scaler for data with outliers. The median and the quartile range above move with the bulk of the column: the same five values give StandardScaler a mean of 1002 and a standard deviation of about 1999, so every ordinary value lands near −0.5 and the shape of the data is lost. Here they land at −1, −0.5, 0 and 0.5, and the outlier is visibly an outlier at 2498.5.

The percentiles interpolate linearly, numpy.percentile's default — see RobustScaler.Fit, which is also where the convention is compared against the two other quantiles this repository ships.

Applies to — net10.0, netstandard2.0.

See also — RobustScalerOptions, the feature scaling index, the Python equivalence table.

Members

Member	What it does
RobustScaler.Fit	Fits a scaler on a row-major sample matrix.
RobustScaler.InverseTransform	Undoes Transform.
RobustScaler.Transform	Centres and scales a matrix.