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 — Lodestar.Preprocessing

Four scalers, at sklearn.preprocessing parity: StandardScaler centres on the mean and scales to unit variance, MinMaxScaler maps onto a fixed range, MaxAbsScaler divides by the largest absolute value without ever subtracting, and RobustScaler centres on the median and scales by an interpercentile range.

Which one, in one line each. StandardScaler is the default for data without outliers; RobustScaler is the one for data with them, since a mean and a deviation both move with whatever is furthest away; MaxAbsScaler is the one for data whose zeros are meaningful, because centring turns every zero into a non-zero; MinMaxScaler is the one for a bounded range a downstream model asks for.

Spans in, arrays out. A sample matrix is row-major — FeatureCount values per row — which is the shape Lodestar.Metrics already uses, so a matrix does not have to be reshaped to cross between the two packages. There is no pipeline object to build, no data view to construct, and nothing to adopt beyond the call.

Why this exists when ML.NET has normalizers

ML.NET has every brick this package grows: NormalizeMeanVariance, NormalizeMinMax, OneHotEncoding, ReplaceMissingValues, TrainTestSplit, CrossValidationSplit. Read on Microsoft.ML 5.0.0's exported surface, each one of them is reached through an IDataView or through a catalog naming columns — NormalizeMeanVariance(TransformsCatalog, string inputColumn, string outputColumn, …) never sees a value, and TrainTestSplit(IDataView, double, …) takes and returns data views. The capability is not missing; the array-in, array-out entry point is.

That is the same shape as TfidfVectorizer against FeaturizeText, and the same reason: a caller who holds a double[] and wants a double[] back should not have to adopt a framework to get one.

Types

Type	What it is
StandardScaler	Centres and scales each feature, and reports the statistics it fitted.
StandardScalerOptions	Which of the two steps to apply.
MinMaxScaler	Maps each feature onto a fixed range.
MinMaxScalerOptions	The range, and whether the transform clips to it.
MaxAbsScaler	Divides each feature by its largest absolute value.
MaxAbsScalerOptions	Whether the transform clips to [−1, 1].
RobustScaler	Centres on the median, scales by an interpercentile range.
RobustScalerOptions	Which steps to apply, and between which percentiles.

See also

• scikit-learn → .NET — what is delegated and what is not.
• Python → C# equivalence.