Preprocessing 0.2.0 powermethod

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 transforming

PowerMethod

Which power family PowerTransformer fits.

public enum PowerMethod

Values — YeoJohnson is Yeo and Johnson's family, which admits zero and negative values; scikit-learn's 'yeo-johnson', and the default. BoxCox is Box and Cox's, which needs strictly positive values; 'box-cox'.

Example — the two exponents on the same positive column.

using Lodestar.Preprocessing;

double[] income = [22.0, 25.0, 28.0, 31.0, 35.0, 42.0, 55.0, 78.0, 120.0, 260.0];

double yeoJohnson = Math.Round(PowerTransformer.Fit(income, 1).Lambdas[0], 4);   // => -0.8072

PowerTransformer fitted = PowerTransformer.Fit(
    income, 1, new PowerTransformerOptions { Method = PowerMethod.BoxCox });

double boxCox = Math.Round(fitted.Lambdas[0], 4);                               // => -0.7802

Remarks — the default is Yeo-Johnson because it always applies. Box-Cox refuses a column holding a zero or a negative value, and refusing is the right answer there: its family is (xᶫ − 1) / λ, which is not defined for them. Yeo-Johnson is built to extend it, applying the Box-Cox form to the positive side and a reflected one to the negative, so a column of temperature anomalies or profit-and-loss figures can be transformed at all.

On a strictly positive column the two are different fits, not the same fit twice. They agree on the direction and disagree on the number, as above, because the likelihoods being maximised are different functions — Yeo-Johnson's shifts each positive value by one before raising it. Prefer Box-Cox when the column is positive by construction and the exponent will be reported, since it is the family a reader will expect; prefer the default otherwise.

Applies to — net10.0, netstandard2.0.

See also — PowerTransformerOptions, PowerTransformer.Fit.