Ridge-penalized logistic regression by iteratively reweighted least squares
Source:R/nre.R
irls_logistic.Rdstats::glm.fit() would do this, but it warns and wanders off when the two
classes are separable, which a sharp ratio makes easy to hit. The ridge term
keeps the normal equations solvable and the coefficients finite, the same
role it plays in fit_linear_gaussian(), and it is measured against each
column's own scale for the same reason: under standardize = FALSE the
quadratic features carry the fourth power of the data's units, so an
absolute 1e-6 is either nothing at all or the only thing left. On a
simulator whose output has sd 5e-4 the absolute version shrank the fit to
noise; the relative one leaves it alone.