1.2.4 Results🔗ℹ

A fit returns a transparent struct. The fields follow one pattern across the families:

• intercept — a real, or a vector of one per class or response (multinomial-result-intercepts, mgaussian-result-intercepts). Cox results have none. As in R, the multinomial intercepts are centred to sum to zero; adding the same constant to each would not change any probability.

• coefficients — a dense vector with one entry per predictor, on the original (unstandardized) scale; a predictor the penalty dropped is exactly 0.0. Multinomial and multi-response results hold a vector of such vectors, one per class or response.

• r-squared or dev-ratio — the fraction of the null deviance explained. For the Gaussian families this is R².

• lambda — the penalty the solver used.

• num-passes — the number of coordinate-descent passes.

A result prints as a one-line summary: its family, λ, deviance ratio, and how many of the predictors have a nonzero coefficient. The fields are read with the struct’s accessors:

> (define fit (ridge X y #:lambda 0.1))

> fit

#<glmnet:gaussian λ=0.1 dev=0.9619 nz=2/2>

> (elnet-result-r-squared fit)

0.9618813537709814

Because the structs are transparent, equal? compares them field by field and match destructures them:

> (require racket/match)

> (match-define (elnet-result a0 beta _ _ _) fit)

> (list a0 beta)

'(1.2148264276558987 #(1.5916761644299766 -0.6843296627955713))