glmnet

Lasso, ridge and elastic-net regularized models for Racket, in R glmnet's six families: Gaussian, binomial, multinomial, Poisson, Cox and multi-response Gaussian. The solver is R glmnet's own Fortran.

Install

raco pkg install glmnet

Requires Racket 9.3 or later.

Prebuilt native libraries exist for Linux x86-64 and macOS arm64 only: the package's own libglmnetcompat, and the library that its dependency rkt-polars ships. On Intel macOS the install stages the arm64 libraries, so it completes but they fail to load; on any other platform the install fails. CI tests Racket 9.3 under Nix, and the current stable release by installing from the sources, as the package catalog does, on Ubuntu 22.04, the latest Ubuntu and macOS arm64.

Example

Fuel economy from weight and horsepower, in R's mtcars: fit, read the coefficients and predict a new car; then a lasso fit, and the same lasso from a formula.

(require glmnet glmnet/datasets)
(define fit (ols mtcars "mpg" #:predictors '("wt" "hp")))
(coef fit)
(predict fit '((wt 3.0) (hp 150)))
(define lasso-fit (lasso mtcars "mpg" #:predictors '("wt" "hp") #:lambda 4.0))
(define named-fit (formula-fit (~ mpg (+ wt hp)) mtcars #:lambda 4.0))
(equal? (predict named-fit mtcars) (predict lasso-fit mtcars))

Documentation

The manual is at https://docs.racket-lang.org/glmnet/, and raco docs glmnet opens the installed copy:

• Getting started: a first regression on mtcars, cross-validation and prediction.
• Concepts: data, the penalty, the families, paths, prediction and cross-validation.
• Data: example datasets, tables, CSV files, math/matrix and Polars.
• Formulas and Plots.
• Examples: one worked example per family.
• Reference.

Status

All six of R glmnet 4.1's families fit at a single λ, along a regularization path and with cross-validation, from matrices, lists, tables and Polars dataframes or through R-style formulas, and glmnet/plot draws R's path and cross-validation plots. Parity tests check the numbers against R. Not bound yet: observation weights, penalty factors, coefficient limits, offsets and exclude (#12), and sparse predictor matrices (#11).

Development

AGENTS.md describes the architecture and the workflow. With Nix:

nix develop                     # builds the native library, link-installs the package
raco test ./glmnet/             # unit tests and the examples' harnesses
bash scripts/run-examples.sh    # runs every example in glmnet/examples/
nix flake check                 # builds everything, runs the tests, renders the manual

With a local toolchain (gfortran, CMake and Racket):

cmake -S fortran -B fortran/build -DBUILD_TESTING=ON
cmake --build fortran/build
ctest --test-dir fortran/build --output-on-failure
cp fortran/build/libglmnetcompat.* glmnet/native-libs/
raco pkg install --batch --auto --link --name glmnet ./glmnet
raco test ./glmnet/

The package shares its architecture with the Racket bindings bkc39/rkt-polars, bkc39/scs and bkc39/xgboost.

License, acknowledgements and AI disclosure

License. This package is distributed under GPL-2.0-or-later.

Acknowledgements. The solver and its algorithms are the work of Jerome Friedman, Trevor Hastie, Rob Tibshirani and the other authors of the R glmnet package, whose vignettes shaped this manual.

AI Disclosure. This package and its documentation were created with the use of AI tools.