Skip to contents

Engine for regularized regression models

Source: R/engine_glmnet.R
engine_glmnet.Rd

This engine allows the estimation of linear coefficients using either ridge, lasso or elastic net regressions techniques. Backbone of this engine is the glmnet R-package which is commonly used in SDMs, including the popular 'maxnet' (e.g. Maxent) package. Ultimately this engine is an equivalent of engine_breg, but in a "frequentist" setting. If user aim to emulate a model that most closely resembles maxent within the ibis.iSDM modelling framework, then this package is the best way of doing so. Compared to the 'maxnet' R-package, a number of efficiency settings are implemented in particular for cross-validation of alpha and lambda values.

Limited amount of prior information can be specified for this engine, specifically via offsets or as GLMNETPrior, which allow to specify priors as regularization constants.

Usage

engine_glmnet(
  x,
  alpha = 0,
  nlambda = 100,
  lambda = NULL,
  type = "response",
  ...
)

Arguments

x

distribution() (i.e. BiodiversityDistribution) object.

alpha

A numeric giving the elasticnet mixing parameter, which has to be between 0 and 1. alpha=1 is the lasso penalty, and alpha=0 the ridge penalty (Default: 0).

nlambda

A numeric giving the number of lambda values to be used (Default: 100).

lambda

A numeric with a user supplied estimate of lambda. Usually best to let this parameter be determined deterministically (Default: NULL).

type

The mode used for creating posterior predictions. Either making "link" or "response" (Default: "response").

...

Other parameters passed on to glmnet.

Value

An Engine.

Details

Regularized regressions are effectively GLMs that are fitted with ridge, lasso or elastic-net regularization. Which of them is chosen is critical dependent on the alpha value: * For alpha equal to 0 a ridge regularization is used. Ridge regularization has the property that it doesn't remove variables entirely, but instead sets their coefficients to 0. * For alpha equal to 1 a lasso regularization is used. Lassos tend to remove those coefficients fully from the final model that do not improve the loss function. * For alpha values between 0 and 1 a elastic-net regularization is used, which is essentially a combination of the two. The optimal lambda parameter can be determined via cross-validation. For this option set "varsel" in train() to "reg".

References

  • Jerome Friedman, Trevor Hastie, Robert Tibshirani (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1-22. URL https://www.jstatsoft.org/v33/i01/.

  • Renner, I.W., Elith, J., Baddeley, A., Fithian, W., Hastie, T., Phillips, S.J., Popovic, G. and Warton, D.I., 2015. Point process models for presence‐only analysis. Methods in Ecology and Evolution, 6(4), pp.366-379.

  • Fithian, W. & Hastie, T. (2013) Finite-sample equivalence in statistical models for presence-only data. The Annals of Applied Statistics 7, 1917–1939

See also

Other engine: engine_bart(), engine_breg(), engine_gdb(), engine_glm(), engine_inla(), engine_inlabru(), engine_stan(), engine_xgboost()

Examples

if (FALSE) {
# Add GLMNET as an engine
x <- distribution(background) |> engine_glmnet(iter = 1000)
}