Add regr.fuser learner (fused Lasso regression)

[EngineerDanny]

Summary

Adds a new regression learner regr.fuser that wraps the fused Lasso model from fuser and requires a single grouping column via the group role.

Implementation Details

• New learner: LearnerRegrFuser
• Backend: fuser::fusedL2DescentGLMNet(X, y, groups, lambda, G, gamma, scaling)
• Requires exactly one grouping column via the group role
• Defaults: lambda = NULL, gamma = 1, scaling = FALSE, G = NULL (defaults to all-ones K×K)
• Errors on unseen groups at prediction time
• Validates scalar lambda, gamma, and G dimensions
• Tests for happy path + validation cases

Example

library(mlr3)
library(mlr3extralearners)

data = data.frame(
  y = rnorm(20),
  x1 = rnorm(20),
  x2 = rnorm(20),
  group = rep(c("A", "B"), each = 10)
)

task = TaskRegr$new(id = "fuser", backend = data, target = "y")
task$set_col_roles("group", roles = "group")

learner = lrn("regr.fuser", lambda = 0.01, gamma = 0.01, scaling = FALSE)
learner$train(task)
pred = learner$predict(task)

Tests

• Rscript -e 'testthat::test_file("tests/testthat/test_fuser_regr_fuser.R")'

Notes

• Requires fuser (GitHub fork): remotes::install_github("EngineerDanny/fuser")

Closes #426