How to apply search space to AutoFSelector

[ShixiangWang]

Hi,

I am new to mlr3. After reading the content of basics and model optimization in mlr3 book, I am trying to apply a xgboost model to my data with mlr3. When I use AutoFSelector to determine important features, I cannot find any place to apply search space and the search space can not be assigned to learner.

Here is some code,

  task = mlr3::as_task_classif(data, target = target, id = "auto-feature-selector", positive = as.character(positive))
  rm(data)
  invisible(gc())

  learner = mlr3::lrn("classif.xgboost", predict_type = "prob")
  terminator = mlr3tuning::trm("evals", n_evals = 10)
  fselector = mlr3fselect::fs("random_search")
  resampling = mlr3::rsmp("holdout", ratio = 0.9)
  measure = mlr3::msr("classif.prauc")

  # rewrite parameters
  #learner$param_set$values$nrounds <- paradox::to_tune(c(1, 100))

  at = AutoFSelector$new(
    learner = learner,
    resampling = resampling,
    measure = measure,
    terminator = terminator,
    fselector = fselector
  )

I am trying to define a search space so the AutoFSelector can run xgboost with different parameters. e.g., set nrounds to 1 and 100.

This doesn't work:

learner$param_set$values$nrounds <- paradox::to_tune(c(1, 100))

Is there an easy way to do this?

Best,

Shixiang

[pat-s]

cc @be-marc

[be-marc]

library(mlr3tuning)
library(mlr3fselect)
library(mlr3learners)

task = tsk("pima")

# feature selection on pima data set
fselect_instance = fselect(
    method = "random_search",
    task = task,
    learner = lrn("classif.xgboost", nrounds = 100),
    resampling = rsmp("cv", folds = 3),
    measure = msr("classif.ce"),
    term_evals = 10
)


# subset task to selected features
task$select(fselect_instance$result_feature_set)

# tune rpart on pima data set
tuning_instance = tune(
  method = "grid_search",
  task = task,
  learner = lrn("classif.xgboost", nrounds = to_tune(100, 1000)),
  resampling = rsmp("cv", folds = 3),
  measure = msr("classif.ce"),
  term_evals = 10,
  resolution = 5
)

Note that the tuning result depends on the previously selected features.