Variable importance fails in superlearner containing interactions or feature screener

[acousticus]

Description:

sl3::importance() works with a simple SL, but fails when using interactions and/or feature screeners. Using sl3_1.4.4.

Steps to reproduce:

library(sl3)
data(cpp_imputed)

covars <- c("apgar1", "apgar5", "parity", "gagebrth", "mage", "meducyrs", "sexn")
outcome <- "haz"
task <- sl3_Task$new(cpp_imputed, covariates = covars, outcome = outcome)

interactions <- list(c("apgar1", "parity"), c("apgar5", "parity"))
keepme <- c("mage", "meducyrs")

# BASE LEARNERS
lrnr_glm <- make_learner(Lrnr_glm)
lrnr_mean <- make_learner(Lrnr_mean)
learners <- c(lrnr_glm, lrnr_mean)
stack <- make_learner(Stack, learners)

# RF SCREENER
miniforest <- Lrnr_ranger$new(importance = "impurity")
screen_rf <- Lrnr_screener_importance$new(learner = miniforest)

# INTERACTIONS
lrn_interaction <- make_learner(Lrnr_define_interactions, interactions)


### 1. SL WITHOUT SCREEN OR INTERACTIONS
sl <- make_learner(Lrnr_sl, learners)
sl_fit <- sl$train(task)
importance(sl_fit)

## Example output:
#   covariate   MSE_difference
# 1:  gagebrth    3.173805e-02
# 2:    parity    1.323431e-02
# 3:      mage    5.495150e-03
# 4:      sexn   -2.788994e-05
# 5:    apgar1   -2.859870e-04
# 6:  meducyrs   -1.242081e-03
# 7:    apgar5   -3.459809e-03


### 2. SL + INTERACTIONS + SCREENER
ints_screen_rf <- make_learner(Pipeline, lrn_interaction, screen_rf, stack)
sl <- make_learner(Lrnr_sl, learners = ints_screen_rf)
sl_fit <- sl$train(task)
importance(sl_fit)
# Error : Passed a vector of type 'list'. Needs to be type 'character'.
# Failed on chain
# Error in self$compute_step() :
#   Error : Passed a vector of type 'list'. Needs to be type 'character'.


### 3. SL + INTERACTIONS
ints_only <- make_learner(Pipeline, lrn_interaction, stack)
sl <- make_learner(Lrnr_sl, learners = ints_only)
sl_fit <- sl$train(task)
importance(sl_fit)
# Error : Passed a vector of type 'list'. Needs to be type 'character'.
# Failed on chain
# Error in self$compute_step() :
#   Error : Passed a vector of type 'list'. Needs to be type 'character'.


### 4. SL + SCREENER
ints_only <- make_learner(Pipeline, lrn_interaction, stack)
sl <- make_learner(Lrnr_sl, learners = ints_only)
sl_fit <- sl$train(task)
importance(sl_fit)
# Error : Passed a vector of type 'list'. Needs to be type 'character'.
# Failed on chain
# Error in self$compute_step() :
#   Error : Passed a vector of type 'list'. Needs to be type 'character'.

Question

Is this an issue with the way I am running importance() on the fits, with how I code the different superlearner variants, or with sl3 itself? Many thanks