Supervised tasks with missing data in the target

[be-marc]

Hi, I’m a member of the mlr3 team. We’re currently considering disallowing supervised tasks with missing values in the target column. With that change, your vimpute.Rmd vignette no longer runs.

library(reactable)
library(VIM)
data(iris)

# Create complete copy before introducing NAs
complete_data <- iris
colnames(complete_data) <- c("S.Length","S.Width","P.Length","P.Width","Species")
df <- copy(complete_data)

# Randomly produce missing values
set.seed(1)
nbr_missing <- 50
y <- data.frame(row = sample(nrow(df), size = nbr_missing, replace = TRUE),
                col = sample(ncol(df), size = nbr_missing, replace = TRUE))
y <- y[!duplicated(y),]
df[as.matrix(y)] <- NA

# Perform imputation with proper method specification
result <- vimpute(
  data = df,
  method = setNames(lapply(names(df), function(x) "xgboost"),names(df)),
  pred_history = TRUE
)

#>Variables with Missing Data: S.Length,S.Width,P.Length,P.Width,Species
#>data is data.table
#> Precheck done.
#> Error: 
#> ✖ Target column 'S.Length' must not contain missing values
#> → Class: Mlr3ErrorInput

Would this mlr3 change be a general issue for your package, or would it only affect this particular example?

[alexkowa]

Hi, thank you for the warning.
yes, it is a general design issue for a subset of methods/data inside vimpute.
I did not get that error with the dev version of mlr3, is the currently in a specific branch?

[alexkowa]

@eileenva I think this is reproducing the current message, where there will be an error in the future, please take a look:
(from test_xgboostImpute.R)
# factor regressor used reasonably", { df2 <- df df2$x[1:10] <- NA idx <- !is.na(df$fac) df.out <- xgboostImpute(x ~ fac, df2)

[be-marc]

Hey, it's this line https://github.com/mlr-org/mlr3/blob/4e9dd10a52ebf1d2b2bd53752c661cb3218d50e9/R/as_task_classif.R#L59-L62. Install with pak::pak("mlr-org/mlr3@issue_vim_98"). However, we might not activate it when it's an issue for you.

[eileenva]

I’ve adjusted vimpute so all mlr3 tasks are built with NA‑free targets:

• Training tasks use only complete cases in the target.
• Prediction tasks use the full data, but target NAs are temporarily filled (median/mode) just for task creation.
  So from our side now it should be okay to activate it.

[be-marc]

Great. Let me know when you the CRAN update is done. Then I can upload the new mlr3 version. Thanks!