forge fails with non-standard roles

[topepo]

forge fails with non-standard roles

[EmilHvitfeldt]

If you are using the dev version you need to set the bake_dependant_roles argument in the recipes blueprint to make sure that the ID variables are passed through. https://hardhat.tidymodels.org/dev/reference/default_recipe_blueprint.html#arguments

with

id_blueprint <- hardhat::default_recipe_blueprint(bake_dependent_roles = "ID")

library(tidymodels)
library(survival)

colon$sex <- ifelse(colon$sex==1,"male","female")
colon$obstruct <- ifelse(colon$obstruct ==1,"yes","no")
colon$perfor <- ifelse(colon$perfor ==1,"yes","no")
colon$adhere <- ifelse(colon$adhere ==1,"yes","no")
colon$status <- ifelse(colon$status ==1,"death","alive")
colon$node4 <- ifelse(colon$node4 ==1,"yes","no")

colon <- select(colon,id,age,rx,sex,age,obstruct,perfor,adhere,nodes,status)
colon <- na.omit(colon)

data_split<- initial_split(colon,
                           prop = 3/4,
                           strata = status)

train_data <- training(data_split)
test_data<- testing(data_split)

train_rec <-
  recipe(status ~., data = train_data) %>%
  update_role(id, new_role = "ID")%>%
  step_zv(all_numeric(),-all_outcomes()) %>%
  step_normalize(all_numeric(),-all_outcomes())%>%
  step_novel(all_nominal(),-all_outcomes()) %>%
  step_dummy(all_nominal(),-all_outcomes())

set.seed(100)

cv_folds <-
  vfold_cv(train_data,
           v = 5,
           strata = status)

log_spec <- # your model specification
  logistic_reg() %>% # model type
  set_engine(engine = "glm") %>% # model engine
  set_mode("classification") # model mode

id_blueprint <- hardhat::default_recipe_blueprint(bake_dependent_roles = "ID")

log_wflow <- # new workflow object
  workflow() %>% # use workflow function
  add_recipe(train_rec, blueprint = id_blueprint) %>% # use the new recipe
  add_model(log_spec) # add your model spec

log_res <-
  log_wflow %>%
  fit_resamples(
    resamples = cv_folds,
    metrics = metric_set(
      precision, f_meas,
      accuracy, kap,
      roc_auc, sens, spec),
    control = control_resamples(
      save_pred = TRUE)
  )
#> ! Fold1: preprocessor 1/1, model 1/1 (predictions): prediction from a rank-defici...
#> ! Fold2: preprocessor 1/1, model 1/1 (predictions): prediction from a rank-defici...
#> ! Fold3: preprocessor 1/1, model 1/1 (predictions): prediction from a rank-defici...
#> ! Fold4: preprocessor 1/1, model 1/1 (predictions): prediction from a rank-defici...
#> ! Fold5: preprocessor 1/1, model 1/1 (predictions): prediction from a rank-defici...


log_res %>% 
  collect_notes() %>% 
  pluck("note") %>% 
  cat(sep = "\n")
#> prediction from a rank-deficient fit may be misleading
#> prediction from a rank-deficient fit may be misleading
#> prediction from a rank-deficient fit may be misleading
#> prediction from a rank-deficient fit may be misleading
#> prediction from a rank-deficient fit may be misleading

^{Created on 2022-05-18 by the reprex package (v2.0.1)}

[DavisVaughan]

This is exactly as expected. This is how it should have worked from the beginning and is required for case weights to work. The only role that forge() looks for by default is the "predictor" role.

If you really require the "ID" role in the recipe then you should use bake_dependent_roles as Emil suggested.

But I think that the user doesn't actually want id to be involved in the recipe at all, so they should have used all_numeric_predictors() and all_nominal_predictors(), like this:

train_rec <-
  recipe(status ~., data = train_data) %>%
  update_role(id, new_role = "ID")%>%
  step_zv(all_numeric_predictors()) %>%
  step_normalize(all_numeric_predictors())%>%
  step_novel(all_nominal_predictors()) %>%
  step_dummy(all_nominal_predictors())

Then it works as expected

[EmilHvitfeldt]

Yes, -all_outcomes() is definitely an anti-pattern.

[github-actions]

This issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.