When to use step_* (and when not to...)?

[athos]

Is there any advise, rule of thumb or guide telling what should and what should not be done using step_* functions in dataprep stage?

Often I found tutorials showing datapreps mixing things been done inside recipes and outside recipes.

example: from this issue

#...
titanic <- titanic %>% 
  select(-PassengerId, -Name, -Ticket, -Cabin) %>% 
  mutate(Survived = case_when(Survived == 0 ~ "No", 
                              Survived == 1 ~ "Yes"))

# Survived and Sex should be factors
titanic <- titanic %>%
  mutate(
    Survived = as.factor(Survived),
    Sex = as.factor(Sex),
    Embarked = as.factor(Embarked)
  )

# imputation stuff
imputed_titanic <- mice(titanic, m = 1, maxit = 30, seed = 1234, printFlag = FALSE)
imputed_titanic <- complete(imputed_titanic)

# splits
train_test_split <- initial_split(imputed_titanic, prop = 0.9)
titanic_train <- training(train_test_split)
titanic_test <- testing(train_test_split)

# Don't dummy the outcome!
rec <- recipe(Survived ~ ., data = titanic_train) %>%
  step_dummy(all_nominal(), -Survived)

#...

[Max]

I would not do imputation first since it involves the test data. If you have a lot of missing data, it will optimistically bias your performance results.

For the other operations, there are step_mutate() and step_rm().