tidymodel error, when calling predict function is asking for target variable

Question

I have trained a churn tidymodel with customer data (more than 200 columns). Got a fairly good metrics using xgbboost but the issue is when tryng to predict on new data.

Predict function asks for target variable (churn) and I am a bit confused as this variable is not supposed to be present on real scenario data as this is the variable I want to predict.

sample code below, maybe I missed the point on procedure. Some questions arised:

1. should I execute prep() at the end of recipe?

2. should I execute recipe on my new data prior to predict?

3. why removing lines from recipe regarding target variable makes predict work?

4. why is asking for my target variable?

       churn_recipe <- recipes::recipe(churn ~ ., data = churn_train) %>%
          recipes::step_naomit(everything(), skip = TRUE) %>% 
          recipes::step_rm(c(v1, v2, v3, v4, v5, v6)) %>%
       #  removing/commenting the next 2 lines makes predict() work
          recipes::step_string2factor(churn) %>%  
          themis::step_downsample(churn) %>%
          recipes::step_dummy(all_nominal_predictors()) %>% 
          recipes::step_novel(all_nominal(), -all_outcomes())  ### %>% prep()
   
           xgboost_model <-
              parsnip::boost_tree(
                mode = "classification",
                trees = 100
              ) %>%
              set_engine("xgboost") %>% 
              set_mode("classification")
   
            xgboost_workflow <-
              workflows::workflow() %>%
              add_recipe(churn_recipe) %>% 
              add_model(xgboost_model) 
   
              my_fit <- last_fit(xgboost_workflow, churn_split)
   
              collect_metrics(my_fit)
   
   
              churn_wf_model <- my_fit$.workflow[[1]]
   
            predict(churn_wf_model, new_data[1,])
            Error: Can't subset columns that don't exist.
            x Column `churn` doesn't exist.
   

I am pretty sure some misconceptions on my side, but unable to solve this issue.

I am stuck in moving my model into production. Tidymodels documentation lack of such topic is enormous.

Answer 1

You are getting this error because of recipes::step_string2factor(churn)

This step works fine when you are training the data. But when it is time to apply the same transformation to the training set, then step_string2factor() complains because it is asked to turn churn from a string to a factor but the dataset doesn't include the churn variable. You can deal with this in two ways.

skip = FALSE in step_string2factor() (less favorable)

By setting skip = FALSE in step_string2factor() you are telling the step o only be applied to when prepping/training the recipe. This is not favorable as this approach can produce errors in certain resampling scenarios using {tune} when the response is expected to be a factor instead of a string.

library(tidymodels)

data("mlc_churn")

set.seed(1234)
churn_split <- initial_split(mlc_churn)

churn_train <- training(churn_split)
churn_test <- testing(churn_split)


churn_recipe <- recipes::recipe(churn ~ ., data = churn_train) %>%
   recipes::step_naomit(everything(), skip = TRUE) %>% 
   recipes::step_string2factor(churn, skip = TRUE) %>%  
   themis::step_downsample(churn) %>%
   recipes::step_dummy(all_nominal_predictors()) %>% 
   recipes::step_novel(all_nominal(), -all_outcomes())

xgboost_model <-
  parsnip::boost_tree(
    mode = "classification",
    trees = 100
  ) %>%
  set_engine("xgboost") %>% 
  set_mode("classification")

xgboost_workflow <-
  workflows::workflow() %>%
  add_recipe(churn_recipe) %>% 
  add_model(xgboost_model) 

my_fit <- last_fit(xgboost_workflow, churn_split)

churn_wf_model <- my_fit$.workflow[[1]]

predict(churn_wf_model, churn_test)
#> # A tibble: 1,250 x 1
#>    .pred_class
#>    <fct>      
#>  1 no         
#>  2 no         
#>  3 no         
#>  4 no         
#>  5 no         
#>  6 no         
#>  7 no         
#>  8 no         
#>  9 no         
#> 10 yes        
#> # … with 1,240 more rows

^{Created on 2021-06-10 by the reprex package (v2.0.0)}

Make response a factor before splitting (recommended)

The recommended way to fix this issue is to make sure that your response churn is a factor before you pass it into {recipes}. I find it easiest to do it as I create the validation split with initial_split() like so. Then you don't need to use step_string2factor() on your response in your recipe

churn_split <- mlc_churn %>%
  mutate(churn = factor(churn)) %>%
  initial_split()