Predict with step_naomit and retain ID using tidymodels - r

Question

I am trying to retain an ID on the row when predicting using a Random Forest model to merge back on to the original dataframe. I am using step_naomit in the recipe that removes the rows with missing data when I bake the training data, but also removes the records with missing data on the testing data. Unfortunately, I don't have an ID to easily know which records were removed so I can accurately merge back on the predictions.
I have tried to add an ID column to the original data, but bake will remove any variable not included in the formula (and I don't want to include ID in the formula). I also thought I may be able to retain the row.names from the original table to merge on, but it appears the row.name is reset upon baking as well.
I realize I can remove the NA values prior to the recipe to solve this problem, but then what is the point of step_naomit in the recipe? I also tried skip=TRUE in the step_naomit, but then I get an error for missing data when fitting the model (only for random forest). I feel I am missing something here in tidymodels that would allow me to retain all the rows prior to baking?
See example:
## R 3.6.1 ON WINDOWS 10 MACHINE
require(tidyverse)
require(tidymodels)
require(ranger)
set.seed(123)
temp <- iris %>%
dplyr::mutate(Petal.Width = case_when(
round(Sepal.Width) %% 2 == 0 ~ NA_real_, ## INTRODUCE NA VALUES
TRUE ~ Petal.Width))
mySplit <- rsample::initial_split(temp, prop = 0.8)
myRecipe <- function(dataFrame) {
recipes::recipe(Petal.Width ~ ., data = dataFrame) %>%
step_naomit(all_numeric()) %>%
prep(data = dataFrame)
}
myPred <- function(mySplit,myRecipe) {
train_set <- training(mySplit)
test_set <- testing(mySplit)
train_prep <- myRecipe(train_set)
analysis_processed <- bake(train_prep, new_data = train_set)
model <- rand_forest(
mode = "regression",
mtry = 3,
trees = 50) %>%
set_engine("ranger", importance = 'impurity') %>%
fit(Sepal.Width ~ ., data=analysis_processed)
test_processed <- bake(train_prep, new_data = test_set)
test_processed %>%
bind_cols(myPrediction = unlist(predict(model,new_data=test_processed)))
}
getPredictions <- myPred(mySplit,myRecipe)
nrow(getPredictions)
## 21 ROWS
max(as.numeric(row.names(getPredictions)))
## 21
nrow(testing(mySplit))
## 29 ROWS
max(as.numeric(row.names(testing(mySplit))))
## 150

Answer 1

To be able to keep track of which observations were removed we need to give the original dataset an id variable.
temp <- iris %>%
dplyr::mutate(Petal.Width = case_when(
round(Sepal.Width) %% 2 == 0 ~ NA_real_, ## INTRODUCE NA VALUES
TRUE ~ Petal.Width),
id = row_number()) #<<<<
Then we use update_role() to first designate it as an "id variable", then remove it as a predictor so it doesn't become part of the modeling process. And that is it. Everything else should work like before. Below is the fully updated code with #<<<< to denote my changes.
require(tidyverse)
#> Loading required package: tidyverse
require(tidymodels)
#> Loading required package: tidymodels
#> Registered S3 method overwritten by 'xts':
#> method from
#> as.zoo.xts zoo
#> ── Attaching packages ───────────────────── tidymodels 0.0.3 ──
#> ✔ broom 0.5.2 ✔ recipes 0.1.7
#> ✔ dials 0.0.3 ✔ rsample 0.0.5
#> ✔ infer 0.5.0 ✔ yardstick 0.0.4
#> ✔ parsnip 0.0.4
#> ── Conflicts ──────────────────────── tidymodels_conflicts() ──
#> ✖ scales::discard() masks purrr::discard()
#> ✖ dplyr::filter() masks stats::filter()
#> ✖ recipes::fixed() masks stringr::fixed()
#> ✖ dplyr::lag() masks stats::lag()
#> ✖ dials::margin() masks ggplot2::margin()
#> ✖ dials::offset() masks stats::offset()
#> ✖ yardstick::spec() masks readr::spec()
#> ✖ recipes::step() masks stats::step()
require(ranger)
#> Loading required package: ranger
set.seed(1234)
temp <- iris %>%
dplyr::mutate(Petal.Width = case_when(
round(Sepal.Width) %% 2 == 0 ~ NA_real_, ## INTRODUCE NA VALUES
TRUE ~ Petal.Width),
id = row_number()) #<<<<
mySplit <- rsample::initial_split(temp, prop = 0.8)
myRecipe <- function(dataFrame) {
recipes::recipe(Petal.Width ~ ., data = dataFrame) %>%
update_role(id, new_role = "id variable") %>% #<<<<
update_role(-id, new_role = 'predictor') %>% #<<<<
step_naomit(all_numeric()) %>%
prep(data = dataFrame)
}
myPred <- function(mySplit,myRecipe) {
train_set <- training(mySplit)
test_set <- testing(mySplit)
train_prep <- myRecipe(train_set)
analysis_processed <- bake(train_prep, new_data = train_set)
model <- rand_forest(
mode = "regression",
mtry = 3,
trees = 50) %>%
set_engine("ranger", importance = 'impurity') %>%
fit(Sepal.Width ~ ., data=analysis_processed)
test_processed <- bake(train_prep, new_data = test_set)
test_processed %>%
bind_cols(myPrediction = unlist(predict(model,new_data=test_processed)))
}
getPredictions <- myPred(mySplit, myRecipe)
getPredictions
#> # A tibble: 23 x 7
#> Sepal.Length Sepal.Width Petal.Length Petal.Width Species id myPrediction
#> <dbl> <dbl> <dbl> <dbl> <fct> <int> <dbl>
#> 1 4.6 3.1 1.5 0.2 setosa 4 3.24
#> 2 4.3 3 1.1 0.1 setosa 14 3.04
#> 3 5.1 3.4 1.5 0.2 setosa 40 3.22
#> 4 5.9 3 4.2 1.5 versico… 62 2.98
#> 5 6.7 3.1 4.4 1.4 versico… 66 2.92
#> 6 6 2.9 4.5 1.5 versico… 79 3.03
#> 7 5.7 2.6 3.5 1 versico… 80 2.79
#> 8 6 2.7 5.1 1.6 versico… 84 3.12
#> 9 5.8 2.6 4 1.2 versico… 93 2.79
#> 10 6.2 2.9 4.3 1.3 versico… 98 2.88
#> # … with 13 more rows
# removed ids
setdiff(testing(mySplit)$id, getPredictions$id)
#> [1] 5 28 47 70 90 132
Created on 2019-11-26 by the reprex package (v0.3.0)

Answer 2

Using skip = TRUE in the step_naomit() recipe specification, and then including the recipe in a workflow might be the proper solution. For example,
myRecipe <- recipe(Petal.Width ~ ., data = dataFrame) %>%
step_naomit(all_numeric(), step = FALSE)`
# don't include the prep()
wflow <- workflow() %>%
add_model(model) %>%
add_recipe(myRecipe)
wflow_fit <- wflow %>%
fit(train_set)
preds <- predict(wflow_fit, new_data = (test_set))