tidymodels Novel levels found in column - r

Question

I am using tidymodels to create a Random Forrest prediction. I have test data that contains a new factor level not present in the training data which results in the error:
1: Novel levels found in column 'Siblings': '4'. The levels have been removed, and values have been coerced to 'NA'.
2: There are new levels in a factor: NA
> test_predict
Fehler: Objekt 'test_predict' nicht gefunden
I tried to include a step_novel and step_dummy on the "Siblings" column but this does not resolve the error. How should I deal with new factors not present in training data?
library(tidyverse)
library(tidymodels)
data <-
data.frame(
Survived = as.factor(c(0,1,1,1,0,0,0,0,0,1,1,1,0,0,0,0)),
Siblings = as.factor(c(1,1,0,1,0,0,0,3,1,1,0,1,0,0,0,3)),
Class = as.factor(c(0,1,0,1,0,1,0,0,0,1,0,1,0,1,0,0)),
Embarked = as.factor(c("s","c","m","m","s","c","s","m","m","s","s","s","s","s","s","s"))
)
test <-
data.frame(
Siblings = as.factor(c(1,1,0,1,0,0,0,3,1,1,0,1,0,0,0,4)), #New factor level
Class = as.factor(c(0,1,0,1,0,1,0,0,0,1,0,1,0,1,0,0)),
Embarked = as.factor(c("s","c","m","m","s","c","s","m","m","s","s","s","s","s","s","s"))
)
#Model
rf_model <-
rand_forest() %>%
set_args(
mtry = 3,
trees = 1000,
min_n = 15
) %>%
set_engine("ranger",
importance = "impurity") %>%
set_mode("classification")
#Recipe
data_recipe <-
recipe(Survived ~Siblings + Class + Embarked, data=data) %>%
step_novel(Siblings) %>%
step_dummy(Siblings)
#Workflow
rf_workflow <-
workflow() %>%
add_recipe(data_recipe) %>%
add_model(rf_model)
final_model <- fit(rf_workflow, data)
final_model
test_predict <- predict(final_model, test)
test_predict

Answer 1

If you notice in the documentation for step_novel(), it says:
When fitting a model that can deal with new factor levels, consider using workflows::add_recipe() with allow_novel_levels = TRUE set in hardhat::default_recipe_blueprint(). This will allow your model to handle new levels at prediction time, instead of throwing warnings or errors.
So you want to do that:
library(tidyverse)
library(tidymodels)
#> Registered S3 method overwritten by 'tune':
#> method from
#> required_pkgs.model_spec parsnip
data <-
data.frame(
Survived = as.factor(c(0,1,1,1,0,0,0,0,0,1,1,1,0,0,0,0)),
Siblings = as.factor(c(1,1,0,1,0,0,0,3,1,1,0,1,0,0,0,3)),
Class = as.factor(c(0,1,0,1,0,1,0,0,0,1,0,1,0,1,0,0)),
Embarked = as.factor(c("s","c","m","m","s","c","s","m","m","s","s","s","s","s","s","s"))
)
test <-
data.frame(
Siblings = as.factor(c(1,1,0,1,0,0,0,3,1,1,0,1,0,0,0,4)), #New factor level
Class = as.factor(c(0,1,0,1,0,1,0,0,0,1,0,1,0,1,0,0)),
Embarked = as.factor(c("s","c","m","m","s","c","s","m","m","s","s","s","s","s","s","s"))
)
#Model
rf_model <-
rand_forest() %>%
set_args(
mtry = 3,
trees = 1000,
min_n = 15
) %>%
set_engine("ranger",
importance = "impurity") %>%
set_mode("classification")
#Recipe
data_recipe <-
recipe(Survived ~Siblings + Class + Embarked, data=data) %>%
step_novel(Siblings) %>%
step_dummy(Siblings)
#Workflow
rf_workflow <-
workflow() %>%
add_recipe(data_recipe,
blueprint = hardhat::default_recipe_blueprint(allow_novel_levels = TRUE)) %>%
add_model(rf_model)
final_model <- fit(rf_workflow, data)
final_model
#> ══ Workflow [trained] ══════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: rand_forest()
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> 2 Recipe Steps
#>
#> • step_novel()
#> • step_dummy()
#>
#> ── Model ───────────────────────────────────────────────────────────────────────
#> Ranger result
#>
#> Call:
#> ranger::ranger(x = maybe_data_frame(x), y = y, mtry = min_cols(~3, x), num.trees = ~1000, min.node.size = min_rows(~15, x), importance = ~"impurity", num.threads = 1, verbose = FALSE, seed = sample.int(10^5, 1), probability = TRUE)
#>
#> Type: Probability estimation
#> Number of trees: 1000
#> Sample size: 16
#> Number of independent variables: 5
#> Mtry: 3
#> Target node size: 15
#> Variable importance mode: impurity
#> Splitrule: gini
#> OOB prediction error (Brier s.): 0.254242
test_predict <- predict(final_model, test)
test_predict
#> # A tibble: 16 x 1
#> .pred_class
#> <fct>
#> 1 0
#> 2 1
#> 3 0
#> 4 1
#> 5 0
#> 6 0
#> 7 0
#> 8 0
#> 9 0
#> 10 1
#> 11 0
#> 12 1
#> 13 0
#> 14 0
#> 15 0
#> 16 0
Created on 2021-07-09 by the reprex package (v2.0.0)
The workflows functions are very strict about factor levels and other aspects of the new data, ensuring that they match up with the training data.

Answer 2

To answer my own question:
We need to apply step_novel followed by step_unknown. As far as I understand from the documentation step_novel labels any new factors occurring in the data with "new". This can be used to easily identify such factors when the data is inspected after applying the recipe. step_unknown removes any such factors from the data and converts the values to NA when the model is applied:
data_recipe <-
recipe(Survived ~Siblings + Class + Embarked, data=data) %>%
step_novel(Siblings) %>%
step_unknown(Siblings)