tidymodels 列中发现的新颖级别

Question

我正在用来tidymodels创建随机福雷斯特预测。我的测试数据包含训练数据中不存在的新因子级别，这会导致错误：

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

我尝试在“兄弟姐妹”列中包含“step_novel和” step_dummy，但这并不能解决错误。我应该如何处理训练数据中不存在的新因素？

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

如果您在文档中注意到step_novel()，它说：

\n

\n

当拟合可以处理新因子水平的模型时，请考虑使用workflows::add_recipe()with allow_novel_levels = TRUEset inhardhat::default_recipe_blueprint()。这将使您的模型能够在预测时处理新的级别，而不是抛出警告或错误。

\n

\n

所以你想这样做：

\n

library(tidyverse)\nlibrary(tidymodels)\n#> Registered S3 method overwritten by \'tune\':\n#>   method                   from   \n#>   required_pkgs.model_spec parsnip\n\ndata <-\n  data.frame(\n    Survived = as.factor(c(0,1,1,1,0,0,0,0,0,1,1,1,0,0,0,0)),\n    Siblings = as.factor(c(1,1,0,1,0,0,0,3,1,1,0,1,0,0,0,3)),\n    Class = as.factor(c(0,1,0,1,0,1,0,0,0,1,0,1,0,1,0,0)),\n    Embarked = as.factor(c("s","c","m","m","s","c","s","m","m","s","s","s","s","s","s","s")) \n  )\n\ntest <-\n  data.frame(\n    Siblings = as.factor(c(1,1,0,1,0,0,0,3,1,1,0,1,0,0,0,4)), #New factor level\n    Class = as.factor(c(0,1,0,1,0,1,0,0,0,1,0,1,0,1,0,0)),\n    Embarked = as.factor(c("s","c","m","m","s","c","s","m","m","s","s","s","s","s","s","s")) \n  )\n\n#Model\nrf_model <-\n  rand_forest() %>%\n  set_args(\n    mtry = 3,\n    trees = 1000,\n    min_n = 15\n  ) %>%\n  set_engine("ranger", \n             importance = "impurity") %>%\n  set_mode("classification")\n\n#Recipe\ndata_recipe <- \n  recipe(Survived ~Siblings + Class + Embarked, data=data) %>%\n  step_novel(Siblings) %>%\n  step_dummy(Siblings)\n\n#Workflow\nrf_workflow <- \n  workflow() %>%\n  add_recipe(data_recipe, \n             blueprint = hardhat::default_recipe_blueprint(allow_novel_levels = TRUE)) %>%\n  add_model(rf_model)\n\nfinal_model <- fit(rf_workflow, data)\nfinal_model\n#> \xe2\x95\x90\xe2\x95\x90 Workflow [trained] \xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\xe2\x95\x90\n#> Preprocessor: Recipe\n#> Model: rand_forest()\n#> \n#> \xe2\x94\x80\xe2\x94\x80 Preprocessor \xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\n#> 2 Recipe Steps\n#> \n#> \xe2\x80\xa2 step_novel()\n#> \xe2\x80\xa2 step_dummy()\n#> \n#> \xe2\x94\x80\xe2\x94\x80 Model \xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\xe2\x94\x80\n#> Ranger result\n#> \n#> Call:\n#>  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) \n#> \n#> Type:                             Probability estimation \n#> Number of trees:                  1000 \n#> Sample size:                      16 \n#> Number of independent variables:  5 \n#> Mtry:                             3 \n#> Target node size:                 15 \n#> Variable importance mode:         impurity \n#> Splitrule:                        gini \n#> OOB prediction error (Brier s.):  0.254242\n\ntest_predict <- predict(final_model, test)\ntest_predict\n#> # A tibble: 16 x 1\n#>    .pred_class\n#>    <fct>      \n#>  1 0          \n#>  2 1          \n#>  3 0          \n#>  4 1          \n#>  5 0          \n#>  6 0          \n#>  7 0          \n#>  8 0          \n#>  9 0          \n#> 10 1          \n#> 11 0          \n#> 12 1          \n#> 13 0          \n#> 14 0          \n#> 15 0          \n#> 16 0\n

\n

^{由reprex 包(v2.0.0)创建于 2021-07-09}

\n

工作流程功能对新数据的因子级别和其他方面非常严格，以确保它们与训练数据匹配。

\n