fit_resamples 中 step_naomit 后丢失数据

问题描述

我目前正在应用以下配方和工作流程，以便使用 fit_resamples 进行 5 折交叉验证来拟合随机森林。工作流程如下所示：

library(tidymodels)

# import data and convert response to factor
train <- read.csv('https://pastebin.com/raw/LJQqdEEE')
train$accepted <- as.factor(train$accepted)

# Train/test split
new_split <- initial_split(train,prop = 0.7)
new_train <- training(new_split)
new_test <- testing(new_split)

# Feature engineering and data prep
admission_rec <- 
  recipe(accepted ~ .,data = new_train) %>% 
  step_impute_median(sat) %>% 
  step_mutate(
    ap_scores = strsplit(as.character(ap_scores),';'),ap_score_max = max(as.numeric(unlist(ap_scores))),ap_score_avg = mean(as.numeric(unlist(ap_scores))),ap_score_min = min(as.numeric(unlist(ap_scores))),ap_score_med = median(as.numeric(unlist(ap_scores)))
  ) %>% 
  step_dummy(ethnicity,one_hot = T) %>% 
  step_center(c(essay_strength,family_income,sat),skip = T) %>%
  step_scale(c(essay_strength,skip = T) %>%
  step_naomit(everything(),skip = T) %>% 
  step_rm(ap_scores)

# Random forest model and workflow
rf_spec <- 
  rand_forest() %>% 
  set_engine('ranger') %>% 
  set_mode('classification')

rf_workflow <- 
  workflow() %>% 
  add_recipe(admission_rec) %>% 
  add_model(rf_spec)

# Cross validation
cv_folds <- 
  vfold_cv(new_train,v = 5)

# Fit model
rf_res <- rf_workflow %>%
  fit_resamples(
    resamples = cv_folds,metrics = metric_set(
      recall,precision,f_meas,accuracy,kap,roc_auc,sens,spec
    )
  )

在拟合模型时，我收到以下失败消息的提示：

preprocessor 1/1: There are new levels in a factor: NA
preprocessor 1/1,model 1/1 (predictions): Missing data in columns: ethnicity_Asian ...

这看起来仅限于一个热编码列，甚至是 step_naomit(skip = TRUE)。出于这个原因，我错误地认为将 step_naomit 放在 step_mutate 之后会解决这个问题。

我可能在这里忽略了一些相当简单的事情，这是我在长时间的 R hyathus 之后第一次尝试 {tidymodels}。

解决方法

您走在正确的轨道上。不幸的是，step_naomit() 不是答案，错误出现在 step_dummy() 中，因为它包含丢失的数据并且不知道如何处理它。解决方案是在之前使用 step_unknown()，它将采用因子变量并将“未知”分配给缺失值。

我还建议您不要在 skip = T 和 step_center() 中设置 step_scale()，因为它会在拟合模型时应用居中和缩放，但如果以后使用模型时会跳过上，例如在预测中。这会产生奇怪和不希望的结果。

library(tidymodels)
train <- read.csv('https://pastebin.com/raw/LJQqdEEE')
train$accepted <- as.factor(train$accepted)

# Train/test split
new_split <- initial_split(train,prop = 0.7)
new_train <- training(new_split)
new_test <- testing(new_split)

# Feature engineering and data prep
admission_rec <- 
  recipe(accepted ~ .,data = new_train) %>% 
  step_impute_median(sat) %>% 
  step_mutate(
    ap_scores = strsplit(as.character(ap_scores),';'),ap_score_max = max(as.numeric(unlist(ap_scores))),ap_score_avg = mean(as.numeric(unlist(ap_scores))),ap_score_min = min(as.numeric(unlist(ap_scores))),ap_score_med = median(as.numeric(unlist(ap_scores)))
  ) %>%
  step_unknown(ethnicity) %>%
  step_dummy(ethnicity,one_hot = T) %>%
  step_center(c(essay_strength,family_income,sat)) %>%
  step_scale(c(essay_strength,sat)) %>%
  step_rm(ap_scores)

# Random forest model and workflow
rf_spec <- 
  rand_forest() %>% 
  set_engine('ranger') %>% 
  set_mode('classification')

rf_workflow <- 
  workflow() %>% 
  add_recipe(admission_rec) %>% 
  add_model(rf_spec)

# Cross validation
cv_folds <- 
  vfold_cv(new_train,v = 5)

# Fit model
rf_res <- rf_workflow %>%
  fit_resamples(
    resamples = cv_folds,metrics = metric_set(
      recall,precision,f_meas,accuracy,kap,roc_auc,sens,spec
    )
  )

rf_res
#> # Resampling results
#> # 5-fold cross-validation 
#> # A tibble: 5 x 4
#>   splits            id    .metrics         .notes          
#>   <list>            <chr> <list>           <list>          
#> 1 <split [560/140]> Fold1 <tibble [8 × 4]> <tibble [0 × 1]>
#> 2 <split [560/140]> Fold2 <tibble [8 × 4]> <tibble [0 × 1]>
#> 3 <split [560/140]> Fold3 <tibble [8 × 4]> <tibble [0 × 1]>
#> 4 <split [560/140]> Fold4 <tibble [8 × 4]> <tibble [0 × 1]>
#> 5 <split [560/140]> Fold5 <tibble [8 × 4]> <tibble [0 × 1]>

^{由 reprex package (v2.0.0) 于 2021 年 6 月 22 日创建}