如何将XGBoost调优元数据转为tidymodels的tune对象用于tune_bayes？

问题背景 #

多次运行XGBoost调优后，仅保存了包含模型参数与性能的文本元数据，结构如下：

str(p)
'data.frame': 130 obs. of 10 variables:
$ mtry : int 922 1046 512 1317 675 1303 518 1029 1345 1180 ...
$ min_n : int 34 36 73 89 91 32 73 52 75 93 ...
$ tree_depth : int 44 33 43 37 34 48 25 19 38 41 ...
$ learn_rate : num 0.0236 0.0257 0.0292 0.0254 0.0271 0.023 0.025 0.0226 0.0281 0.0641 ...
$ loss_reduction: num 0.0268 0.745 0.148 0.171 0.0275 ...
$ sample_size : num 0.967 0.947 0.789 0.825 0.973 0.521 0.798 0.813 0.993 0.959 ...
$ .metric : chr "mn_log_loss" "mn_log_loss" "mn_log_loss" "mn_log_loss" ...
$ .estimator : chr "binary" "binary" "binary" "binary" ...
$ mean : num 0.423 0.424 0.424 0.424 0.424 0.425 0.425 0.426 0.427 0.427 ...
$ std_err : num 0.000382 0.000439 0.000408 0.000344 0.000368 0.000407 0.000386 0.000398 0.000392 0.000441 ...

尝试将该元数据作为tune_bayes的初始输入时，触发报错：

Error in check_initial():
! initial should be a positive integer or the results of [tune_grid()]
Run rlang::last_trace() to see where the error occurred.

可复现代码示例（param为已有元数据）：

library(tidyverse)
library(tidymodels) 

data(cells)
set.seed(123)
df_split <- initial_split(cells %>% select(-case))
df_train <- training(df_split)
df_test  <- testing(df_split)

set.seed(123)
df_train_folds <- vfold_cv(df_train, v = 5)

# 预处理配方
recipe_df <- recipe(class ~ ., data=df_train) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_numeric_predictors())

xgb_prep<- prep(recipe_df,verbose=T)

# 定义XGBoost模型
xgb_spec <- boost_tree(
  trees = 100,
  tree_depth = tune(),
  min_n = tune(),
  mtry = tune(),
  loss_reduction = tune(),
  sample_size = tune(),
  learn_rate = tune(),
) %>%
  set_engine('xgboost') %>%
  set_mode('classification')

# 构建工作流
xgb_wf <- workflow() %>% 
  add_model(xgb_spec) %>% 
  add_recipe(xgb_prep)

# 设置参数范围
xgb_params <-
  parameters(xgb_wf) %>%
  update(learn_rate = learn_rate(c(0.01, 0.3), trans=NULL),
         tree_depth = tree_depth(c(3, 100)),
         min_n = min_n(c(1, 10)),
         loss_reduction = loss_reduction(c(1, 5),trans=NULL),
         sample_size= sample_prop(c(0.5,1)),
         mtry = finalize(mtry(), df_train))

# 已有元数据
param <- tibble(mtry = c(36,27), min_n = c(8,25), tree_depth = c(4,6), learn_rate = c(0.116,0.000000111), loss_reduction = c(11.5406445593368101,0.0000001105066459), sample_size = c(0.306 ,0.984 ), metric = "mn_log_loss", estimator = "binary", mean = c(0.404,0.420), std_err = c(0.0137,0.0171))

# 尝试运行tune_bayes
options(tidymodels.dark = TRUE)
set.seed(123)

xgb_res <-
  tune_bayes(
    xgb_wf,
    resamples = df_train_folds,
    param_info = xgb_params,
    initial = param,
    iter =5,
    metrics = metric_set(mn_log_loss),
    control = control_bayes(verbose = TRUE,
                            parallel_over = "everything",
                            no_improve = 10)
  )  

解决方案 #

tune_bayes要求initial参数为tune_grid的输出格式（带tune_results类的交叉验证结果表），需按以下步骤转换已有元数据：

1. 修正列名与结构 #

确保元数据列名与tidymodels输出一致，并补充交叉验证所需的关键列：

# 修正列名，匹配tidymodels规范
param_fixed <- param %>%
  rename(.metric = metric, .estimator = estimator)

# 生成每个参数组合对应所有交叉验证折的行
initial_results <- param_fixed %>%
  # 关联所有 Fold ID
  crossing(id = df_train_folds$id) %>%
  mutate(
    .iter = 0,  # 初始迭代标记为0
    .config = paste0("Pre-set", row_number()),  # 每个参数组合的唯一标识
    .estimate = mean  # 用已有平均性能作为单折性能（若有单折数据可替换）
  ) %>%
  # 按tidymodels结果格式排序列
  select(id, .iter, .config, mtry, min_n, tree_depth, learn_rate, loss_reduction, sample_size, .metric, .estimator, .estimate)

# 标记为tune_results类
class(initial_results) <- c("tune_results", "tbl_df", "tbl", "data.frame")

2. 重新运行tune_bayes #

将转换后的initial_results传入initial参数：

set.seed(123)
xgb_res <-
  tune_bayes(
    xgb_wf,
    resamples = df_train_folds,
    param_info = xgb_params,
    initial = initial_results,
    iter =5,
    metrics = metric_set(mn_log_loss),
    control = control_bayes(verbose = TRUE,
                            parallel_over = "everything",
                            no_improve = 10)
  )  

关键说明 #

• 必须为每个参数组合补充所有交叉验证折的行，tune_bayes依赖完整的折级性能数据
• 若原始元数据有单折的详细性能值，替换.estimate为对应数值可提升结果准确性
• 确保参数列名与模型定义的调优参数完全一致，避免识别失败

内容的提问来源于stack exchange，提问作者arne maa