colino的step_select_*与workflow_map/tune_grid联用报错问题

在使用workflow_set拟合多模型前，尝试将colino包的各类step_select_*函数作为预处理步骤加入recipe，但无法与workflow_map正常联用；直接调用包含step_select_*函数的workflow执行tune_grid时也会触发相同报错。报错提示purrr::map执行时出现tibble列尺寸不兼容问题，但无法定位涉事tibble。

复现代码：

library(parsnip)
library(dplyr)
library(fastDummies)
library(colino)
library(workflowsets)
library(tune)

data(iris)

logit_model <- logistic_reg(engine = "glm")

iris2 <- iris %>% 
  as_tibble() %>% 
  filter(Species != "setosa") 

iris_cv <- vfold_cv(iris2)

recipe <- recipe(Species ~ ., data = iris2) %>% 
  step_nzv(all_predictors()) %>% 
  step_select_vip(all_predictors(), model = logit_model,
                  threshold = 0.9, outcome = "Species") %>%
  step_normalize(all_numeric_predictors())

logit_vip2 <- 
  workflow_set(
    preproc = list(rec = recipe),
    models = list(logit = logistic_reg(engine = "glm")),
    cross = TRUE) 

res_logit_vip2 <- 
  logit_vip2 %>% 
  workflow_map(fn = "tune_grid", resamples =iris_cv,
               verbose = TRUE, control =control_grid(verbose =TRUE)) 

问题原因 #

colino的step_select_vip这类动态特征选择步骤，在交叉验证的不同折中，会基于当前折的训练数据计算特征重要性，进而选择满足threshold的特征。由于不同折的数据分布存在差异，最终选出的特征数量/名称可能不一致，导致每折处理后的数据集列数不同。而tune_grid和workflow_map要求交叉验证的所有折输出的数据集结构必须完全一致，因此触发了列尺寸不兼容的错误。

解决方案 #

方案1：提前在全量训练集上完成特征选择 #

先基于整个训练数据集确定固定的特征列表，再将这些特征加入recipe，避免交叉验证折内的动态选择导致结构不一致：

# 基于全量训练数据计算VIP特征并筛选
full_vip_results <- logit_model %>%
  fit(Species ~ ., data = iris2) %>%
  vip::vi() %>%
  arrange(desc(Importance)) %>%
  filter(cumsum(Importance) <= 0.9) %>%
  pull(Variable)

# 使用固定特征列表构建recipe
fixed_recipe <- recipe(Species ~ ., data = iris2) %>%
  step_nzv(all_predictors()) %>%
  step_select(all_of(full_vip_results)) %>%
  step_normalize(all_numeric_predictors())

# 正常执行workflow_set和workflow_map
logit_vip_fixed <- 
  workflow_set(
    preproc = list(rec = fixed_recipe),
    models = list(logit = logistic_reg(engine = "glm")),
    cross = TRUE) 

res_logit_vip_fixed <- 
  logit_vip_fixed %>% 
  workflow_map(fn = "tune_grid", resamples = iris_cv,
               verbose = TRUE, control = control_grid(verbose = TRUE)) 

方案2：固定选择的特征数量 #

如果必须在折内动态选择特征，可将step_select_vip的参数从threshold改为top_p，固定选择指定数量的特征，确保每折输出的数据集列数一致：

# 使用top_p固定选择特征数量
recipe_fixed_p <- recipe(Species ~ ., data = iris2) %>% 
  step_nzv(all_predictors()) %>% 
  step_select_vip(all_predictors(), model = logit_model,
                  top_p = 3, outcome = "Species") %>%  # 固定选择top3特征
  step_normalize(all_numeric_predictors())

# 正常执行workflow_set和workflow_map
logit_vip_p <- 
  workflow_set(
    preproc = list(rec = recipe_fixed_p),
    models = list(logit = logistic_reg(engine = "glm")),
    cross = TRUE) 

res_logit_vip_p <- 
  logit_vip_p %>% 
  workflow_map(fn = "tune_grid", resamples = iris_cv,
               verbose = TRUE, control = control_grid(verbose = TRUE)) 

内容的提问来源于stack exchange，提问作者Louie-david Desachy