如何在tidymodels中用step_interact()生成与Caret一致的全交互项

核心问题 #

你当前的tidymodels配方中，step_interact(terms = ~.^2)没有正确生成与Caret mpg ~ .^2完全一致的交互项集合。Caret的~.^2会包含所有主效应 + 所有两两变量的交互项，而直接在step_interact中使用~.^2无法正确触发这一行为。

解决方案 #

将step_interact的terms参数改为~ (all_predictors())^2，这个语法会明确指定基于所有预测变量生成全双向交互项，同时保留主效应，与Caret的~.^2行为完全匹配。

另外确认：

• 标准化步骤放在交互项生成之前，这和Caret的preProcess行为一致（先标准化主变量，再计算交互项）
• 交叉验证折的复用逻辑已经正确，无需调整

修正后的含交互项模型代码 #

### Create Folds
set.seed(123)
train_index <- createDataPartition(mtcars$mpg, p = 0.8, list = FALSE)
train <- mtcars[train_index, ]
test <- mtcars[-train_index, ]

# Create model in Caret
set.seed(123)
my_grid = 10^seq(-5,5,length.out=1000)
model_caret <- train(mpg ~ .^2, data = train,
                     method = "glmnet",
                     trControl = trainControl(method = "cv", number = 5),
                     preProcess=c("center","scale"),
                     tuneGrid=expand.grid(alpha=1,lambda=my_grid))
model_caret
cat("The final values used for the model were alpha = 1 and lambda
 = 0.9014776")

### Create the same folds as in Caret
rsplit <- map2(model_caret$control$index,
               model_caret$control$indexOut,
               function(x,y) list(analysis = x, assessment = y))

splits <- lapply(rsplit, make_splits, data = train)
splits <- manual_rset(splits, names(splits))

### Create Model
my_model = linear_reg(mode = "regression",
                      engine = "glmnet",
                      penalty = tune(),
                      mixture = 1)

# 修正后的recipe：使用(all_predictors())^2生成全双向交互
my_recipe <- recipe(mpg ~ ., data = train) %>%
  step_normalize(all_predictors()) %>% 
  step_interact(terms = ~ (all_predictors())^2)

### Create Grid
my_grid = tibble(penalty=10^seq(-5,5,length.out=1000))

#### Create Workflow
my_wf = workflow() %>% 
  add_recipe(my_recipe) %>% 
  add_model(my_model)

### Tuning hyperparameter
model_tuning = tune_grid(my_wf,resamples = splits,grid = my_grid)

### Show Best
model_tuning %>% select_best("rmse")

验证说明 #

修正后，tidymodels生成的特征集合（主变量+全双向交互）会和Caret完全一致，结合已复用的交叉验证折，最终的超参数调优结果（lambda值）应该会和Caret的0.9014776一致。

内容的提问来源于stack exchange，提问作者ACN