如何在MLR3管道中为仅用于分层的分类变量使用SMOTE上采样？

我用MLR3管道构建二分类模型，数据集存在类别不平衡，因此采用SMOTE做过采样补偿。所有预测变量均为数值型，但有两个分类变量仅用于和目标变量一同完成训练/测试拆分、嵌套交叉验证拆分的分层操作，不会作为预测变量使用。

遇到的问题：即便SMOTE不使用这些分层变量，只要它们存在于任务中，SMOTE就无法插入新生成的数据，报错信息如下：

Error: Cannot rbind data to task 'diagnostic', missing the following mandatory columns: site, hiv
This happened PipeOp smote's $train()
Execution halted

想请教：有没有无需完全移除分层变量的解决办法？还是这是MLR3中SMOTE实现的固有局限？另外我试过在SMOTE PipeOp前仅选择预测变量，但没有效果，问题似乎出在SMOTE尝试将新数据合并回原有训练数据的环节。

示例代码 #

library(mlr3verse)
library(data.table)
library(doFuture)

plan(multisession, workers = availableCores()-1)

dt <- data.table(
  record_id = 1:100,
  outcome = sample(c('negative','positive'),100,T,c(0.7,0.3)),
  group = sample(c('A','B','C'),100,T,c(0.4,0.3,0.3)),
  x = rnorm(100),
  y = rnorm(100, 3, 1.5),
  z = rnorm(100, -5, 2)
)

task <- as_task_classif(dt, target='outcome', id='test', positive='positive')

task$set_col_roles('record_id', roles = 'name')
task$set_col_roles('outcome', roles = c('target','stratum'))
task$set_col_roles('group', roles = 'stratum')

set.seed(682)
splits <- partition(task, ratio = 0.75, stratify = T)

po_scaler <- po("scale", center = TRUE, scale = TRUE)

po_smote <- po("smote")

lrn_glmnet <- lrn("classif.glmnet",
                  predict_type = "prob",
                  fallback = lrn("classif.featureless", predict_type = "prob"),
                  encapsulate = c(train = "evaluate", predict = "evaluate"),
                  id = "glmnet"
)

graph <- po_scaler %>>% po_smote %>>% lrn_glmnet

search_space <- ps(
  glmnet.alpha = p_dbl(0, 1),
  glmnet.lambda = p_dbl(0.01, 10, logscale = TRUE)
)

glearner = as_learner(graph)

glearner$id = "glearner"
glearner$predict_type = "prob"

inner_resampling = rsmp("repeated_cv", folds = 3, repeats = 10)


at = auto_tuner(
  tuner = tnr("mbo"),
  learner = glearner,
  resampling = inner_resampling,
  measure = msr("classif.auc"),
  search_space = search_space,
  terminator =trm("combo", 
                  list(
                    trm("stagnation", iters = 10, threshold = 0.0001),
                    trm("clock_time", stop_time = Sys.time() + 60^2)
                  ),
                  any = TRUE))

rsmp_rcv <- rsmp("cv", folds = 5)

rr = resample(task, at, rsmp_rcv)

这个问题的核心原因是：MLR3的PipeOpSmote仅基于标记为feature的列生成新样本，新样本中不会包含stratum（分层）角色的列；但这些分层列属于任务的强制列，SMOTE尝试将新样本合并回原任务时，就会因缺少这些列而报错。

无需移除分层变量的解决思路是拆分管道分支：一个分支处理SMOTE过采样，另一个分支保留分层变量，最后合并两个分支的结果，这样既保留了分层变量用于拆分，又能正常执行SMOTE。

修改后的完整管道代码如下：

library(mlr3verse)
library(data.table)
library(doFuture)

plan(multisession, workers = availableCores()-1)

dt <- data.table(
  record_id = 1:100,
  outcome = sample(c('negative','positive'),100,T,c(0.7,0.3)),
  group = sample(c('A','B','C'),100,T,c(0.4,0.3,0.3)),
  x = rnorm(100),
  y = rnorm(100, 3, 1.5),
  z = rnorm(100, -5, 2)
)

task <- as_task_classif(dt, target='outcome', id='test', positive='positive')

task$set_col_roles('record_id', roles = 'name')
task$set_col_roles('outcome', roles = c('target','stratum'))
task$set_col_roles('group', roles = 'stratum')

set.seed(682)
splits <- partition(task, ratio = 0.75, stratify = T)

po_scaler <- po("scale", center = TRUE, scale = TRUE)
po_smote <- po("smote")

lrn_glmnet <- lrn("classif.glmnet",
                  predict_type = "prob",
                  fallback = lrn("classif.featureless", predict_type = "prob"),
                  encapsulate = c(train = "evaluate", predict = "evaluate"),
                  id = "glmnet"
)

# -------------------------- 关键修改部分 --------------------------
# 1. 复制任务，生成两个分支
po_copy <- po("copy", outnum = 2)
# 分支1：仅保留特征+目标变量，执行标准化和SMOTE
branch_smote <- po("select", selector = selector_invert(c("group"))) %>>% po_scaler %>>% po_smote
# 分支2：仅保留分层变量和目标变量（确保后续合并时匹配样本）
branch_stratum <- po("select", selector = c("group", "outcome"))
# 2. 合并两个分支的结果
po_merge <- po("cbind", param_vals = list(keep_original_cols = TRUE))
# 3. 构建完整管道
graph <- po_copy %>>% gunion(list(branch_smote, branch_stratum)) %>>% po_merge %>>% lrn_glmnet
# -----------------------------------------------------------------

search_space <- ps(
  glmnet.alpha = p_dbl(0, 1),
  glmnet.lambda = p_dbl(0.01, 10, logscale = TRUE)
)

glearner = as_learner(graph)
glearner$id = "glearner"
glearner$predict_type = "prob"

inner_resampling = rsmp("repeated_cv", folds = 3, repeats = 10)

at = auto_tuner(
  tuner = tnr("mbo"),
  learner = glearner,
  resampling = inner_resampling,
  measure = msr("classif.auc"),
  search_space = search_space,
  terminator =trm("combo", 
                  list(
                    trm("stagnation", iters = 10, threshold = 0.0001),
                    trm("clock_time", stop_time = Sys.time() + 60^2)
                  ),
                  any = TRUE))

rsmp_rcv <- rsmp("cv", folds = 5)

rr = resample(task, at, rsmp_rcv)

逻辑说明 #

• po_copy将输入任务复制为两份，分别进入不同分支；
• branch_smote分支移除分层变量，仅处理用于建模的特征列，避免SMOTE生成样本时丢失强制列的问题；
• branch_stratum分支保留分层变量和目标变量，确保后续合并后任务仍包含分层信息；
• po_merge将两个分支的结果合并，得到包含所有必要列（特征、目标、分层变量）的任务，供后续模型训练使用。

这种方式既保留了分层变量用于交叉验证的分层拆分，又能正常执行SMOTE过采样，无需移除分层变量。

内容的提问来源于stack exchange，提问作者Gian