如何用mlr3pipelines移除高比例NA值的列？我的尝试未生效

问题原因 #

你当前的代码中，通过参数转换（trafo）动态生成的selector函数无法正确捕获na_cutoff变量的值，导致筛选逻辑未按预期执行——本该被移除的Sepal.Width列（NA占比约33%，超过默认阈值0.2）没有被过滤掉。

修复方案1：修正作用域绑定问题 #

通过force()函数强制绑定na_cutoff的当前值，避免延迟解析导致的作用域问题：

library(mlr3)
library(mlr3pipelines)

task = tsk("iris")
dt = task$data()
dt[1:50, Sepal.Width := NA]
task_ = as_task_classif(dt, target = "Species")

graph = po("removeconstants", id = "removeconstants", ratio = 0.01) %>>%
  po("select", id = "drop_na_cols")
ps = ParamSet$new(list(ParamDbl$new("na_cutoff", lower = 0, upper = 1, default = 0.2)))
graph$param_set$add(ps)

graph$param_set$trafo = function(x, param_set) {
  na_cutoff = x$na_cutoff
  # 强制绑定当前na_cutoff的值，避免作用域延迟解析
  force(na_cutoff)
  x$drop_na_cols.selector = function(task) {
    fn = task$feature_names
    data = task$data(cols = fn)
    na_ratios = colMeans(is.na(data))
    # 保留NA占比不超过阈值的列
    fn[na_ratios <= na_cutoff]
  }
  x$na_cutoff = NULL
  x
}

train_res = graph$train(task_)
# 验证结果：Sepal.Width已被移除
train_res$drop_na_cols.output$data()

修复方案2：自定义PipeOp（更直观可靠） #

直接创建一个专门用于移除高NA占比列的PipeOp，逻辑清晰且避免作用域问题：

library(mlr3)
library(mlr3pipelines)

# 自定义移除高NA列的PipeOp
PipeOpRemoveHighNA = R6::R6Class("PipeOpRemoveHighNA",
  inherit = PipeOpTaskPreproc,
  public = list(
    initialize = function(id = "remove_high_na", param_vals = list()) {
      ps = ParamSet$new(list(
        ParamDbl$new("na_cutoff", lower = 0, upper = 1, default = 0.2)
      ))
      super$initialize(id, param_set = ps, param_vals = param_vals)
    }
  ),
  private = list(
    .train_task = function(task) {
      na_cutoff = self$param_set$values$na_cutoff
      fn = task$feature_names
      na_ratios = colMeans(is.na(task$data(cols = fn)))
      self$state$keep_cols = fn[na_ratios <= na_cutoff]
      task$select(self$state$keep_cols)
    },
    .predict_task = function(task) {
      # 预测阶段复用训练时确定的保留列，保证前后一致
      task$select(self$state$keep_cols)
    }
  )
)

# 测试代码
task = tsk("iris")
dt = task$data()
dt[1:50, Sepal.Width := NA]
task_ = as_task_classif(dt, target = "Species")

# 构建管道
graph = po("removeconstants", ratio = 0.01) %>>%
  PipeOpRemoveHighNA$new(param_vals = list(na_cutoff = 0.2))

train_res = graph$train(task_)
# 验证结果
train_res[[1]]$data()

内容的提问来源于stack exchange，提问作者Mislav Sagovac