如何在mlr包中的makeClassifTask()中包含阻塞因子?

时间:2016-11-04 12:11:41

标签: r mlr

在某些分类任务中,使用mlr包,我需要处理与此类似的data.frame

set.seed(pi)
# Dummy data frame
df <- data.frame(
   # Repeated values ID
   ID = sort(sample(c(0:20), 100, replace = TRUE)),
   # Some variables
   X1 = runif(10, 1, 10),
   # Some Label
   Label = sample(c(0,1), 100, replace = TRUE)
   )
df 

我需要交叉验证模型,将值与ID保持在一起,我从教程中了解到:

https://mlr-org.github.io/mlr-tutorial/release/html/task/index.html#further-settings

  

我们可以在任务中包含阻塞因子。这表明某些观察结果“属于一起”,在将数据拆分为训练和测试集进行重采样时不应分开。

问题是如何在makeClassifTask

中包含此阻止因素

不幸的是,我找不到任何例子。

2 个答案:

答案 0 :(得分:4)

你有什么版本的mlr?阻塞应该是它的一部分,因为一段时间。您可以在makeClassifTask

中直接将其作为参数找到

以下是您的数据示例:

df$ID = as.factor(df$ID)
df2 = df
df2$ID = NULL
df2$Label = as.factor(df$Label)
tsk = makeClassifTask(data = df2, target = "Label", blocking = df$ID)
res = resample("classif.rpart", tsk, resampling = cv10)

# to prove-check that blocking worked
lapply(1:10, function(i) {
  blocks.training = df$ID[res$pred$instance$train.inds[[i]]]
  blocks.testing = df$ID[res$pred$instance$test.inds[[i]]]
  intersect(blocks.testing, blocks.training)
})
#all entries are empty, blocking indeed works! 

答案 1 :(得分:0)

@ jakob-r的答案不再起作用。我的猜测是cv10有所改变。

较小的修改以使用“ blocking.cv = TRUE”

完整的工作示例:

set.seed(pi)
# Dummy data frame
df <- data.frame(
   # Repeated values ID
   ID = sort(sample(c(0:20), 100, replace = TRUE)),
   # Some variables
   X1 = runif(10, 1, 10),
   # Some Label
   Label = sample(c(0,1), 100, replace = TRUE)
   )
df 

df$ID = as.factor(df$ID)
df2 = df
df2$ID = NULL
df2$Label = as.factor(df$Label)
resDesc <- makeResampleDesc("CV",iters=10,blocking.cv = TRUE)
tsk = makeClassifTask(data = df2, target = "Label", blocking = df$ID)
res = resample("classif.rpart", tsk, resampling = resDesc)

# to prove-check that blocking worked
lapply(1:10, function(i) {
  blocks.training = df$ID[res$pred$instance$train.inds[[i]]]
  blocks.testing = df$ID[res$pred$instance$test.inds[[i]]]
  intersect(blocks.testing, blocks.training)
})