R符号:列车与rfe的区别

Question

由于在没有内置的特性选择的插入符号-ml中有一些可用的方法，所以我正在试验rfe。

，有人能解释一下， train()和rfe()之间的区别是什么?除了训练时间和排列特征列表之外？是否只有train()才能做到，反之亦然？

下面是我的例子来说明这个问题：

# worker functions
useRFE <- function(trainctrl) {
    control <- rfeControl(functions=caretFuncs, 
    method="cv", 
    repeats=5, 
    number=5, 
    returnResamp="final", 
    verbose = TRUE)


 # run the RFE algorithm
 resultsRFE <-
    rfe(
       training[,1:ncol(iris) - 1], 
       training$Species, 
       sizes = subsets, 
       rfeControl = control,
       method = "svmRadial",
       metric = "Accuracy",
       trControl = trainctrl
    )

  # test
  predictedClasses <- predict(resultsRFE, testing)
  confmat <- confusionMatrix(data = predictedClasses$pred, reference = testing$Species)
  confmat
  # Balanced Accuracy = 100%
}

useTrain <-  function(trainctrl) {      
  # train
  resultsTrain <- train(
       Species ~ .,
       data = training,
       method = "svmRadial",
       metric = "Accuracy",
       trControl = trainctrl
  )

  # test
  predictedClasses <- predict(resultsTrain, testing)
  confmat <- confusionMatrix(data = predictedClasses, 
       reference = testing$Species)
  confmat
  # Balanced Accuracy = 100%
}
# load library
library(caret)

# load iris data
data(iris)

# create datasets
inTrain <- createDataPartition(y = iris$Species,
    p = .75,
    list = FALSE)

## The output is a set of integers for the rows of Sonar
## that belong in the training set.
training <<- iris[ inTrain,]
testing <<- iris[-inTrain,]

# select count of features
subsets <- c(1:ncol(iris)-1)

# set controls
trainctrl <- trainControl(
    classProbs= TRUE, 
    method="cv", 
    repeats=5, 
    number=5, 
    verbose = TRUE)

# compare
ptm <- proc.time()
useTrain(trainctrl = trainctrl)
proc.time() - ptm
# Accuracy : 0.9167    
# User      System     passed 
# 1.54      0.06       10.27

ptm <- proc.time()
useRFE(trainctrl = trainctrl)
proc.time() - ptm
# Accuracy : 0.9167
# User      System     passed 
# 0.86      0.04       24.44  

Answer 1

train 建筑与曲调 a更多的是以所有的预测器作为输入。rfe进行递归特性消除，其中不同的预测器被移除。试图优化模型性能。您可以将rfe看作train的一个可能的包装器。

对这些方法的解释都在网站上。