训练和测试数据的偏差图

Question

我为Decision树构建了一个deviance / splits图，如下所示。

  data(mtcars)
  cars <- mtcars[,1:4]
  smp_size <- floor(0.75 * nrow(cars))
  set.seed(100)
  train_ind <- sample(seq_len(nrow(LoanData)), size = smp_size)
  train <- LoanData[train_ind, ]
  test <- LoanData[-train_ind, ]

  #Deciscion tree model
  library(tree)
  car_tree <- tree(mpg ~., data=cars, mindev=0.003,mincut=2,minsize = 6)

  #Check deviance
  cv_tree <- cv.tree(car_tree,FUN=prune.tree, K=10)
  plot(cv_tree$size,cv_tree$dev,type="b",xlab="splits",ylab="deviance",
  main="deviance by splits")

我想在此图上叠加测试数据的偏差，以便观察之后的偏差开始再次增加。你能告诉我怎么做吗？

Answer 1

要查看树的不同深度的准确性，需要修剪树，预测训练和测试结果，并评估训练和测试结果的准确性。

这是数据，提取了训练和测试子集

data(mtcars)
cars <- mtcars
smp_size <- floor(0.75 * nrow(cars))
set.seed(100)
train_ind <- sample(seq_len(nrow(cars)), size = smp_size)
train <- cars[train_ind, ]
test <- cars[-train_ind, ]

这是一个辅助函数，用于确定给定模型的训练和测试数据的准确性。您可能希望修改此项以包含其他验证估算值。

compare<-function(tr, train, test, dpth, rst=NULL) {
  est.train <- predict(tr,train)
  est.test <- predict(tr,test)
  delta.train = est.train - train$mpg
  delta.test = est.test - test$mpg
  df <- data.frame(cor.train = cor(train$mpg,est.train),
                   cor.test = cor(test$mpg,est.test),
                   sd.train = sd(delta.train),
                   sd.test = sd(delta.test),
                   depth = dpth)
  return(rbind(rst,df))
}

创建树

#Deciscion tree model
library(tree)
car_tree <- tree(mpg ~., data=train, mindev=0.003,mincut=2,minsize = 6)

打印树，确定深度（3）和最深的分裂节点（4：7）

car_tree
# Depth is 3 and the 3rd level nodes are 4:7

获得深度为3的结果

rslts<-compare(car_tree,train,test,3)

现在修剪树，然后打印它。注意下一个深度是2，最深的节点是2：3

(car_tree_sn_1 <- snip.tree(car_tree,c(4:7)))
# Depth is 2 and the 2nd level nodes are 2:3

获得深度为2的结果

rslts<-compare(car_tree_sn_1,train,test,2,rslts)

现在修剪树，然后打印它。请注意，下一个深度为1，并且没有拆分节点

(car_tree_sn_2 <- snip.tree(car_tree,c(2:3)))
# Depth is 1 and there are no split nodes

获得深度为1的结果

rslts<-compare(car_tree_sn_2,train,test,1,rslts)

绘制准确度估算值

plot(rslts$depth,rslts$cor.train,type="b",xlab="splits",ylab="Correlation Coefficient",
     main="Correlation by splits",log="y",ylim=c(.5,1))
lines(rslts$depth,rslts$cor.test,type="b",col="red")

plot(rslts$depth,rslts$sd.train,type="b",xlab="splits",ylab="Standard Deviation",
     main="Correlation by splits",log="y",ylim=c(.5,5))
lines(rslts$depth,rslts$sd.test,type="b",col="red")

还有其他树模型。这是一个rpart的例子。

# Regression Tree Example
data(mtcars)
cars <- mtcars[,1:6]
library(rpart)

# grow tree 
fit <- rpart(mpg ~., data=cars, control=list(minsplit = 1))

printcp(fit) # display the results 
plotcp(fit) # visualize cross-validation results 
summary(fit) # detailed summary of splits

# create additional plots 
par(mfrow=c(1,2)) # two plots on one page 
rsq.rpart(fit) # visualize cross-validation results    

# plot tree 
library(rpart.plot)
prp(fit,extra=101,branch.type=3)
plot(fit, uniform=TRUE, 
     main="Regression Tree for Mileage ")
text(fit, use.n=TRUE, all=TRUE, cex=.8)

参考：cart tree，prp plotting，model prediction