Question

对于每个级别的因子，我需要提取在data.frame的所有子集上聚合的值，除了当前的一个。例如，有几个受试者在几天内做了反应时间任务，我需要计算所有受试者和所有日子的平均反应时间，但不包括计算平均值的受试者。目前，我这样做：

 library(lme4)
 ddply(sleepstudy, .(Subject, Days), summarise, 
       avg_rt = mean(sleepstudy[sleepstudy$Subject != Subject &
                   sleepstudy$Days == Days,"Reaction"]), .progress="text")

它适用于小型数据集，但对于大型数据集，它可能非常慢。有没有办法更快地完成它？

Answer 1

#create big dataset
n <- 1e4
set.seed(1)
sleepstudy <- data.frame(Reaction=rnorm(n),Subject=1:4,Days=sort(rep((1:(n/4)),4)))


library(plyr)
system.time(
  res <- ddply(sleepstudy, .(Subject, Days), summarise, 
               avg_rt = mean(sleepstudy[sleepstudy$Subject != Subject &
                 sleepstudy$Days == Days,"Reaction"]))
)
#User      System      elapsed 
#6.532       0.013       6.556  

#use data.table for big datasets
library(data.table)

dt<- as.data.table(sleepstudy)
system.time(
 {dt[,avg_rt:=mean(Reaction),by=Days];
  dt[,n:=.N,by=Days];
  dt[,avg_rt:=(avg_rt*n-Reaction)/(n-1)]}
)
#User      System      elapsed 
#0.005       0.001       0.005 


#test if results are equal
dt2 <- as.data.table(res)
setkey(dt2,Subject,Days)
setkey(dt,Subject,Days)
all.equal(dt[,avg_rt],dt2[,avg_rt])
#[1] TRUE

对于非常大的数据集，速度增益应该更加明显。我无法与较大的数据集进行比较，因为ddply非常慢。

Answer 2

使用lapply和aggregate可能会更快：

do.call("rbind", (lapply(unique(sleepstudy$Subject),
                         function(x)
                           cbind(Subject = x,
                                 aggregate(Reaction ~ Days,
                                           subset(sleepstudy, Subject != x),
                                           mean)))))

<强>更新

我将这两个命令与system.time进行了比较，看起来原来的速度较慢。

library(lme4)
library(plyr)

system.time(
ddply(sleepstudy, .(Subject, Days), summarise, 
      avg_rt = mean(sleepstudy[sleepstudy$Subject != Subject &
                    sleepstudy$Days == Days,"Reaction"]))
)

   # user  system elapsed 
   # 0.17    0.00    0.22 

system.time(
do.call("rbind", (lapply(unique(sleepstudy$Subject),
                         function(x) 
                           cbind(Subject = x,
                                 aggregate(Reaction ~ Days,
                                           subset(sleepstudy, Subject != x),
                                           mean)))))
)


   # user  system elapsed 
   # 0.12    0.00    0.12

对于除当前级别之外的所有级别的每个级别的因子聚合值

2 个答案: