R使用ddply或聚合

时间:2012-12-27 02:37:31

标签: r aggregate plyr

我有一个包含3列的数据框:custId,saleDate,DelivDateTime。

> head(events22)
     custId            saleDate      DelivDate
1 280356593 2012-11-14 14:04:59 11/14/12 17:29
2 280367076 2012-11-14 17:04:44 11/14/12 20:48
3 280380097 2012-11-14 17:38:34 11/14/12 20:45
4 280380095 2012-11-14 20:45:44 11/14/12 23:59
5 280380095 2012-11-14 20:31:39 11/14/12 23:49
6 280380095 2012-11-14 19:58:32 11/15/12 00:10

这是dput:

> dput(events22)
structure(list(custId = c(280356593L, 280367076L, 280380097L, 
280380095L, 280380095L, 280380095L, 280364279L, 280364279L, 280398506L, 
280336395L, 280364376L, 280368458L, 280368458L, 280368456L, 280368456L, 
280364225L, 280391721L, 280353458L, 280387607L, 280387607L), 
    saleDate = structure(c(1352901899.215, 1352912684.484, 1352914714.971, 
    1352925944.429, 1352925099.247, 1352923112.636, 1352922476.55, 
    1352920666.968, 1352915226.534, 1352911135.077, 1352921349.592, 
    1352911494.975, 1352910529.86, 1352924755.295, 1352907511.476, 
    1352920108.577, 1352906160.883, 1352905925.134, 1352916810.309, 
    1352916025.673), class = c("POSIXct", "POSIXt"), tzone = "UTC"), 
    DelivDate = c("11/14/12 17:29", "11/14/12 20:48", "11/14/12 20:45", 
    "11/14/12 23:59", "11/14/12 23:49", "11/15/12 00:10", "11/14/12 23:35", 
    "11/14/12 22:59", "11/14/12 20:53", "11/14/12 19:52", "11/14/12 23:01", 
    "11/14/12 19:47", "11/14/12 19:42", "11/14/12 23:31", "11/14/12 23:33", 
    "11/14/12 22:45", "11/14/12 18:11", "11/14/12 18:12", "11/14/12 19:17", 
    "11/14/12 19:19")), .Names = c("custId", "saleDate", "DelivDate"
), row.names = c("1", "2", "3", "4", "5", "6", "7", "8", "9", 
"10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20"
), class = "data.frame")

我正在为每个DelivDate找到最新saleDate的{​​{1}}。

我可以像这样使用plyr :: ddply来做到这一点:

custId

我的问题是,是否有更快的方法来执行此操作,因为ddply方法有点耗时(完整数据集约为400k行)。我已经看过使用dd1 <-ddply(events22, .(custId),.inform = T, function(x){ x[x$saleDate == max(x$saleDate),"DelivDate"] }) ,但不知道如何获得除我正在排序的值之外的值。

有什么建议吗?

编辑:

以下是10k行@ 10次迭代的基准测试结果:

aggregate()

EDIT2: 虽然最快的AGG2()没有给出正确的答案。

      test replications elapsed relative user.self
2   AGG2()           10    5.96    1.000      5.93
1   AGG1()           10   20.87    3.502     20.75
5 DATATABLE()        10   61.32        1     60.31
3  DDPLY()           10   80.04   13.430     79.63
4 DOCALL()           10   90.43   15.173     88.39

4 个答案:

答案 0 :(得分:10)

我也会在此推荐data.table,但由于您要求aggregate解决方案,因此这里有一个结合aggregatemerge来获取所有列的解决方案:

merge(events22, aggregate(saleDate ~ custId, events22, max))

如果您只想要“custId”和“DelivDate”列,则只需aggregate

aggregate(list(DelivDate = events22$saleDate), 
          list(custId = events22$custId),
          function(x) events22[["DelivDate"]][which.max(x)])

最后,这是使用sqldf的选项:

library(sqldf)
sqldf("select custId, DelivDate, max(saleDate) `saleDate` 
      from events22 group by custId")

基准

我不是基准测试或data.table专家,但令我感到惊讶的是data.table在这里并不快。 我怀疑在较大的数据集上的结果会有很大不同,例如,你的400k第一行。无论如何,这里有一些基准代码modeled after @mnel's answer here,因此您可以对实际数据集进行一些测试以供将来参考。

library(rbenchmark)

首先,根据您想要的基准设置您的功能。

DDPLY <- function() { 
  x <- ddply(events22, .(custId), .inform = T, 
             function(x) {
               x[x$saleDate == max(x$saleDate),"DelivDate"]}) 
}
DATATABLE <- function() { x <- dt[, .SD[which.max(saleDate), ], by = custId] }
AGG1 <- function() { 
  x <- merge(events22, aggregate(saleDate ~ custId, events22, max)) }
AGG2 <- function() { 
  x <- aggregate(list(DelivDate = events22$saleDate), 
                 list(custId = events22$custId),
                 function(x) events22[["DelivDate"]][which.max(x)]) }
SQLDF <- function() { 
  x <- sqldf("select custId, DelivDate, max(saleDate) `saleDate` 
             from events22 group by custId") }
DOCALL <- function() {
  do.call(rbind, 
          lapply(split(events22, events22$custId), function(x){
            x[which.max(x$saleDate), ]
          })
  )
}

其次,进行基准测试。

benchmark(DDPLY(), DATATABLE(), AGG1(), AGG2(), SQLDF(), DOCALL(), 
          order = "elapsed")[1:5]
#          test replications elapsed relative user.self
# 4      AGG2()          100   0.285    1.000     0.284
# 3      AGG1()          100   0.891    3.126     0.896
# 6    DOCALL()          100   1.202    4.218     1.204
# 2 DATATABLE()          100   1.251    4.389     1.248
# 1     DDPLY()          100   1.254    4.400     1.252
# 5     SQLDF()          100   2.109    7.400     2.108

答案 1 :(得分:7)

ddplyaggregate之间的最快速度,我认为会aggregate,特别是对于您拥有的大量数据。但是,最快的是data.table

require(data.table)
dt <- data.table(events22)
dt[, .SD[which.max(saleDate),], by=custId]

来自?data.table.SDdata.table,其中包含x的子集           每组的数据,不包括组列。

答案 2 :(得分:3)

这应该非常快,但data.table可能更快:

do.call(rbind, 
    lapply(split(events22, events22$custId), function(x){
        x[which.max(x$saleDate), ]
    })
)

答案 3 :(得分:2)

这里有一个更快data.table的功能:

DATATABLE <- function() { 
  dt <- data.table(events, key=c('custId', 'saleDate'))
  dt[, maxrow := 1:.N==.N, by = custId]
  return(dt[maxrow==TRUE, list(custId, DelivDate)])
}

请注意,此功能会创建data.table并对数据进行排序,这是您只需执行一次的步骤。如果您删除此步骤(可能您有一个多步骤数据处理管道,并创建data.table一次,作为第一步),该功能的速度是原来的两倍。

我还修改了所有以前的函数以返回结果,以便于比较:

DDPLY <- function() { 
  return(ddply(events, .(custId), .inform = T, 
               function(x) {
                 x[x$saleDate == max(x$saleDate),"DelivDate"]}))
}
AGG1 <- function() { 
  return(merge(events, aggregate(saleDate ~ custId, events, max)))}

SQLDF <- function() { 
  return(sqldf("select custId, DelivDate, max(saleDate) `saleDate` 
             from events group by custId"))}
DOCALL <- function() {
  return(do.call(rbind, 
                 lapply(split(events, events$custId), function(x){
                   x[which.max(x$saleDate), ]
                 })
  ))
}

这里是10k行的结果,重复10次:

library(rbenchmark)
library(plyr)
library(data.table)
library(sqldf)
events <- do.call(rbind, lapply(1:500, function(x) events22))
events$custId <- sample(1:nrow(events), nrow(events))

benchmark(a <- DDPLY(), b <- DATATABLE(), c <- AGG1(), d <- SQLDF(),
 e <- DOCALL(), order = "elapsed", replications=10)[1:5]

              test replications elapsed relative user.self
2 b <- DATATABLE()           10    0.13    1.000      0.13
4     d <- SQLDF()           10    0.42    3.231      0.41
3      c <- AGG1()           10   12.11   93.154     12.03
1     a <- DDPLY()           10   32.17  247.462     32.01
5    e <- DOCALL()           10   56.05  431.154     55.85

由于所有函数都返回结果,我们可以验证它们都返回相同的答案:

c <- c[order(c$custId),]
dim(a); dim(b); dim(c); dim(d); dim(e)
all(a$V1==b$DelivDate)
all(a$V1==c$DelivDate)
all(a$V1==d$DelivDate)
all(a$V1==e$DelivDate)

/编辑:在较小的20行数据集上,data.table仍然是最快的,但是更薄的边距:

              test replications elapsed relative user.self
2 b <- DATATABLE()          100    0.22    1.000      0.22
3      c <- AGG1()          100    0.42    1.909      0.42
5    e <- DOCALL()          100    0.48    2.182      0.49
1     a <- DDPLY()          100    0.55    2.500      0.55
4     d <- SQLDF()          100    1.00    4.545      0.98

/ Edit2:如果我们从函数中删除data.table创建,我们会得到以下结果:

dt <- data.table(events, key=c('custId', 'saleDate'))
DATATABLE2 <- function() { 
  dt[, maxrow := 1:.N==.N, by = custId]
  return(dt[maxrow==TRUE, list(custId, DelivDate)])
}
benchmark(a <- DDPLY(), b <- DATATABLE2(), c <- AGG1(), d <- SQLDF(),
           e <- DOCALL(), order = "elapsed", replications=10)[1:5]
              test replications elapsed relative user.self
2 b <- DATATABLE()           10    0.09    1.000      0.08
4     d <- SQLDF()           10    0.41    4.556      0.39
3      c <- AGG1()           10   11.73  130.333     11.67
1     a <- DDPLY()           10   31.59  351.000     31.50
5    e <- DOCALL()           10   55.05  611.667     54.91