1. 技术文章
  2. R use ddply or aggregate

R use ddply or aggregate

由 被刻印的时光 ゝ 提交于 2019-11-27 09:15:57
A5C1D2H2I1M1N2O1R2T1

I, too, would recommend data.table here, but since you asked for an aggregate solution, here is one which combines aggregate and merge to get all the columns:

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

Or just aggregate if you only want the "custId" and "DelivDate" columns:

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

Finally, here's an option using sqldf:

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

Benchmarks

I'm not a benchmarking or data.table expert, but it surprised me that data.table is not faster here. My suspicion is that the results would be quite different on a larger dataset, say for instance, your 400k lines one. Anyway, here's some benchmarking code modeled after @mnel's answer here so you can do some tests on your actual dataset for future reference.

library(rbenchmark)

First, set up your functions for what you want to benchmark.

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), ]
          })
  )
}

Second, do the benchmarking.

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

The fastest between ddply and aggregate, I suppose would be aggregate, especially on huge data as you have. However, the fastest would be data.table.

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

From ?data.table: .SD is a data.table containing the subset of x's Data for each group, excluding the group column(s).

This should be pretty fast but data.table is likely faster:

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

Here's a much faster data.table function:

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

Note that this function creates a data.table and sorts the data, which is a step you'd only need to perform once. If you remove this step (perhaps you have a multi-step data processing pipeline, and create the data.table once, as a first step), the function is more than twice as fast.

I also modified all the previous functions to return the result, for easier comparison:

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), ]
                 })
  ))
}

Here's the results for 10k rows, repeated 10 times:

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

Since all the functions return their results, we can verify they all return the same answer:

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)

/Edit: On the smaller, 20 row dataset, data.table is still the fastest, but by a thinner margin:

              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: If we remove the data.table creation from the function we get the following results:

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
标签
r
aggregate
plyr
易学教程内所有资源均来自网络或用户发布的内容，如有违反法律规定的内容欢迎反馈！
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!