Aggregating values in a data frame based on key

问题:

I've got a piece of aggregation code that works well enough but runs a bit slow against a data frame with 10e6 rows. I'm not that experienced in R so apologies for my cringe worthy code!

I just want to do a basic roll up and sum of values for a common key...

eg go from...

  key val
1   a   5
2   b   7
3   a   6

to...

  key val
1   a   11
2   b   7

the best i can manage is...

keys = unique(inp$key)
vals = sapply(keys, function(x) { sum(inp[inp$key==x,]$val) })
out = data.frame(key=keys, val=vals)

I have this gut feel that the inp[inp$key==x,] is not the best way. Is there an obvious speed up i'm missing? I can do it in Hadoop (since the 10e6 dataset is actually already a rollup from a 2e9 row dataset) but I'm trying to improve my R.

Cheers, Mat

回答1:

Another option using tapply:

dat <- data.frame(key = c('a', 'b', 'a'), val = c(5,7,6))

> with(dat, tapply(val, key, FUN = sum))
 a  b 
11  7

My tests indicate this is the fastest method for this particular exercise, obviously your miles may vary:

fn.tapply <- function(daters) with(daters, tapply(val, key, FUN = sum))
fn.aggregate <- function(daters) aggregate(val~key, sum, data = daters)
fn.ddply <- function(daters) ddply(daters, .(key), summarize, val = sum(val))


library(rbenchmark)

benchmark(fn.tapply(dat), fn.aggregate(dat), fn.ddply(dat)
          , columns = c("test", "elapsed", "relative")
          , order = "relative"
          , replications = 100
          )


               test elapsed  relative
1    fn.tapply(dat)    0.03  1.000000
2 fn.aggregate(dat)    0.20  6.666667
3     fn.ddply(dat)    0.30 10.000000

Note that converting the tapply solution into a data.frame cut this difference in half by ~40% for a true apples to apples comparison to the first two.

Using a 1M row dataset as indicated in the comments does seem to change things a bit:

 dat2 <- data.frame(key = rep(letters[1:5], each = 200000), val = runif(1e6))
> benchmark(fn.tapply(dat2), fn.aggregate(dat2), fn.ddply(dat2)
+           , columns = c("test", "elapsed", "relative")
+           , order = "relative"
+           , replications = 100
+           )
               test elapsed relative
1   fn.tapply(dat2)  39.114 1.000000
3     fn.ddply(dat2)  62.178 1.589661
2 fn.aggregate(dat2) 157.463 4.025745

回答2:

you can use aggregate:

> d
  key val
1   a   5
2   b   7
3   a   6
> aggregate(val~key, sum, data=d)
  key val
1   a  11
2   b   7

you can also use ddply from hadley's plyr package:

> ddply(d, .(key), summarize, val=sum(val))
  key val
1   a  11
2   b   7

回答3:

Using sapply and split is another option. I'll extend via the data and benchmarks from @Chase's excellent answer.

fn.tapply <- function(daters) with(daters, tapply(val, key, FUN = sum))
fn.split <- function(daters) with(daters, sapply(split(val, key), sum))

str(dat)
# 'data.frame': 1000000 obs. of  2 variables:
#  $ key: Factor w/ 5 levels "a","b","c","d",..: 1 1 1 1 1 1 1 1 1 1 ...
#  $ val: num  0.186 0.875 0.42 0.294 0.878 ...

benchmark(fn.tapply(dat), fn.split(dat)
          , columns = c("test", "elapsed", "relative")
          , order = "relative"
          , replications = 100
          )
#             test elapsed relative
# 2  fn.split(dat)   4.106  1.00000
# 1 fn.tapply(dat)  69.982 17.04384

回答4:

Technically you said "data frame", but another much faster option (over 22x faster) is to use the data.table package.

Here's a benchmark for 10 reps of 10e6 rows.

library(rbenchmark)
library(plyr)
library(data.table)

key <- sample(letters,10e6, replace = T)
val <- sample.int(10,10e6, replace = T)
big_df <- data.frame(key,val)
rm(key)
rm(val)
big_dtable <- data.table(big_df)
setkey(big_dtable,key)

fn.data_table <- function(mydata) mydata[,list(sum = sum(val)), by = 'key']
fn.tapply <- function(mydata) with(mydata, tapply(val, key, FUN = sum))
fn.aggregate <- function(mydata) aggregate(val~key, sum, data = mydata)
fn.ddply <- function(mydata) ddply(mydata, .(key), summarize, val = sum(val))

And now the benchmark....

benchmark(fn.data_table(big_dtable)
          , fn.tapply(big_df)
          , fn.aggregate(big_df)
          , fn.ddply(big_df)
          , fn.ddply(big_dtable)
          , columns = c("test","elapsed","relative")
          , order = "relative"
          , replications = 10         
          )

And the result....

                       test elapsed  relative
1 fn.data_table(big_dtable)    1.98   1.00000
5      fn.ddply(big_dtable)   37.59  18.98485
4          fn.ddply(big_df)   44.36  22.40404
2         fn.tapply(big_df)   51.03  25.77273
3      fn.aggregate(big_df)  238.52 120.46465