Using dplyr to make sample from data frame

Question

I have a very large data frame (150.000.000 rows) with a format like this:

df = data.frame(pnr = rep(500+2*(1:15),each=3), x = runif(3*15))

pnr is person id and x is some data. I would like to sample 10% of the persons. Is there a fast way to do this in dplyr?

The following is a solution, but it is slow because of the merge-statement

prns = as.data.frame(unique(df$prn))
names(prns)[1] = "prn"
prns$s = rbinom(nrow(prns),1,0.1)

df = merge(df,prns)
df2 = df[df$s==1,]

Answer 1

I would actually suggest the "data.table" package over "dplyr" for this. Here's an example with some big-ish sample data (not much smaller than your own 15 million rows).

I'll also show some right and wrong ways to do things :-)

Here's the sample data.

library(data.table)
library(dplyr)
library(microbenchmark)
set.seed(1)
mydf <- DT <- data.frame(person = sample(10000, 1e7, TRUE),
                   value = runif(1e7))

We'll also create a "data.table" and set the key to "person". Creating the "data.table" takes no significant time, but setting the key can.

system.time(setDT(DT))
#    user  system elapsed 
#   0.001   0.000   0.001 

## Setting the key takes some time, but is worth it
system.time(setkey(DT, person)) 
#    user  system elapsed 
#   0.620   0.025   0.646

I can't think of a more efficient way to select your "person" values than the following, so I've removed these from the benchmarks--they are common to all approaches.

## Common to all tests...
A <- unique(mydf$person)
B <- sample(A, ceiling(.1 * length(A)), FALSE)

For convenience, the different tests are presented as functions...

## Base R #1
fun1a <- function() {
  mydf[mydf$person %in% B, ]
}

## Base R #2--sometimes using `which` makes things quicker
fun1b <- function() {
  mydf[which(mydf$person %in% B), ]
}

## `filter` from "dplyr"
fun2 <- function() {
  filter(mydf, person %in% B)
}

## The "wrong" way to do this with "data.table"
fun3a <- function() {
  DT[which(person %in% B)]
}

## The "right" (I think) way to do this with "data.table"
fun3b <- function() {
  DT[J(B)]
}

Now, we can benchmark:

## The benchmarking
microbenchmark(fun1a(), fun1b(), fun2(), fun3a(), fun3b(), times = 20)
# Unit: milliseconds
#     expr       min        lq    median        uq       max neval
#  fun1a() 382.37534 394.27968 396.76076 406.92431 494.32220    20
#  fun1b() 401.91530 413.04710 416.38470 425.90150 503.83169    20
#   fun2() 381.78909 394.16716 395.49341 399.01202 417.79044    20
#  fun3a() 387.35363 397.02220 399.18113 406.23515 413.56128    20
#  fun3b()  28.77801  28.91648  29.01535  29.37596  42.34043    20

Look at the performance we get from using "data.table" the right way! All the other approaches are impressively fast though.

---

summary shows the results to be the same. (The row order for the "data.table" solution would be different since it has been sorted.)

summary(fun1a())
#      person         value         
#  Min.   :  16   Min.   :0.000002  
#  1st Qu.:2424   1st Qu.:0.250988  
#  Median :5075   Median :0.500259  
#  Mean   :4958   Mean   :0.500349  
#  3rd Qu.:7434   3rd Qu.:0.749601  
#  Max.   :9973   Max.   :1.000000  

summary(fun2())
#      person         value         
#  Min.   :  16   Min.   :0.000002  
#  1st Qu.:2424   1st Qu.:0.250988  
#  Median :5075   Median :0.500259  
#  Mean   :4958   Mean   :0.500349  
#  3rd Qu.:7434   3rd Qu.:0.749601  
#  Max.   :9973   Max.   :1.000000  

summary(fun3b())
#      person         value         
#  Min.   :  16   Min.   :0.000002  
#  1st Qu.:2424   1st Qu.:0.250988  
#  Median :5075   Median :0.500259  
#  Mean   :4958   Mean   :0.500349  
#  3rd Qu.:7434   3rd Qu.:0.749601  
#  Max.   :9973   Max.   :1.000000 

Answer 2

If you don't necessarily want a thoroughly random sample, then you could do

filter(df, pnr %% 10 ==0).

Which would take every 10th person (you could get 10 different samples by changing to ==1,...). You could make this random by re-allocating IDs randomly - fairly trivial to do this using sample(15)[(df$pnr-500)/2] for your toy example - reversing the mapping of pnr onto a set that's suitable for sample might be less easy for the real-world case.

Answer 3

In base R, to sample 10% of the rows, rounding up to the next row

> df[sample(nrow(df), ceiling(0.1*nrow(df)), FALSE), ]
##    pnr         x
## 16 512 0.9781232
## 21 514 0.5279925
## 33 522 0.8332834
## 14 510 0.7989481
## 4  504 0.7825318

or rounding down to the next row

> df[sample(nrow(df), floor(0.1*nrow(df)), FALSE), ]
##    pnr           x
## 43 530 0.449985180
## 35 524 0.996350657
## 2  502 0.499871966
## 25 518 0.005199058

or sample 10% of the pnr column, rounding up

> sample(df$pnr, ceiling(0.1*length(df$pnr)), FALSE)
## [1] 530 516 526 518 514 

ADD:

If you're looking to sample 10% of the people (unique pnr ID), and return those people and their respective data, I think you want

> S <- sample(unique(df$pnr), ceiling(0.1*length(unique(df$pnr))), FALSE)
> df[df$pnr %in% S, ]
##    pnr         x
## 1  502 0.7630667
## 2  502 0.4998720
## 3  502 0.4839460
## 22 516 0.8248153
## 23 516 0.5795991
## 24 516 0.1572472

PS: I would wait for a dplyr answer. It will likely be quicker on 15mil rows.