Fastest way to remove all duplicates in R

Question

I\'d like to remove all items that appear more than once in a vector. Specifically, this includes character, numeric and integer vectors. Currently, I\'m using duplica

Answer 1

You could use a set operation:

d <- c(1,2,3,4,1,5,6,4,2,1)
duplicates = d[duplicated(d)]
setdiff(d, duplicates)
[1] 3 5 6

(Not certain if that is more efficient than the above code but it does seem conceptually cleaner)

Answer 2

Some timings:

set.seed(1001)
d <- sample(1:100000, 100000, replace=T)
d <- c(d, sample(d, 20000, replace=T))  # ensure many duplicates
mb <- microbenchmark::microbenchmark(
  d[!(duplicated(d) | duplicated(d, fromLast=TRUE))],
  setdiff(d, d[duplicated(d)]),
  {tmp <- rle(sort(d)); tmp$values[tmp$lengths == 1]},
  as.integer(names(table(d)[table(d)==1])),
  d[!(duplicated.default(d) | duplicated.default(d, fromLast=TRUE))],
  d[!(d %in% d[duplicated(d)])],
  { ud = unique(d); ud[tabulate(match(d, ud)) == 1L] },
  d[!(.Internal(duplicated(d, F, F, NA)) | .Internal(duplicated(d, F, T, NA)))]
)
summary(mb)[, c(1, 4)]  # in milliseconds
#                                                                                expr      mean
#1                               d[!(duplicated(d) | duplicated(d, fromLast = TRUE))]  18.34692
#2                                                       setdiff(d, d[duplicated(d)])  24.84984
#3                       {     tmp <- rle(sort(d))     tmp$values[tmp$lengths == 1] }   9.53831
#4                                         as.integer(names(table(d)[table(d) == 1])) 255.76300
#5               d[!(duplicated.default(d) | duplicated.default(d, fromLast = TRUE))]  18.35360
#6                                                      d[!(d %in% d[duplicated(d)])]  24.01009
#7                        {     ud = unique(d)     ud[tabulate(match(d, ud)) == 1L] }  32.10166
#8 d[!(.Internal(duplicated(d, F, F, NA)) | .Internal(duplicated(d,      F, T, NA)))]  18.33475

Given the comments let's see if they are all correct?

 results <- list(d[!(duplicated(d) | duplicated(d, fromLast=TRUE))],
         setdiff(d, d[duplicated(d)]),
         {tmp <- rle(sort(d)); tmp$values[tmp$lengths == 1]},
         as.integer(names(table(d)[table(d)==1])),
         d[!(duplicated.default(d) | duplicated.default(d, fromLast=TRUE))],
         d[!(d %in% d[duplicated(d)])],
         { ud = unique(d); ud[tabulate(match(d, ud)) == 1L] },
         d[!(.Internal(duplicated(d, F, F, NA)) | .Internal(duplicated(d, F, T, NA)))])
 all(sapply(ls, all.equal, c(3, 5, 6)))
 # TRUE

Answer 3

You can do this with rle function:

tmp <- rle(sort(d))
res <- tmp$values[tmp$lengths == 1]

The idea is to find the count of same values in the vector.

There are plenty of alternatives here: Counting the number of elements with the values of x in a vector

Edit

After looking at the benchmarks, @NBATrends I got suspicious. In theory counting items with a single pass through must be ~2x faster compared to original duplicated logic.

I tried doing this with data.table:

library(data.table)
dt <- data.table(d)
res <-  dt[, count:= .N, by = d][count == 1]$d

And here are the benchmarks on different sample sizes for three solutions (I have reduced it to fast unique approaches):

You can see that with the growth of the sample data.table begins to outperform other methods (2x).

Here is the code to reproduce:

set.seed(1001)
N <- c(3, 4, 5, 6 ,7)
n <- 10^N
res <- lapply(n, function(x) {
d <- sample(1:x/10, 5 * x, replace=T)
d <- c(d, sample(d, x, replace=T))  # ensure many duplicates
dt <- data.table(d)
mb <- microbenchmark::microbenchmark(
  "duplicated(original)" = d[!(duplicated(d) | duplicated(d, fromLast=TRUE))],
  "tabulate" = { ud = unique(d); ud[tabulate(match(d, ud)) == 1L] },
  "data.table" = dt[, count:= .N, by = d][count == 1]$d,
  times = 1,unit = "ms")
sm <- summary(mb)[, c(1, 4, 8)]
sm$size = x
return(sm)

})

res <- do.call("rbind", res)

require(ggplot2)
##The values Year, Value, School_ID are
##inherited by the geoms
ggplot(res, aes(x = res$size, y = res$mean, colour=res$exp)) + 
geom_line() + scale_x_log10() + scale_y_log10() +
geom_point()