Search code examples
rstringsplitr-faq

Split comma-separated strings in a column into separate rows

I have a data frame, like so:

data.frame(director = c("Aaron Blaise,Bob Walker", "Akira Kurosawa", 
                        "Alan J. Pakula", "Alan Parker", "Alejandro Amenabar", "Alejandro Gonzalez Inarritu", 
                        "Alejandro Gonzalez Inarritu,Benicio Del Toro", "Alejandro González Iñárritu", 
                        "Alex Proyas", "Alexander Hall", "Alfonso Cuaron", "Alfred Hitchcock", 
                        "Anatole Litvak", "Andrew Adamson,Marilyn Fox", "Andrew Dominik", 
                        "Andrew Stanton", "Andrew Stanton,Lee Unkrich", "Angelina Jolie,John Stevenson", 
                        "Anne Fontaine", "Anthony Harvey"), AB = c('A', 'B', 'A', 'A', 'B', 'B', 'B', 'A', 'B', 'A', 'B', 'A', 'A', 'B', 'B', 'B', 'B', 'B', 'B', 'A'))

As you can see, some entries in the director column are multiple names separated by commas. I would like to split these entries up into separate rows while maintaining the values of the other column. As an example, the first row in the data frame above should be split into two rows, with a single name each in the director column and 'A' in the AB column.

Solution

  • This old question frequently is being used as dupe target (tagged with r-faq). As of today, it has been answered three times offering 6 different approaches but is lacking a benchmark as guidance which of the approaches is the fastest1.

    The benchmarked solutions include

    Overall 8 different methods were benchmarked on 6 different sizes of data frames using the microbenchmark package (see code below).

    The sample data given by the OP consists only of 20 rows. To create larger data frames, these 20 rows are simply repeated 1, 10, 100, 1000, 10000, and 100000 times which give problem sizes of up to 2 million rows.

    Benchmark results

    enter image description here

    The benchmark results show that for sufficiently large data frames all data.table methods are faster than any other method. For data frames with more than about 5000 rows, Jaap's data.table method 2 and the variant DT3 are the fastest, magnitudes faster than the slowest methods.

    Remarkably, the timings of the two tidyverse methods and the splistackshape solution are so similar that it's difficult to distiguish the curves in the chart. They are the slowest of the benchmarked methods across all data frame sizes.

    For smaller data frames, Matt's base R solution and data.table method 4 seem to have less overhead than the other methods.

    Code

    director <- 
  c("Aaron Blaise,Bob Walker", "Akira Kurosawa", "Alan J. Pakula", 
    "Alan Parker", "Alejandro Amenabar", "Alejandro Gonzalez Inarritu", 
    "Alejandro Gonzalez Inarritu,Benicio Del Toro", "Alejandro González Iñárritu", 
    "Alex Proyas", "Alexander Hall", "Alfonso Cuaron", "Alfred Hitchcock", 
    "Anatole Litvak", "Andrew Adamson,Marilyn Fox", "Andrew Dominik", 
    "Andrew Stanton", "Andrew Stanton,Lee Unkrich", "Angelina Jolie,John Stevenson", 
    "Anne Fontaine", "Anthony Harvey")
AB <- c("A", "B", "A", "A", "B", "B", "B", "A", "B", "A", "B", "A", 
        "A", "B", "B", "B", "B", "B", "B", "A")

library(data.table)
library(magrittr)

    Define function for benchmark runs of problem size n

    run_mb <- function(n) {
  # compute number of benchmark runs depending on problem size `n`
  mb_times <- scales::squish(10000L / n , c(3L, 100L)) 
  cat(n, " ", mb_times, "\n")
  # create data
  DF <- data.frame(director = rep(director, n), AB = rep(AB, n))
  DT <- as.data.table(DF)
  # start benchmarks
  microbenchmark::microbenchmark(
    matt_mod = {
      s <- strsplit(as.character(DF$director), ',')
      data.frame(director=unlist(s), AB=rep(DF$AB, lengths(s)))},
    jaap_DT1 = {
      DT[, lapply(.SD, function(x) unlist(tstrsplit(x, ",", fixed=TRUE))), by = AB
         ][!is.na(director)]},
    jaap_DT2 = {
      DT[, strsplit(as.character(director), ",", fixed=TRUE), 
         by = .(AB, director)][,.(director = V1, AB)]},
    jaap_dplyr = {
      DF %>% 
        dplyr::mutate(director = strsplit(as.character(director), ",")) %>%
        tidyr::unnest(director)},
    jaap_tidyr = {
      tidyr::separate_rows(DF, director, sep = ",")},
    cSplit = {
      splitstackshape::cSplit(DF, "director", ",", direction = "long")},
    DT3 = {
      DT[, strsplit(as.character(director), ",", fixed=TRUE),
         by = .(AB, director)][, director := NULL][
           , setnames(.SD, "V1", "director")]},
    DT4 = {
      DT[, .(director = unlist(strsplit(as.character(director), ",", fixed = TRUE))), 
         by = .(AB)]},
    times = mb_times
  )
}

    Run benchmark for different problem sizes

    # define vector of problem sizes
n_rep <- 10L^(0:5)
# run benchmark for different problem sizes
mb <- lapply(n_rep, run_mb)

    Prepare data for plotting

    mbl <- rbindlist(mb, idcol = "N")
mbl[, n_row := NROW(director) * n_rep[N]]
mba <- mbl[, .(median_time = median(time), N = .N), by = .(n_row, expr)]
mba[, expr := forcats::fct_reorder(expr, -median_time)]

    Create chart

    library(ggplot2)
ggplot(mba, aes(n_row, median_time*1e-6, group = expr, colour = expr)) + 
  geom_point() + geom_smooth(se = FALSE) + 
  scale_x_log10(breaks = NROW(director) * n_rep) + scale_y_log10() + 
  xlab("number of rows") + ylab("median of execution time [ms]") +
  ggtitle("microbenchmark results") + theme_bw()

    Session info & package versions (excerpt)

    devtools::session_info()
#Session info
# version  R version 3.3.2 (2016-10-31)
# system   x86_64, mingw32
#Packages
# data.table      * 1.10.4  2017-02-01 CRAN (R 3.3.2)
# dplyr             0.5.0   2016-06-24 CRAN (R 3.3.1)
# forcats           0.2.0   2017-01-23 CRAN (R 3.3.2)
# ggplot2         * 2.2.1   2016-12-30 CRAN (R 3.3.2)
# magrittr        * 1.5     2014-11-22 CRAN (R 3.3.0)
# microbenchmark    1.4-2.1 2015-11-25 CRAN (R 3.3.3)
# scales            0.4.1   2016-11-09 CRAN (R 3.3.2)
# splitstackshape   1.4.2   2014-10-23 CRAN (R 3.3.3)
# tidyr             0.6.1   2017-01-10 CRAN (R 3.3.2)

    1My curiosity was piqued by this exuberant comment Brilliant! Orders of magnitude faster! to a tidyverse answer of a question which was closed as a duplicate of this question.