将一列中以逗号分隔的字符串拆分为单独的行

我有一个数据框，如下所示：

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'))

如您所见，该director列中的某些条目是多个名称，以逗号分隔。我想将这些条目拆分为单独的行，同时保持另一列的值。例如，上面数据框中的第一行应分为两行，在该director列中各有一个名称，在该列中有一个“ A” AB。

这个老问题经常被用作欺骗对象（标记为r-faq）。截止到今天，已经提供了6种不同的方法进行了三次回答，但是缺乏基准作为指导，其中哪种方法是最快的¹。

基准测试解决方案包括

• Matthew Lundberg的基本R方法，但根据Rich Scriven的评论进行了修改，
• Jaap的两种data.table方法和两种dplyr/ tidyr方法，
• 阿南达的splitstackshape解决方案，
• 以及Jaap data.table方法的两个其他变体。

使用该microbenchmark软件包，对6种不同大小的数据帧总共进行了8种不同方法的基准测试（请参见下面的代码）。

OP提供的样本数据仅包含20行。为了创建更大的数据帧，将这20行简单重复1、10、100、1000、10000和100000次，这样问题大小最多可达200​​万行。

基准结果

基准测试结果表明，对于足够大的数据帧，所有data.table方法都比其他任何方法都快。对于具有约5000行以上的数据帧，Jaap的data.table方法2及其变体DT3是最快的，幅度比最慢的方法快。

值得注意的是，这两种tidyverse方法的时间和splistackshape解决方案是如此相似，以至于很难消除图表中的曲线。在所有数据帧大小中，它们是基准测试方法中最慢的。

对于较小的数据帧，Matt的基本R解决方案和data.table方法4似乎比其他方法具有更少的开销。

码

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)

为问题规模的基准运行定义函数 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
  )
}

针对不同问题规模运行基准测试

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

准备绘图数据

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)]

创建图表

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()

会话信息和软件包版本（节选）

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)

¹ _{我的好奇心被这种热烈的评论 激怒了！快几个数量级！一个问题的tidyverse答案，该问题已作为该问题的副本被关闭。}

几种选择：

1）两种方式 数据表：

library(data.table)
# method 1 (preferred)
setDT(v)[, lapply(.SD, function(x) unlist(tstrsplit(x, ",", fixed=TRUE))), by = AB
         ][!is.na(director)]
# method 2
setDT(v)[, strsplit(as.character(director), ",", fixed=TRUE), by = .(AB, director)
         ][,.(director = V1, AB)]

2） dplyr / 提迪尔 组合：

library(dplyr)
library(tidyr)
v %>% 
  mutate(director = strsplit(as.character(director), ",")) %>%
  unnest(director)

3）与 提迪尔only：使用tidyr 0.5.0（及更高版本），您还可以使用separate_rows：

separate_rows(v, director, sep = ",")

您可以使用该convert = TRUE参数将数字自动转换为数字列。

4）以R为基数：

# if 'director' is a character-column:
stack(setNames(strsplit(df$director,','), df$AB))

# if 'director' is a factor-column:
stack(setNames(strsplit(as.character(df$director),','), df$AB))

命名您的原始data.frame v，我们有：

> s <- strsplit(as.character(v$director), ',')
> data.frame(director=unlist(s), AB=rep(v$AB, sapply(s, FUN=length)))
                      director AB
1                 Aaron Blaise  A
2                   Bob Walker  A
3               Akira Kurosawa  B
4               Alan J. Pakula  A
5                  Alan Parker  A
6           Alejandro Amenabar  B
7  Alejandro Gonzalez Inarritu  B
8  Alejandro Gonzalez Inarritu  B
9             Benicio Del Toro  B
10 Alejandro González Iñárritu  A
11                 Alex Proyas  B
12              Alexander Hall  A
13              Alfonso Cuaron  B
14            Alfred Hitchcock  A
15              Anatole Litvak  A
16              Andrew Adamson  B
17                 Marilyn Fox  B
18              Andrew Dominik  B
19              Andrew Stanton  B
20              Andrew Stanton  B
21                 Lee Unkrich  B
22              Angelina Jolie  B
23              John Stevenson  B
24               Anne Fontaine  B
25              Anthony Harvey  A

请注意使用rep来构建新的AB列。在此，sapply返回每个原始行中名称的数量。

晚了，但是另一个通用的替代方法是使用cSplit我的“ splitstackshape”程序包中的一个direction参数。设置它以"long"获得您指定的结果：

library(splitstackshape)
head(cSplit(mydf, "director", ",", direction = "long"))
#              director AB
# 1:       Aaron Blaise  A
# 2:         Bob Walker  A
# 3:     Akira Kurosawa  B
# 4:     Alan J. Pakula  A
# 5:        Alan Parker  A
# 6: Alejandro Amenabar  B

devtools::install_github("yikeshu0611/onetree")

library(onetree)

dd=spread_byonecolumn(data=mydata,bycolumn="director",joint=",")

head(dd)
            director AB
1       Aaron Blaise  A
2         Bob Walker  A
3     Akira Kurosawa  B
4     Alan J. Pakula  A
5        Alan Parker  A
6 Alejandro Amenabar  B

使用所得的另一基准strsplit从基部可以目前被推荐给分割用逗号分隔的字符串在列到不同的行，因为它是最快的在宽范围的尺寸：

s <- strsplit(v$director, ",", fixed=TRUE)
s <- data.frame(director=unlist(s), AB=rep(v$AB, lengths(s)))

请注意，使用fixed=TRUE时间对计时有重大影响。

比较方法：

met <- alist(base = {s <- strsplit(v$director, ",") #Matthew Lundberg
   s <- data.frame(director=unlist(s), AB=rep(v$AB, sapply(s, FUN=length)))}
 , baseLength = {s <- strsplit(v$director, ",") #Rich Scriven
   s <- data.frame(director=unlist(s), AB=rep(v$AB, lengths(s)))}
 , baseLeFix = {s <- strsplit(v$director, ",", fixed=TRUE)
   s <- data.frame(director=unlist(s), AB=rep(v$AB, lengths(s)))}
 , cSplit = s <- cSplit(v, "director", ",", direction = "long") #A5C1D2H2I1M1N2O1R2T1
 , dt = s <- setDT(v)[, lapply(.SD, function(x) unlist(tstrsplit(x, "," #Jaap
   , fixed=TRUE))), by = AB][!is.na(director)]
#, dt2 = s <- setDT(v)[, strsplit(director, "," #Jaap #Only Unique
#  , fixed=TRUE), by = .(AB, director)][,.(director = V1, AB)]
 , dplyr = {s <- v %>%  #Jaap
    mutate(director = strsplit(director, ",", fixed=TRUE)) %>%
    unnest(director)}
 , tidyr = s <- separate_rows(v, director, sep = ",") #Jaap
 , stack = s <- stack(setNames(strsplit(v$director, ",", fixed=TRUE), v$AB)) #Jaap
#, dt3 = {s <- setDT(v)[, strsplit(director, ",", fixed=TRUE), #Uwe #Only Unique
#  by = .(AB, director)][, director := NULL][, setnames(.SD, "V1", "director")]}
 , dt4 = {s <- setDT(v)[, .(director = unlist(strsplit(director, "," #Uwe
   , fixed = TRUE))), by = .(AB)]}
 , dt5 = {s <- vT[, .(director = unlist(strsplit(director, "," #Uwe
   , fixed = TRUE))), by = .(AB)]}
   )

库：

library(microbenchmark)
library(splitstackshape) #cSplit
library(data.table) #dt, dt2, dt3, dt4
#setDTthreads(1) #Looks like it has here minor effect
library(dplyr) #dplyr
library(tidyr) #dplyr, tidyr

数据：

v0 <- 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'))

计算和计时结果：

n <- 10^(0:5)
x <- lapply(n, function(n) {v <- v0[rep(seq_len(nrow(v0)), n),]
  vT <- setDT(v)
  ti <- min(100, max(3, 1e4/n))
  microbenchmark(list = met, times = ti, control=list(order="block"))})

y <- do.call(cbind, lapply(x, function(y) aggregate(time ~ expr, y, median)))
y <- cbind(y[1], y[-1][c(TRUE, FALSE)])
y[-1] <- y[-1] / 1e6 #ms
names(y)[-1] <- paste("n:", n * nrow(v0))
y #Time in ms
#         expr     n: 20    n: 200    n: 2000   n: 20000   n: 2e+05   n: 2e+06
#1        base 0.2989945 0.6002820  4.8751170  46.270246  455.89578  4508.1646
#2  baseLength 0.2754675 0.5278900  3.8066300  37.131410  442.96475  3066.8275
#3   baseLeFix 0.2160340 0.2424550  0.6674545   4.745179   52.11997   555.8610
#4      cSplit 1.7350820 2.5329525 11.6978975  99.060448 1053.53698 11338.9942
#5          dt 0.7777790 0.8420540  1.6112620   8.724586  114.22840  1037.9405
#6       dplyr 6.2425970 7.9942780 35.1920280 334.924354 4589.99796 38187.5967
#7       tidyr 4.0323765 4.5933730 14.7568235 119.790239 1294.26959 11764.1592
#8       stack 0.2931135 0.4672095  2.2264155  22.426373  289.44488  2145.8174
#9         dt4 0.5822910 0.6414900  1.2214470   6.816942   70.20041   787.9639
#10        dt5 0.5015235 0.5621240  1.1329110   6.625901   82.80803   636.1899

注意，方法如

(v <- rbind(v0[1:2,], v0[1,]))
#                 director AB
#1 Aaron Blaise,Bob Walker  A
#2          Akira Kurosawa  B
#3 Aaron Blaise,Bob Walker  A

setDT(v)[, strsplit(director, "," #Jaap #Only Unique
  , fixed=TRUE), by = .(AB, director)][,.(director = V1, AB)]
#         director AB
#1:   Aaron Blaise  A
#2:     Bob Walker  A
#3: Akira Kurosawa  B

返回strsplit为unique 董事，并可能与可比

tmp <- unique(v)
s <- strsplit(tmp$director, ",", fixed=TRUE)
s <- data.frame(director=unlist(s), AB=rep(tmp$AB, lengths(s)))

但是据我了解，这不是必需的。