data.table根据条件更新组中的最后一个元素

时间:2016-03-03 00:29:57

标签: r data.table

我有一个包含3列的data.table:id,time和status。对于每个id,我想找到具有最大时间的记录 - 然后如果对于该记录,状态为真,如果时间是>我想将其设置为假。 7(例如)。我是按照以下方式进行的。

x <- data.table(id=c(1,1,2,2),time=c(5,6,7,8),status=c(FALSE,TRUE,FALSE,TRUE))
setkey(x,id,time)
y <- x[,.SD[.N],by=id]
x[y,status:=status & time > 7]

我正在使用大量数据,并希望加快此操作。任何建议将不胜感激!

4 个答案:

答案 0 :(得分:8)

x[x[,.N, by=id][,cumsum(N)], status := status * time <=7]

如果我没有弄错的话,这不是连接,因为x[,.N, by=id][,cumsum(N)]返回每组最后一个元素的行索引。

<强>更新 在看到速度比较之后,这个似乎是赢家,应该首先列出

这是我最初的尝试,结果证明是所有建议解决方案中最慢的

x[,status := c(.SD[-.N, status], .SD[.N, status * time <=7]), by=id]

答案 1 :(得分:7)

一种data.table方法

x[ x[order(time), .I[.N], by=id]$V1 , status := ifelse(time > 7, FALSE, TRUE)]

> x
#   id time status
#1:  1    5  FALSE
#2:  1    6   TRUE
#3:  2    7  FALSE
#4:  2    8  FALSE

x[order(time), .I[.N], by=id]$V1为我们提供了每个组的最大time行标记(id

而且,借用@ Floo0的答案,我们可以稍微简化一下

x[ x[order(time), .I[.N], by=id]$V1 , status := status * time <= 7]

速度比较

各种答案的速度测试(并保持数据上的键)

set.seed(123)
x <- data.table(id=c(rep(seq(1:10000), each=10)),
                time=c(rep(seq(1:10000), 10)),
                status=c(sample(c(TRUE, FALSE), 10000*10, replace=T)))
setkey(x,id,time)
x1 <- copy(x); x2 <- copy(x); x3 <- copy(x); x4 <- copy(x); x5 <- copy(x); x6 <- copy(x)

library(microbenchmark)

microbenchmark(

    Symbolix = {x1[ x1[order(time), .I[.N], by=id]$V1 , status := status * time < 7 ] },

    Floo0_1 = {x2[,status := c(.SD[-.N, status], .SD[.N, status * time > 7]), by=id]},

    Floo0_2 = {x3[x3[,.N, by=id][,cumsum(N)], status := status * time > 7]},

    Original = { 
                y <- x4[,.SD[.N],by=id]
                x4[y,status:=status & time > 7]
               },

    Frank = {
             y <- x5[, .SD[.N, .(time, status)], by=id][time > 7 & status]
             x5[y, status := FALSE]
             },

    thelatemail = {x6[ x6[,.I==.I[which.max(time)], by=id]$V1 & time > 7, status := FALSE]}
)

Unit: milliseconds
        expr         min          lq        mean      median          uq         max neval cld
    Symbolix    5.419768    5.857477    6.514111    6.222118    6.936000   11.284580   100 a  
     Floo0_1 4550.314775 4710.679867 4787.086279 4776.794072 4850.334011 5097.136148   100   c
     Floo0_2    1.653419    1.792378    1.945203    1.881609    2.014325    4.096006   100 a  
    Original   10.052947   10.986294   12.541595   11.431182   12.391287   89.494783   100 a  
       Frank 4609.115061 4697.687642 4743.886186 4735.086113 4785.212543 4932.270602   100  b 
 thelatemail   10.300864   11.594972   12.421889   12.315852   12.984146   17.630736   100 a  

答案 2 :(得分:5)

另一次尝试:

x[ x[,.I==.I[which.max(time)], by=id]$V1 & time > 7, status := FALSE]
x

#   id time status
#1:  1    5  FALSE
#2:  1    6   TRUE
#3:  2    7  FALSE
#4:  2    8  FALSE

答案 3 :(得分:3)

这是另一种方式,类似于OP:

y = unique(x[,c("id","time"), with=FALSE], by="id", fromLast=TRUE)
x[y[time > 7], status := FALSE]

这是另一个基准:

n_id = 1e3; n_col = 100; n_draw  = 5

set.seed(1)
X = data.table(id = 1:n_id)[, .(
    time    = sample(10,n_draw), 
    status  = sample(c(T,F), n_draw, replace=TRUE)
), by=id][, paste0("V",1:n_col) := 0]
setkey(X,id,time)

X1 = copy(X); X2 = copy(X); X3 = copy(X); X4 = copy(X)
X5 = copy(X); X6 = copy(X); X7 = copy(X); X8 = copy(X)

library(microbenchmark)
library(multcomp)

microbenchmark(
unique = {
    Y = unique(X1[,c("id","time"), with=FALSE], by="id", fromLast=TRUE)
    X1[Y[time > 7], status := FALSE]
},
OP = {
    y <- X2[,.SD[.N],by=id]
    X2[y,status:=status & time > 7]
},
Floo0a = X3[,status := c(.SD[-.N, status], .SD[.N, status * time >7]), by=id],
Floo0b = X4[X4[,.N, by=id][,cumsum(N)], status := status * time >7],
tlm = X5[ X5[,.I==.I[which.max(time)], by=id]$V1 & time > 7, status := FALSE],
Symbolix=X6[ X6[order(time), .I[.N], by=id]$V1 , status := ifelse(time > 7, FALSE, TRUE)],
Frank1 = {
    y <- X7[, .SD[.N, .(time, status)], by=id][time > 7 & status]
    X7[y, status := FALSE]
},
Frank2 = {
    y <- X8[, .SD[.N], by=id][time > 7 & status]
    X8[y, status := FALSE]
}, times = 1, unit = "relative")

结果:

     expr        min         lq       mean     median         uq        max neval
   unique   1.348592   1.348592   1.348592   1.348592   1.348592   1.348592     1
       OP  35.048724  35.048724  35.048724  35.048724  35.048724  35.048724     1
   Floo0a 416.175654 416.175654 416.175654 416.175654 416.175654 416.175654     1
   Floo0b   1.000000   1.000000   1.000000   1.000000   1.000000   1.000000     1
      tlm   2.151996   2.151996   2.151996   2.151996   2.151996   2.151996     1
 Symbolix   1.770835   1.770835   1.770835   1.770835   1.770835   1.770835     1
   Frank1 404.045660 404.045660 404.045660 404.045660 404.045660 404.045660     1
   Frank2  36.603303  36.603303  36.603303  36.603303  36.603303  36.603303     1