假设我有data.table
看起来像这样:
dt <- data.table(
a = c( "A", "B", "C", "C" ),
b = c( "U", "V", "W", "X" ),
c = c( 0.1, 0.2, 0.3, 0.4 ),
min = c( 0, 1, 2, 3 ),
max = c( 11, 12, 13, 14 ),
val = c( 100, 200, 300, 400 ),
key = "a"
)
我实际的data.table
有更多的列,最多有几百万行。大约10%的行具有重复的键a
。我想将这些行与一个看起来像这样的函数聚合:
comb <- function( x ){
k <- which.max( x[ ,c ] )
list( b = x[ k, b ], c = x[ k, c ], min = min( x[ , min ] ), max = max( x[ , max ] ), val = sum( x[ ,val ] ) )
}
但是,打电话
dt <- dt[ , comb(.SD), by = a ]
非常慢,我想知道如何改善这一点。任何帮助表示赞赏。
通过将c
放入键中并使用.N
来获得最大值,我们可以避免which.max
(未经测试):
setkey(dt, a, c)
dt[, c(.SD[.N], min = min[1], val = sum(val)), by = a][, -c(4, 6)]
添加:或此变体:
dt[, c(.SD[.N, c(1:2, 4)], min = min[1], val = sum(val)), by = a]
添加2:我们仅使用.SD
,因为您表示您有很多列,但是如果您愿意将它们写出来,则可以编写以上内容:
dt[, list(b = b[.N], c = c[.N], min = min[1], max = max[.N], val = sum(val)), by = a]
添加3:另一种变化:
dt[, c("min", "val") := list(min[1], sum(val)), by = a][, .SD[.N], by = a]
对这四个解决方案进行微基准测试,得出以下箱线图(n = 10):
<< img src =“ https://image.soinside.com/eyJ1cmwiOiAiaHR0cHM6Ly9pLnN0YWNrLmltZ3VyLmNvbS81ckx6Wi5wbmcifQ==” alt =“在此处输入图像描述”>“ >>