我在一段时间内进行了一系列测量。测量次数通常为4.在任何测量中可以捕获的数字范围是1-5(在现实生活中,给定测试集,范围可以高达100或低至20)。
我想每天计算每个价值在当天之前发生了多少。
让我解释一些样本数据:
# test data creation
d1 = list(as.Date("2013-5-4"), 4,2)
d2 = list(as.Date("2013-5-9"), 2,5)
d3 = list(as.Date("2013-5-16"), 3,2)
d4 = list(as.Date("2013-5-30"), 1,4)
d = rbind(d1,d2,d3,d4)
colnames(d) <- c("Date", "V1", "V2")
tt = as.data.table(d)
我想运行一个可以添加5列的函数(在可能的值范围内每个值可能有1列)。在每个列中,我想要在测试日期之前出现该值的COUNT。
例如,2013-5-30的函数输出为C1=0, C2=3, C3=1, C4=1, C5=1
。
它计算了多少次:
1出现在之前,不包括5/30,这是零 2出现之前,不包括5/30,这是三个 3出现在之前,不包括5/30,这是一个 等。
此外,它还应包括一列,显示该数字显示的总测量值的百分比。例如在5/30
上,在5/30之前有6次测量,所以
pc1 =(0/6),pc2 = 3/6,pc3 = 1/6,pc4 = 1/6,pc5 = 1/6
我想使用data.table赋值表示法(:=)一次性添加这些多列。我正在寻找的输出格式为:
Date V1 V2 C1 PC1 C2 PC2 C3 PC3 C4 PC4 C5 PC5
答案 0 :(得分:4)
<强> 1。 data.table 强>
首先用更常见的方法替换问题中t
的奇怪构造:
library(data.table)
t <- data.table(
Date = as.Date(c("2013-5-4", "2013-5-9", "2013-5-16", "2013-5-30")),
V1 = c(4, 2, 3, 1),
V2 = c(2, 5, 2, 4)
)
现在每行tabulate
并使用cumsum
累积前一行。 perm
是一个置换向量,用于重新排列C列(nc + 1:n)和PC列(nc + n + 1:n)的列号。
nc <- ncol(t) # 3
n <- t[, max(V1, V2)] # 5
Cnames <- paste0("C", 1:n)
PCnames <- paste0("PC", 1:n)
perm <- c(1:nc, rbind(nc + 1:n, nc + n + 1:n))
t[, (Cnames) := as.list(tabulate(c(V1, V2), n)), by = 1:nrow(t)][,
(Cnames):=lapply(.SD, function(x) cumsum(x) - x), .SDcol=Cnames][,
(PCnames):=lapply(.SD, function(x) x/seq(0,len=.N,by=nc-1)), .SDcols=Cnames][,
perm, with = FALSE]
最后一行给出:
Date V1 V2 C1 PC1 C2 PC2 C3 PC3 C4 PC4 C5 PC5
1: 2013-05-04 4 2 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN
2: 2013-05-09 2 5 0 0 1 0.5 0 0.0000000 1 0.5000000 0 0.0000000
3: 2013-05-16 3 2 0 0 2 0.5 0 0.0000000 1 0.2500000 1 0.2500000
4: 2013-05-30 1 4 0 0 3 0.5 1 0.1666667 1 0.1666667 1 0.1666667
1a.data.table备选
如果可以省略第一个日期的行(由于第一个日期之前没有日期,因此不是很有用),那么我们可以执行以下繁琐但直接的自我加入:
t <- data.table(
Date = as.Date(c("2013-5-4", "2013-5-9", "2013-5-16", "2013-5-30")),
V1 = c(4, 2, 3, 1),
V2 = c(2, 5, 2, 4)
)
tt <- t[, one := 1]
setkey(tt, one)
tt[tt,,allow.cartesian=TRUE][Date > Date.1, list(
C1 = sum(.SD == 1), PC1 = mean(.SD == 1),
C2 = sum(.SD == 2), PC2 = mean(.SD == 2),
C3 = sum(.SD == 3), PC3 = mean(.SD == 3),
C4 = sum(.SD == 4), PC4 = mean(.SD == 4),
C5 = sum(.SD == 5), PC5 = mean(.SD == 5)
), by = list(Date, V1, V2), .SDcols = c("V1.1", "V2.1")]
<强> 1b中。 data.table替代
或者我们可以更紧凑地重写1a(tt
,n
,Cnames
和PCnames
来自上方):
tt[tt,,allow.cartesian=TRUE][Date > Date.1, setNames(as.list(rbind(
sapply(1:n, function(i, .SD) sum(.SD==i), .SD=.SD),
sapply(1:n, function(i, .SD) mean(.SD==i), .SD=.SD)
)), c(rbind(Cnames, PCnames))),
by = list(Date, V1, V2), .SDcols = c("V1.1", "V2.1")]
<强> 2。 sqldf 强>
data.table的替代方法是将SQL与这种类似的单调但直接的自联接一起使用:
library(sqldf)
sqldf("select a.Date, a.V1, a.V2,
sum(((b.V1 = 1) + (b.V2 = 1)) * (a.Date > b.Date)) C1,
sum(((b.V1 = 1) + (b.V2 = 1)) * (a.Date > b.Date)) /
cast (2 * count(*) - 2 as real) PC1,
sum(((b.V1 = 2) + (b.V2 = 2)) * (a.Date > b.Date)) C2,
sum(((b.V1 = 2) + (b.V2 = 2)) * (a.Date > b.Date)) /
cast (2 * count(*) - 2 as real) PC2,
sum(((b.V1 = 3) + (b.V2 = 3)) * (a.Date > b.Date)) C3,
sum(((b.V1 = 3) + (b.V2 = 3)) * (a.Date > b.Date)) /
cast (2 * count(*) - 2 as real) PC3,
sum(((b.V1 = 4) + (b.V2 = 4)) * (a.Date > b.Date)) C4,
sum(((b.V1 = 4) + (b.V2 = 4)) * (a.Date > b.Date)) /
cast (2 * count(*) - 2 as real) PC4,
sum(((b.V1 = 5) + (b.V2 = 5)) * (a.Date > b.Date)) C5,
sum(((b.V1 = 5) + (b.V2 = 5)) * (a.Date > b.Date)) /
cast (2 * count(*) - 2 as real) PC5
from t a, t b where a.Date >= b.Date
group by a.Date")
<强> 2a上。 sqldf替代
另一种方法是使用字符串操作来创建上面的sql字符串,如下所示:
f <- function(i) {
s <- fn$identity("sum(((b.V1 = $i) + (b.V2 = $i)) * (a.Date > b.Date))")
fn$identity("$s C$i,\n $s /\ncast (2 * count(*) - 2 as real) PC$i")
}
s <- fn$identity("select a.Date, a.V1, a.V2, `toString(sapply(1:5, f))`
from t a, t b where a.Date >= b.Date
group by a.Date")
sqldf(s)
<强> 2B。第二个sqldf替代
如果我们愿意在没有第一个日期的输出行的情况下,可以大大简化sql解决方案。这可能是有道理的,因为第一个日期没有先前的日期列表:
sqldf("select a.Date, a.V1, a.V2,
sum((b.V1 = 1) + (b.V2 = 1)) C1,
avg((b.V1 = 1) + (b.V2 = 1)) PC1,
sum((b.V1 = 2) + (b.V2 = 2)) C2,
avg((b.V1 = 2) + (b.V2 = 2)) PC2,
sum((b.V1 = 3) + (b.V2 = 3)) C3,
avg((b.V1 = 3) + (b.V2 = 3)) PC3,
sum((b.V1 = 4) + (b.V2 = 4)) C4,
avg((b.V1 = 4) + (b.V2 = 4)) PC4,
sum((b.V1 = 5) + (b.V2 = 5)) C5,
avg((b.V1 = 5) + (b.V2 = 5)) PC5
from t a, t b where a.Date > b.Date
group by a.Date")
同样可以创建sql字符串以避免重复,方法与先前解决方案中显示的方式相同。
更新:添加了PC列和一些简化
更新2:添加了其他解决方案
答案 1 :(得分:1)
这是一个开始。我没有理由“一举一动”这样做。这可能是可能的。试试吧。
library(data.table)
DT = as.data.table(d)
DT[,i:=as.numeric(Date)]
setkey(DT,"i")
uv <- 1:max(unlist(DT[,2:3,with=FALSE]))
DT[,paste0("C",uv):=lapply(uv,function(x) x %in% unlist(.SD)),.SDcols=2:3,by=i]
DT[,paste0("C",uv):=lapply(.SD,function(x) c(NA,head(cumsum(x),-1))),.SDcols=paste0("C",uv)]
DT[,paste0("PC",uv):=lapply(.SD,function(x) x/(2*.I-2)),.SDcols=paste0("C",uv)]
# Date V1 V2 i C1 C2 C3 C4 C5 PC1 PC2 PC3 PC4 PC5
# 1: 2013-05-04 4 2 15829 NA NA NA NA NA NA NA NA NA NA
# 2: 2013-05-09 2 5 15834 0 1 0 1 0 0 0.5 0.0000000 0.5000000 0.0000000
# 3: 2013-05-16 3 2 15841 0 2 0 1 1 0 0.5 0.0000000 0.2500000 0.2500000
# 4: 2013-05-30 1 4 15855 0 3 1 1 1 0 0.5 0.1666667 0.1666667 0.1666667
答案 2 :(得分:0)
您可能想要%in%
运算符。
> foo<-sample(1:10,4)
> bar<-sample(1:10,3)
> foo
[1] 5 3 9 6
> bar
[1] 1 7 2
> bar2<-sample(1:10,5)
> bar2
[1] 2 9 4 8 5
> which(bar2%in%foo)
[1] 2 5 #those are the indices of the values in bar2 which appear in foo
> which(bar%in%foo)
integer(0)