data.tables和扫描功能

Question

Using a data.table, which would be the fastest way to "sweep" out a statistic across a selection of columns?

Starting with (considerably larger versions of ) DT

p <- 3
DT <- data.table(id=c("A","B","C"),x1=c(10,20,30),x2=c(20,30,10))
DT.totals <- DT[, list(id,total = x1+x2) ]

I'd like to get to the following data.table result by indexing the target columns (2:p) in order to skip the key:

    id  x1  x2
[1,]    A   0.33    0.67
[2,]    B   0.40    0.60
[3,]    C   0.75    0.25

Answer 1

I believe that something close to the following (which uses the relatively new set() function) will be quickest:

DT <- data.table(id = c("A","B","C"), x1 = c(10,20,30), x2 = c(20,30,10))
total <- DT[ , x1 + x2]

rr <- seq_len(nrow(DT))
for(j in 2:3) set(DT, rr, j, DT[[j]]/total) 
DT
#      id        x1        x2
# [1,]  A 0.3333333 0.6666667
# [2,]  B 0.4000000 0.6000000
# [3,]  C 0.7500000 0.2500000

FWIW, calls to set() takes the following form:

# set(x, i, j, value), where: 
#     x is a data.table 
#     i contains row indices
#     j contains column indices 
#     value is the value to be assigned into the specified cells

---

My suspicion about the relative speed of this, compared to other solutions, is based on this passage from data.table's NEWS file , in the section on changes in Version 1.8.0:

 o New function set(DT,i,j,value) allows fast assignment to elements of DT. Similar to := but avoids the overhead of [.data.table, so is much faster inside a loop. Less flexible than :=, but as flexible as matrix subassignment. Similar in spirit to setnames(), setcolorder(), setkey() and setattr(); ie, assigns by reference with no copy at all. M = matrix(1,nrow=100000,ncol=100) DF = as.data.frame(M) DT = as.data.table(M) system.time(for (i in 1:1000) DF[i,1L] <- i) # 591.000s system.time(for (i in 1:1000) DT[i,V1:=i]) # 1.158s system.time(for (i in 1:1000) M[i,1L] <- i) # 0.016s system.time(for (i in 1:1000) set(DT,i,1L,i)) # 0.027s