Improve performance of data.table date+time pasting?

问题

I am not sure that I can ask this question here, let me know if I should do it somewhere else.

I have a data.table with 1e6 rows having this structure:

        V1       V2     V3
1: 03/09/2011 08:05:40 1145.0
2: 03/09/2011 08:06:01 1207.3
3: 03/09/2011 08:06:17 1198.8
4: 03/09/2011 08:06:20 1158.4
5: 03/09/2011 08:06:40 1112.2
6: 03/09/2011 08:06:59 1199.3

I am converting the V1 and V2 variables to a unique datetime variable, using this code:

 system.time(DT[,`:=`(index= as.POSIXct(paste(V1,V2),
                         format='%d/%m/%Y %H:%M:%S'),
                     V1=NULL,V2=NULL)])

   user  system elapsed 
   47.47    0.16   50.27

Is there any method to improve performance of this transformation?

Here the dput(head(DT)):

DT <- structure(list(V1 = c("03/09/2011", "03/09/2011", "03/09/2011", 
"03/09/2011", "03/09/2011", "03/09/2011"), V2 = c("08:05:40", 
"08:06:01", "08:06:17", "08:06:20", "08:06:40", "08:06:59"), 
    V3 = c(1145, 1207.3, 1198.8, 1158.4, 1112.2, 1199.3)), .Names = c("V1", 
"V2", "V3"), class = c("data.table", "data.frame"), row.names = c(NA, 
-6L), .internal.selfref = <pointer: 0x00000000002a0788>)

回答1:

This approach, which appears to run ~40X faster than OP's, uses lookup tables and takes advantage of the extremely fast data table joins. Also, it takes advantage of the fact that, while there may be 1e6 combinations of date and time, there can be at most 86400 unique times, and probably even fewer dates. Finally, it avoids the use of paste(...) altogether.

library(data.table)
library(stringr)

# create a dataset with 1MM rows
set.seed(1)
x  <- 1000*sample(1:1e5,1e6,replace=T)
dt <- data.table(id=1:1e6,
                 V1=format(as.POSIXct(x,origin="2011-01-01"),"%d/%m/%Y"),
                 V2=format(as.POSIXct(x,origin="2011-01-01"),"%H:%M:%S"),
                 V3=x)
DT <- dt

index.date <- function(dt) {
  # Edit: this change processes only times from the dataset; slightly more efficient
  V2 <- unique(dt$V2)
  dt.time <- data.table(char.time=V2,
                        int.time=as.integer(substr(V2,7,8))+
                          60*(as.integer(substr(V2,4,5))+
                                60*as.integer(substr(V2,1,2))))
  setkey(dt.time,char.time)
  # all dates from dataset
  dt.date <- data.table(char.date=unique(dt$V1), int.date=as.integer(as.POSIXct(unique(dt$V1),format="%d/%m/%Y")))
  setkey(dt.date,char.date)
  # join the dates
  setkey(dt,V1)
  dt <- dt[dt.date]
  # join the times
  setkey(dt,V2)
  dt <- dt[dt.time, nomatch=0]
  # numerical index
  dt[,int.index:=int.date+int.time]
  # POSIX date index
  dt[,index:=as.POSIXct(int.index,origin='1970-01-01')]
  # get back original order
  setkey(dt,id)
  return(dt)
}
# new approach
system.time(dt<-index.date(dt))
#   user  system elapsed 
#   2.26    0.00    2.26 


# original approach
DT <- dt
system.time(DT[,`:=`(index= as.POSIXct(paste(V1,V2),
                                       format='%d/%m/%Y %H:%M:%S'),
                     V1=NULL,V2=NULL)])
#   user  system elapsed 
#  84.33    0.06   84.52

Note that performance does depend on how many unique dates there are. In the test case there were ~1200 unique dates.

EDIT proposition to write the function in more data.table-sugar syntax and avoid "$" for subsetting:

index.date <- function(dt,fmt="%d/%m/%Y") {
    dt.time <- data.table(char.time = dt[,unique(V2)],key='char.time')
    dt.time[,int.time :=as.integer(substr(char.time,7,8))+
                                            60*(as.integer(substr(char.time,4,5))+
                                                        60*as.integer(substr(char.time,1,2)))]
    # all dates from dataset
    dt.date <- data.table(char.date = dt[,unique(V1)],key='char.date')
    dt.date[,int.date:=as.integer(as.POSIXct(char.date,format=fmt))]
    # join the dates
    setkey(dt,V1)
    dt <- dt[dt.date]
    # join the times
    setkey(dt,V2)
    dt <- dt[dt.time, nomatch=0]
    # numerical index
    dt[,int.index:=int.date+int.time]
    # POSIX date index
    dt[,index:=as.POSIXct.numeric(int.index,origin='1970-01-01')]
    # remove extra/temporary variables
    dt[,`:=`(int.index=NULL,int.date=NULL,int.time=NULL)]
}

回答2:

If there are many time stamps that will be repeated in your data, you can try adding ,by=list(V1, V2), but there would have to be enough repetition to pay for the cost of splitting.

The bottle neck here is the paste & conversion, so that leads me to think that the answer is no. (Unless you use an alternate method of converting to POSIX)

来源：https://stackoverflow.com/questions/20483217/improve-performance-of-data-table-datetime-pasting

标签

performance

data.table

posixct