Optimization of Data.Table in R?

Question

I have two tables in R that look as the following:

DT.Purchase <- data.frame( ID = c(1,1,1,2,2,3,3,3,3,3,4,4,4,4),
                          CDS = c("0389","0389", "3298", "4545", "1282", "4545", 
                                  "0389","0389", "5685", "4545", "1282", "0389", 
                                  "1282", "1282")
                         Date = c("5/28/2016","5/26/2016","8/9/2016","2/2/2015", 
                                  "2/24/2015", "9/27/2015", "9/27/2015", "9/5/2015", 
                                  "3/3/2016", "4/9/2014", "5/1/2014", "5/4/2014",
                                  "6/9/2014", "7/7/2014"),   
                          JFK = c(T,F,F,F,T,T,F,F,T,F,T,T,T,F),
                          RFK = c(F,T,T,F,T,F,F,F,F,T,T,T,T,T), 
                          RUG = c(T,F,T,F,T,F,F,F,F,T,F,F,T,T),
                          LPG = c(T,T,T,F,F,T,T,F,F,F,F,F,T,F))


DT.Purchase$Date <- as.Date(DT.Purchase$Date, format = "%m/%d/%Y")
DT.Purchase      <- data.table(DT.Purchase)
ID  CDS     Date    JFK RFK RUG LPG
1   0389    5/28/2016   T   F   T   T
1   0389    5/26/2016   F   T   F   T
1   3298    8/9/2016    F   T   T   T
2   4545    2/2/2015    F   F   F   F
2   1282    2/24/2015   T   T   T   F
3   4545    9/27/2015   T   F   F   T
3   0389    9/27/2015   F   F   F   T
3   0389    9/5/2015    F   F   F   F
3   5685    3/3/2016    T   F   F   F
3   4545    4/9/2014    F   T   T   F
4   1282    5/1/2014    T   T   F   F
4   0389    5/4/2014    T   T   F   F
4   1282    6/9/2014    T   T   T   T
4   1282    7/7/2014    F   T   T   F

DT.Stay <- data.frame(Stay.ID = c(1,2,3,5,6,9,10,11), 
                            ID = c(1,1,2,3,3,3,4,4), 
                    Start.Date = c('5/26/2016','8/1/2016', '2/1/2015', '3/1/2016', 
                                   '9/1/2015', '4/9/2014', '4/7/2014','6/1/2014'),
                    End.Date   = c('6/6/2016','9/1/2016','3/1/2015','3/7/2016',
                                   '9/30/2015','4/14/2014','5/9/2014','7/11/2014'))
DT.Stay$Start.Date <- as.Date(DT.Stay$Start.Date, format = "%m/%d/%Y")
DT.Stay$End.Date <- as.Date(DT.Stay$End.Date, format = "%m/%d/%Y")

DT.Stay <- data.table(DT.Stay)

Stay.ID ID  Start.Date  End.Date
1   1   5/26/2016   6/6/2016
2   1   8/1/2016    9/1/2016
3   2   2/1/2015    3/1/2015
5   3   3/1/2016    3/7/2016
6   3   9/1/2015    9/30/2015
9   3   4/9/2014    4/14/2014
10  4   4/7/2014    5/9/2014
11  4   6/1/2014    7/11/2014

Now in reality, DT.Purchase is much larger (10 million observations) and DT.Stay is over 50000 observations. There are weights for DT.Purchase[ ,.(JFK, RFK, DUG, LPG)] which equals c.weights = c(1,2,1,3) . These weights represent an in-house cost weight that we are considering. The goal of this is to determine purchasing deals based on previous purchases and the weights. What I would like to do is determine the previous cost.index which is the sum of all previous purchases based on the weights and previous to each End.Date for each Stay.ID . So the final data.table should look like

Stay.ID cost.index
1        10
2        16
3        4
5        11
6        10
9        3
10       6
11       10

The way I have done this involves merging according to ID ( allow.cartesian = TRUE ) the two datasets and checking to see if Date <= End.Date . Then I substitute the weights in and sum for each Stay.ID . It works but I am looking for a quicker way to do this. With 10 million and 50000 observations the merging becomes time and resource consuming.

Answer 1

With the latest devel version (1.9.7+), smth like this will work:

DT.Purchase[DT.Stay, on = .(ID = ID, Date >= Start.Date, Date <= End.Date),
            .(Stay.ID, sum(as.matrix(.SD) %*% c.weights)),
            by = .EACHI, .SDcols = JFK:LPG]

Assuming your dates are in either Date or IDate format.

In 1.9.6 you can use foverlaps instead:

foverlaps(setkey(DT.Purchase[, Date2 := Date], ID, Date, Date2),
          setkey(DT.Stay, ID, Start.Date, End.Date))[,
  sum(as.matrix(.SD) %*% c.weights), keyby = Stay.ID, .SDcols = JFK:LPG]