Question

If I have some multivariate irregular time series, such as zoo or xts objects with:

> clicks
           user item
2003-01-02    a    i
2003-01-03    a    i
2003-01-08    b    i
2003-01-09    a    j
2003-01-09    c    j
2003-01-10    b    j
> downloads
           user file
2003-01-08    a    f
2003-01-11    b    g
2003-01-11    b    f
> purchases
           user
2003-01-10    a
2003-01-16    b

I can write some code to produce a simple featurization of the above data into a data frame with a row per (user, day) for all days (up to the day of the user’s first purchase), and with these columns:

• # clicks of item i in past 7 days
• # clicks of item i between 7 and 31 days ago
• # total past clicks
• same for item j
• same for downloads f, g
• whether a purchase occurs in the next 7 days

However, I’m curious whether there are easy, elegant, and not-painfully-slow ways to accomplish this using any of the various time series manipulation packages. I looked around at things in zoo and xts but I didn’t find anything promising.

Answer 1

You can represent each type of event (e.g., “user A clicks item i”)
as a time series x, with value 1 each time it occurs.
The quantities you are interested can be computed from cumsum(x)
(the number of events until today) and its translations
(the number of events until k days in the past or the future).

# Sample data
set.seed(0)
k <- 100
users <- LETTERS[1:4]
files <- letters[1:4]
items <- letters[24:26]
clicks <- data.frame( 
  time = Sys.time() + runif(k, 0, k * 24 * 3600),
  user = sample( users, k, replace=TRUE ),
  item = sample( items, k, replace=TRUE )
)
clicks <- unique(clicks)

For a single time series:

x <- subset( clicks, user=="C" & item=="x" )
xts( rep(1,nrow(x)), x$time )
x <- xts( rep(1,nrow(x)), x$time )

y <- xts( coredata(x), index(x)+7*3600*24 )
z <- cbind(y, x)
z[ is.na(z) ] <- 0
cumsum(z[,2]) - cumsum(z[,1])
# cbind(x,z,cumsum(z[,2]) - cumsum(z[,1]))

For the whole dataset, you can use ddply.