Efficient way to perform running total in the last

2019-02-14 08:02发布

问题:

This is what my data frame looks like:

library(data.table)

df <- fread('
                Name  EventType  Date  SalesAmount RunningTotal Runningtotal(prior365Days)
                John    Email      1/1/2014      0          0            0
                John    Sale       2/1/2014     10          10           10
                John    Sale       7/1/2014     20          30           30
                John    Sale       4/1/2015     30          60           50 
                John    Webinar    5/1/2015      0          60           50
                Tom     Email      1/1/2014      0          0            0
                Tom     Sale       2/1/2014     15          15           15
                Tom     Sale       7/1/2014     10          25           25
                Tom     Sale       4/1/2015     25          50           35 
                Tom     Webinar    5/1/2015      0          50           35
                ')
    df[,Date:= as.Date(Date, format="%m/%d/%Y")]

The last column was my desired column which is the cumulative sum of SalesAmount(for each Name) in the last 365 days rolling window and I performed this with the help of @6pool. His solution was:

df$EventDate <- as.Date(df$EventDate, format="%d/%m/%Y")
df <- df %>%
   group_by (Name) %>%
   arrange(EventDate) %>% 
   mutate(day = EventDate - EventDate[1])

f <- Vectorize(function(i)
    sum(df[df$Name[i] == df$Name & df$day[i] - df$day >= 0 & 
             df$day[i] - df$day <= 365, "SalesAmount"]), vec="i")
df$RunningTotal365 <- f(1:nrow(df))

However,df$RunningTotal365 <- f(1:nrow(df)) is taking a long time(over 1.5 days so far) as my dataframe is over 1.5 million rows. I was suggested "rollapply" in my initial question but I have struggled to figure out how to use it in this instance. Kindly help.

回答1:

Give this a try:

DF <- read.table(text = "Name  EventType  EventDate  SalesAmount RunningTotal Runningtotal(prior365Days)
John    Email      1/1/2014      0          0            0
John    Sale       2/1/2014     10          10           10
John    Sale       7/1/2014     20          30           30
John    Sale       4/1/2015     30          60           50 
John    Webinar    5/1/2015      0          60           50
Tom     Email      1/1/2014      0          0            0
Tom     Sale       2/1/2014     15          15           15
Tom     Sale       7/1/2014     10          25           25
Tom     Sale       4/1/2015     25          50           35 
Tom     Webinar    5/1/2015      0          50           35", header = TRUE)


fun <- function(x, date, thresh) {
  D <- as.matrix(dist(date)) #distance matrix between dates
  D <- D <= thresh
  D[lower.tri(D)] <- FALSE #don't sum to future
  R <- D * x #FALSE is treated as 0
  colSums(R)
}


library(data.table)
setDT(DF)
DF[, EventDate := as.Date(EventDate, format = "%m/%d/%Y")]
setkey(DF, Name, EventDate)

DF[, RT365 := fun(SalesAmount, EventDate, 365), by = Name]

#    Name EventType  EventDate SalesAmount RunningTotal Runningtotal.prior365Days. RT365
# 1: John     Email 2014-01-01           0            0                          0     0
# 2: John      Sale 2014-02-01          10           10                         10    10
# 3: John      Sale 2014-07-01          20           30                         30    30
# 4: John      Sale 2015-04-01          30           60                         50    50
# 5: John   Webinar 2015-05-01           0           60                         50    50
# 6:  Tom     Email 2014-01-01           0            0                          0     0
# 7:  Tom      Sale 2014-02-01          15           15                         15    15
# 8:  Tom      Sale 2014-07-01          10           25                         25    25
# 9:  Tom      Sale 2015-04-01          25           50                         35    35
#10:  Tom   Webinar 2015-05-01           0           50                         35    35


回答2:

Here's an approach using foverlaps function from data.table package:

require(data.table)
setDT(df)[, end := as.Date(EventDate, format="%d/%m/%Y")
        ][, start := end - 365L]
setkey(df, Name, start, end)
olaps = foverlaps(df, df, nomatch=0L, which=TRUE)
olaps = olaps[xid >= yid, .(ans = sum(dt$SalesAmount[yid])), by=xid]

df[olaps$xid, Runningtotal := olaps$ans]

You can remove the start and end columns, if necessary, by doing:

df[, c("start", "end") := NULL]

Would be nice to know how fast/slow it is..



回答3:

Using newer non-equi joins feature in data.table:

    df1 = df[.(iName=Name,start = Date - 365L, end = Date),
    on=.(Name=iName,Date >= start, Date <= end),nomatch = 0, allow.cart=TRUE][,
  .(MyTotal = sum(SalesAmount)), by=.(Name,Date = Date.1)]


    df[df1, on = .(Name,Date)]