Transforming subsets of data in R with by, ddply and data.table

7 comments
Transforming data sets with R is usually the starting point of my data analysis work. Here is a scenario which comes up from time to time: transform subsets of a data frame, based on context given in one or a combination of columns.

As an example I use a data set which shows sales figures by product for a number of years:
df <- data.frame(Product=gl(3,10,labels=c("A","B", "C")), 
                 Year=factor(rep(2002:2011,3)), 
                 Sales=1:30)
head(df)
##   Product Year Sales
## 1       A 2002     1
## 2       A 2003     2
## 3       A 2004     3
## 4       A 2005     4
## 5       A 2006     5
## 6       A 2007     6
I am interested in absolute and relative sales developments by product over time. Hence, I would like to add a column to my data frame that shows the sales figures divided by the total sum of sales in each year, so I can create a chart which looks like this:

There are lots of ways of doing this transformation in R. Here are three approaches using:
  • base R with by,
  • ddply of the plyr package,
  • data.table of the package with the same name.

by

The idea here is to use by to split the data for each year and to apply the transform function to each subset to calculate the share of sales for each product with the following function: fn <- function(x) x/sum(x). Having defined the function fn I can apply it in a by statement, and as its output will be a list, I wrap it into a do.call command to row-bind (rbind) the list elements:
R1 <- do.call("rbind", as.list(
  by(df, df["Year"], transform, Share=fn(Sales))
))

head(R1)
##         Product Year Sales      Share
## 2002.1        A 2002     1 0.03030303
## 2002.11       B 2002    11 0.33333333
## 2002.21       C 2002    21 0.63636364
## 2003.2        A 2003     2 0.05555556
## 2003.12       B 2003    12 0.33333333
## 2003.22       C 2003    22 0.61111111

ddply

Hadely's plyr package provides an elegant wrapper for this job with the ddply function. Again I use the transform function with my self defined fn function:

library(plyr)
R2 <- ddply(df, "Year", transform, Share=fn(Sales))

head(R2)
##   Product Year Sales      Share
## 1       A 2002     1 0.03030303
## 2       B 2002    11 0.33333333
## 3       C 2002    21 0.63636364
## 4       A 2003     2 0.05555556
## 5       B 2003    12 0.33333333
## 6       C 2003    22 0.61111111

data.table

With data.table I have to do a little bit more legwork, in particular I have to think about the indices I need to use. Yet, it is still straight forward:

library(data.table)
## Convert df into a data.table
dt <- data.table(df) 
## Set Year as a key
setkey(dt, "Year") 
## Calculate the sum of sales per year(=key(dt))
X <- dt[, list(SUM=sum(Sales)), by=key(dt)] 
## Join X and dt, both have the same key and
## add the share of sales as an additional column
R3 <- dt[X, list(Sales, Product, Share=Sales/SUM)]

head(R3)
##      Year Sales Product      Share
## [1,] 2002     1       A 0.03030303
## [2,] 2002    11       B 0.33333333
## [3,] 2002    21       C 0.63636364
## [4,] 2003     2       A 0.05555556
## [5,] 2003    12       B 0.33333333
## [6,] 2003    22       C 0.61111111
Although data.table may look cumbersome compared to ddply and by, I will show below that it is actually a lot faster than the two other approaches.

Plotting the results

With any of the three outputs I can create the chart from above with latticeExtra:
library(latticeExtra)
asTheEconomist(
 xyplot(Sales + Share ~ Year, groups=Product, 
  data=R3, t="b", 
  scales=list(relation="free",x=list(rot=45)), 
  auto.key=list(space="top", column=3),
  main="Product information")
)

Comparing performance of by, ddply and data.table

Let me move on to a more real life example with 100 companies, each with 20 products and a 10 year history:
set.seed(1)
df <- data.frame(Company=rep(paste("Company", 1:100),200),
                 Product=gl(20,100,labels=LETTERS[1:20]), 
                 Year=sort(rep(2002:2011,2000)), 
                 Sales=rnorm(20000, 100,10))
I use the same three approaches to calculate the share of sales by product for each year and company, but this time I will measure the execution time on my old iBook G4, running R-2.15.0:
r1 <- system.time(
 R1 <- do.call("rbind", as.list(
   by(df, df[c("Year", "Company")], 
      transform, Share=fn(Sales))
 ))
)

r2 <- system.time(
 R2 <- ddply(df, c("Company", "Year"), 
             transform, Share=fn(Sales))
)

r3 <- system.time({
 dt <- data.table(df)
 setkey(dt, "Year", "Company")
 X <- dt[, list(SUM=sum(Sales)), by=key(dt)]
 R3 <- dt[X, list(Company, Sales, Product, Share=Sales/SUM)]
})
And here are the results:
r1 # by
##  user  system elapsed 
## 13.690   4.178  42.118 
r2 # ddply 
##  user  system elapsed 
## 18.215   6.873  53.061
r3 # data.table 
##  user  system elapsed 
## 0.171   0.036   0.442
It is quite astonishing to see the speed of data.table in comparison to by and ddply, but maybe it shouldn't be surprise that the elegance of ddply comes with a price as well.

Addition (13 June 2012): See also Matt's comments below. I completely missed ave from base R, which is rather simple and quick as well. Additionally his link to a stackoverflow discussion provides further examples and benchmarks.

Finally my session info:
> sessionInfo() # iBook G4 800 MHZ, 640 MB RAM
R version 2.15.0 Patched (2012-06-03 r59505)
Platform: powerpc-apple-darwin8.11.0 (32-bit)

locale:
[1] C

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] latticeExtra_0.6-19 lattice_0.20-6      RColorBrewer_1.0-5 
[4] data.table_1.8.0    plyr_1.7.1         

loaded via a namespace (and not attached):
[1] grid_2.15.0

7 comments :

FC said...

In your data.table example, why not simply do:
R3 <- dt[, list(Company, Sales, Product, Share=Sales/sum(Sales)), by=key(dt)]

It should shaved off some time vs doing the merge.

Matthew Dowle said...

First thought is that there are faster ways in base and plyr, so this is being a little too nice to data.table. Using `ave()` you can get the sum repeated through each group, cbind that on, then divide Sales by sum(Sales) as a vectorized op rather than by group.  It's transform() which kills performance really by copying all those groups separately, copying again to add the column on, and rbind'ing them all afterwards.

Matthew Dowle said...

For this use case, the real advantage of data.table comes in v1.8.1 (on R-Forge, not yet on CRAN). There you can add a column by reference, by group. The syntax is one line :

    `DT[, Share := Sales/sum(Sales), by=list(Year,Company)]`

That adds a new column `Share` to `DT`, and populates it within each group. The aim with this syntax is that it reads more than English, whilst being significantly faster than `ave()`, which is the next fastest solution I know.

Markus Gesmann said...

Many thanks for your comments. I had tried something like this with v1.8.0. This will be a great addition to the functionality of the package. I am looking forward to your data.table talk at next week's LondonR meeting.

Markus Gesmann said...

Here are the results with 'ave,' which looks indeed quite impressive. 

system.time(  R4 <- cbind(df,Share=ave(df[["Sales"]], df[["Year"]], df[["Company"]],                                          FUN=function(x) x/sum(x)) )# user  system elapsed #  0.118   0.023   0.280 

How could I have missed this function for all those years?

Matthew Dowle said...

That's more like it, right scale now.

The r3 time for data.table appears slower (0.442s) because that timing includes a call to data.table(), to setkey() and the join back. None of those are actually needed. Cutting straight to v1.8.1, you should find the direct := by group is much faster than the 0.280 for ave()+cbind().

However, 2e4 is a very tiny dataset. Time differences of under 1 second hardly matter. Try scaling it up to 1e6, 1e7, 1e8 rows and you should see data.table significantly faster on significant times (e.g. hours down to minutes have been reported on datatable-help).

Here's a related benchmark (but needs N increasing for data.table to shine, see comments) :
http://stackoverflow.com/questions/10748253/idiomatic-r-code-for-partitioning-a-vector-by-an-index-and-performing-an-operati

SF said...

"...but maybe it shouldn't be surprise that the elegance of ddply comes with a price as well"
why should elegance come with a price in terms of performance?

Post a Comment