Efficient reshaping using data.tables

2020-07-22

This vignette discusses the default usage of reshaping functions melt (wide to long) and dcast (long to wide) for data.tables as well as the new extended functionalities of melting and casting on multiple columns available from v1.9.6.


Data

We will load the data sets directly within sections.

Introduction

The melt and dcast functions for data.tables are for reshaping wide-to-long and long-to-wide, respectively; the implementations are specifically designed with large in-memory data (e.g. 10Gb) in mind.

In this vignette, we will

  1. First briefly look at the default melting and dcasting of data.tables to convert them from wide to long format and vice versa

  2. Look at scenarios where the current functionalities become cumbersome and inefficient

  3. Finally look at the new improvements to both melt and dcast methods for data.tables to handle multiple columns simultaneously.

The extended functionalities are in line with data.table’s philosophy of performing operations efficiently and in a straightforward manner.

1. Default functionality

a) melting data.tables (wide to long)

Suppose we have a data.table (artificial data) as shown below:

  • measure.vars specify the set of columns we would like to collapse (or combine) together.

  • We can also specify column indices instead of names.

  • By default, variable column is of type factor. Set variable.factor argument to FALSE if you’d like to return a character vector instead.

  • By default, the molten columns are automatically named variable and value.

  • melt preserves column attributes in result.

  • By default, when one of id.vars or measure.vars is missing, the rest of the columns are automatically assigned to the missing argument.

  • When neither id.vars nor measure.vars are specified, as mentioned under ?melt, all non-numeric, integer, logical columns will be assigned to id.vars.

    In addition, a warning message is issued highlighting the columns that are automatically considered to be id.vars.

b) dcasting data.tables (long to wide)

In the previous section, we saw how to get from wide form to long form. Let’s see the reverse operation in this section.

- How can we get back to the original data table DT from DT.m?

That is, we’d like to collect all child observations corresponding to each family_id, age_mother together under the same row. We can accomplish it using dcast as follows:

  • dcast uses formula interface. The variables on the LHS of formula represents the id vars and RHS the measure vars.

  • value.var denotes the column to be filled in with while casting to wide format.

  • dcast also tries to preserve attributes in result wherever possible.

- Starting from DT.m, how can we get the number of children in each family?

You can also pass a function to aggregate by in dcast with the argument fun.aggregate. This is particularly essential when the formula provided does not identify single observation for each cell.

Check ?dcast for other useful arguments and additional examples.

2. Limitations in current melt/dcast approaches

So far we’ve seen features of melt and dcast that are implemented efficiently for data.tables, using internal data.table machinery (fast radix ordering, binary search etc..).

However, there are situations we might run into where the desired operation is not expressed in a straightforward manner. For example, consider the data.table shown below:

s2 <- "family_id age_mother dob_child1 dob_child2 dob_child3 gender_child1 gender_child2 gender_child3
1         30 1998-11-26 2000-01-29         NA             1             2            NA
2         27 1996-06-22         NA         NA             2            NA            NA
3         26 2002-07-11 2004-04-05 2007-09-02             2             2             1
4         32 2004-10-10 2009-08-27 2012-07-21             1             1             1
5         29 2000-12-05 2005-02-28         NA             2             1            NA"
DT <- fread(s2)
DT
#    family_id age_mother dob_child1 dob_child2 dob_child3 gender_child1 gender_child2 gender_child3
# 1:         1         30 1998-11-26 2000-01-29       <NA>             1             2            NA
# 2:         2         27 1996-06-22       <NA>       <NA>             2            NA            NA
# 3:         3         26 2002-07-11 2004-04-05 2007-09-02             2             2             1
# 4:         4         32 2004-10-10 2009-08-27 2012-07-21             1             1             1
# 5:         5         29 2000-12-05 2005-02-28       <NA>             2             1            NA
## 1 = female, 2 = male

And you’d like to combine (melt) all the dob columns together, and gender columns together. Using the current functionality, we can do something like this:

DT.m1 = melt(DT, id = c("family_id", "age_mother"))
DT.m1[, c("variable", "child") := tstrsplit(variable, "_", fixed = TRUE)]
DT.c1 = dcast(DT.m1, family_id + age_mother + child ~ variable, value.var = "value")
DT.c1
#     family_id age_mother  child        dob     gender
#  1:         1         30 child1 1998-11-26 1970-01-02
#  2:         1         30 child2 2000-01-29 1970-01-03
#  3:         1         30 child3       <NA>       <NA>
#  4:         2         27 child1 1996-06-22 1970-01-03
#  5:         2         27 child2       <NA>       <NA>
#  6:         2         27 child3       <NA>       <NA>
#  7:         3         26 child1 2002-07-11 1970-01-03
#  8:         3         26 child2 2004-04-05 1970-01-03
#  9:         3         26 child3 2007-09-02 1970-01-02
# 10:         4         32 child1 2004-10-10 1970-01-02
# 11:         4         32 child2 2009-08-27 1970-01-02
# 12:         4         32 child3 2012-07-21 1970-01-02
# 13:         5         29 child1 2000-12-05 1970-01-03
# 14:         5         29 child2 2005-02-28 1970-01-02
# 15:         5         29 child3       <NA>       <NA>

str(DT.c1) ## gender column is character type now!
# Classes 'data.table' and 'data.frame':    15 obs. of  5 variables:
#  $ family_id : int  1 1 1 2 2 2 3 3 3 4 ...
#  $ age_mother: int  30 30 30 27 27 27 26 26 26 32 ...
#  $ child     : chr  "child1" "child2" "child3" "child1" ...
#  $ dob       : IDate, format: "1998-11-26" "2000-01-29" NA ...
#  $ gender    : IDate, format: "1970-01-02" "1970-01-03" NA ...
#  - attr(*, ".internal.selfref")=<externalptr> 
#  - attr(*, "sorted")= chr [1:3] "family_id" "age_mother" "child"

Issues

  1. What we wanted to do was to combine all the dob and gender type columns together respectively. Instead we are combining everything together, and then splitting them again. I think it’s easy to see that it’s quite roundabout (and inefficient).

    As an analogy, imagine you’ve a closet with four shelves of clothes and you’d like to put together the clothes from shelves 1 and 2 together (in 1), and 3 and 4 together (in 3). What we are doing is more or less to combine all the clothes together, and then split them back on to shelves 1 and 3!

  2. The columns to melt may be of different types, as in this case (character and integer types). By melting them all together, the columns will be coerced in result, as explained by the warning message above and shown from output of str(DT.c1), where gender has been converted to character type.

  3. We are generating an additional column by splitting the variable column into two columns, whose purpose is quite cryptic. We do it because we need it for casting in the next step.

  4. Finally, we cast the data set. But the issue is it’s a much more computationally involved operation than melt. Specifically, it requires computing the order of the variables in formula, and that’s costly.

In fact, stats::reshape is capable of performing this operation in a very straightforward manner. It is an extremely useful and often underrated function. You should definitely give it a try!

3. Enhanced (new) functionality

a) Enhanced melt

Since we’d like for data.tables to perform this operation straightforward and efficient using the same interface, we went ahead and implemented an additional functionality, where we can melt to multiple columns simultaneously.

- Using patterns()

Usually in these problems, the columns we’d like to melt can be distinguished by a common pattern. We can use the function patterns(), implemented for convenience, to provide regular expressions for the columns to be combined together. The above operation can be rewritten as:

That’s it!

  • We can remove the variable column if necessary.

  • The functionality is implemented entirely in C, and is therefore both fast and memory efficient in addition to being straightforward.

b) Enhanced dcast

Okay great! We can now melt into multiple columns simultaneously. Now given the data set DT.m2 as shown above, how can we get back to the same format as the original data we started with?

If we use the current functionality of dcast, then we’d have to cast twice and bind the results together. But that’s once again verbose, not straightforward and is also inefficient.

  • Attributes are preserved in result wherever possible.

  • Everything is taken care of internally, and efficiently. In addition to being fast, it is also very memory efficient.

Multiple functions to fun.aggregate:

You can also provide multiple functions to fun.aggregate to dcast for data.tables. Check the examples in ?dcast which illustrates this functionality.