Best way to store variable-length data in an R data.frame?

I have some mixed-type data that I would like to store in an R data structure of some sort. Each data point has a set of fixed attributes which may be 1-d numeric, factors, or characters, and also a set of variable length data. For exampl开发者_StackOverflow社区e:

id  phrase                    num_tokens  token_lengths
1   "hello world"             2           5 5
2   "greetings"               1           9
3   "take me to your leader"  4           4 2 2 4 6

The actual values are not all computable from one another, but that's the flavor of the data. The operations I'm going to want to do include subsetting the data based on boolean functions (e.g. something like nchar(data$phrase) > 10 or lapply(data$token_lengths, length) > 2). I'd also like to index and average values in the variable length portion by index. This doesn't work, but something like: mean(data$token_lengths[1], na.rm=TRUE))

I've found I can shoehorn "token_lengths" into a data.frame by making it an array:

d <- data.frame(id=c(1,2,3), ..., token_lengths=as.array(list(c(5,5), 9, c(4,2,2,4,6)))

But is this the best way?

Trying to shoehorn the data into a data frame seems hackish to me. Far better to consider each row as an individual object, then think of the dataset as an array of these objects.

This function converts your data strings to an appropriate format. (This is S3 style code; you may prefer to use one of the 'proper' object oriented systems.)

as.mydata <- function(x)
{
   UseMethod("as.mydata")
}

as.mydata.character <- function(x)
{
   convert <- function(x)
   {
      md <- list()
      md$phrase = x
      spl <- strsplit(x, " ")[[1]]
      md$num_words <- length(spl)
      md$token_lengths <- nchar(spl)
      class(md) <- "mydata"
      md
   }
   lapply(x, convert)
}

Now your whole dataset looks like

mydataset <- as.mydata(c("hello world", "greetings", "take me to your leader"))

mydataset
[[1]]
$phrase
[1] "hello world"

$num_words
[1] 2

$token_lengths
[1] 5 5

attr(,"class")
[1] "mydata"

[[2]]
$phrase
[1] "greetings"

$num_words
[1] 1

$token_lengths
[1] 9

attr(,"class")
[1] "mydata"

[[3]]
$phrase
[1] "take me to your leader"

$num_words
[1] 5

$token_lengths
[1] 4 2 2 4 6

attr(,"class")
[1] "mydata"

You can define a print method to make this look prettier.

print.mydata <- function(x)
{
   cat(x$phrase, "consists of", x$num_words, "words, with", paste(x$token_lengths, collapse=", "), "letters.")
}
mydataset
[[1]]
hello world consists of 2 words, with 5, 5 letters.
[[2]]
greetings consists of 1 words, with 9 letters.
[[3]]
take me to your leader consists of 5 words, with 4, 2, 2, 4, 6 letters.

The sample operations you wanted to do are fairly straightforward with data in this format.

sapply(mydataset, function(x) nchar(x$phrase) > 10)
[1]  TRUE FALSE  TRUE

I would just use the data in the "long" format.

E.g.

> d1 <- data.frame(id=1:3, num_words=c(2,1,4), phrase=c("hello world", "greetings", "take me to your leader"))
> d2 <- data.frame(id=c(rep(1,2), rep(2,1), rep(3,5)), token_length=c(5,5,9,4,2,2,4,6))
> d2$tokenid <- with(d2, ave(token_length, id, FUN=seq_along))
> d <- merge(d1,d2)
> subset(d, nchar(phrase) > 10)
  id num_words                 phrase token_length tokenid
1  1         2            hello world            5       1
2  1         2            hello world            5       2
4  3         4 take me to your leader            4       1
5  3         4 take me to your leader            2       2
6  3         4 take me to your leader            2       3
7  3         4 take me to your leader            4       4
8  3         4 take me to your leader            6       5
> with(d, tapply(token_length, id, mean))
  1   2   3 
5.0 9.0 3.6 

Once the data is in the long format, you can use sqldf or plyr to extract what you want from it.

Another option would be to convert your data frame into a matrix of mode list - each element of the matrix would be a list. standard array operations (slicing with [, apply(), etc. would be applicable).

> d <- data.frame(id=c(1,2,3), num_tokens=c(2,1,4), token_lengths=as.array(list(c(5,5), 9, c(4,2,2,4,6))))
> m <- as.matrix(d)
> mode(m)
[1] "list"
> m[,"token_lengths"]
[[1]]
[1] 5 5

[[2]]
[1] 9

[[3]]
[1] 4 2 2 4 6

> m[3,]
$id
[1] 3

$num_tokens
[1] 4

$token_lengths
[1] 4 2 2 4 6

Since the R data frame structure is based loosely on the SQL table, having each element of the data frame be anything other than an atomic data type is uncommon. However, it can be done, as you've shown, and this linked post describes such an application implemented on a larger scale.

An alternative is to store your data as a string and have a function to retrieve it, or create a separate function to which the data is attached and extract it using indices stored in your data frame.

> ## alternative 1
> tokens <- function(x,i=TRUE) Map(as.numeric,strsplit(x[i],","))
> d <- data.frame(id=c(1,2,3), token_lengths=c("5,5", "9", "4,2,2,4,6"))
> 
> tokens(d$token_lengths)
[[1]]
[1] 5 5

[[2]]
[1] 9

[[3]]
[1] 4 2 2 4 6

> tokens(d$token_lengths,2:3)
[[1]]
[1] 9

[[2]]
[1] 4 2 2 4 6

> 
> ## alternative 2
> retrieve <- local({
+   token_lengths <- list(c(5,5), 9, c(4,2,2,4,6))
+   function(i) token_lengths[i]
+ })
> 
> d <- data.frame(id=c(1,2,3), token_lengths=1:3)
> retrieve(d$token_lengths[2:3])
[[1]]
[1] 9

[[2]]
[1] 4 2 2 4 6

I would also use strings for the variable length data, but as in the following example: "c(5,5)" for the first phrase. One needs to use eval(parse(text=...)) to carry out computations.

For example, the mean can be computed as follows:

sapply(data$token_lengths,function(str) mean(eval(parse(text=str))))