vectorization in fuzz_m_ratio()

[luketudge]

The documentation for fuzz_m_ratio() says:

Arguments
a A character vector of items to match to b.

b A character vector of items to match to a.

method The method to use for fuzzy matching.

Value
Returns a score of same length as b, giving the proportional dissimilarity between a and b.

Problem

I am not sure how to interpret the fact that a is permitted to be a character vector with more than a single item.

I guess that the most common use case of the fuzz_ functions is with a as a single item. In this case, fuzz_m_ratio() works as expected and recycles a over all items in b:

b <- c("a", "b", "ab")
fuzz_m_ratio("a", b)

[1] 0.0000000 1.0000000 0.3333333

According to the standard R vectorization rules (see for example R for Data Science), if a is a vector with the same number of items as b, then the items from a and b should be paired up and the result should be the pairwise dissimilarity.

But this isn't what happens. For example:

fuzz_m_ratio(c("a", "b", "ab"), b)

[1] 1 1 1

Here, each item in a is identical to the corresponding item in b so we should get 0 0 0 dissimilarity.

It isn't the case that the function is simply reporting similarity instead of dissimilarity here. For example:

fuzz_m_ratio(c("a", "b", "x"), b)

[1] 1 1 1

Here some items from a match their counterparts in b and some don't, so we should get both 0s and 1s in the answer.

It looks as though the same inconsistencies occur for at least some of the other fuzz_ functions too.

Suggestions

I am not sure what is intended for the case of multiple items in a, and from the source code I can't quite work out what is actually being done. Possibly the function is only intended for single-item a. If this is the case, then maybe the documentation needs amending, and perhaps also an error or warning message if a has multiple items, since some R users will probably expect the standard vectorization rules.

On the other hand, if vectorization is wanted here, then maybe one general solution is to first create a non-vectorized version of the function that calculates the distance for a single pair of strings, for example:

fmr_nonvectorized <- function(str1, str2){
  lengths <- nchar(c(str1, str2))
  len_shortest <- min(lengths)
  char_matches <- unlist(strsplit(str1, ""))[1:len_shortest] == unlist(strsplit(str2, ""))[1:len_shortest]
  similarity <- 2 * sum(char_matches) / sum(lengths)
  return(1 - similarity)
}

fmr_nonvectorized("a", "ab")

[1] 0.3333333

And then apply it to two vectors using mapply(), for example:

fmr_vectorized <- function(a, b){
  distances <- mapply(fmr_nonvectorized, a, b, USE.NAMES = FALSE)
  return(distances)
}

This gives the standard R-style vectorization rules for the examples above:

fmr_vectorized("a", b)

[1] 0.0000000 1.0000000 0.3333333

fmr_vectorized(c("a", "b", "ab"), b)

[1] 0 0 0

fmr_vectorized(c("a", "b", "x"), b)

[1] 0 0 1

fmr_vectorized(c("x", "x", "x"), b)

[1] 1 1 1

And as a bonus this will give the standard warning if the longer of a and b is not an integer multiple of the other:

fmr_vectorized(c("a", "b"), b)

[1] 0.0000000 0.0000000 0.3333333
Warning message:
In mapply(fmr_nonvectorized, a, b, USE.NAMES = FALSE) :
  longer argument not a multiple of length of shorter

[mjwestgate]

Hi Luke,

Thanks for picking this up and apologies for not getting back to you sooner. You are indeed correct that this function is intended for the case where length(a) == 1. It honestly hadn't occurred to me to try and make it work for cases where both a and b have length > 1, but I can see why that would be a useful feature.

As for the error - that's a shame - obviously something got missed when we ported this from revtools to synthesisr. I'll have a look at that soon.

Martin

[luketudge]

Hi @mjwestgate

Thanks for taking a look at this! I hadn't realized this came from revtools. I see now that the same issue is already present there (though the revtools documentation for fuzz_functions does specify that a should be a string).

Let me know if I can be of any help.