Multiple dispatch based on dataframes

library(interfacer)

Rationale

The S3 type system allows for dispatch based on the first argument of a function. In the situation where we are developing functions that use dataframes as input selecting a dispatch function needs to be based on the structure of the input rather than its class. interfacer can use iface specifications to associate

Dispatch

Dispatching to one of a number of functions based on the nature of a dataframe input is enabled by idispatch(...). This emulates the behaviour of S3 classes but for dataframes, based on their columns and also their grouping. Consider the following iface specifications:

i_test = iface(
  id = integer ~ "an integer ID",
  test = logical ~ "the test result"
)

# Extends the i_test to include an additional column
i_test_extn = iface(
  i_test,
  extra = character ~ "a new value",
  .groups = FALSE
)

We can create specific handlers for each type of data and decide which function to dispatch to at runtime based on the input dataframe.


# The generic function
disp_example = function(x, ...) {
  idispatch(x,
    disp_example.extn = i_test_extn,
    disp_example.no_extn = i_test
  )
}

# The handler for extended input dataframe types
disp_example.extn = function(x = i_test_extn, ...) {
  message("extended data function")
  return(colnames(x))
}

# The handler for non-extended input dataframe types
disp_example.no_extn = function(x = i_test, ...) {
  message("not extended data function")
  return(colnames(x))
}

If we call disp_example() with data that matches the i_test_extn specification we get one type of behaviour:


tmp = tibble::tibble(
    id=c("1","2","3"),
    test = c(TRUE,FALSE,TRUE),
    extra = 1.1
)

tmp %>% disp_example()
#> extended data function
#> [1] "id"    "test"  "extra"

But if we call disp_example() with data that only matches the i_test specification we get different behaviour:

# this matches the i_test_extn specification:
tmp2 = tibble::tibble(
    id=c("1","2","3"),
    test = c(TRUE,FALSE,TRUE)
)

tmp2 %>% disp_example()
#> not extended data function
#> [1] "id"   "test"

I’ve used this mechanism, for example, to configure how plots are produced depending on the input.

Recursive dispatch

The order of the rules is important. In general the more detailed specifications needing to be provided first, and the more generic specifications last. We can leverage this to create a recursive functional pattern of dataframe processing that allows multiple inputs to converge on a single output, this also demonstrates the use of itest() which simply checks an input conforms to an iface specification:


# generic type 1 input
i_input_1 = iface(
  x = integer ~ "the positives",
  n = default(100) + integer ~ "the total"
)

# generic type 2 input
i_input_2 = iface(
  p = proportion ~ "the proportion",
  n = default(100) + integer ~ "the total"
)

# more detailed combined type 1 and 2 input
i_interim = iface(
  i_input_1,
  i_input_2
)

# most detailed input format
i_final = iface(
  i_interim,
  lower = double ~ "wilson lower CI",
  upper = double ~ "wilson lower CI",
  mean = double ~ "wilson mean"
)

# final target output format
i_target = iface(
  i_final,
  label = character ~ "a printable label"
)

# processes input of type 1 and 
process.input_1 = function(x = i_input_1,...) {
  message("process input 1")
  ireturn(x %>% dplyr::mutate(p = x/n), iface = i_interim)
}

process.input_2 = function(x = i_input_2,...) {
  message("process input 2")
  ireturn(x %>% dplyr::mutate(x = floor(p*n)), iface = i_interim)
}

process.interim = function(x) {
  message("process interim")
  ireturn(x %>% dplyr::mutate(binom::binom.wilson(x,n)), iface = i_final)
}

process.final = function(x) {
  message("process final")
  ireturn(x %>% dplyr::mutate(label = sprintf("%1.1f%% [%1.1f%% - %1.1f%%] (%d/%d)", 
    mean*100, lower*100, upper*100, x, n)), iface = i_target)
}

process = function(x,...) {
  # this test must be at the front to prevent infinite recursion
  if (itest(x, i_target)) return(x)
  out = idispatch(x,
    process.final = i_final,
    process.interim = i_interim,
    process.input_2 = i_input_2,
    process.input_1 = i_input_1
  )
  return(process(out))
}

Processing an input of type 1 results in one path through the data pipeline:

# tibble::tibble(x=c(10,30), n=c(NA,50)) %>% itest(i_input_1)
process(tibble::tibble(x=c(10,30), n=c(NA,50))) %>% dplyr::glimpse()
#> process input 1
#> process interim
#> process final
#> Rows: 2
#> Columns: 8
#> $ x      <int> 10, 30
#> $ n      <int> 100, 50
#> $ p      <dbl> 0.1, 0.6
#> $ method <chr> "wilson", "wilson"
#> $ mean   <dbl> 0.1, 0.6
#> $ lower  <dbl> 0.05522914, 0.46181438
#> $ upper  <dbl> 0.1743657, 0.7239161
#> $ label  <chr> "10.0% [5.5% - 17.4%] (10/100)", "60.0% [46.2% - 72.4%] (30/50)"

Processing an input of type 2, results in a different path through the data pipeline, but the same outcome:

# tibble::tibble(p=0.15,n=1000) %>% itest(i_input_2)
process(tibble::tibble(p=0.15,n=1000)) %>% dplyr::glimpse()
#> process input 2
#> process interim
#> process final
#> Rows: 1
#> Columns: 8
#> $ p      <dbl> 0.15
#> $ n      <int> 1000
#> $ x      <int> 150
#> $ method <chr> "wilson"
#> $ mean   <dbl> 0.15
#> $ lower  <dbl> 0.1292101
#> $ upper  <dbl> 0.1734687
#> $ label  <chr> "15.0% [12.9% - 17.3%] (150/1000)"

Care must be taken though in this pattern, particularly if you are re-using column names,as data-type coercion could result in some column types being switched backwards and forwards, and other infinite loop problems.

Grouping based dispatch

It is often useful to have a function that can expects a specific grouping but can handle additional groups. One way of handling these is to use purrr and nest columns extensively. Nesting data in the unexpected groups and repeatedly applying the function you want. An alternative dplyr solution is to use a group_modify. interfacer leverages this second option to automatically determine a grouping necessary for a pipeline function from the stated grouping requirements and automatically handle them without additional coding in the package.

For example if we have the following iface the input for a function must be grouped only by the color column:

i_diamond_price = interfacer::iface(
  color = enum(`D`,`E`,`F`,`G`,`H`,`I`,`J`, .ordered=TRUE) ~ "the color column",
  price = integer ~ "the price column",
  .groups = ~ color
)

A package developer writing a pipeline function may use this fact to handle possible additional grouping by using a igroup_process(df, ...)

 # exported function in package
 # at param can use `r idocument(ex_mean, df)` for documentation
 ex_mean = function(df = i_diamond_price, extra_param = ".") {

   # dispatch based on groupings:
   igroup_process(df,

     # the real work of this function is provided as an anonymous inner
     # function (but can be any other function e.g. package private function)
     # or a purrr style lambda.

     function(df, extra_param) {
       message(extra_param, appendLF = FALSE)
       return(df %>% dplyr::summarise(mean_price = mean(price)))
     }

   )
 }

If we pass this to correctly grouped data conforming to i_diamond_price the inner function is executed once transparently, after the input has been validated:

# The correctly grouped dataframe
ggplot2::diamonds %>%
  dplyr::group_by(color) %>%
  ex_mean(extra_param = "without additional groups...") %>%
  dplyr::glimpse()
#> without additional groups...
#> Rows: 7
#> Columns: 2
#> $ color      <ord> D, E, F, G, H, I, J
#> $ mean_price <dbl> 3169.954, 3076.752, 3724.886, 3999.136, 4486.669, 5091.875,…

If on the other hand additional groups are present the inner function is executed once for each of the additional groups. Data validation happens once per group, which affects interpretation of uniqueness.

# The incorrectly grouped dataframe
ggplot2::diamonds %>%
  dplyr::group_by(cut, color) %>%
  ex_mean() %>%
  dplyr::glimpse()
#> .....
#> Rows: 35
#> Columns: 3
#> Groups: cut [5]
#> $ cut        <ord> Fair, Fair, Fair, Fair, Fair, Fair, Fair, Good, Good, Good,…
#> $ color      <ord> D, E, F, G, H, I, J, D, E, F, G, H, I, J, D, E, F, G, H, I,…
#> $ mean_price <dbl> 4291.061, 3682.312, 3827.003, 4239.255, 5135.683, 4685.446,…

The output of this is actually grouped by cut as the color column grouping is consumed by the nested function in igroup_process.