GRP: Fast Grouping / collapse Grouping Objects

Description

GRP performs fast, ordered and unordered, groupings of vectors and data frames (or lists of vectors) using radixorder or group. The output is a list-like object of class 'GRP' which can be printed, plotted and used as an efficient input to all of collapse's fast statistical and transformation functions and operators (see macros .FAST_FUN and .OPERATOR_FUN), as well as to collap, BY and TRA.

fgroup_by is similar to dplyr::group_by but faster and class-agnostic. It creates a grouped data frame with a 'GRP' object attached - for fast dplyr-like programming with collapse's fast functions.

There are also several conversion methods to and from 'GRP' objects. Notable among these is GRP.grouped_df, which returns a 'GRP' object from a grouped data frame created with dplyr::group_by or fgroup_by, and the duo GRP.factor and as_factor_GRP.

gsplit efficiently splits a vector based on a 'GRP' object, and greorder helps to recombine the results. These are the workhorses behind functions like BY, and collap, fsummarise and fmutate when evaluated with base R and user-defined functions.

Usage

GRP(X, ...)
# S3 method for default
GRP(X, by = NULL, sort = .op[["sort"]], decreasing = FALSE, na.last = TRUE,
    return.groups = TRUE, return.order = sort, method = "auto",
    call = TRUE, ...)
# S3 method for factor
GRP(X, ..., group.sizes = TRUE, drop = FALSE, return.groups = TRUE,
    call = TRUE)
# S3 method for qG
GRP(X, ..., group.sizes = TRUE, return.groups = TRUE, call = TRUE)
# S3 method for pseries
GRP(X, effect = 1L, ..., group.sizes = TRUE, return.groups = TRUE,
    call = TRUE)
# S3 method for pdata.frame
GRP(X, effect = 1L, ..., group.sizes = TRUE, return.groups = TRUE,
    call = TRUE)
# S3 method for grouped_df
GRP(X, ..., return.groups = TRUE, call = TRUE)
# Identify 'GRP' objects
is_GRP(x)
# S3 method for GRP
length(x)                          # Length of data being grouped
GRPN(x, expand = TRUE, ...)        # Group sizes (default: expanded to match data length)
GRPid(x, sort = FALSE, ...)        # Group id (data length, same as GRP(.)$group.id)
GRPnames(x, force.char = TRUE, sep = ".")  # Group names
as_factor_GRP(x, ordered = FALSE, sep = ".") # 'GRP'-object to (ordered) factor conversion
# Efficiently split a vector using a 'GRP' object
gsplit(x, g, use.g.names = FALSE, ...)
# Efficiently reorder y = unlist(gsplit(x, g)) such that identical(greorder(y, g), x)
greorder(x, g, ...)
# Fast, class-agnostic pendant to dplyr::group_by for use with fast functions, see details
fgroup_by(.X, ..., sort = .op[["sort"]], decreasing = FALSE, na.last = TRUE,
          return.groups = TRUE, return.order = sort, method = "auto")
# Standard-evaluation analogue (slim wrapper around GRP.default(), for programming)
group_by_vars(X, by = NULL, ...)
# Shorthand for fgroup_by
gby(.X, ..., sort = .op[["sort"]], decreasing = FALSE, na.last = TRUE,
    return.groups = TRUE, return.order = sort, method = "auto")
# Get grouping columns from a grouped data frame created with dplyr::group_by or fgroup_by
fgroup_vars(X, return = "data")
# Ungroup grouped data frame created with dplyr::group_by or fgroup_by
fungroup(X, ...)
# S3 method for GRP
print(x, n = 6, ...)
# S3 method for GRP
plot(x, breaks = "auto", type = "l", horizontal = FALSE, ...)

Value

A list-like object of class `GRP' containing information about the number of groups, the observations (rows) belonging to each group, the size of each group, the unique group names / definitions, whether the groups are ordered and data grouped is sorted or not, the ordering vector used to perform the ordering and the group start positions. The object is structured as follows:

List-index		Element-name		Content type		Content description
	[[1]]		N.groups		`integer(1)`
Number of Groups		[[2]]		group.id		`integer(NROW(X))`
	An integer group-identifier		[[3]]		group.sizes
`integer(N.groups)`		Vector of group sizes		[[4]]		groups
	`unique(X)` or `NULL`		Unique groups (same format as input, except for `fgroup_by` which uses a plain list, sorted if `sort = TRUE`), or `NULL` if `return.groups = FALSE`		[[5]]
group.vars		`character`		The names of the grouping variables		[[6]]
	ordered		`logical(2)`		`[1]` Whether the groups are ordered: equal to the `sort` argument in the default method, or `TRUE` if converted objects inherit a class `"ordered"` and `NA` otherwise, `[2]` Whether the data (`X`) is already sorted: the result of `!is.unsorted(group.id)`. If `sort = FALSE` (default method) the second entry is `NA`.
[[7]]		order		`integer(NROW(X))` or `NULL`		Ordering vector from `radixorder` (with `"starts"` attribute), or `NULL` if `return.order = FALSE`
	[[8]]		group.starts		`integer(N.groups)` or `NULL`
The first-occurrence positions/rows of the groups. Useful e.g. with `ffirst(x, g, na.rm = FALSE)`. `NULL` if `return.groups = FALSE`.		[[9]]		call		`match.call()` or `NULL`

Arguments

X

a vector, list of columns or data frame (default method), or a suitable object (conversion / extractor methods).

.X

a data frame or list.

x, g

a 'GRP' object. For gsplit/greorder, x can be a vector of any type, or NULL to return the integer indices of the groups. gsplit/greorder/GRPN/GRPid also support vectors or data frames to be passed to g/x.

by

if X is a data frame or list, by can indicate columns to use for the grouping (by default all columns are used). Columns must be passed using a vector of column names, indices, a one-sided formula i.e. ~ col1 + col2, a logical vector (converted to indices) or a selector function e.g. is_categorical.

sort

logical. If FALSE, groups are not ordered but simply grouped in the order of first appearance of unique elements / rows. This often provides a performance gain if the data was not sorted beforehand. See also method.

ordered

logical. TRUE adds a class 'ordered' i.e. generates an ordered factor.

decreasing

logical. Should the sort order be increasing or decreasing? Can be a vector of length equal to the number of arguments in X / by (argument passed to radixorder).

na.last

logical. If missing values are encountered in grouping vector/columns, assign them to the last group (argument passed to radixorder).

return.groups

logical. Include the unique groups in the created GRP object.

return.order

logical. If sort = TRUE, include the output from radixorder in the created GRP object. This brings performance improvements in gsplit (and thus also benefits grouped execution of base R functions).

method

character. The algorithm to use for grouping: either "radix", "hash" or "auto". "auto" will chose "radix" when sort = TRUE, yielding ordered grouping via radixorder, and "hash"-based grouping in first-appearance order via group otherwise. It is possibly to put method = "radix" and sort = FALSE, which will group character data in first appearance order but sort numeric data (a good hybrid option). method = "hash" currently does not support any sorting, thus putting sort = TRUE will simply be ignored.

group.sizes

logical. TRUE tabulates factor levels using tabulate to create a vector of group sizes; FALSE leaves that slot empty when converting from factors.

drop

logical. TRUE efficiently drops unused factor levels beforehand using fdroplevels.

call

logical. TRUE calls match.call and saves it in the final slot of the GRP object.

expand

logical. TRUE returns a vector the same length as the data. FALSE returns the group sizes (computed in first-appearance-order of groups if x is not already a 'GRP' object).

force.char

logical. Always output group names as character vector, even if a single numeric vector was passed to GRP.default.

sep

character. The separator passed to paste when creating group names from multiple grouping variables by pasting them together.

effect

plm / indexed data methods: Select which panel identifier should be used as grouping variable. 1L takes the first variable in the index, 2L the second etc., identifiers can also be passed as a character string. More than one variable can be supplied.

return

an integer or string specifying what fgroup_vars should return. The options are:

Int.	String	Description
1	"data"	full grouping columns (default)
2	"unique"	unique rows of grouping columns
3	"names"	names of grouping columns
4	"indices"	integer indices of grouping columns
5	"named_indices"	named integer indices of grouping columns
6	"logical"	logical selection vector of grouping columns
7	"named_logical"	named logical selection vector of grouping columns

use.g.names

logical. TRUE returns a named list, like split. FALSE is slightly more efficient.

n: integer. Number of groups to print out.
breaks: integer. Number of breaks in the histogram of group-sizes.
type: linetype for plot.
horizontal: logical. TRUE arranges plots next to each other, instead of above each other. Note that the size of each group is only plotted for objects with less than 10,000 groups.
...: for fgroup_by: unquoted comma-separated column names, sequences of columns, expressions involving columns, and column names, indices, logical vectors or selector functions. See Examples. For group_by_vars, gsplit, greorder, GRPN and GRPid: further arguments passed to GRP (if g/x is not already a 'GRP' object). For example the by argument could be used if a data frame is passed.

Details

GRP is a central function in the collapse package because it provides, in the form of integer vectors, some key pieces of information to efficiently perform grouped operations at the C/C++ level.

Most statistical function require information about (1) the number of groups (2) an integer group-id indicating which values / rows belong to which group and (3) information about the size of each group. Provided with these, collapse's Fast Statistical Functions pre-allocate intermediate and result vectors of the right sizes and (in most cases) perform grouped statistical computations in a single pass through the data.

The sorting functionality of GRP.default lets groups receive different integer-id's depending on whether the groups are sorted sort = TRUE (FALSE gives first-appearance order), and in which order (argument decreasing). This affects the order of values/rows in the output whenever an aggregation is performed.

Other elements in the object provide information about whether the data was sorted by the variables defining the grouping (6) and the ordering vector (7). These also feed into optimizations in gsplit/greorder that benefit the execution of base R functions across groups.

Complimentary to GRP, the function fgroup_by is a significantly faster and class-agnostic alternative to dplyr::group_by for programming with collapse. It creates a grouped data frame with a 'GRP' object attached in a "groups" attribute. This data frame has classes 'GRP_df', ..., 'grouped_df' and 'data.frame', where ... stands for any other classes the input frame inherits such as 'data.table', 'sf', 'tbl_df', 'indexed_frame' etc.. collapse functions with a 'grouped_df' method respond to 'grouped_df' objects created with either fgroup_by or dplyr::group_by. The method GRP.grouped_df takes the "groups" attribute from a 'grouped_df' and converts it to a 'GRP' object if created with dplyr::group_by.

The 'GRP_df' class in front responds to print.GRP_df which first calls print(fungroup(x), ...) and prints one line below the object indicating the grouping variables, followed, in square brackets, by some statistics on the group sizes: [N | Mean (SD) Min-Max]. The mean is rounded to a full number and the standard deviation (SD) to one digit. Minimum and maximum are only displayed if the SD is non-zero. There also exist a method [.GRP_df which calls NextMethod but makes sure that the grouping information is preserved or dropped depending on the dimensions of the result (subsetting rows or aggregation with data.table drops the grouping object).

GRP.default supports vector and list input and will also return 'GRP' objects if passed. There is also a hidden method GRP.GRP which simply returns grouping objects (no re-grouping functionality is offered).

Apart from GRP.grouped_df there are several further conversion methods:

The conversion of factors to 'GRP' objects by GRP.factor involves obtaining the number of groups calling ng <- fnlevels(f) and then computing the count of each level using tabulate(f, ng). The integer group-id (2) is already given by the factor itself after removing the levels and class attributes and replacing any missing values with ng + 1L. The levels are put in a list and moved to position (4) in the 'GRP' object, which is reserved for the unique groups. Finally, a sortedness check !is.unsorted(id) is run on the group-id to check if the data represented by the factor was sorted (6). GRP.qG works similarly (see also qG), and the 'pseries' and 'pdata.frame' methods simply group one or more factors in the index (selected using the effect argument) .

Creating a factor from a 'GRP' object using as_factor_GRP does not involve any computations, but may involve interacting multiple grouping columns using the paste function to produce unique factor levels.

Examples

Run this code

## default method
GRP(mtcars$cyl)
GRP(mtcars, ~ cyl + vs + am)       # Or GRP(mtcars, c("cyl","vs","am")) or GRP(mtcars, c(2,8:9))
g <- GRP(mtcars, ~ cyl + vs + am)  # Saving the object
print(g)                           # Printing it
plot(g)                            # Plotting it
GRPnames(g)                        # Retain group names
GRPid(g)                           # Retain group id (same as g$group.id), useful inside fmutate()
fsum(mtcars, g)                    # Compute the sum of mtcars, grouped by variables cyl, vs and am
gsplit(mtcars$mpg, g)              # Use the object to split a vector
gsplit(NULL, g)                    # The indices of the groups
identical(mtcars$mpg,              # greorder and unlist undo the effect of gsplit
          greorder(unlist(gsplit(mtcars$mpg, g)), g))

## Convert factor to GRP object and vice-versa
GRP(iris$Species)
as_factor_GRP(g)
 
## dplyr integration
library(dplyr)
mtcars |> group_by(cyl,vs,am) |> GRP()    # Get GRP object from a dplyr grouped tibble
mtcars |> group_by(cyl,vs,am) |> fmean()  # Grouped mean using dplyr grouping
mtcars |> fgroup_by(cyl,vs,am) |> fmean() # Faster alternative with collapse grouping

mtcars |> fgroup_by(cyl,vs,am)            # Print method for grouped data frame

## Adding a column of group sizes.
mtcars |> fgroup_by(cyl,vs,am) |> fsummarise(Sizes = GRPN())
# Note: can also set_collapse(mask = "n") to use n() instead, see help("collapse-options")
# Other usage modes:
mtcars |> fgroup_by(cyl,vs,am) |> fmutate(Sizes = GRPN())
mtcars |> fmutate(Sizes = GRPN(list(cyl,vs,am)))  # Same thing, slightly more efficient

## Various options for programming and interactive use
fgroup_by(GGDC10S, Variable, Decade = floor(Year / 10) * 10) |> head(3)
fgroup_by(GGDC10S, 1:3, 5) |> head(3)
fgroup_by(GGDC10S, c("Variable", "Country")) |> head(3)
fgroup_by(GGDC10S, is.character) |> head(3)
fgroup_by(GGDC10S, Country:Variable, Year) |> head(3)
fgroup_by(GGDC10S, Country:Region, Var = Variable, Year) |> head(3)

## Note that you can create a grouped data frame without materializing the unique grouping columns
fgroup_by(GGDC10S, Variable, Country, return.groups = FALSE) |> fmutate(across(AGR:SUM, fscale))
fgroup_by(GGDC10S, Variable, Country, return.groups = FALSE) |> fselect(AGR:SUM) |> fmean()

## Note also that setting sort = FALSE on unsorted data can be much faster... if not required...
library(microbenchmark)
microbenchmark(gby(GGDC10S, Variable, Country), gby(GGDC10S, Variable, Country, sort = FALSE))