sum_by {gustave}

Efficient by-group (weighted) summation

Description

sum_by performs an efficient and optionally weighted by-group summation by using linear algebra and the Matrix package capabilities. The by-group summation is performed through matrix cross-product of the y parameter (coerced to a matrix if needed) with a (very) sparse matrix built up using the by and the (optional) w parameters.

Compared to base R, dplyr or data.table alternatives, this implementation aims at being easier to use in a matrix-oriented context and can yield efficiency gains when the number of columns becomes high.

Usage

sum_by(y, by, w = NULL, na_rm = TRUE, keep_sparse = FALSE)

Arguments

y

A (sparse) vector, a (sparse) matrix or a data.frame. The object to perform by-group summation on.

by

The factor variable defining the by-groups. Character variables are coerced to factors.

w

The optional row weights to be used in the summation.

na_rm

Should NA values in y be removed (ie treated as 0 in the summation) ? Similar to na.rm argument in sum, but TRUE by default. If FALSE, NA values in y produce NA values in the result.

keep_sparse

When y is a sparse vector or a sparse matrix, should the result also be sparse ? FALSE by default. As sparseVector-class does not have a name attribute, when y is a sparseVector the result does not have any name (and a warning is cast).

Value

A vector, a matrix or a data.frame depending on the type of y. If y is sparse and keep_sparse = TRUE, then the result is also sparse (without names when it is a sparse vector, see keep_sparse argument for details).

Author(s)

Martin Chevalier

Examples

# Data generation
set.seed(1)
n <- 100
p <- 10
H <- 3
y <- matrix(rnorm(n*p), ncol = p, dimnames = list(NULL, paste0("var", 1:10)))
y[1, 1] <- NA
by <- letters[sample.int(H, n, replace = TRUE)]
w <- rep(1, n)
w[by == "a"] <- 2

# Standard use
sum_by(y, by)

# Keeping the NAs
sum_by(y, by, na_rm = FALSE)

# With a weight
sum_by(y, by, w = w)