rKhatriRao: `reduced' Khatri-Rao product (sparse matrices)

Description

This function performs a Khatri-Rao product (`column-wise Kronecker product', see KhatriRao for more info) on two sparse matrices. However, the result of such a product on sparse matrices normally results in very many empty rows. This function removes those empty rows, and, most importantly, it produces row names only for the remaining rows. For large sparse matrices this is much more efficient than first producing all rownames, and then removing the one with the empty rows.

Usage

rKhatriRao(X, Y, 
	rownamesX = rownames(X), rownamesY = rownames(Y), 
	simplify = FALSE, binder = ":", FUN = "*")

Arguments

X,Y

matrices of with the same number of columns.

rownamesX, rownamesY

row names of matrices X and Y. These can be specified separately, but they default to the row names of the matrices.

simplify

by default, the names of rows and columns are not included into the matrix to keep the matrix as lean as possible: the row names are returned separately. Using include.dimnames=T adds the row names into the matrix. The column names are directly taken from X.

binder

symbol to include between the row names of X and Y for the resulting matrix

FUN

function to be used in the KhatriRao product, passed internally to the workhorse KhatriRao

Value

By default, the result is a list of two items:

resulting sparse product matrix with empty rows removed

rownames

a vector with the resulting row names for the non-empty rows

When simplify=T, then the matrix is return with the row names included.

Details

Up to 1e6 row names to be produced goes reasonably quick with the basic KhatriRao function. However, larger amounts of pasting of row names becomes very slow, and the row names take an enormous amount of RAM. This function solves that problem by only producing row names for the non-empty rows.

Examples

Run this code

# NOT RUN {
# two sparse matrices with row names

X <- rSparseMatrix(1e4, 1e3, 1e4)
Y <- rSparseMatrix(1e4, 1e3, 1e4)

rownames(X) <- 1:nrow(X)
rownames(Y) <- 1:nrow(Y)

# the basic KhatriRao product is very fast
# but almost all rows are empty

system.time(M <- KhatriRao(X, Y))
# }
# NOT RUN {
sum(rowSums(M)==0)/nrow(M) # 99.9% empty rows
# }
# NOT RUN {
# To produce all row names takes a long time
# with 1e8 row names it took half an hour on my laptop
# so: don't try the following, except on a very large machine!
# }
# NOT RUN {
system.time(M <- KhatriRao(X, Y, make.dimnames = TRUE))
# }
# NOT RUN {
# Using the current special version works just fine and is reasonably quick
system.time(M <- rKhatriRao(X, Y))

# }