Type: Package
Title: Computes Proximity in Large Sparse Matrices
Version: 0.5.2
Description: Computes proximity between rows or columns of large matrices efficiently in C++. Functions are optimised for large sparse matrices using the Armadillo and Intel TBB libraries. Among various built-in similarity/distance measures, computation of correlation, cosine similarity, Dice coefficient and Euclidean distance is particularly fast.
URL: https://github.com/koheiw/proxyC, https://koheiw.github.io/proxyC/
BugReports: https://github.com/koheiw/proxyC/issues
License: GPL-3
Depends: R (≥ 3.1.0), methods
Imports: Matrix (≥ 1.2), Rcpp (≥ 0.12.12)
Suggests: testthat, entropy, proxy, knitr, rmarkdown
LinkingTo: Rcpp, RcppArmadillo (≥ 0.7.600.1.0)
NeedsCompilation: yes
Encoding: UTF-8
RoxygenNote: 7.3.2
VignetteBuilder: knitr
Config/Needs/website: coop, Rfast, parallelDist, distances, proxy, microbenchmark, ggplot2
Packaged: 2025-04-25 10:25:52 UTC; watan
Author: Kohei Watanabe ORCID iD [cre, aut, cph], Robrecht Cannoodt ORCID iD [aut]
Maintainer: Kohei Watanabe <watanabe.kohei@gmail.com>
Repository: CRAN
Date/Publication: 2025-04-25 12:40:06 UTC

Standard deviation of columns and rows of large matrices

Description

Produces the same result as apply(x, 1, sd) or apply(x, 2, sd) without coercing matrix to dense matrix. Values are not identical to sd() because of the floating point precision issue in C++.

Usage

colSds(x)

rowSds(x)

Arguments

x

a base::matrix or Matrix::Matrix object.

Examples

mt <- Matrix::rsparsematrix(100, 100, 0.01)
colSds(mt)
apply(mt, 2, sd) # the same

Number of zeros in columns and rows of large matrices

Description

Produces the same result as applying sum(x == 0) to each row or column.

Usage

colZeros(x)

rowZeros(x)

Arguments

x

a base::matrix or Matrix::Matrix object.

Examples

mt <- Matrix::rsparsematrix(100, 100, 0.01)
colZeros(mt)
apply(mt, 2, function(x) sum(x == 0)) # the same

Cross-product of large sparse matrices

Description

Compute the (transposed) cross-product of large sparse matrices using the same infrastructure as simil() and dist().

Usage

crossprod(x, y = NULL, min_prod = NULL, digits = 14)

tcrossprod(x, y = NULL, min_prod = NULL, digits = 14)

Arguments

x

a base::matrix or Matrix::Matrix object. Dense matrices are covered to the Matrix::CsparseMatrix internally.

y

if a base::matrix or Matrix::Matrix object is provided, proximity between documents or features in x and y is computed.

min_prod

the minimum product to be recorded.

digits

determines rounding of small values towards zero. Use primarily to correct floating point errors. Rounding is performed in C++ in a similar way as base::zapsmall.

Create a pattern matrix for masking

Description

Create a pattern matrix for simil() or dist() to enable masked similarity computation. If the matrix is passed to the function, it computes similarity scores only for cells with TRUE.

Usage

mask(x, y = NULL)

Arguments

x

a numeric or character vector matched against each other.

y

a numeric or character vector matched against x if provided.

Value

a sparse logical matrix with TRUE for matched pairs.

Examples

mt1 <- Matrix::rsparsematrix(100, 6, 1.0)
colnames(mt1) <- c("a", "a", "d", "d", "e", "e")
mt2 <- Matrix::rsparsematrix(100, 5, 1.0)
colnames(mt2) <- c("a", "b", "c", "d", "e")

(msk <- mask(colnames(mt1), colnames(mt2)))
simil(mt1, mt2, margin = 2, mask = msk, drop0 = TRUE)

Compute similarity/distance between rows or columns of large matrices

Description

Fast similarity/distance computation function for large sparse matrices. You can floor small similarity value to to save computation time and storage space by an arbitrary threshold (min_simil) or rank (rank). You can specify the number of threads for parallel computing via options(proxyC.threads).

Usage

simil(
  x,
  y = NULL,
  margin = 1,
  method = c("cosine", "correlation", "dice", "edice", "jaccard", "ejaccard", "fjaccard",
    "hamann", "faith", "simple matching"),
  mask = NULL,
  min_simil = NULL,
  rank = NULL,
  drop0 = FALSE,
  diag = FALSE,
  use_nan = NULL,
  sparse = TRUE,
  digits = 14
)

dist(
  x,
  y = NULL,
  margin = 1,
  method = c("euclidean", "chisquared", "kullback", "jeffreys", "jensen", "manhattan",
    "maximum", "canberra", "minkowski", "hamming"),
  mask = NULL,
  p = 2,
  smooth = 0,
  drop0 = FALSE,
  diag = FALSE,
  use_nan = NULL,
  sparse = TRUE,
  digits = 14
)

Arguments

x

a base::matrix or Matrix::Matrix object. Dense matrices are covered to the Matrix::CsparseMatrix internally.

y

if a base::matrix or Matrix::Matrix object is provided, proximity between documents or features in x and y is computed.

margin

integer indicating margin of similarity/distance computation. 1 indicates rows or 2 indicates columns.

method

method to compute similarity or distance

mask

a pattern matrix created using mask() for masked similarity/distance computation. The shape of the matrix must be the same as the resulting matrix.

min_simil

the minimum similarity value to be recorded.

rank

an integer value specifying top-n most similarity values to be recorded.

drop0

if TRUE, removes zero values to make the similarity/distance matrix sparse. It has no effect when dense = TRUE.

diag

if TRUE, only compute diagonal elements of the similarity/distance matrix; useful when comparing corresponding rows or columns of x and y.

use_nan

if TRUE, returns NaN if the standard deviation of a vector is zero when method is "correlation"; if all the values are zero in a vector when method is "cosine", "chisquared", "kullback", "jeffreys" or "jensen". Note that use of NaN makes the similarity/distance matrix denser and therefore larger in RAM. If FALSE, return zero in same use situations as above. If NULL, will also return zero but also generate a warning (default).

sparse

if TRUE, returns Matrix::sparseMatrix object. When neither min_simil nor rank is used, dense matrices require less space in RAM.

digits

determines rounding of small values towards zero. Use primarily to correct floating point errors. Rounding is performed in C++ in a similar way as base::zapsmall.

p

weight for Minkowski distance.

smooth

adds a fixed value to all the cells to avoid division by zero. Only used when method is "chisquared", "kullback", "jeffreys" or "jensen".

Details

Available Methods

Similarity:

Distance:

See the vignette for how the similarity and distance are computed: vignette("measures", package = "proxyC")

Parallel Computing

It performs parallel computing using Intel oneAPI Threads Building Blocks. The number of threads for parallel computing should be specified via options(proxyC.threads) before calling the functions. If the value is -1, all the available threads will be used. Unless the option is used, the number of threads will be limited by the environmental variables (OMP_THREAD_LIMIT or RCPP_PARALLEL_NUM_THREADS) to comply with CRAN policy and offer backward compatibility.

See Also

zapsmall

Examples

mt <- Matrix::rsparsematrix(100, 100, 0.01)
simil(mt, method = "cosine")[1:5, 1:5]
mt <- Matrix::rsparsematrix(100, 100, 0.01)
dist(mt, method = "euclidean")[1:5, 1:5]