fastCor: Fast correlation for large matrices

Description Usage Arguments Details Value Author(s) References See Also Examples

View source: R/fastCor.R

Description

fastCor is a helper function that compute Pearson correlation matrix for HiClimR and validClimR functions. It is similar to cor function in R but uses a faster implementation on 64-bit machines (an optimized BLAS library is highly recommended). fastCor also uses a memory-efficient algorithm that allows for splitting the data matrix and only compute the upper-triangular part of the correlation matrix. It can be used to compute correlation matrix for the columns of any data matrix.

Usage

1
fastCor(xt, nSplit = 1, upperTri = FALSE, verbose = TRUE)

Arguments

xt

an (M rows by N columns) matrix of 'double' values: N objects (spatial points or stations) to be clustered by M observations (temporal points or years). It is the transpose of the imput matrix x required for HiClimR and validClimR functions.

nSplit

integer number greater than or equal to one, to split the data matrix into nSplit splits of the total number of columns ncol(xt). If nSplit = 1, the default method will be used to compute correlation matrix for the full data matrix (no splits). If nSplit > 1, the correlation matrix (or the upper-triangular part if upperTri = TRUE) will be allocated and filled with the computed correlation sub-matrix for each split. the first n-1 splits have equal size while the last split may include any remaining columns. This is used with upperTri = TRUE to compute only the upper-triangular part of the correlation matrix. The maximum number of splits nSplitMax = floor(N / 2) makes splits with 2 columns; if nSplit > nSplitMax, nSplitMax will be used. Very large number of splits nSplit makes computation slower but it could handle big data or if the available memory is not enough to allocate the correlation matrix, which helps in solving the “Error: cannot allocate vector of size...” memory limitation problem. It is recommended to start with a small number of splits. If the data is very large compared to the physical memory, it is highly recommended to use a 64-Bit machine with enough memory resources and/or use coarsening feature for gridded data by setting lonStep > 1 and latStep > 1.

upperTri

logical to compute only the upper-triangular half of the correlation matrix if upperTri = TRUE and nSplit > 1., which includes all required info since the correlation/dissimilarity matrix is symmetric. This almost halves memory use, which can be very important for big data.

verbose

logical to print processing information if verbose = TRUE.

Details

The fastCor function computes the correlation matrix by calling the cross product function in the Basic Linear Algebra Subroutines (BLAS) library used by R. A significant performance improvement can be achieved when building R on 64-bit machines with an optimized BLAS library, such as ATLAS, OpenBLAS, or the commercial Intel MKL. For big data, the memory required to allocate the square matrix of correlations may exceed the total amount of physical memory available resulting in “Error: cannot allocate vector of size...”. fastCor allows for splitting the data matrix into nSplit splits and only computes the upper-triangular part of the correlation matrix with upperTri = TRUE. This almost halves memory use, which can be very important for big data. If nSplit > 1, the correlation matrix (or the upper-triangular part if upperTri = TRUE) will be allocated and filled with computed correlation sub-matrix for each split. the first n-1 splits have equal size while the last split may include any remaining columns.

Value

An (N rows by N columns) correlation matrix.

Author(s)

Hamada Badr <[email protected]>, Ben Zaitchik <[email protected]>, and Amin Dezfuli <[email protected]>.

References

Hamada S. Badr, Zaitchik, B. F. and Dezfuli, A. K. (2015): A Tool for Hierarchical Climate Regionalization, Earth Science Informatics, 1-10, http://dx.doi.org/10.1007/s12145-015-0221-7.

Hamada S. Badr, Zaitchik, B. F. and Dezfuli, A. K. (2014): Hierarchical Climate Regionalization, CRAN, http://cran.r-project.org/package=HiClimR.

bigcor: Large correlation matrices in R, https://rmazing.wordpress.com.

See Also

HiClimR, validClimR, geogMask, coarseR, fastCor, grid2D, and minSigCor.

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
require(HiClimR)

## Load test case data
x <- TestCase$x

## Use fastCor function to compute the correlation matrix
t0 <- proc.time() ; xcor <- fastCor(t(x)) ; proc.time() - t0
## compare with cor function
t0 <- proc.time() ; xcor0 <- cor(t(x)) ; proc.time() - t0

## Not run: 

## Split the data into 10 splits and return upper-triangular half only
xcor10 <- fastCor(t(x), nSplit = 10, upperTri = TRUE)


## End(Not run)

HiClimR documentation built on May 29, 2017, 8:45 p.m.