The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.
The coop package does co-operations: covariance, correlation, and cosine similarity. And it does them very quickly. If you can do this faster, I’d love to know how.
The package is optimized for high performance, and has different implementations for dense matrix inputs, dense vector-vector inputs, and sparse matrix inputs. Note that to get good performance with this package (as seen in these benchmarks), you will need to use a good BLAS library. See the package vignette for details.
For more information, including algorithmic details, see the package vignettes.
To install the R package, run:
install.packages("coop")
The development version is maintained on GitHub, and can easily be installed by any of the packages that offer installations from GitHub:
### Pick your preference
::install_github("wrathematics/coop")
devtools::install_github("wrathematics/coop")
ghit::install_github("wrathematics/coop") remotes
The C internals are completely separated from the R wrapper code. So
if you prefer, you can easily build this as a C shared library after
removing the file src/wrapper.c
.
The package has functions for covariance and pearson correlation with
interfaces that mimic base R’s, with the addition of a
cosine()
function. At this time, the basic interface looks
like this:
### matrix input
covar(x) # like cov(x)
pcor(x) # like cor(x)
cosine(x) # like lsa::cosine(x)
### vector input
covar(x, y) # like cov(x, y)
pcor(x, y) # like cor(x, y)
cosine(x, y) # like lsa::cosine(x, y)
There are also t
versions of the functions which operate
on the transposed data (without producing a copy). So
tcovar(x)
will do the same computation as as
cov(t(x))
(but much more efficiently).
The functions also have an additional argument inverse
,
which will return the matrix inverse of the specified operation. So
covar(x, inverse=TRUE)
will return the inverted covariance
matrix.
For more details, see the package vignette.
Here we provide some benchmarks for dense matrices. The package also has vector-vector methods for each operation, and a sparse method for cosine similarity. These also perform quite well, but in the case of the former are generally not performance intensive, and in the case of the latter, I am not aware of any other sparse cosine similarity implementations available to R.
All of these benchmarks can be found in the source tree of this
package, under coop/inst/benchmarks
. Implementation details
can be found in the package vignette. All tests performed using:
library(coop)
library(rbenchmark)
<- c("test", "replications", "elapsed", "relative")
cols <- 25
reps
<- 10000
m <- 250
n <- matrix(rnorm(m*n), m, n) x
benchmark(cov(x), covar(x), replications=reps, columns=cols)
## test replications elapsed relative
## 2 covar(x) 25 0.431 1.000
## 1 cov(x) 25 8.656 20.084
benchmark(cor(x), pcor(x), replications=reps, columns=cols)
## test replications elapsed relative
## 1 cor(x) 25 8.695 16.818
## 2 pcor(x) 25 0.517 1.000
benchmark(lsa::cosine(x), cosine(x), replications=reps, columns=cols)
## test replications elapsed relative
## 2 cosine(x) 25 0.341 1.000
## 1 lsa::cosine(x) 25 181.045 530.924
We can outperform the lsa implementation using just R code as follows:
<- function(x)
cosine_R
{<- crossprod(x)
cp <- sqrt(diag(cp))
rtdg <- cp / tcrossprod(rtdg)
cos return(cos)
}
We note that while this implementation is reasonably “clock efficient”, it is very memory wasteful compared to the implementation in coop (and for much larger data sizes, the implementation in coop will dominate).
library(compiler)
<- cmpfun(cosine_R)
cosine_R
benchmark(cosine_R(x), cosine(x), replications=reps, columns=cols)
## test replications elapsed relative
## 1 cosine_R(x) 25 0.287 1.079
## 2 cosine(x) 25 0.266 1.000
In fact, similar tricks can be played with both covariance and correlation. One major reason coop is so much faster is because of its careful use of the BLAS. See the vignette for more information.
These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.