This is a README file of the R package matSPACE. Our package
fits sparse partial correlation networks for matrix-variate data,
i.e. observations that are themselves p x q matrices rather than a
single vector of q measurements. We extend the SPACE (Sparse PArtial
Correlation Estimation) joint estimation framework to a
Kronecker-product precision structure, Omega = V %x% U,
where V (q x q) is the column-wise precision matrix and
U (p x p) is the row-wise precision matrix. All partial
correlations are estimated jointly via an L1-penalized (lasso) shooting
algorithm, which preserves symmetry of the estimated network and avoids
the tuning-parameter selection difficulties of separate node-wise
regressions.
To install our package, you may simply execute the following codes:
install.packages("matSPACE")
# development version
# install.packages("pak")
pak::pak("kimhyew1/matSPACE")We give a toy example to apply the main function
matSPACE, which estimates the Kronecker-structured
row/column precision matrices U and V.
We first generate simulated matrix-variate data. Columns 1-2 and
columns 3-4 are correlated by construction, so there is real sparse
structure for matSPACE to recover (pure noise data has no
structure to penalize away, so the BIC-optimal fit is always the fully
sparse one).
set.seed(1)
p = 5
q = 4
n = 20
library(matSPACE)
data = replicate(n, {mat = matrix(rnorm(p * q), p, q)
mat[, 2] = mat[, 2] + 0.8 * mat[, 1]
mat[, 4] = mat[, 4] + 0.8 * mat[, 3]
mat}, simplify = FALSE)Applying matSPACE to the simulated data, a lasso penalty
path is searched for V and U separately, and
the BIC-minimizing fit is returned for each. By default the columns are
weighted by their estimated residual precisions
(f_type_V = f_type_U = "variance"), and at every penalty
the reweighting is iterated until convergence (iter = 500,
outer_tol = 1e-8) over a path of K = 80
penalties.
fit = matSPACE(data)
fit$V # q x q column-wise precision matrix
#> [,1] [,2] [,3] [,4]
#> [1,] 1.0000000 -0.49410136 0.12114557 -0.1096305
#> [2,] -0.4941014 0.66081614 -0.06264098 0.0000000
#> [3,] 0.1211456 -0.06264098 1.03791174 -0.4934104
#> [4,] -0.1096305 0.00000000 -0.49341039 0.8061834
fit$U # p x p row-wise precision matrix
#> [,1] [,2] [,3] [,4] [,5]
#> [1,] 1.3444220 0.2122121 -0.3425416 0.0000000 -0.1820373
#> [2,] 0.2122121 1.6067464 0.0000000 0.2654804 0.1063301
#> [3,] -0.3425416 0.0000000 1.8489166 0.0000000 -0.1215758
#> [4,] 0.0000000 0.2654804 0.0000000 1.5704397 0.1701304
#> [5,] -0.1820373 0.1063301 -0.1215758 0.1701304 1.6366881The full lambda path behind this selection is attached as a
"path" attribute, which lets us plot BIC against lambda
without refitting.
path = attr(fit, "path")
plot(log(path$V$lambda), path$V$bic, type = "b",
xlab = "log(lambda)", ylab = "BIC", main = "Column (V) lambda path")
i_min = which.min(path$V$bic)
points(log(path$V$lambda[i_min]), path$V$bic[i_min], col = "red", pch = 19)
If you already know the penalty you want, space fits a
single lasso penalty directly, without searching a path.
fit_single = space(data, lam = 0.5)
fit_single$ParCor
#> [,1] [,2] [,3] [,4]
#> [1,] 1.0000000 0.62837905 -0.1669241 0.16732558
#> [2,] 0.6283791 1.00000000 0.1215284 -0.04533202
#> [3,] -0.1669241 0.12152840 1.0000000 0.56279861
#> [4,] 0.1673256 -0.04533202 0.5627986 1.00000000We are happy to troubleshoot any issue with the package;