cv_ggb: Cross Validation to Select Lambda for GGB Estimator

In unix9999/insurance: Graph-Guided Banding of the Covariance Matrix

Description

Performs nfold-fold cross validation to select the tuning parameter for GGB. Takes fit object created by ggb and computes CV error at the sequence of lambda values from that object.

Usage

cv_ggb(x, fit, g, errfun = NULL, nfolds = 5, ...)

Arguments

x	n-by-p data matrix
fit	object of class ggb (as outputted by ggb)
g	seed graph (should match that used to compute fit)
errfun	a user-specified function measuring the loss incurred by estimating est (first argument) when the true covariance matrix is true (second argument). Default: Squared Frobenius norm.
nfolds	number of folds for cross-validation (default: 5)
...	additional arguments (passed to ggb)

Value

errs:

A nlam-by-nfolds matrix of errors. errs[l, j] is error incurred in using lamlist[l] on fold j

m:

CV error error for each value of lambda.

se:

Standard error (estimated over folds) for each value of lambda

lambda_best:

Value of lambda minimizing CV error.

ibest:

Index of lambda minimizing CV error.

lambda_1se:

Selected value of lambda using the one-standard-error rule, a common heuristic that favors a sparser model when there isn't strong evidence against it.

i1se:

Index of lambda of one-standard-error rule.

See Also

ggb

Examples

set.seed(123)
n <- 20
p <- 10
g <- igraph::graph.lattice(c(5, 2))
x <- matrix(rnorm(n * p), n, p)
S <- stats::cov(x)
fit <- ggb(S, g, type = "local")
cv <- cv_ggb(x, fit, g)
plot(cv)
Sighat <- fit$Sig[[cv$i1se]]

unix9999/insurance documentation built on May 26, 2019, 5:34 a.m.