GFA (version 1.0.5)

robustComponents: Robust GFA components

Description

robustComponents analyzes a collection of sampling chains and returns robust components.

Usage

robustComponents(models, corThr = 0.9, matchThr = 0.5)

Value

A list with the following elements (when input data are paired in two modes, the returned list is of length 2, containing the following elements for each mode):

Krobust

The number of robust components found with the given thresholds.

effect

The component effect in the data space; and array of size \(N \times \sum_{m=1}^M D_m \times Krobust.\)

indices

The corresponding component indices; a \(length(models) \times Krobust\) matrix. Negative indices denote the closest component in the corresponding repetition, having no components above the threshold.

cor

The correlations of the components matched to this robust component; a matrix of size \(length(models) \times Krobust\). The correlations are reported relative to the first repetition with this component observed.

Arguments

models

Either a vector containing the file names, where the models are saved as 'res', or a list containing the models.

corThr

How close two components are required to be, in terms of correlation, in order to match them.

matchThr

How big proportion of the chains need to contain the component to include it in the robust components.

Details

The function returns the effects (i.e. reconstructions) of robust components to the data level. It is useful for a thorough model interpretation, accumulating power over several sampling chains by comparing them in the observation space (as opposed to the latent space). The function is needed for this task, as the extreme multi-modality of factor analysis prohibits efficient sampling techniques that would result in a posterior estimate converging to the true posterior in practice. The function uses a heuristic correlation-based procedure to analyze which components occur frequently in GFA sampling chains.

Examples

X <- matrix(rnorm(10*2),10,2)
W <- matrix(rnorm(15*2),15,2)
Y <- tcrossprod(X,W) + matrix(rnorm(10*15),10,15)
opts <- getDefaultOpts() #Default options
#Fast runs for the demo, default options recommended in general
opts[c("iter.burnin", "iter.max")] <- c(500, 1000)
res <- list()
for(i in 1:4) res[[i]] <- gfa(list(Y[,1:6],Y[,7:15]),opts=opts,K=3)
rob <- robustComponents(res)