getDefaultOpts: A function for generating the default priors of GFA model

In GFA: Group Factor Analysis

A function for generating the default priors of GFA model

Description

getDefaultOpts returns the priors of GFA

Usage

getDefaultOpts(bicluster = FALSE)

Arguments

bicluster

Use binary sparsity priors in both the data modes? If FALSE (default), the components will be dense in the data sources, but group-sparse, i.e., each component is active in a (potentially different) subset of the data sources. If TRUE, binary sparsity is inferred for each data sample and feature, resulting in each component to be interpretable as a multi-source bicluster.

Details

This function returns options defining the model's high-level structure (sparsity priors) and the model's hyperparameters, defining uninformative priors. We recommend keeping these as provided, with one exception: if the uninformative prior of the noise residual (tau) seems to result in an overly complex model (no components become shut down even if the initial K is set high), risking overfitting, we recommend using function informativeNoisePrior to adjust the priors.

Value

A list with the following model options:

tauGrouped	If TRUE (default), data views have separate noise precisions, otherwise each feature has.
normalLatents	If TRUE, X will have a normal prior; if FALSE X, will have a spike-and-slab prior.
spikeW	Sparsity prior of W. "group"=group sparsity, "element"= element-wise sparsity with shared hyperparameter across views, "shared"= element-wise sparsity with no grouping.
ARDW	ARD prior type for W, determining the scale of the inferred components. "shared"=same scale for all the data sources, "grouped" (default)= separate scale for each data source, "element"=separate scale for each feature.
ARDLatent	ARD prior type for X: "shared" (default)=shared scale for all the samples, "element"=separate scale for each sample.
imputation	Missing value imputation type: "Bayesian" (default)=proper Bayesian handling of missing values. "conservative"=missing values result in smaller parameter scale, which can be useful if tricky missing value structure causes exaggerated imputed values with the default setting (which can also be dealt with informative priors for alpha and beta).
iter.max	The total number of Gibbs sampling steps (default 5000).
iter.saved	The number of saved posterior samples (default 100).
iter.burnin	The number of burn-in samples (default 2500).
init.tau	The initial noise precision. High values imply initializing the model with an adequate number of components. Default 1000.
sampleZ	When to start sampling spike and slab parameters (default: Gibbs sample 1).
prior.alpha_0t	The shape parameter of tau's prior (default 10).
prior.beta_0t	The rate parameter of tau's prior (default 10).
prior.alpha_0	The shape parameter of alpha's prior (default 10).
prior.beta_0	The rate parameter of alpha's prior (default 01).
prior.alpha_0X	The shape parameter of beta's prior (default 10).
prior.beta_0X	The rate parameter of beta's prior (default 1).
prior.beta	Bernoulli prior for the spike-and-slab prior of W (counts for 1s and 0s; default c(1,1)).
prior.betaX	Bernoulli prior for the possible spike-and-slab prior of X (default c(1,1)).
verbose	The verbosity level. 0=no printing, 1=moderate printing, 2=maximal printing (default 1).
convergenceCheck	Check for the convergence of the data reconstruction, based on the Geweke diagnostic (default FALSE).
save.posterior	A list determining which parameters' posterior samples are saved (default: X, W and tau).

Examples

#Given pre-specified data collection Y and component number K
opts <- getDefaultOpts(bicluster=FALSE)
opts$normalLatents <- FALSE #Binary sparsity for each sample and data source
 ## Not run: model <- gfa(Y,opts,K)

GFA documentation built on Oct. 21, 2023, 5:06 p.m.