GFA-package {GFA}

Group factor analysis.

Description

GFA does factor analysis for multiple data sets having matched observations, for exploratory or predictive data analysis.

Details

The posterior distribution of GFA model parameters can be inferred with function gfa, once the priors have been defined with getDefaultOpts. The priors are widely customizable, with two recommended setups: (i) dense group-sparse components (default; similar to package CCAGFA that provides variational Bayesian inference for the same model) and (ii) components interpretable as biclusters. It is recommended to preprocess the data with function normalizeData. Functions are provided for predicting missing data, choosing a prior for the residual noise, identifying robust components and visualizing the inferred model. A simple toy example of the pipeline is provided as demo(GFApipeline), and a more elaborate one as demo(GFAexample). Finally, the experiment presented in (Bunte, Leppaaho, Saarinen and Kaski: Sparse group factor analysis for biclustering of multiple data sources, Bioinformatics, 32(16):2457–2463, 2016) can be replicated with demo(GFAdream). Most of the computational complexity of the package is related to matrix operations, which can be parallelized inherently by using e.g. OpenBLAS libraries.

Examples

#Data generation
X <- matrix(rnorm(20*3),20,3)                     #Latent variables
W <- matrix(rnorm(30*3),30,3)                     #Projection matrix
Y <- tcrossprod(X,W) + matrix(rnorm(20*30),20,30) #Observations
Y <- sweep(Y, MARGIN=2, runif(30), "+")           #Feature means
Y <- list(Y[,1:10], Y[,11:30])                    #Data grouping
#Model inference and visualization
norm <- normalizeData(Y, type="center")           #Centering
opts <- getDefaultOpts()                          #Model options
#Fast runs for the demo, default options recommended in general
opts[c("iter.burnin", "iter.max")] <- c(500, 1000)
res <- gfa(norm$train, K=5, opts=opts)            #Model inference
rec <- reconstruction(res)                        #Reconstruction
recOrig <- undoNormalizeData(rec, norm)           #... to original space
vis <- visualizeComponents(res, Y, norm)          #Visualization