CCAGFA (version 1.0.8)

GFA: Estimate a Bayesian IBFA/CCA/GFA model

Description

Estimates the parameters of a Bayesian group factor analysis (GFA), canonical correlation analysis (BCCA), or inter-battery factor analysis (BIBFA).

GFA is a latent variable model for explaining relationships between multiple data matrices with co-occurring samples. The model finds linear factors that explain dependencies between these matrices, similarly to how factor analysis explains dependencies between individual variables. BIBFA is a special case of GFA for two data matrices. It finds factors explaining the relationship between them, as well as factors explaining the residual variation in each matrix. The solution of BIBFA equals that of CCA, with additional factors for explaining the data-specific noise.

Usage

CCA(Y, K, opts)
GFA(Y, K, opts)
CCAexperiment(Y, K, opts, Nrep=10)
GFAexperiment(Y, K, opts, Nrep=10)

Arguments

Y

A list containing matrices with N rows (samples) and D[m] columns (features). Must have exactly two matrices for CCA and any number of co-occurring matrices for GFA.

K

The number of components.

opts

A list of parameters and options to be used when learning the model. See getDefaultOpts.

Nrep

The number of random initializations used for learning the model; only used for CCAexperiment and GFAexperiment.

Value

The methods return a list that contains all the model parameters and other details.

Z

The mean of the latent variables; N times K matrix

covZ

The covariance of the latent variables; K times K matrix

ZZ

The second moments Z^TZ; K times K matrix

W

List of the mean projections; D_i times K matrices

covW

List of the covariances of the projections; K times K matrices

WW

List of the second moments W^TW; K times K matrices

tau

The mean precisions (inverse variance, so 1/tau gives the variances denoted by sigma in the paper); M-element vector

alpha

The mean precisions of the projection weights, used in the ARD prior; M times K matrix

cost

Vector collecting the variational lower bounds for each iteration

D

Data dimensionalities; M-element vector

K

The number of latent factors

datavar

The total variance in the data sets, needed for GFAtrim

R

The rank of alpha

U

The group factor loadings; M times R matrix

V

The latent group factors; K times R matrix

u.mu

The mean of group factor loadings U; M-element vector

v.mu

The mean of latent group factors V; K-element vector

Details

The recommended strategy is to use GFAexperiment for learning a Bayesian group factor analysis model. It simply calls GFA Nrep times and returns the model with the best variational lower bound for the marginal likelihood.

CCAexperiment and CCA are simple wrappers for the corresponding GFA functions, to be used for the case of M=2 data sets. CCA is a special case of GFA with exactly two co-occurring matrices, and these functions are provided for convenience only.

References

Virtanen, S. and Klami, A., and Kaski, S.: Bayesian CCA via group-wise sparsity. In Proceedings of the 28th International Conference on Machine Learning (ICML), pages 457-464, 2011.

Virtanen, S. and Klami, A., and Khan, S.A. and Kaski, S.: Baysian group factor analysis. In Proceedings of the 15th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 22 of JMLR W&CP, pages 1269-1277, 2012.

Klami, A. and Virtanen, S., and Kaski, S.:Bayesian Canonical Correlation Analysis. Journal of Machine Learning Research, 2013.

Klami, A. and Virtanen, S., Leppaaho, E., and Kaski, S.:Group Factor Analysis. Submitted to a journal, 2014.

See Also

getDefaultOpts

Examples

#
# Create simple random data
#

N <- 50; D <- c(4,6)       # 50 samples with 4 and 6 dimensions
tau <- c(3,3)              # residual noise precision

K <- 3                          # K real components (1 shared, 1+1 private)
Z <- matrix(rnorm(N*K,0,1),N,K) # drawn from the prior
alpha <- matrix(c(1,1,1e6,1,1e6,1),2,3) 

Y <- vector("list",length=2)
W <- vector("list",length=2)

for(view in 1:2) {
  W[[view]] <- matrix(0,D[view],K)
  for(k in 1:K) {
    W[[view]][,k] <- rnorm(D[view],0,1/sqrt(alpha[view,k]))
  }
  Y[[view]] <- Z %*% t(W[[view]]) +
    matrix(rnorm(N*D[view],0,1/sqrt(tau[view])),N,D[view])
}

#
# Run the model
#
opts <- getDefaultOpts()
opts$iter.max <- 10       # Terminate early for fast testing

# Only tries two random initializations for faster testing
model <- CCAexperiment(Y,K,opts,Nrep=2)