NSFA_gibbs: Gibbs Sampler for the Non-parametric Sparse Factor Analysis...

In miguelbiron/SBGLM: Sparse Bayes Generalized Linear Models

Description

This function runs a Gibbs sampler to obtain posterior samples from the NSFA model by Knowles & Ghahramani (2011). This is my personal implementation, which contains some corrections with respect to the model presented in the original article.

Usage

NSFA_gibbs(S, Y, hp = NULL, Y_test = NULL, R = 1L, verbose = 0L)

Arguments

S	number of iterations for the chain.
Y	matrix with the observed data, of size D x N.
hp	list of hyperparameters. See Details for a description and default values used.
Y_test	matrix with test set for assessing out-of-sample performance. Number of columns must be same as Y (N).
R	function returns last R values in the chain. Default is 1.
verbose	integer. Function prints every verbose iterations. The default value of 0 prints no output.

Details

The model was originally proposed in Knowles & Ghahramani (2011). However, you can read my own report evaluating this model here.

The following hyperparameters can be supplied by the user through hp. Default values are in parenthesis and tend to work well.

• e: shape parameter of prior for alpha (1).

• f: rate parameter of prior for alpha (1).

• a: shape parameter of prior for psi^-1 (1).

• a_0: shape parameter of prior for b (1).

• b_0: rate parameter of prior for b (1).

• c: shape parameter of prior for lambda (1).

• c_0: shape parameter of prior for d (1).

• d_0: rate parameter of prior for d (1).

• lambda_MH: lambda parameter for kappa_d proposal (10).

• pi_MH: pi parameter for kappa_d proposal (0.1).

If chain keeps restarting, try increasing lambda_MH and/or pi_MH to get more aggressive proposals for growing Z.

Value

List with two lists:

chain

list with last R samples of the chain.

stats

list of traces of train set log-likelihood (ll), test set ll, and K.

References

Knowles, D., & Ghahramani, Z. (2011). Nonparametric Bayesian sparse factor models with application to gene expression modeling. The Annals of Applied Statistics, 1534-1552.

miguelbiron/SBGLM documentation built on May 29, 2019, 8:23 p.m.