exponential_family_harmonium: Exponential Family Harmoniums

In andland/generalizedPCA: Generalized PCA

Description

Welling et al. (2005)'s Exponential Family Harmoniums

Usage

exponential_family_harmonium(x, k = 2, family = c("gaussian", "binomial",
  "poisson"), family_hidden = c("gaussian", "binomial", "poisson"),
  cd_iters = 1, learning_rate = 0.001, max_iters = 100,
  rms_prop = FALSE, quiet = TRUE, random_start = TRUE, start_W, mu,
  main_effects = TRUE)

Arguments

x	matrix of either binary, count, or continuous data
k	dimension (number of hidden units)
family	exponential family distribution of data
family_hidden	exponential family distribution of hidden units
cd_iters	number of iterations for contrastive divergence (CD) at each iteration. It should be a vector of two integers, which is the range of CD interations that the algorithm will perform from the beginning until the end, linearly interpolated
learning_rate	learning rate used for gradient descent
max_iters	maximum number of iterations
rms_prop	logical; whether to use RMS prop for optimization. Default is FALSE.
quiet	logical; whether the calculation should give feedback
random_start	whether to randomly initialize W
start_W	initial value for W
mu	specific value for mu, the mean (bias) vector of x
main_effects	logical; whether to include main effects (bias terms) in the model

Value

An S3 object of class efh which is a list with the following components:

mu	the main effects (bias terms) for dimensionality reduction
hidden_bias	the bias for the hidden units (currently hard coded to 0)
W	the dxk-dimentional matrix with the loadings
family	the exponential family of the data
family_hidden	the exponential family of the hidden units
iters	number of iterations required for convergence
loss_trace	the trace of the average deviance of the algorithm. Should be non-increasing
prop_deviance_expl	the proportion of deviance explained by this model. If main_effects = TRUE, the null model is just the main effects, otherwise the null model estimates 0 for all natural parameters.

References

Welling, Max, Michal Rosen-Zvi, and Geoffrey E. Hinton. "Exponential family harmoniums with an application to information retrieval." Advances in neural information processing systems. 2005.

Examples

rows = 100
cols = 10
set.seed(1)
mat_np = outer(rnorm(rows), rnorm(cols))

# generate a count matrix and binary response
mat = matrix(rpois(rows * cols, c(exp(mat_np))), rows, cols)
mat[1, 1] <- NA

modp = exponential_family_harmonium(mat, k = 2, family = "poisson", quiet = FALSE,
                                    learning_rate = 0.001, rms_prop = FALSE, max_iters = 250)
gmf = generalizedMF(mat, k = 1, family = "poisson", quiet = FALSE)

andland/generalizedPCA documentation built on May 12, 2019, 2:42 a.m.