genSupMF: Supervised matrix factorization for exponential family data

In andland/genSupPCA: Supervised Dimensionality Reduction for Exponential Family Data

Description

Supervised matrix factorization for exponential family data

Usage

genSupMF(x, y, k = 2, alpha = NULL, family_x = c("gaussian", "binomial",
  "poisson"), family_y = c("gaussian", "binomial", "poisson"), quiet = TRUE,
  max_iters = 1000, conv_criteria = 1e-05, random_start = FALSE, start_A,
  start_B, start_beta, mu, update_A = TRUE, update_beta = TRUE,
  lambda = 0.01)

Arguments

x	covariate matrix
y	response vector
k	dimension
alpha	balance between dimensionality reduction of x and prediction of y
family_x	exponential family distribution of covariates
family_y	exponential family distribution of response
quiet	logical; whether the calculation should give feedback
max_iters	maximum number of iterations
conv_criteria	convergence criteria
random_start	whether to randomly initialize A and B
start_A	initial value for A
start_B	initial value for B
start_beta	initial value for beta. Mainly effective if update_beta is TRUE
mu	specific value for mu, the mean vector of x
update_A	logical; whether to update the scores A
update_beta	logical; whether to update the coefficients beta
lambda	an L2 penalty on the parameters to prevent numerical issues. This was reportedly used in Rish's implementation

Value

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

mu	the main effects for dimensionality reduction
A	the nxk-dimentional matrix with the scores
B	the dxk-dimentional matrix with the loadings
beta	the k + 1 length vector of the coefficients
family_x	the exponential family of covariates
family_y	the exponential family of response
iters	number of iterations required for convergence
loss_trace	the trace of the average negative log likelihood of the algorithm. Should be non-increasing

References

Rish, Irina, et al. "Closed-form supervised dimensionality reduction with generalized linear models." Proceedings of the 25th international conference on Machine learning. ACM, 2008.

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)
response = rbinom(rows, 1, rowSums(mat) / max(rowSums(mat)))

mod = genSupMF(mat, response, k = 1, alpha = 1, family_x = "poisson", family_y = "binomial",
               quiet = FALSE)

plot(inv.logit.mat(cbind(1, mod$A) %*% mod$beta), response)
plot(rowSums(mat), response)
## Not run: 
ggplot(data.frame(PC = mod$PCs[, 1], y = response), aes(PC, y)) + stat_summary_bin(bins = 10)

## End(Not run)

andland/genSupPCA documentation built on May 30, 2019, 11:43 a.m.