All notable changes to glmpca will be documented in this file. The format is based on Keep a Changelog
This release should be backwards-compatible even though some
arguments are now deprecated. ### Added - AvaGrad optimizer which is
faster and more numerically stable than Fisher scoring. - Minibatch
methods (memoized and stochastic gradients) to support large sparse
Matrices - Automatic tuning of optimization hyperparameters (learning
rate or penalty) - Negative binomial likelihood with feature-specific
overdispersion parameters - print and predict methods for fitted glmpca
object ### Changed - more detailed documentation of main
glmpca
function - split off large blocks of code in the
main glmpca
function to separate functions - organize
results as an S3 class instead of a list - moved arguments
penalty
and verbose
into the ctl
parameters list - deprecated family mult
and
bern
, replaced with family binom
- control
parameter eps
has now been renamed tol
Initial release of the GLM-PCA method of dimension reduction. The primary focus is on count data (Poisson and negative binomial likelihoods). Bernoulli likelihood has also been implemented for binary data but this has not been tested as thoroughly. ### Added - basic checking for error conditions such as the data matrix having rows or columns that are all zero - L2 (ridge) penalty on latent variables to improve numerical stability - Arbitrary row and/or column covariates may be included to remove nuisance variation from latent factors. - negative binomial shape parameter automatically estimated. - latent factors automatically projected, scaled, and rotated to remove correlation with covariates and maintain orthogonality and improve interpretability.