cpp.fit.random.block.sgd: Fit a GMF model using the adaptive SGD with block-wise minibatch subsampling

Description

Fit a GMF model using the adaptive SGD with block-wise minibatch subsampling

Usage

cpp.fit.random.block.sgd(
  Y,
  X,
  B,
  A,
  Z,
  U,
  V,
  O,
  W,
  familyname,
  linkname,
  varfname,
  ncomp,
  lambda,
  maxiter = 1000L,
  eps = 0.01,
  nafill = 10L,
  tol = 1e-08,
  size1 = 100L,
  size2 = 100L,
  burn = 0.75,
  rate0 = 0.01,
  decay = 0.01,
  damping = 0.001,
  rate1 = 0.95,
  rate2 = 0.99,
  parallel = FALSE,
  nthreads = 1L,
  verbose = TRUE,
  frequency = 250L,
  progress = FALSE
)

Arguments

Y: matrix of responses (\(n \times m\))
X: matrix of row fixed effects (\(n \times p\))
B: initial row-effect matrix (\(n \times p\))
A: initial column-effect matrix (\(n \times q\))
Z: matrix of column fixed effects (\(m \times q\))
U: initial factor matrix (\(n \times d\))
V: initial loading matrix (\(m \times d\))
O: matrix of constant offset (\(n \times m\))
W: matrix of constant weights (\(n \times m\))
familyname: a glm model family name
linkname: a glm link function name
varfname: variance function name
ncomp: rank of the latent matrix factorization
lambda: penalization parameters
maxiter: maximum number of iterations
eps: shrinkage factor for extreme predictions
nafill: how often the missing values are updated
tol: tolerance threshold for the stopping criterion
size1: row-minibatch dimension
size2: column-minibatch dimension
burn: burn-in period in which the learning late is not decreased
rate0: initial learning rate
decay: decay rate of the learning rate
damping: diagonal dumping factor for the Hessian matrix
rate1: decay rate of the first moment estimate of the gradient
rate2: decay rate of the second moment estimate of the gradient
parallel: if TRUE, allows for parallel computing
nthreads: number of cores to be used in parallel
verbose: if TRUE, print the optimization status
frequency: how often the optimization status is printed
progress: if TRUE, print an progress bar