ZITNMF: Zero-Inflated Tweedie Non-negative Matrix Factorization...

In nnTensor: Non-Negative Tensor Decomposition

Zero-Inflated Tweedie Non-negative Matrix Factorization (ZITNMF)

Description

The input data is assumed to be a non-negative matrix with excess zeros. ZITNMF decomposes the matrix into two low-dimensional factor matrices while accounting for zero inflation using a Tweedie distribution model.

Usage

ZITNMF(X, Z=NULL, pseudocount=1e-10,
  initF=NULL, initA=NULL, initW=0.5, fixF=FALSE, fixA=FALSE,
  initializer=c("NMF", "ALS", "Random"),
  J=3, Beta=2, phi=1, thr=1e-10, num.iter=100, verbose=FALSE)

Arguments

X	The input non-negative matrix which has N-rows and M-columns.
Z	The zero-inflation indicator matrix which has N-rows and M-columns (Default: NULL). If NULL, it is estimated from the data.
pseudocount	The pseudo count to avoid zero division (Default: 1e-10).
initF	The initial values of factor matrix F, which has N-rows and J-columns (Default: NULL).
initA	The initial values of factor matrix A, which has M-rows and J-columns (Default: NULL).
initW	The initial value of zero-inflation probability w, a scalar between 0 and 1 (Default: 0.5).
fixF	Whether the factor matrix F is updated in each iteration step (Default: FALSE).
fixA	Whether the factor matrix A is updated in each iteration step (Default: FALSE).
initializer	The initialization method for factor matrices (Default: "NMF"). "NMF" uses Beta-divergence NMF, "ALS" uses SVD-based initialization, and "Random" uses random uniform values.
J	The number of low-dimension (Default: 3).
Beta	The Beta parameter for Beta-divergence. Beta=2 corresponds to Frobenius norm, Beta=1 to KL divergence, Beta=0 to IS divergence (Default: 2). Note that Beta < 1 requires the tweedie package.
phi	The dispersion parameter for the Tweedie distribution (Default: 1).
thr	The threshold for convergence based on relative change of Beta-divergence (Default: 1e-10). If NULL, no convergence check is performed and all num.iter iterations are run.
num.iter	The number of maximum iterations (Default: 100).
verbose	If verbose is TRUE, the iteration number and Beta-divergence value are printed (Default: FALSE).

Value

F : The factor matrix (N-rows and J-columns)

A : The factor matrix (M-rows and J-columns)

Z : The estimated zero-inflation indicator matrix (N-rows and M-columns)

w : The estimated zero-inflation probability (scalar)

RecError : The Beta-divergence values at each iteration

RelChange : The relative change of Beta-divergence at each iteration

Author(s)

Koki Tsuyuzaki

References

Zero-inflated models for count data and their extensions.

Examples

if(interactive()){
  # Test data
  X <- toyModel("NMF")

  # Frobenius
  out <- ZITNMF(X, J=3, Beta=2, num.iter=10, initializer="Random")

  # KL
  out <- ZITNMF(X, J=3, Beta=1, num.iter=10, initializer="Random")
}

nnTensor documentation built on May 8, 2026, 5:07 p.m.