ndNMF: a new discriminant Non-Negative Matrix Factorization (dNMF)

In DNMF: Discriminant Non-Negative Matrix Factorization

a new discriminant Non-Negative Matrix Factorization (dNMF)

Description

The ndNMF algorithm with the additional Fisher criterion on the cost function of conventional NMF was designed to increase class-related discriminating power.

This algorithm is based on articles.

1. Kim, Bo-Kyeong, and Soo-Young Lee. "Spectral Feature Extraction Using dNMF for Emotion Recognition in Vowel Sounds." Neural Information Processing. Springer Berlin Heidelberg, 2013.

2. Lee, Soo-Young, Hyun-Ah Song, and Shun-ichi Amari. "A new discriminant NMF algorithm and its application to the extraction of subtle emotional differences in speech." Cognitive neurodynamics 6.6 (2012): 525-535.

Usage

ndNMF(
  dat,
  trainlabel,
  r = 2,
  lambada = 0.1,
  maxIter = 1000,
  tol = 1e-07,
  log = TRUE,
  plotit = FALSE,
  verbose = FALSE,
  ...
)

Arguments

dat	a matrix with gene in row and sample in column
trainlabel	the label of sample, like c(1,1,2,2,2)
r	the dimension of expected reduction dimension, with the default value 2
lambada	a relative weighting factor for the discriminant. Default 0.1
maxIter	the maximum iteration of update rules, with the default value 1000
tol	the toleration of coverange, with the default value 1e-7
log	log2 data. Default is TRUE.
plotit	whether plot H (V=WH). Default: FALSE.
verbose	TRUE
...	to gplots::heatmap.2

Author(s)

Zhilong Jia and Xiang Zhang

Examples

dat <- rbind(matrix(c(rep(3, 16), rep(8, 24)), ncol=5), 
matrix(c(rep(5, 16), rep(5, 24)), ncol=5), 
matrix(c(rep(18, 16), rep(7, 24)), ncol=5)) + 
matrix(runif(120,-1,1), ncol=5)
trainlabel <- c(1,1,2,2,2)

res <- ndNMF(dat, trainlabel, r=2, lambada = 0.1)
res$H
res$rnk

DNMF documentation built on May 10, 2022, 5:12 p.m.