Multi-channel Non-negative Matrix Factorization

Multi-channel Non-negative Matrix Factorization by multiplicative update written in Octave.

Features

You can execute Multi-channel NMF easily on Octave(of course MATLAB).

Background of Multi-channel NMF

You should read the following article, http://www.kecl.ntt.co.jp/people/kameoka.hirokazu/publications/Sawada2013IEEETrans05-published.pdf (I'm not an author of Multi-channel NMF.)

Usage

./exc_mnmf.m

Default Setting

• This default program execute a simulation by real number.

If you want to deal with real data, you can change codes, for example, "parameter_setting.m", "input_data.m" etc., for your use.

• Size of the observation matrix: 12 × 20

• Channel number: 2

• Numbers of basis vectors: 4

• Numbers of iterations: 100

• Initial value setting method of update rules: random non-negative values for T & V, and identity matrix for H.

You can change these default settings on "parameter_setting.m" like the following,

% parameters
% Number of channel
M = 2;
% size of observation matrix
I = 12;
J = 20;
% basis vectors
K = 2; 
% numbers of iterations
itr = 100;

Data that you want to use can be also changed on "input_data.m"

Initial value setting can be changed on each functions, "mnmf_Frb.m& mnmf_IS.m"

% initialization
T = abs(randn( I, K ));
V = abs(randn( K, J ));
H = zeros( M, M, I, K );
Id = eye( M, M ) / M;
for ii=1:I
 for kk=1:K
   H(:,:,ii,kk) = Id;
 end
end

• I do NOT implement the NMF basis clustering and top-down clustering using the latent variables Z and bottom-up clustering in the reference paper.

Caution

Please read Issue.

This code is provided without liability and warranty.