| |
Non-negative Matrix Factorization object
|
A=nmf(H) returns a Non-negative Matrix Factorization object.
Given matrix X>0 creates matrices W, H > 0 with X ~ X*=W*H, such that
the original features X are represented by a new positive basis W and new
positive features H. W can be interpreted as components or topics, while H
are the weights of these components in the original samples.
Usage: [r n]=train(nmf,d);
Result: r.X = The reduced feature weights, r.Y = original X. The basis is
given in n.H.
Hyperparameters, and their defaults
N=5; -- number of final basis functions
maxIteration=500; -- maximum number of iterations per cycle
nrofrestarts=1; -- nr of restarts to overcome local minima.
eps=1e-3; -- convergence epsilon (on objective function).
Model
W = The (sparse and independant) basis vectors.
H = The reduced feature vectors.
Methods:
train
testing
Example:
d=data(abs(rand(20,10))); produce random data
[r,n]=train(nmf('N=2'),d); reduce 10 elements to 2 components
reconstructed=(n.W*r.X')'; component basis * weights
The example is illustrative only, as random data is unlikely to contain any
meaningful components.
Note: This code is based on Patrik Hoyer's code for simple MSE NMF.
| |
Reference : Algorithms for Non-negative Matrix Factorization |
| |
Author : Daniel D. Lee, H. Sebastian Seung |