20.02.2019 Views

CLC-Conference-Proceeding-2018

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

e obtained as the sum of matrices of rank 1,<br />

ajbj T . If the decomposition is exact (E = 0), then<br />

it is named Non-negative Rank Factorization<br />

(NRF) xii . The rank J or internal dimension is also<br />

called a nonnegative rank and is denoted by<br />

rank+(Y) xiii . The rank J satisfies rank(Y) ≤<br />

rank+(Y) ≤ min{I, T}.<br />

In general, the rank J usually satisfies<br />

.<br />

Figure 1: Classic model of NMF<br />

In Figure 1 taken from xiv , page 9, the structure of<br />

the factorization can be observed.<br />

It is not difficult to notice that an NMF does not<br />

have to be an NRF. Non-negative matrices<br />

cannot always be found such that factorize the Y<br />

matrix in an exactly way or that the matrix E = 0.<br />

This is due, among others, to the approximation<br />

errors (rounding) that are generated and<br />

propagated in the calculations in a finite<br />

precision arithmetic.<br />

To start the factorization process three<br />

fundamental questions must be answered. The<br />

first is related to the NMF model that will be<br />

used, this depends on the characteristics of the<br />

problem being solved or the type of the data.<br />

NMF have been adapted to many applications.<br />

Choosing the model involves selecting<br />

the cost function to be minimized and the<br />

additional constraints to be imposed to the<br />

matrices, as we will see later. In this work we<br />

present some of the applications in which the<br />

Images Group of the University of Havana is<br />

currently working and the models that are being<br />

studied.<br />

The second question is related to the<br />

dimension or rank of the matrix J, J ≤ min(I, T),<br />

or also known as internal dimension. Of course,<br />

this choice has a lot to do with the a priori<br />

knowledge of the problem to be treated. NMF<br />

are considered methods of dimensionality<br />

reduction so that this choice is crucial both for<br />

the appropriate modeling of the problem and for<br />

obtaining the solutions. The rank should be set to<br />

address the solution of the problem. Several<br />

approaches to determine the rank of the matrix<br />

appear in the literature. Several authors prefer<br />

the use the SVD xv . Another approaches suggest<br />

to obtain the graph of the singular values and<br />

select the range after observing when the graph<br />

has an elbow. This fact is related to the theory of<br />

discrete ill-posed problems and the regularization<br />

of solutions that is very well studied in xvi . The<br />

Images Team of the University of Havana is<br />

studying a criterion proposed in xvii to compare<br />

the solutions with the use of the SVD. Other<br />

authors have agreed that a suitable variant is the<br />

use of heuristics based on populations as in xviii .<br />

The third question is related to the<br />

initialization of the matrices depending on the<br />

model to be used. It has been shown that the<br />

algorithms that are used to find the matrices are<br />

sensitive to the initialization. The convergence<br />

and the speed of convergence of the algorithms<br />

may be affected by the initialization and with<br />

this, of course, the quality of the solutions. To<br />

answer this question we could also cite many<br />

works. The most intuitive ideas are random<br />

initializations. These in general do not take into<br />

account the characteristics of the problem. Other<br />

works propose to use population-based heuristics

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!