20.02.2019 Views

CLC-Conference-Proceeding-2018

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

solution is unique for any vector y and can be<br />

obtained in polynomial time.<br />

Another kind of methods can be used to<br />

solve NMF, the descent algorithms. These<br />

algorithms choose a descent direction to move<br />

looking for a next solution that successively<br />

improves the values of the objective function.<br />

The methods of the projected gradient are<br />

descent methods. This algorithms update the<br />

approximation to the solution x k building the<br />

x k +1<br />

as follows:<br />

x k +1<br />

= x k + s k (̅x̅k̅<br />

− x k ), where ̅x̅k̅ = P[x k − α k p k ]<br />

here P[ ] denotes the projection in the set of<br />

solutions, s k ∈ (0,1] is a step size, α k is a positive<br />

scalar and p k is the direction of descent. Using<br />

the ideas several methods can be constructed<br />

taking the appropriate descent directions and the<br />

rules to obtain the step size. For more<br />

information on this type of methods see xxx .<br />

The Updating rules can be obtained<br />

considering an objective function and trying to<br />

solve the problem xxxi .<br />

For example, if the objective function is<br />

minimize the Euclidean norm between Y and AX,<br />

i.e.<br />

min‖Y − AX‖ 2 = min∑ yij − (AX| ij) 2<br />

ij<br />

and the rules that derive from that are:<br />

1.2 Initialization of the NMF<br />

The initialization of the matrices in the<br />

NMF have a direct influence in the solution and<br />

that is why the solutions depend strongly on that.<br />

Bad initializations produce slow convergence.<br />

As already explained the objective function of<br />

the ALS is not convex in both variables although<br />

it is strictly convex in one variable, for example<br />

in A or X and the algorithm can stop in a local<br />

minimum.<br />

The problem of initialization can be even<br />

more complex if there is a particular structure of<br />

the factorization, for example, in the symNMF or<br />

the triNMF or if the matrix is large.<br />

In xxxii it is proposed to follow the<br />

following procedure:<br />

Build a search algorithm in which the<br />

best initialization is found in a space of R pairs<br />

of matrices. Frequently the beginning of this<br />

procedure can be with randomly generated<br />

matrices or take them as the output of a simple<br />

ALS algorithm. The R parameter usually<br />

constitutes the necessary iterations in which 10<br />

to 20 are considered enough.<br />

Run a specific NMF algorithm for each<br />

pair of matrices with a fixed number of iterations<br />

(10 to 20 are also considered enough). As a<br />

result, we obtain R pairs<br />

initial matrices.<br />

, from estimates of<br />

Select the pair A Rmin ,X Rmin , that is, those with the<br />

lowest evaluation of the objective function as<br />

initialization for the factorization.<br />

In xxxiii an implementation of this<br />

procedure is presented, the Algorithm 1.2. To<br />

read about the use of genetic algorithms used to<br />

perform the initialization in NMF, see xxxiv<br />

and xxxv .<br />

1.3 Stopping criteria<br />

Different stopping criteria can be considered for<br />

the algorithms used to calculate the NMF. Some<br />

of them are listed:

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

Saved successfully!

Ooh no, something went wrong!