CLC-Conference-Proceeding-2018
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
xix or strategies based on the SVD of the known<br />
data matrix xx .<br />
A very general scheme of a solution<br />
strategy would be:<br />
Model selection and cost function.<br />
Determination of the internal rank or dimension.<br />
Initialization of the matrices according to 1.<br />
Choose a solution method.<br />
Estimation of the approximation error.<br />
, which also refers to the Euclidean<br />
distance xxi . This measure is the most simply and<br />
frequently used. The Frobenius norm,<br />
written as<br />
the columns of A.<br />
, can be equivalently<br />
, with aj<br />
1.2 Algorithmic ideas for the solution of NMF<br />
The general classical problem to be solved as it<br />
appears in 1.1 is:<br />
Known Y ∈ R IxT + the data and a rank J, J ≤<br />
min(I, T), find A = [a1,a2, … , aJ ] ∈ R IxJ + , and X<br />
= B T =<br />
T JxT T + E where E ∈ R IxT + represents the<br />
approximation error [b1,b2,… , bJ ] ∈ R+ ,Y = AX<br />
+ E = AB<br />
This is a nonlinear optimization problem with<br />
unknowns A and X and nonnegative constraints<br />
(inequality constraints). It is not always possible<br />
to find the matrices that exactly factor Y nor to<br />
know the error matrix E. In general, the problem<br />
of finding two matrices A and X is raised. That<br />
minimizes a measure of<br />
"distance" (according to the function of cost that<br />
is used) between the data matrix and the product<br />
AX.<br />
The approximate problem of NMF, Y ≈ AX,<br />
corresponds to minimizing some measure of<br />
distance, divergence or measure of dissimilarity.<br />
This could be formulated as, find A and X, such<br />
that they satisfy<br />
, where ‖ ‖F<br />
represents the Frobenius norm, AIxJ, XJxT and A ≥<br />
0, X ≥ 0 or as it is usually denoted DF(Y‖AX) =<br />
Factors A and X are sought in other matrix<br />
spaces, in particular in and in . For this<br />
reason, the determination of the rank has a<br />
special influence on the solution, since it<br />
determines the search space. It is important to<br />
emphasize that the function to be minimized<br />
(objective or cost function) is convex with<br />
respect to the elements of matrix A or with<br />
respect to the elements of matrix X but not with<br />
respect to both at the same time.<br />
Another widely used measure is the generalized<br />
Kullback-Leibler divergence (I-divergence):<br />
From<br />
p ln(p) ≥ p − 1 follows that the I- Divergence<br />
takes non-negative values and it vanish if and<br />
only if p =<br />
q.<br />
As a side effect, very often, the matrices A and X<br />
are sparse. An interesting idea could be explicitly<br />
control the sparsity of the matrices. In xxii some<br />
sparseness criteria are presented that can be<br />
imposed as constraints in the model.<br />
Three main strategies are used to obtain the<br />
matrices in the NMF.<br />
Alternating Least Squares (ALS).<br />
Descent methods.<br />
Updating rules.