Nonnegative matrix factorization (NMF) is a dimension-reduction technique based on a low-rank approximation of the feature space. Besides providing a reduction in the number of features, NMF guarantees that the features are nonnegative, producing additive models that respect, for example, the nonnegativity of physical quantities.
Given a nonnegative m-by-n matrix X and a positive integer k < min(m,n), NMF finds nonnegative m-by-k and k-by-n matrices W and H, respectively, that minimize the norm of the difference X – WH. W and H are thus approximate nonnegative factors of X.
The k columns of W represent transformations of the variables in X; the k rows of H represent the coefficients of the linear combinations of the original n variables in X that produce the transformed variables in W. Since k is generally smaller than the rank of X, the product WH provides a compressed approximation of the data in X. A range of possible values for k is often suggested by the modeling context.
The Statistics and Machine Learning Toolbox™ function
nnmf carries out nonnegative matrix factorization.
one of two iterative algorithms that begin with random initial values
for W and H. Because the norm
of the residual X – WH may
have local minima, repeated calls to
yield different factorizations. Sometimes the algorithm converges
to a solution of lower rank than k, which may indicate
that the result is not optimal.