Non-negative matrix factorization with fixed row and column sums
Given a non-negative \(m\times n\) matrix \(A\), find two non-negative matrices \(U\) (\(m\times k\)) and \(V\) (\(n \times k\)) with \(k\ll m, n\) that minimize \(F (A, U V^T)\), where \[ F (A, U V^T)=D(A|| UV^T) :=\sum_{ij}A_{ij} \log\frac{A_{ij}}{[UV^T]_{ij}} -A_{ij}+[UV^T]_{ij}. \] The problem \(\min_{U\geq 0,V\geq 0} ED(A||U V^T)\), where \(A\geq 0\), is called non-negative matrix factorization using the generalized Kullback-Leibler (KL) divergence. The authors show: Let \(A_{m\times n}\) be a non-negative matrix. Then every stationary point \((U, V)\) of the cost function \(D(A||U V^T)\) preserves the column sums of \(A\), the row sums of \(A\), and the matrix sum of \(A\). Every stationary point \((U_{m\times k}, V_{n\times k})\) of the KL minimization problem has the form \(U V^T = P_{m\times k} D_{k\times k} Q_{n\times k}^T\), where \(P, Q\) are column stochastic, \(D\) is diagonal non-negative, and \(\sum_i D_{ii} = \sum_{ij} A_{ij}\). Furthermore, if \(A\) is column stochastic (or row stochastic), then the matrix \(DQ^T\) (or \(P D\)) is also column stochastic (or row stochastic). The authors give an application in stochastic matrix approximation.
- Nonnegative matrix factorization
- On nonnegative factorization of matrices
- Non-negative matrix factorization with sparseness constraints
- Sparse non-negative tensor factorization using columnwise coordinate descent
- On nonnegative matrices with given row and column sums
- A mixed iteration for nonnegative matrix factorizations
- The nonnegative matrix factorization: regularization and complexity
- Nonnegative matrix factorization of a correlation matrix
- scientific article; zbMATH DE number 961607 (Why is no real title available?)
- Learning the parts of objects by non-negative matrix factorization
- Nonnegative matrix factorization and I-divergence alternating minimization
- On reduced rank nonnegative matrix factorization for symmetric nonnegative matrices
- Efficient histogram dictionary learning for text/image modeling and classification
- Algorithms for nonnegative matrix factorization with the Kullback-Leibler divergence
- Non-negative matrix factorization under equality constraints -- a study of industrial source identification
- Symmetric nonnegative matrix trifactorization
- The Kullback-Leibler Divergence and Nonnegative Matrices
- A partial proximal point algorithm for nuclear norm regularized matrix least squares problems
- Non-negative low-rank approximations for multi-dimensional arrays on statistical manifold
- Structured nonnegative matrix factorization with applications to hidden Markov realization and clustering
- Probability matrices, non-negative rank, and parameterization of mixture models
This page was built for publication: Non-negative matrix factorization with fixed row and column sums
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q935373)
