When is non-negative matrix decomposition unique?
暂无分享,去创建一个
In this paper, we discuss why non-negative matrix factorization (NMF) potentially works for zero-grounded non-negative components and why it fails when the components are not zero-grounded. We show the demixing process is not uniquely defined (up to the usual permutation/scaling ambiguity) when the original matrices are not zero-grounded. If fact, zero-groundedness alone is not enough. The key observation is that if each component has at least one point for which it is the only active component, the solution is unique. When the non-negative matrices are not zero-grounded, no such point exists and the solution space contains demixtures which are linear combinations of the original components. Thus, the NMF problem has a unique solution for matrices with disjoint components, a condition we call Subset Monomial Disjoint (SMD). The SMD condition is sufficient, but not necessary for NMF to have a unique decomposition, whereas the zero-grounded condition is necessary, but not sufficient.
[1] Mark D. Plumbley. Conditions for nonnegative independent component analysis , 2002, IEEE Signal Processing Letters.
[2] Michael R. Lyu,et al. Nonnegative independent component analysis based on minimizing mutual information technique , 2006, Neurocomputing.
[3] Özgür Yõlmaz,et al. Blind Separation of Speech Mixtures via , 2004 .
[4] Victoria Stodden,et al. When Does Non-Negative Matrix Factorization Give a Correct Decomposition into Parts? , 2003, NIPS.