Lodestar.Decomposition 0.2.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

NmfInitialization

Where a non-negative matrix factorization starts from, and — because multiplicative updates are multiplicative — most of what decides the shape of where it ends up.

An update multiplies every entry of W and H by a non-negative ratio, so a zero stays a zero for as long as the loop runs. The initialisation is therefore the only thing that can put a zero in the answer, and choosing between the two members below is choosing between a sparse factorization and a dense one, not between two ways of seeding the same result.

Both members are NNDSVD — the non-negative double SVD of Boutsidis and Gallopoulos, which takes the leading singular triplets of the matrix and splits each one into its positive and its negative part, keeping whichever carries more energy. Nothing about it is random once Ω is fixed, which is why an NMF fit is reproducible in a way its Python counterpart's random initialisation is not.

Members

Member	Value	What it does
NndSvd	0	Leaves the zeros NNDSVD produces in place. scikit-learn's nndsvd, and the choice when a sparse factorization is the point.
NndSvda	1	Fills each zero with the mean of every cell of the matrix, zeros included. scikit-learn's nndsvda: denser, and it converges in fewer iterations because no entry is frozen at zero.

Neither is a tuning knob to sweep. NndSvd is the one that answers "which words does this component not contain"; NndSvda is the one to reach for when the factorization is a compression rather than a description, and the zeros are an accident of the corpus.

Why nndsvdar is not here

scikit-learn offers a third member, nndsvdar, which fills the zeros with small draws from numpy's own normal stream instead of the mean. This package does not, because it does not reproduce RandomState — see ADR 0072 for the rule Ω follows and why a seed is not portable across the two ecosystems. An initialisation that could not be compared entry by entry against the reference is one no oracle can pin, and this package ships nothing it cannot check.

NndSvda is what nndsvdar approximates: the same fill, without the noise. If you were reaching for nndsvdar to break ties between identical rows, the noise is doing that job and a different Ω does it too.