Home › Decomposition › Factorization

A fitted non-negative matrix factorization of a sparse matrix: X ≈ W H, with every entry of both factors at or above zero. There is no unfitted state — Nmf.Fit is the only way to reach an instance — and no property has to decide what to do when it is read too early.

The non-negativity is the whole point. A truncated SVD's components are signed and orthogonal, so a "topic" may subtract a word as readily as add one; these components only ever add, which is what makes a row of H readable as a set of terms and a row of W readable as a mixture of them. TruncatedSvd is the answer when the subspace matters and the signs do not.

Property	What it holds
ComponentCount	int — how many components were fitted, the rank W and H share.
FeatureCount	int — how many columns the factorized matrix had.
Iterations	int — how many multiplicative updates ran, scikit-learn's n_iter_. Equal to MaxIterations when the tolerance is zero.
ReconstructionError	double — the beta divergence between X and W H at the end, square-rooted: scikit-learn's reconstruction_err_.
Weights	IReadOnlyList<double> — W, row-major rows × ComponentCount. Row i's mixture starts at i * ComponentCount. This is what scikit-learn's fit_transform returns.
Components	IReadOnlyList<double> — H, row-major ComponentCount × FeatureCount, scikit-learn's components_. Component c starts at c * FeatureCount.

ReconstructionError is comparable only between fits that minimised the same NmfBetaLoss: the Frobenius number is a distance and the Kullback–Leibler one is a divergence, and they are not on one scale.

Transform is a factorization, not a projection. Applying an unseen row to a non-negative basis runs the same multiplicative loop with H held fixed, rather than the product a name borrowed from the SVD would suggest — so it costs what a small fit costs, and it honours the loss, the cap and the tolerance this fit was given. It leaves Components untouched, which is what makes one fitted model usable across as many batches as a caller has.

Member	What it does
Nmf.Fit	Factorizes a sparse matrix, from an initialisation it computes or one you supply.
Nmf.Transform	W for an unseen matrix, with this fit's H held fixed.