NMF

API/Reference

ADI.NMF — Type

NMF(;ncomps=nothing)
NMF(ncomps)

Use non-negative matrix factorization (NMF, NNMF) to form a non-negative, low-rank, and orthonormal basis of the input. The implementation of the underlying NMF is provided by NMF.jl. The implementation uses a non-negative SVD for initialization and a coordinate-descent solver to fit.

If ncomps is nothing, it will be set to the number of frames in the reference cube when processed.

Non-negativity constraint

NMF is not designed to fit negative values. This algorithm will warn you (but will not error) if a target or reference cube contains negative values. The output may seem reasonable, but it is not well-defined with respect to the NMF algorithm. To overcome this, rescaling the data by its minimum before processing is necessary

target = cube .- minimum(cube)
S = reconstruct(NMF(), target, angles)

When doing full-frame reduction (e.g. NMF()(cube, angles)) this is handled automatically, so this constraint only applies to the lower-level API and methods which rely on those, such as GreeDS. In general, if you see warnings, heed them.

References

Ren et al. 2018 Non-negative Matrix Factorization: Robust Extraction of Extended Structures

source