search

linxihui / NNLM

Nonnegative Linear Models (NNLM) and Nonnegative Matrix Factorization (NMF or NNMF).
Other
75 stars 22 forks source link

Support for symmetric A factorization #3

Open aparamon opened 6 years ago

aparamon commented 6 years ago

Currently, nnmf() doesn't specifically handle the case of symmetric A, resulting asymmetric W, H. Please consider adding such feature: that would be extremely useful for factorizing pairwise distance and correlation matrices.

See also: https://github.com/andybaoxv/symnmf

linxihui commented 6 years ago

@aparamon Hmm, good suggestion. I will take a look and consider to implement it. Thanks!

zdebruine commented 3 years ago

@aparamon Alternating least squares requires non-identical W and H initializations, and during the factorization W and H converge towards one another, after linear transformation. W and H will never exactly equal one another without convergence having been called.

Basically, after convergence, convert the factorization of the form A = WH to the form A = WDH (like in SVD), and then W and H will be nearly identical. You can average the two to find a near-exact symmetric factorization solution.