Instability in rsparse::WRMF convergence and loss function about rsparse

[dselivanov]

@ZacharyDeBruine are you referring to the CRAN version of the package? May we know your sessionInfo() ?

Stopping criterion is following: https://github.com/rexyai/rsparse/blob/0ea378bc8c1cbe3e0ee2af3dbd9762e2a843655e/R/model_WRMF.R#L237
ALS optimization step by design should decrease loss. If loss starting to grow - there is no point to continue. Sudden loss increase seems related to the way NMF was initially implemented (somewhat similar to "projected gradient" method) - see discussion in #36. You can install github version of the pkg from master branch and test if it works for you.

However, if your algorithm really is not meant to handle missing values, then that should be in big bold letters at the top of the package documentation so the theoretical implications are understood by all users immediately.

First of all I would definitely phrase this differently as it sounds really rude. Secondly algorithm treats missing values as missing values but not zeros as in your dense.nnls example. So it tries to minimize error only for observed values and ignores missing. Anyway there was a bug in implementation - see discussion in #35. Now it is fixed in master branch.

[zdebruine]

Thanks for the quick reply! Ah, yes the discussion and fix in #36 is exactly what I was after. I was using the CRAN version. Opened up the GitHub latest commit and it's great now.

Thanks for all your hard work on this! Sorry I was a bit overstated on my frustration. Now I get the concept of minimizing error only for observed values but not for missing values.