This tries to use the approach in Blondel et al and Zhu, 16 for using a weight/mask matrix to ignore missing values while doing the Lee & Seung multiplicative update. The latter implementation can be found in a git repo here
It initially seemed to work well for some data where out-of-sample performance increased, but for other data it performed horrendously where predictions weren't even on the same scale as the user ratings. I got a little excited and merged it into master in #31, but given how finicky it seems to be, I've unmerged it now and I'm leaving it in this branch/PR for posterity.
This tries to use the approach in Blondel et al and Zhu, 16 for using a weight/mask matrix to ignore missing values while doing the Lee & Seung multiplicative update. The latter implementation can be found in a git repo here
It initially seemed to work well for some data where out-of-sample performance increased, but for other data it performed horrendously where predictions weren't even on the same scale as the user ratings. I got a little excited and merged it into master in #31, but given how finicky it seems to be, I've unmerged it now and I'm leaving it in this branch/PR for posterity.