updates to recommenders

[ajschumacher]

Ben's slides: http://www.slideshare.net/BenjaminBengfort/non-negative-matrix-factorization

Ben recommends:

co-occurrence linear log likelihood ratio (in tase for mahout?)

combining item stuff and user rating stuff:

• ensembling or
• add in item-item similarity as another feature? (not sure how this works)

[ajschumacher]

https://github.com/JohnLangford/vowpal_wabbit/wiki/Matrix-factorization-example

[ajschumacher]

https://piazza.com/class/hywzo7yre651ik

[ajschumacher]

https://github.com/DistrictDataLabs/science-bookclub

[ajschumacher]

flippin' flabbergasted that there doesn't seem to be anything built in anywhere for NNMF or ALS...

http://www.quuxlabs.com/blog/2010/09/matrix-factorization-a-simple-tutorial-and-implementation-in-python/ maybe this is the post that Ben worked from?

http://stackoverflow.com/questions/22767695/python-non-negative-matrix-factorization-that-handles-both-zeros-and-missing-dat http://stackoverflow.com/questions/17982931/matrix-completion-in-python could be helpful?

[ajschumacher]

https://github.com/malemi/cold-start-recommender

and maybe check this out too? http://mymedialite.net/documentation/use_from_python.html

and: django-recommends.readthedocs.org/

loglikelihood filter on the co-occurence matrix

http://csc.media.mit.edu/docs/divisi2/tutorial_aspace.html#singular-value-decomposition