Comments on: Maximum Margin Matrix Factorization

By: Charles Martin

Charles Martin — Fri, 06 Mar 2009 08:55:13 +0000

I had some good success at eBay/Shopping.com using a ranking SVMs. So let me as a few questions?
Is there now an open source implementation that others can use, or do I need formulate this myself. I assume you have a high performance implementation…did you end up solving the dual or the primal? Also, how would you compare using MMMF vs, say, using a combination of SVD + a Transductive SVM (that is, guess the labels, then use a transductive SVM to solve…and, yes, I know MMMF is convex and this is not…but I can do this with existing code). Overall–nice to see this–it is encouraging.

By: Jason Rennie

Jason Rennie — Tue, 27 May 2008 19:09:26 +0000

Hi Manu,

Yes, MMMF certainly can be used for binary data. MMMF can be applied to data with any number of ordinal labels >= 2. In both the original NIPS MMMF paper, as well as the ICML Fast MMMF paper, you may have noticed that we used `R’ to represent the number of ordinal labels. You can simply set R=2 for binary data. This should simplify many of the expressions. For my MMMF matlab code, pass 2 for the value of l for the gradient/objective function.

Jason

By: manu

manu — Wed, 16 Apr 2008 15:10:04 +0000

Hi,
Can MMMF be used for recommendation based on binary data (ie: i like/dislike this item) instead of rating data ? If so, could you point out any relevant publication.
I have trouble to figure that out.
Thank you.

By: Jason Rennie

Jason Rennie — Mon, 18 Feb 2008 21:01:14 +0000

Hi Ruth,

MMMF can certainly be used for search. Here at StyleFeeder, we use it to provide personalized search results, based on the ratings and bookmarks the user has provided.

By “vector space model”, I’m guessing that you mean a dimensionality-reduction of the term-document matrix. We don’t use MMMF for that, but you certainly could. You’d need a different loss function (e.g. one based on the multinomial model). With that and the trace norm regularization term, you’d get a convex objective which yields a low-rank solution given a sufficiently large weight on the regularization term. Do contact me if you have further interest/questions.

Jason

By: ruth

ruth — Sun, 10 Feb 2008 10:54:03 +0000

Hi,

Can MMMF be used for vector space model search engine, such as SVD or NNMF (non-negative matrix factorisation) ? IF so, could you please point out any publication that MMMF is used for search engine? I have Google, but to no luck at all.

By: armynavy

armynavy — Thu, 22 Nov 2007 03:23:22 +0000

My Style twins don’t appear at all similar in tastes?
Why aren’t these updated with our own recommendations and knowledge?

By: Whirlycott / Philip Jacob » StyleFeeder Tech Blog

Whirlycott / Philip Jacob » StyleFeeder Tech Blog — Mon, 30 Jul 2007 20:55:20 +0000

[...] Blog for those of you who are interested in the black magic that goes on behind the scenes here. Jason’s been doing all of the writing so far, mainly about our CF recommendation engine and Fast [...]