Those look very cool, I'd love to see how those compare with doing plain-old Lanczos for SVD on Hadoop. Speaking of which, I've got an implementation of that which I wrote up for my own matrix library ( http://decomposer.googlecode.com ) a while back, and I noticed that we still don't have any large-scale SVD impls in Mahout. Is there any interest by the community for me to try and port that / contribute this to Mahout? It's Apache-licensed, but I'm currently using mostly my own sparse and dense vector writables for use on Hadoop (designed specifically for things like Lanczos and AGHA), so I'd need to port them over to use whichever vector impls Mahout is using.
-jake On Fri, Sep 25, 2009 at 11:25 AM, Ted Dunning <[email protected]> wrote: > Isabel, > > Very interesting post. Here are more accessible resources: > > http://arxiv.org/abs/0909.4061 > http://www.pnas.org/content/104/51/20167 > > THese provide a very interesting and solid link between random indexing and > SVD algorithms. They also definitely provide a fantastic way to implement > large scale SVD using map-reduce. > > Nice pointer! > > 2009/9/25 Michael Brückner <[email protected]> > > > > > year's NIPS (http://nips.cc/Conferences/2009/Program/event.php?ID=1491) > > > > > -- > Ted Dunning, CTO > DeepDyve >
