Thanks to everyone for the great replies. I am beginning to synthesize, from these and other responses, some needs:
1) "Just Work": Provide a tool that works out of the box in as many situations as possible without having to write code, pick algorithms, etc. 2) Work Better: Produce better recommendations. 3) Work Faster: Run faster / scale better 4) Explore My Data: Explore clusters of users, clusters of items, why items are recommended, etc., perhaps with a visual tool? 5) Specialize for Common Domains: Begin adding some frameworks or pre-made solutions that are appropriate for particular, common domains like advertising, books, etc. 6) Analytics: Now that I've implemented CF, how is it helping my metrics? For my purposes, and at the moment this is just an idea for an assignment, I am trying to separate what needs should be answered by Mahout versus some kind of commercial enterprise system built on top. If you're saying, hey, why not build this all as open-source, I'd just say, I don't think much of this would be interesting for the community to spend time building. It may well be properly the province of commercial products (think MySQL or JBoss and their enterprise components.) I think 2 and 3 are general goals that the core, open library should continue to strive for. The rest seem like good beginnings of problems to solve for enterprises. 6 seems more appropriate for analytics tools; doesn't seem like something one necessarily needs from a CF system itself. 5 is absolutely right -- I think a specialized recommender will always do better than a general one, and performance is so crucial, that enterprises need custom solutions. I can see a need for specializations of the general algorithms, which may not be of general interest. The question is, which industries to tackle? and, how to even begin to gain experience in them? But that would be some real value-add a product could bring to an enterprise. 1 and 4 seem like the more immediate, solvable issues. Yes, lots of folks understand this idea of recommending books and such based on ratings. Some people have a business with ratings. The next logical step should be just that simple: run a demo with minimal set up against your existing data and see what happens -- maybe show off user clusters and example recommendations and comparisons of recommendations versus actual preferences. Assuming the results are decent, I think that sort of plug-and-play access to this technology is really valuable. Right now one really has to be a developer and a bit of a specialist to begin to apply the framework. Where there is value... there is a product idea! More on this as it develops. On Thu, Oct 9, 2008 at 1:34 PM, Sean Owen <[EMAIL PROTECTED]> wrote: > I mentioned this to Otis but thought I'd expand it -- > > As some of you may know I am actually an MBA student. I have worked on > the collaborative filtering portion of Mahout (formerly called Taste) > for several years. I have always been curious to know what, exactly, > enterprises need out of collaborative filtering. What product would > they like to see? > > Now I have the opportunity to use this as the basis of a business > school project. So I am wondering if anyone can comment on any of > these questions, without naming names if it's confidential -- > > - What do enterprises imagine they can get out of collaborative filtering? > - At what stage is the interest -- barely know about it, ready to use it, > other? > - What do people use to solve these needs today, custom code, other products? > - What's the biggest need enterprises have from collaborative > filtering products or services? > - What's the one product or feature an enterprise would love to see in > the market, in terms of CF? > > Thanks for any thoughts, folks! > Sean >
