|
Suppose we have news reader like Google reader. each document has its topics. each topics has its followers. each item when seen by user has mark - is it interest for him or not. Is the are approaches to ranking document feed based on those data? |
|
This is exactly the problem we are trying to solve at the moment at FUSELabs Project Emporia. We are crawling Twitter (and many other less beefy sources) for links and try to rank them according to your personal taste (you tell us whether you like the story or not in the UI). We are currently using Matchbox to do the recommendation part. It scales very well to large sets of users and we can do online learning using expectation propagation. I assume you mean whether Matchbox is open source (the rest of the code is making it run on the Azure cloud platform)? Matchbox is published (see link above) so it contains all the details of how to build the factor graph and message passing. It is fairly straightforward to implement Matchbox on top of Infer.NET (http://research.microsoft.com/en-us/um/cambridge/projects/infernet/). Definitely ask on the forum if you need any help constructing it (Infer.NET runs fine on Linux and Apple using Mono I believe).
(Oct 11 '10 at 17:11)
Jurgen
|
|
You can try to use a standard learning to rank method, such as RankSVM (a good implementation is google's sofia-ml, decribed in the Sculley paper Large scale learing to rank in NIPS 2010) with features such as how many times a user has voted on that feed, how many times a user has voted on that topic, given that we're ranking a document of topic y how many times the user has voted another topic (to cover the fact that topic x might be a good predictor of topic y), etc. |
