This is actually a really useful discussion to have, since the tools we can provide depend very much on what the current workflows are (and are made quite challenging by the fact that the workflow essentially goes "off-web" once people open a PDF). - Richard Akerman

is mendeley supposed to do something like this? the key with pandora is that you can tune it by telling it you don't like certain songs. you would want to say look at this folder or this tag and go from there, and then tune it by liking or blocking certain papers that came up in the results. - Christina Pikas

Mendeley is currently just listing popular papers, journals and authors by category. But I understand that the plan is to recommend papers based on your personal library. A first easy step would be an RSS feed of the references you add to your library (similar to Connotea and CiteULike). - Martin Fenner

Getting enough data to do a meaningful recommender is hard. You need to either skim off implicit data from people's actions in their workflow (as MESUR and b2X do) or you need to get explicitly into their workflow and ask for feedback (as CISTI's Recommender experiment does). - Richard Akerman

To accumulate enough data of either kind, you need good traffic volume, so resolvers, bookmarking services (2collab, Connotea, CiteULike), reference managers (e.g. Zotero) and article PDF managers (Mendeley, Papers) are all potential participants in a science recommender network. (We really should build recommender APIs to exchange recommendations and data between the services.) - Richard Akerman