Synthese Recommender December 16, 2008Posted by Andre Vellino in CISTI, Collaborative filtering, Digital library, General, Recommender.
For the recommender experts among you – there’s nothing fundamentally new here that you don’t know about already: the Synthese Recommender applies user-based collaborative filtering (implemented using Taste) with article citations as a substitute for user-data to address the cold-start problem. This has been done before in TechLens. And, I should add, TechLens (now in its third iteration) is quite a bit more full-featured and polished.
My aim was modest: to gather data about how well a simple collaborative filtering recommends articles to researchers in diverse scientific fields. That will give me a baseline from which to repeat the experiment – with a content-based recommender, multi-dimensional ratings, and an explanation feature. I’m hoping this will tell me how much more valuable each element is to the overall user-experience. My hypthesis: a hockey-stick curve in usefulness as more features are added. Explanations, I think, are going to make all the difference.
This is how to use Synthese:
- Search an index of about 1.5 Million BioMed articles
- Add important articles to a “baskets” of favourite articles
- Recommended other articles based on that basket
- Rate how relevant the recommendations are for your research
Users can search on the author / title / keyword / abstract fields (the next version will be on the full-text as well), view the articles from the publisher’s site, create multiple topic-baskets, generate recommendations based only on a given article’s citations and keep a list of previous searches that produced results for each of your baskets.
In previous version, I tried visualizing recommendation results with a prefuse applet, but the initial coments I got back from most people who tried it was that this was worse than useless.
Although I don’t know a lot about biology or medicine it looks to me like recommendations are in the right ballpark. By which I mean, they appear to strike some balance between serendipity, diversity and semantic relevance to the original article basket.
Let me know what you think!