jump to navigation

NextBio Recommender February 15, 2009

Posted by Andre Vellino in CISTI, Information retrieval, Recommender, Recommender service, Search.
trackback

nblogosmThe biosciences search portal NextBio is interesting for several reasons.  According to this interview, the VP of Engineering Satnam Alag (also the author of  Collective Intelligence in Action) says NextBio will shortly be introducing an article recommender

The key point about this particular recommendation engine is its strong use of an ontology, similar in concept to tags, to develop a common vocabulary for items and users. The system then makes use of profile information and user interactions, both short- and long-term, to provide recommendations. The system leverages both item- and user-based approaches.

I am a little too jaded to (completely) believe the enthusiastic assertion that article recommenders will be the next killer-app, but I do hope this prediction comes true.  Recommenders are basically just a feature in portal and they depend on a lot of other things in it – user tracking, content, collections, ratings.  They are killer apps for Amazon and Netflix but only because everything else they do is also done well.  It will take a perfect storm to get everything right for a scientific article recommender.

In addition to a recommender it appears that NextBio also has a feature that Glen Newton came up with for query refinement: “drill clouds

nextbio-drillcloudThe major difference is that, in Glen’s drill clouds, clicking on a term in adds the term to the conjnction of terms in the query and narrows the search to the subset of documents that contains that conjunction. In NextBio the tag-cloud changes the class of things that the original search term applies to – i.e. it narrows the context for the query rather than adding terms to the query.  Which is a little bit counter-intuitive once you’ve tried Glen’s method (you can experiement with it on CISTI Lab).

I think NextBio – which also includes scientific datasets, clinical trials and news – is a science portal to keep tabs on.

Comments»

1. Mr. Gunn - February 15, 2009

Thanks for the link, Andre, and for your comments here and at Friendfeed.

You’re exactly right that recommendation based on user activity is a feature that works well only on sites which already provide value in other ways. Your examples of Amazon and Netflix are apt, as are the examples of Flickr and Youtube. However, recommendation based on user activity is only one way to do it.

The streaming radio service Pandora, which isn’t available outside of the US, doesn’t recommend based on user activity, but rather on a deep understanding of the track or artist used to seed the recommendation algorithm. http://www.pandora.com/mgp.shtml You get good recommendations from the very first time, even for rarely listened to tracks and you would even if there were very few users.

The reason there are so many users is precisely because all you have to do is go to the site, name a track or artist, and sit back and listen. For Pandora, the recommendation engine is the killer app, and the context in which recommendation will be a killer app for the scientific literature is the same.

2. Andre Vellino - February 15, 2009

Thanks for your comment William.

Yes, having different methods for recommendation that require no user input – e.g. content-similarity based recommendations, or in the case of articles, citation-based collaborative filtering – lowers the “barrier to entry” for recommender usage.

So the model where you have “more like this” to recommend off the bat certainly helps. Over time, though, you’ll only get interesting serendipity in recommendations if they are personalized to your interests over all – which requires some kind of implicit or explicit profiling.

It’s a tricky tradeoff between making a recommender “zero touch” and making it high quality. It’s like trying to get high precision and high recall in a search engine that only uses one or two keywords. You’ll get better results if there’s an interactive query refinement process. But that’s more onerous on the user.

3. Mr. Gunn - February 15, 2009

Yeah, it’s definitely a tradeoff. The way Pandora handles it is to allow a thumbs up or thumbs down on recommended items, which personalizes things as you go, but doesn’t depend on this to start with. That’s the kind of hybrid approach which seems to work well, similar to the “Like” on Friendfeed.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: