Can Twitter tell you what books to read next? That's the premise behind BookRx, an online experiment from the Knight Lab, created by Larry Birnbaum, co-director of the Intelligent Information Laboratory at Northwestern University, IL, and Shawn Ryan O'Banion, a PhD student in the Knight Lab and Computer Science at Northwestern.
We asked them how it came about, and how it works:
How did you come up with the idea of BookRx?
We've been interested in social media and content recommendation for a while now, because social media data, and especially Twitter, are publicly available. That means you can build interesting systems even if you're not a big e-commerce vendor or web site with lots of proprietary data about user behavior.
So we've built a couple of systems along these lines already. More than a year ago we built a system called Twitter Profiling (using a somewhat different approach) that recommends news articles to you based on your tweets -- we used Huffington Post stories for one of our demo prototypes! Just before the election, we built a system called Tweetcast Your Vote, that predicted who you would vote for in the 2012 Presidential election based on your tweets -- it works a lot like BookRx.
How does it work?
BookRx works in two phases. In the first phase, it analyzes your tweets (in terms of the words, Twitter usernames, and hashtags you use) and compares them to terms that are correlated with book categories. In the second phase, it looks within those categories to find specific books to recommend, again based on correlations with the terms in your tweets. The first phase is very fast but the second takes a few seconds.
How accurate do you think it is? Does it work for you?
We haven't had a chance to evaluate BookRx formally yet -- we were rushing to get it done before Christmas! Anecdotally, it seems to have enough hits and near-misses to keep people interested.
What can people's Twitter word usage tell us about their personalities?
That's a really interesting question. We're really interested in how Twitter can hold up a mirror to ourselves, and seeing BookRx's recommendations might be one way to do that. That's one of the reasons we show you the terms you used that made the system think you might be interested in a book it's recommending to you -- to make its operation a bit more visible.
If you look at Tweetcast Your Vote you can see some interesting distinctions we found between the terms that Romney/Ryan supporters used in their tweets, and the terms that Obama/Biden supporters used.
Is there something particular about people's book choices that makes them different from other media?
I'm not sure. We're going to do other media too, eventually, and that might help answer that question.
What would you like to do to refine the tool?
A couple of things. First, we'd like to evaluate for real, on a large scale; that'll help us tune it and make it incrementally better. Also right now just because of sheer numbers it tends to recommend current best-sellers. We'd like to bias it a bit away from that at least some of the time so that you get a more diverse set of recommendations.
And of course it would be neat to hook it directly to some booksellers so you could easily get something that caught your interest!
Try BookRx here and let us know in the comments if it worked for you!