Can AI Be Your Guide to the Web?
A startup has licensed heavyweight artificial-intelligence technology for its online recommendation engine.
Today, a startup called TrapIt launched a beta website that recommends content after learning your tastes via an artificial-intelligence engine spun out of research originally funded by the Defense Advanced Research Projects Agency (DARPA). The company hopes this technological pedigree will set its method apart from other ways of finding information, such as searching or receiving recommendations from social-media sources.
TrapIt’s technology has its roots in the CALO project at the independent nonprofit research institution SRI. CALO is an ambitious attempt to help computers understand the intentions of their human users. A previous company spun out of CALO, Siri, developed an intelligent software assistant that could perform simple tasks, such as planning an evening out, when given voice commands. Apple acquired Siri in April 2010, although the technology has yet to appear in any Apple products.
TrapIt relies on a different part of CALO’s intelligence. “Learning from data is the property we’ve got our hands on,” says CEO and cofounder Gary Griffiths. He explains that in the aftermath of September 11, U.S. government agencies felt they’d had access to data that could have predicted the attacks, but they didn’t know where to look for it. DARPA funded CALO in part to work on this problem. The project sought ways to sift through information to find what might be most relevant to a given topic, and to learn from a user’s response to the information offered. It’s this technology that TrapIt is converting into a consumer product.
At first blush, TrapIt might look like any Web 2.0 site. After signing up, the user can select from existing “traps”—collections of articles related to featured or trending topics, such as the golfer Rory McIlroy, who just won the U.S. Open. The user can also create new traps by entering a few keywords and going through one screen of training data.
However, in either case, the traps then belong to the user, and they change according to his or her tastes alone—even if they were originally created by someone else. TrapIt’s algorithms comb through about 50,000 unique sources of content, analyzing articles to classify the types of information they contain. (The 50,000 sources were vetted by humans to filter out content farms and other material of dubious quality.) TrapIt combines this information with machine-learning analysis of what the user has previously clicked on to recommend new information.
TrapIt’s founders argue that this approach provides the perfect balance of serendipity and precision. While search engines recommend popular Web pages for a particular topic, TrapIt is designed to do a better job of surfacing obscure content, Griffiths says. And while social media can provide interesting new links, TrapIt can draw on more content and make sure recommendations stay closely related to a user’s interests.
The company has also put a lot of effort into packaging the site. For every trap, the company’s algorithms extract headlines and images.
TrapIt’s team hopes that users will not only enjoy the site but also understand that they are benefiting from heavyweight technology. A lot of recommendation engines are very manually driven, says Frank Meehan, CEO of the social mobile company INQ and a member of TrapIt’s board of directors. For example, Facebook’s “Like” button relies on user clicks to gain information about people’s tastes. Crucially, in Meehan’s view, TrapIt’s design reveals the considerable artificial intelligence beneath the interface. Showing featured and trending traps will reveal how well the algorithms select content, he says.
TrapIt’s founders envision supporting the service with advertising, but they also have other ideas for how it might make money. For example, they could offer paid premium service to users who want to research niche topics such as law or biomedicine; this could include access to content normally hidden behind pay walls. The company also hopes to license the platform to media companies, who could use the technology to put together personalized packages of content for subscribers and other users.
For now, however, Griffiths says the company is just focused on making the site work well. In early tests, users averaged 18 page views per visit—an unusually high number. (Facebook, an acknowledged leader in this area, averages about 24 page views per visit.) Griffiths says he hopes the site’s ability to capture people’s attention will build a strong user base and attract advertisers.
Daniel Tunkelang, principal data scientist at LinkedIn and an expert in information retrieval, says personalization through data analysis can be complementary to personalization through social analysis. He notes that researchers have long considered developing recommender systems that use both social- and content-based methods to make recommendations.