The recommendation problem is a machine-learning problem, says startup Jybe, and one that it aims to address with its iPhone app that now is in beta. Coming soon (though not immediately) to the iPhone 5, which will require some redesigning to maximize real estate, the mobile app supports earlier iPhones, the iPod touch and iPads running iOS 4.3 or later.
Unlike services such as Yelp, that are more reviews than recommendations, Jybe takes the “serendipitous discovery” approach to real-world goods and services (movies, books, restaurants, and dishes). Founded by CEO Arnab Bhattacharjee, CTO Tim Converse, and chairman of the board Tuoc Vinh Luong, a team with a slate of experience in the search engine industry at names like Yahoo and Powerset, Jybe looks to provide implicit search, i.e. search without query. “The only way to figure out your interests is to figure out who you are, what you like and surface things interesting for you to consume,” says Bhattacharjee.
Jybe strives to narrow the universe of objects that a particular user might like based on all it knows of the user, including location in this age of mobile devices, previous likes, and so on. “The tactical goal is to take clues from as many different levels of information as we can,” says Converse. “We pay attention to how many others have liked this thing, to what extent those you are connected to say [about it], to what people similar to you like that say, and delving into the properties of the object.” For instance, if you like a certain dish at a restaurant, it can do an analysis of it and figure what kind of ingredients you like.
“There are a lot of different features that lead to [the results], but the machine learning algorithm is trying to do a good job on unseen data to predict what you like,” says Converse. The service also supports predictions by breaking people into clusters around their tastes. It doesn’t decide which features for a particular topic are most important in advance: For movies, for instance, machine learning can help it figure out if actors are the most important part of the equation to the movies a user tends to like, vs. genre or director, and so that automatically gets more weight as a result of the modeling, he explains.
When a user is satisfied with a recommendation, she can click to add a movie to her Netflix queue, buy a book from Amazon, or reserve a table at a nearby restaurant. There’s a challenge and opportunity with location, Converse notes: “If we suggest something very nearby you, that is likely to be actionable….But how much do you want to forefront location. Do you want something that is somewhat appealing that is 10 feet away or a fabulous one about a mile away? So there are interesting challenges in updating search results to get those location things in the right way.”
Get Going on Jybe
After a start in which the app starts learning by asking users a few questions about its main categories – the kinds of movies they do or don’t like, for instance – and then going a little more deeply into each one to further plumb perspectives on different flick titles, for instance, they get right into the app and its recommendations. It continues to learn as users like, dislike and share more things, or connect with others they see being active in the categories they’re interested in, so that those persons’ feeds show up more often in their own choices. In fact, those whose likes seem to be affecting the likes of others can become influencers in a particular category.
“Friends [you’re already connected to] don’t necessarily share your taste, but this gives you ways to find those who do for more accurate recommendations. And we can consume how much you and your friends are matched in what you like and dislike, and we can get strong signal here, too,” says Converse.
The social stuff matters, of course, but, as Bhattacharjee points out, “A system like this, if purely based on social signals, it’s not much good for the first user. You need great information. So we leverage our background as web search engineers – we know how to crawl the web, how to extract signals, so intelligent crawling algorithms download content from expert sites, social sites, and are collecting and normalizing and aggregating information so we have a pretty good sense of details about things and what is the web buzz.”
Leisure and entertainment is just the start of where the technology can go. “There is a lot of data available on the web and you could build a great experience – you can imagine events and all kinds of goods and services being recommended,” says Bhattacharjee.