The future of Foursquare is data-fueled recommendations
By Ryan Kim
When Foursquare first appeared on the scene, it looked more like a real-world game, with people checking in to locations to try to secure points and “mayorships.” But from the beginning co-founder Dennis Crowley also had a deeper vision that hinged around tapping into the wealth of data,. That vision became clear with the launch of Foursquare’s Explore feature last year.
Suddenly all of that fun check-in data was put to use as the fuel driving Explore’s very capable recommendation and search engine. Foursquare, it appeared, was a powerful big data company using a catchy front end to feed in more information.
The development, Crowley explained to me in an interview, was a bit how Mr. Miyagi taught Daniel karate in the The Karate Kid: “We asked people to check in, which is like painting the fence. Now we’re teaching karate,” Crowley said, adding, “It all goes into a recommendation engine that knows what you like and what else you’ll like.”
Explore debuted in March 2011 with Foursquare 3.0, fundamentally changing the usage model behind the location-based service. Instead of charting a user’s movements, Foursquare was spitting out recommendations and answers about the best places it thought users would want to visit.
Explore bases its recommendations on the places a user, a user’s friends and similar types of users visit. The platform factors in types of places and time of day when people search for recommendations and tailors responses for each location and time. Increasingly, it is also being fed by other signals like the tips people leave, the lists they construct and the places they are saving to Foursquare to visit later.
Foursquare takes all of this data and returns suggestions within 200 milliseconds. This is all done using the data Foursquare users have entered themselves, without prompting, so it has an air of authenticity that reflects their real tastes.
That unprompted data alone is huge. Foursquare has had well over 1.5 billion check-ins, including 5 million per day from more than 15 million users. There are more than 35 million venues on Foursquare, 750,000 of which have been claimed by business owners.
You and your peers
But what Foursquare is essentially doing with Explore is creating profiles of people based on what it can gather from users and their friends. The company applies machine learning algorithms to the collective movements of its users to figure out what people are doing and how one user fits into the larger group. If it understands a user prefers independent coffee shops, for example, it doesn’t suggest Starbucks or Dunkin’ Donuts. Wherever that person goes, Explore understands what that person and similar people are looking for. And it suggests coffee shops around the time of day a user has previously gone for coffee.
Explore then matches those tastes to the geo-spatial and temporal data it has collected on all the venues in its universe. Over time, a place can develop a profile or fingerprint, just like a user, attracting certain types of people at specific times of the day. Some places prompt even more engagement from users who capture pictures and leave tips and share the venue online or on a Foursquare list.
That data helps Foursquare understand more about that place and similar venues like it, and it helps the engine decide if it should recommend the venue to users that match that profile. It is a complex engine and one that is being tweaked and improved constantly. But it’s an example of how social data, both structured and unstructured, can be put to smart use for real-world recommendations.
Foursquare has increasingly been building up its data team. It hired Justin Moore, a former quantitative analyst and Bear Stearns VP and director of technology in May of 2010 (though it recently lost him to Facebook). The company now has about a tenth of its 100 percent staff on its data team, including Andrew Hogue, the new head of search, an engineer who spent seven years working on search projects at Google. The company has built a big data stack using MongoDB, Hadoop, Amazon S3, Elastic MapReduce and other tools.
Hogue said even as Foursquare becomes known as more of a recommendation and planning tool, it is important to keep getting people to check in. The more people put in, the better and more refined the recommendations that come back. And that also helps support the majority of users, who are less engaged and don’t input as much data.
“We don’t want check-in data to go away because it’s the most direct, visceral signal we have,” Hogue said. “It tells us you’re there, interacting, one to one. It captures the experience in a raw kind of way.”
In the future, Foursquare will look to incorporate and balance more data to find even better recommendations for users. Hogue said, for example, Foursquare may look at giving more weight to knowledgeable people who are creating lists and leaving tips. The service will also look to get more direct input from users about their tastes, so it doesn’t have to just infer it from their actions. The future of Foursquare is very entwined in its big data operation, as it learns to pull out more value from its growing pile of information.