How Foursquare Transformed Itself with Machine Learning
The last time I wrote about Foursquare was over ten years ago, when it was just a location check-in app for consumers. At that point, July 2012, it was trying to take on Yelp in the consumer recommendations market. Several pivots later and Foursquare is now squarely an enterprise service. It markets itself as “the leading cloud-based location technology platform for unlocking the power of places and movement.” Among its customers is Uber, which uses Foursquare’s ML-enhanced data to help pinpoint a user’s exact location.
To learn more about how Foursquare is using machine learning, I talked to the leader of Foursquare’s “Places Data Science” team, Zisis Petrou. He joined the company as a data scientist in 2017, soon after it had pivoted to the enterprise market. That was also the year Foursquare announced its Pilgrim SDK, an enterprise software development kit based on the eight years of data it had accumulated. By this time, Foursquare was harvesting phone sensor data and other automated forms of location tracking (all opt-in for the consumer), in addition to manual check-ins.
“Sometime in the middle of the 2010s, focus switched [to] enterprise users,” Petrou explained. “And then, of course, we decided to make use of the rich [store of] content of places that we have and also the unique potential to identify users — both visits and getting feedback from users.”
As noted above, Uber uses Foursquare data to identify the precise location of places a user might request to go to. Petrou put himself in the shoes of an Uber driver to help explain.
“So if a client asked me to go to a specific place of interest — a specific restaurant or a specific theater or something — it is very possible that under the hood, Uber’s hitting our API or uses data that we provide it as a flat file, and gets the location coordinates from there.”
How Foursquare Does Machine Learning
Foursquare recently announced an enhanced version of its Geosummarizer ML model, which it claims “increases the accuracy of a point of interest by a substantial 20%.” Geosummarizer, the company explained, “is the model that selects a final Lat/Long for a POI [point of interest] based on an analysis of geocodes from various inputs within its cluster.”
Foursquare says it now has over 120 million POIs in its databases. According to Petrou, this data comes from a variety of sources. It continues to get data from users of the apps it has (more on that in a minute), but it also purchases data from third parties, and crawls the web for even more information. This is where the ML comes in — it combines those separate pieces of data to come up with a precise calculation (Lat/Long) for a POI.
“In order to get to the best representation of the attributes of [a] place, like the address or […] the geocodes — the geographical coordinates — we apply a process that we call summarization. Which is, we take into account all the information that comes from different sources for the same POI. We reach what we consider to be the optimal final attribute.”
Under the hood, he continued, is a machine learning pipeline that “takes into account the confidence that we have assigned after careful evaluation of all these sources, [and] takes into account geospatial information that we have as well.” The algorithm, he added, “predicts the optimal lat/long based on all these inputs.”
I asked specifically what type of ML techniques it uses for this? Petrou replied that it’s “a supervised learning algorithm — it’s a problem that we have formulated under the hood as a regression problem.”
So essentially, Foursquare assigns a predicted score to each potential POI, ranks them, and the algorithm ultimately picks “the one with the highest probability as the final representation of the place.”
Petrou says it uses Python libraries like scikit-learn and PyTorch, as well as “libraries that we have developed in-house as well.”
Foursquare’s Consumer Apps Are Still Used
Getting back to the sources of data, I was surprised to hear that Foursquare still gathers significant data from its consumer apps. I’d long ago deleted the original Foursquare check-in app from my phone, but after checking the iOS App Store, I saw that it has two current apps: Foursquare Swarm (the current name for its “lifelogging” check-in app) and Foursquare City Guide (for “restaurants and bars nearby”). The latter is actually the original app, as I discovered when I downloaded it again and saw my old check-ins — most of them over a decade old. Update: I didn’t see all of my old metadata — all photos and text — until I downloaded Swarm.
Re-installing Foursquare’s apps and seeing the photos I’d uploaded brought back memories of places I used to go to regularly — I used to be the “mayor” of a cafe in Petone, New Zealand, called Go Bang Expresso. Sadly, that cafe no longer exists (and also, I’ve moved to the other side of the planet). But actually, Foursquare isn’t just about memories for old timers like me — its current apps are still well-used, said Petrou.
“Yeah, we still get consumer data,” he said, sounding almost a little insulted that I’d asked. “We identified that this is one of the unique elements of the data that Foursquare has — the ability to have people on the ground willingly supporting us with data, willingly finding the place that they want to check in. And if they notice that it is a little outside […] of the coordinates that they have, or [it’s] the building next door, they send us votes (as we call them) for moving the place into the right part of the building or of the block. So this is something unique that we decided to keep focusing on and keep using because it’s something that differentiates us from other [companies] in the industry.”
Foursquare doesn’t officially release usage numbers for its consumer apps, but a CNBC report from last year quoted the company as having “9 billion-plus visits monthly from 500 million unique devices.”
Regardless of whether you still use Foursquare’s check-in apps, there’s no doubt that Foursquare has amassed an incredible store of location data since launching in 2009. Not only that, it’s using machine learning to enhance the underlying data — making it even more valuable to third parties like Uber.
I never thought I’d write this, but Foursquare is a model enterprise technology company in 2023. Certainly, a pivot well executed, plus a good case study for how to use ML in this era.