When a social app reaches a certain level of popularity, it becomes something much bigger than its interface and functionality—indeed, it often balloons beyond the scope of whatever its creators intended (see Twitter). In Instagram’s case, the scale is massive: 800 million users and growing. With numbers like that, what was once a photo-sharing app becomes, among other things, a complex and nuanced visual map of people’s interests and experiences around the globe, updated with a fresh layer of imagery every millisecond.
To keep us addicted (the name of the game in the advertising-fueled, attention-thirsty social networking business), Instagram finds ways to mine this global graph of our interests and curiosities, and then feed us new things–things we’ll tap, follow, and comment on, ideally. Instagram’s Explore tab has become its chief engine for this kind of discovery; there, you’ll find finely personalized (sometimes creepily so) recommendations for images and accounts that you’re likely to enjoy, mixed in with more serendipitous, curated content you never knew you wanted to tap. The methodology used by Instagram to craft these suggestions is a blend of human and machine intelligence that points to machine learning’s greatest potential: augmenting our brains, rather than replacing them.
“Our challenge at Instagram is to understand the nuances of why people follow what they follow and try and help them achieve a more rewarding and enriching experience,” says Dan Toffey, who oversees Instagram’s newly created Community Lab. “We try to drive discovery of what’s happening on our platform, but also create an appetite for discovery. We try to showcase things that broaden people’s understanding of what Instagram is.”
So how does Instagram figure out what kind of bait to leave waiting for us in Explore? Algorithms and machine learning are a core part of the equation. At the heart of Explore’s image recommendation system is a two-phase algorithmic process that attempts to whittle down Instagram’s massive sea of content into images that align with your interests (a process called “sourcing”)—and then fine-tune things further by prioritizing which images to show you first (a process called “ranking”). The system takes cues from your social graph—such as which accounts your friends follow and how they relate to the types of things you already like and follow—as well as from signals found in image captions and account bios of various accounts.
As crucial as machines are to discovery on Instagram—it’s hard to imagine a team of humans weeding through billions of photos and videos and creating customized suggestions for 800 million people—the human side of the equation is arguably just as important.
“The community team spends essentially all of its time looking at stuff on Instagram, in the same way that you might serendipitously go down the rabbit hole or follow or click through on something,” says Toffey. “That’s essentially what we do all day long.”
If you’ve ever spent time tapping around in the Explore tab, you may have spotted this team’s hand-curated work. From mesmerizing slime videos and trippy animations to cake decorating, puppies, and style inspiration, Explore offers up hand-picked channels of videos that may not have been swept up in the algorithm’s content scavenger hunt. In my case, Explore is heavy on musicians and artists doing their thing. There’s also pizza. And science experiments. Who can resist tapping the “Oddly Satisfying” video channel? The name says it all. “I often call these systems Borg systems because they’re human-machine hybrids,” says Thomas Dimson, a software engineer at Instagram who focuses on machine learning.
In many cases, the machine side will pick up on new trends and niche communities on Instagram, allowing the flesh-and-blood researchers and curators to use their more nuanced, culturally aware eyes to assess them. Other times, the community team will happen upon new trends themselves. In the course of digging into Instagram’s trove of imagery, for example, Toffey says he came across the “adaptive athlete” hashtag, which is used by athletes with disabilities to showcase their skills and connect with others. However they’re unearthed, these visual representations of people’s interests make their way into a queue for human eyes to evaluate and then send down the curatorial assembly line—potentially to your account, if you’re into it.
“As the result of [Instagram’s] breadth of sharing, there’s almost infinite niche types of content,” says Layla Amjadi, a product manager at Instagram who works on Explore. “The challenge is: How do we get the right content to the right people 800 million times over in an instant? How can we be hyper-aware of how your tastes and your interests are evolving over time?”
If the system is working properly, you and I should see a well-balanced mix of irresistibly relevant imagery and new things that pique our interest nonetheless. How does Instagram know if it’s working? Like any digital product of its kind, the definition of success is driven by metrics that measure how we respond to what the app serves up. With so many signals and data points to learn from, the team’s method for measuring success can undoubtedly get fairly complex and down-in-the-weeds. But at its heart, it seeks to understand one thing: Are people coming back and tapping around more everyday?
Like the curation of the Explore tab itself, the evaluation of its performance relies on a combination of data and human intelligence. In this case, the people-powered side of the equation includes things like user experience research trips around the world to observe how people use the Instagram app.
“Some of the biggest insights I’ve had have been on these research trips where we go abroad and talk to people,” says Dimson. “There are times when you want to pull out your hair, like, you’re using the product wrong! But nobody’s using it wrong. You’re building it wrong. There are a lot of moments like that.”
Sometimes, Amjadi says, these user experience research inquiries can happen more informally in the wild, like when she’s standing in line at the store and spots somebody swiping and tapping away on Instagram. The questions she asks strangers in public may not be as scientifically rigorous as the more controlled UX studies conducted by the company, but they point back to a familiar theme: As smart as the machines have become at spotting patterns and evaluating success, they still very much need our help to make meaningful sense of it all.