Pinterest is boosting its user and ad-revenue numbers at a much faster pace than most of its social-media peers—and it has artificial intelligence to thank for it.
The image-search site lets people post, or pin, photos and videos of things that interest them into collections called boards. The key to the site’s popularity is that it lets you discover ideas through images saved by people and brands you follow—and suggests other images that might be a good fit for you. You also get ads that follow your interests but carry a “promoted” tag.
If someone is interested in home improvement, for instance, the site will deliver pins and ads showing tools and remodeling ideas, says Brandon Purcell, a principal analyst at tech-research firm Forrester Research Inc.
The process of figuring out what people like and finding appropriate images happens without human intervention. It is handled by an advanced AI called a neural network that makes millions of calculations incredibly quickly to find pins to catch your eye.
Neural networks are driving “nearly 100%” of growth, says Jeremy King, Pinterest’s senior vice president of engineering and a former executive vice president and chief technology officer at Walmart Inc.
“It’s like this giant machine-learning engine that is being powered by the people that are [pinning content],” he says.
Picture of growth
The neural-network formula drew nearly 480 million people to Pinterest in the first quarter looking for style tips, gardening ideas and more. That number is up 30% from last year’s first quarter, while ad sales have more than doubled since 2018. In comparison, U.S. ad sales at YouTube and Facebook, with which Pinterest is often grouped, are up 80% and 56%, respectively, according to eMarketer.
Pinterest is just one of many organizations taking advantage of this technology. For these groups, neural networks are a leap forward in analysis. In the past, it was easier for computers to delve into “structured” data, such as numbers in a spreadsheet, than “unstructured” data such as images, speech and text. With neural networks, AI can tackle unstructured information more easily, bringing advances in medical imaging, automated voice response, text generation and a host of other uses, says Mr. Purcell.
Forrester and International Data Corp. say about a third of companies that use AI in some form have adopted neural networks or will do so within 12 to 18 months.
When it comes to social media, LinkedIn, which is owned by Microsoft Corp., says it uses neural networks to match users and ads, while Twitter Inc. says it doesn’t. Representatives for Facebook Inc. and YouTube, which is owned by Alphabet Inc.’s Google, declined to say what technology they use.
But the growth of the technology has some critics worried. For one thing, neural networks are often fueled by data on users, which is supposed to be anonymized. However, Nader Henein, a research VP at Gartner Inc., says that the networks are incredibly good at finding patterns. They could, with a large enough data set, theoretically make enough connections to identify people whose data is in a repository. The technology, he says, “gets surprisingly better at saying, ‘Oh, this belongs to this person and this belongs to that person.’ ”
And the more layers there are in a neural network, the less is known about how it’s actually making those connections.
This lack of visibility leads consumers to wonder about “black box” decisions—neural-network outcomes that a system’s owner can’t explain, says Stephen Messer, co-founder and vice chairman of AI company Collective[i].
This doesn’t mean organizations should shy away from neural networks, which have many benefits, Mr. Henein says. But organizations need to take responsibility for their data and technology and carefully manage their uses. “We need to be responsible for the decisions this engine is making,” he says.
Pinterest says users share information about themselves and their interests through profiles and pins, knowing the company will, with their permission, use that data to find content that interests them. The company says it takes steps to protect that data, but declined to provide specifics, citing security concerns.
How neural nets work
The power of neural networks lies in how they analyze information, and apply those lessons.
The systems are composed of artificial neurons—a combination of processors and software algorithms that are designed to mimic the way human neurons transmit and process information. They ingest large amounts of data such as images, numbers and text, and use it to learn about the features of a specific object.
For example, neural networks will pick up patterns in the shape of eyes and whiskers that distinguish, say, a cat from a dog. This is very much how humans learn. When babies start noticing animals, they might point to a dog and say “cat.” As adults correct them, the babies keep looking and start picking out common characteristics, or patterns, that distinguish a dog from a cat. With a similar trial-and-error method, a neural network will start recognizing patterns and identifying objects.
A neural network with many layers of neurons, called a deep neural network, is able to make finer distinctions. For instance, it might be asked to analyze a photo of an animal underneath a bed and be able to tell not only that it is a cat, but also what species of cat it is when only its face is showing.
At Pinterest, neural networks crunch mountains of data about users, such as searches, the boards of people they follow and what pins they click on and save. At the same time, the networks look at ad data on users, such as what content gets them to click on ads. The neural networks learn users’ interests and can serve up content that is increasingly relevant.
A single neural-network model focused on Pinterest ads may make more than 30 million predictions per second. Considering the complexity and speed needed to instantly match the right ads with the right users, it would be impossible for a company such as Pinterest to narrowly target ads without neural networks, says Ritu Jyoti, program vice president for artificial-intelligence research at International Data Corp.
So, what does this look like in practice?
Let’s say a homeowner is thinking of installing a bathroom floor and pins flooring ideas, such as images depicting tiles. Pinterest needs to figure out what the user is pinning and what those images signify. A “convolutional” neural network—which analyzes images—will scrutinize the pixels in the tile pins and classify them by, say, material, size and color.
The system will match those attributes to other pinned images of tiles on the site with similar features. Now it’s got some attributes of the tiles figured out. Next up: figuring out what the tiles are for.
The neural network looks at those images of tiles and the other images in those pictures. If the network spots something like a bathtub, it might determine the tiles are for bathroom flooring.
SHARE YOUR THOUGHTS
How much do you care whether the ads you see are personally relevant? Join the conversation below.
The neural network will apply internal labels about objects in an image, such as a bathroom tile’s color and style, as well as other data points, such as the name of a board on which an image was saved.
The network then needs to find images to show the user that have to do with bathroom flooring, as well as related images. A Pinterest “graph” neural network—which looks for different types of relationships on the site—analyzes the user’s behavior. It looks at pins the user searches for and saves, and how he or she organizes them into boards.
The system also analyzes the boards of other people with related interests, to make related recommendations.
From there, the AI may spot a pattern: For instance, people interested in bathroom flooring often shopped for a vanity by using one of Pinterest’s shopping pins.
Likewise, neural networking analyzes its advertisers’ pins, classifies the objects in their ads, and shows the user images with a style and price to match the user. A homeowner may be paired with ads from a small retailer offering a rustic-looking bathroom vanity.
The power of Pinterest’s neural networks is that they make associations a user wouldn’t normally make. Users installing a new bathroom floor might not even think about buying a vanity until an ad popped up showing a unit that exactly fits their interests. Pinterest argues that the fact that the ad was presented to a Pinterest user in the same format as their pins actually enhances the user experience.
Floor & Decor, which sells hard-surface flooring and related accessories, says it likes advertising on Pinterest because it is a place where users plan projects, says Andrea Striebel, vice president of marketing at the company. When Pinterest users look for bathroom-floor tile, a Floor & Decor ad may appear that includes a vanity. Users “may not realize that we carry vanities as well,” Ms. Striebel says. “So, it’s a great way for us to tell the whole picture of everything we have to offer.”