Whether you’re combing the online aisles at Wayfair to find a just-right sectional or tuning in to a Spotify playlist during a morning run, machine learning algorithms are the secret sauce making sure there’s a perfect match.
Wayfair, a digital-first retailer focused on the home category, and Spotify, an audio streaming media service, serve two different markets. However, both are platform-based businesses built around data and customer interactions, employing machine learning to deliver highly personalized user experiences and curated product and content recommendations.
At the recent EmTech Digital conference hosted by MIT Technology Review, machine learning experts from Wayfair and Spotify gave a peek at how they leverage technology to deliver bespoke experiences that foster lasting and profitable customer relationships.
At Wayfair, more than 3,000 technologists work on the platform, which connects customers with suppliers in a marketplace of more than 14 million products spanning five websites. Spotify serves 406 million active users across 184 markets, along with millions of creators that have generated 82 million music tracks and 3.6 million podcasts.
For both, the challenge of making on-point recommendations and personalized experiences boils down to a problem of scale.
“You’ve got tens of millions or hundreds of millions of pieces of content and hundreds of millions of users, so the problem becomes multiplicative — how do you determine those key value connections,” said Tony Jebara, vice president of engineering and head of machine learning at Spotify. At the same time, “when you start recommending one of 100 million things to somebody, you really are helping them go beyond what they could find on their own.”
Wayfair: Helping customers find the right product
Given its focus on the home category, where style and taste vary widely, Wayfair is confident that using technology to deliver personalized experiences gives it a competitive edge, according to Fiona Tan, the company’s chief technology officer.
Tan said the company generally applies two rules to determine the best use cases for automation. The first is identifying areas with large quantities of accessible, relevant data that can be harnessed for insights. The second is seeking out problems with tolerance for uncertainty.
"Tasks with more tolerance for forgiveness of faulty predictions and where we have a clear path to improvements and control are the ones best suited for higher degrees of automation,” she said.
It's a balancing act, and Wayfair is using machine learning in several areas where risk is more tolerable, including:
Advertising. The company has fully optimized and automated the bidding process for online ad auctions, ensuring it secures the right placement to boost its search rankings without overpaying on bids.
Search. With Wayfair’s search functions, the challenge is two-fold: The platform has to properly interpret what customers want even when they’re asking in vague terms, and suppliers need to provide proper product information in order to return optimal results.
Initially, Wayfair focused on lexical search, where if a customer was searching for a red sofa, the platform would only return results classified as red sofas. The company has now adopted semantic search, which considers the meaning of the search terms as opposed to finding a literal match. The query for a red sofa might return burgundy couches because it’s a close enough match to the user’s criteria.
“That’s been really good from a customer matching and satisfaction perspective,” Tan said, explaining that revenue sales have increased as customers are less likely to exit the site after an unsatisfactory search.
Wayfair is also working on efforts to build out its product catalog beyond data collected from suppliers. Its plan is to use machine learning to extract information from explicit data stated by suppliers as well as through implicit data that draws from the specific product descriptions to create rich and accurate product descriptions. Because there’s a fair level of risk — serving up a product match with the wrong dimensions, for example, could erode customer trust — Wayfair augments machine learning with humans in the loop to ensure greater accuracy, Tan said.
Supply chain challenges. Given that Wayfair products are located about 1,000 miles away from a customer, on average, it’s important to be able to connect customers with desired products that are closer in proximity to drive down costs. A new “geo sort” capability leverages machine learning to identify and boost products that are closer to customers, with a 250-mile radius being the target.
“There’s some interesting re-ranking that we do,” Tan said. “We want to make sure we’re balancing — that the product choice is relevant, but the ones that are closer to the customer are boosted.”
Spotify: Aiming for a lifetime of content
At Spotify, machine learning is the key to moving consumers beyond finding and curating familiar content to encouraging exploration and new experiences.
“If you let search decide what people listen to, they don’t diversify as much,” Jebara said. “You have to recommend and nudge users into new things.”
Initially, Spotify used machine learning techniques like collaborative filtering, which is a method of making predictions about what a user will like. Now, Spotify is using machine learning to improve recommendations and personalization.
Relevant data includes playlists, user listening behaviors, information about specific tracks or podcasts, and analytics that illustrate how users browse, what they click on, and even what they skip or like. This data is used to inform models trained to learn about user tastes.
The result allows the company to personalize users’ experiences, including what they see through elements like the homepage, the search page, and suggested playlists. These same models fuel some of Spotify’s newer features like blended playlists where you can blend your music taste with friends or artists.
As recommendations and personalizations gain in sophistication, Spotify doesn’t rely solely on machine learning. Human editorial knowledge is used for “algotorial” playlists, in which lists created by human editors are customized by algorithms.
“It turns out, you can’t do everything in the world with just machine learning,” Jebara said. “We also rely a lot on our editors because they have trillions of real neurons between their ears, and are cutting-edge cultural experts that can tell us there’s a new genre of music coming up.”
Spotify is also using reinforcement learning, a way of training machine learning models to make decisions in a complex, uncertain environment, Jebara said. The company hopes to use this to promote discovery of new genres and artists and encourage listeners to consume new content. Spotify designed simulators to model how users will react to a series of recommendations to more effectively train the machine learning models.
“It gives us a great way to train these systems,” Jebara said. “It turns out, reinforcement learning is already giving us huge wins over machine learning systems and that’s a really exciting next frontier. We think personalization should deliver a lifetime of content rather than optimize for the next click.”