An MIT doctoral student won a $10,000 award in August for developing a new approach to recommendation algorithms that has already been adopted by Tencent WeChat, where it helps determine what content is shown to more than a billion users.
Lei Huang, 30, a final-year doctoral student at MIT’s Sloan School of Management, won Texas A&M University’s Mays Business School AI Dissertation Proposal Competition for his ongoing research, “Designing Self-Sustaining Markets: An Application to Content Platforms,” which seeks to solve a fundamental problem facing content platforms: Giving every user exactly what they most want to see in the moment does not create the healthiest content ecosystem over the long run.
The research, conducted jointly with Huang’s adviser, MIT Sloan professor Juanjuan Zhang, proposes an algorithm that sometimes sacrifices a user’s top choice to steer attention toward creators whose continued presence makes a platform more valuable, potentially giving someone their second-favorite recommendation if the additional views help keep a valuable creator producing.
“The consumption of any single consumer not only affects that consumer, but also affects all other consumers by affecting future supply,” Huang said, describing an economic phenomenon known as a “consumption externality” that affects markets ranging from social media to ride-sharing and e-commerce.
Huang uses an example from his dissertation to explain the phenomenon: Imagine a university offering two courses, coding and branding. Of 60 students, 31 prefer branding and 29 prefer coding, but each course needs 30 students to remain available. A traditional recommendation system would direct everyone to their preferred course, causing coding to disappear. The algorithm would instead recommend coding to a student with only a slight preference for branding, sacrificing a little immediate satisfaction to keep coding available for students who strongly prefer it now and in the future.
The same principle applies to content platforms. A creator who receives little attention may stop producing, shrinking the content available to everyone.
“This is well-documented in the literature,” Huang said. “The real contribution of our project is not to document that externality, it’s to operationalize it and design solutions at massive scale.”
Deciding which creators deserve additional attention requires the algorithm to weigh two factors: “creator sensitivity,” or how responsive a creator is to additional views, and “creator contribution,” or how much the content ecosystem would lose if that creator disappeared.
Popularity matters, but so does whether other creators offer similar content. “Even if the creator is popular, if there are many close substitutes on that platform, it doesn’t really hurt the consumers too much if that creator leaves,” Huang said.
Calculating that contribution becomes enormously complicated across tens of millions of creators and billions of users. To address that problem, the researchers developed PIXSET, which converts information about users’ preferences and creators’ content into visual representations that can be analyzed using machine learning.
In one analysis involving roughly 29,000 creators, calculating creator contribution individually took about seven and a half hours. PIXSET cut that to about 30 seconds, a 99.9% reduction.
Huang began working on the project with his adviser in early 2022. Later that year, the pair met researchers from WeChat at an MIT conference, who invited them to present their research to the company. WeChat is a Chinese “super app” with more than a billion users that combines features Americans might associate with several separate apps, including social media, messaging and PayPal-like digital payments.
When Huang presented the research to WeChat’s content team, “They said, ‘This is exactly the problem that we are working on,’” he recalled. Now, the algorithm is used on WeChat’s content platform, where users can publish “images, videos, articles, whatever,” Huang said.
Huang subsequently spent two years, from late 2023 through late 2025, as a machine-learning research intern at Tencent, working with company teams to adapt, test and deploy the system. In a two-week field experiment involving more than 25,000 creators, the algorithm performed 46% better than WeChat’s existing approach on the company’s “net user benefit” metric, a tool which weighs the immediate cost of less-preferred recommendations against the future benefit of keeping valuable creators producing. Huang’s algorithm also increased the likelihood that creators would continue producing content by 2.6%.
“I just observed the metrics over the experiment window and observed that, wow, the graphs become better, better, better, better,” Huang said. “In the end, statistically significantly better than the other algorithms. I knew that they were going to adopt it.”
Huang called seeing his research adopted across WeChat one of his proudest moments. Asked to describe the feeling, he turned the comparison back on his interviewer.
“Suppose your article is featured by all the big media,” Huang said. “You’d be so proud.”
Huang, who grew up in China’s Hunan province, came to Cambridge after earning his undergraduate and master’s degrees at Tsinghua University in Beijing.
Now in his final year at MIT, Huang believes the approach could eventually extend beyond content platforms to markets such as Amazon and Uber, where sellers and drivers also respond to how demand is distributed. Asked whether the technology could form the basis of a startup, Huang said he and his collaborators had previously considered the possibility but did not pursue it, though he still sees potential for a company that helps platforms design more efficient algorithms.
Huang said he has not yet received the $10,000 prize but expects to put it toward his research, including computing costs associated with training and running machine-learning models.
For now, he is seeking professorships across North America Europe and Asia, including in the Boston area. Whether he remains in Cambridge after completing his doctorate will largely depend on where he finds a job, but regardless Huang said his years at MIT have transformed him.
“It totally changed the way I view the world,” Huang said. “There are super intelligent people all around me. I just love it.”
But what continues to excite Huang most about the research is the underlying idea that inspired the project four years ago: One person’s seemingly individual decision can ripple through a marketplace, changing what gets produced and, ultimately, what remains available to everyone else.
“From the individual consumer’s perspective, our single decision affects other producers’ supply incentive in the future, and therefore affects every other consumer,” Huang said. “I think this chain is just amazing, so elegant.”
A stronger
Please consider making a financial contribution to maintain, expand and improve Cambridge Day.
We are now a 501(c)(3) nonprofit and all donations are tax deductible.
Please consider a recurring contribution.