Limited personal data doesn’t have to limit personalization

Personalization makes experiences feel special. But does personalization require lots of personal data? Khaled Boughanmi, assistant professor at the Samuel Curtis Johnson Graduate School of Management, developed an algorithm called MetaTP that infers an individual’s music preferences and personalizes recommendations with just five data points. His findings were published June 3, 2026 in the Journal of Marketing Research.

“You listen to the song, accept the song, or you reject it, you skip it,” said Boughanmi. “And just from those few observations, we are able to quickly learn the preferences of customers with a degree of accuracy that is extremely surprising to the industry and to us as researchers.”

The technology, built by Boughanmi and the study’s co-authors, addresses the early stage of the customer journey when there is limited data and no established brand loyalty.

To pique interest, the researchers focused their work on music. “Music is something that is fundamental to almost all of us, we all listen to it,” Boughanmi said.

MetaTP analyzes the energetic rhythm and acoustics of the music during a listening session to personalize recommendations.

“Not only we are able to learn quickly, not only we are able to predict dynamics, but also we are fully interpretable,” Boughanmi said. “If I recommend Thriller by Michael Jackson, I can tell you it’s because my model tells me that you are really hype for good tempo.”

While immediately useful for those lukewarm customers, the technology can also aid in quickly adjusting to distributional shifts from an existing customer.

“I’m someone who likes Michael Jackson a lot, I like Adele, I like Taylor Swift. I like that kind of music. The distribution of my preference is around that point…” Boughanmi said. “So that’s who I am from a musical preference standpoint.”

But what happens when a new variable, like a breakup, disrupts someone’s typical listening style?

“The distribution of preference shifts from listening to happy songs to listening to very sad songs. So they start listening to Cry Me a River [by Justin Timberlake] and that distributional shift is extremely difficult to capture,” he explained. “We have this new technology or methodology that is able to capture the distributional shift just from a few observations.”

The technology is able to adjust so quickly, precisely and with limited data because it’s meta-learning, using listening sessions, as opposed to supervised learning. Unlike a supervised learning version that would simply look at the characteristics of the songs for its prediction capabilities, Boughanmi explains that he’s teaching the machine how to learn from the user’s behavior.

Through a meta-learning methodology, MetaTP takes the entire sequence of a listening session and cuts it in half. The first half is the context set that trains the technology on the listener, and the second half is the task completed according to the initial data.

“We do it for like 10,000 sessions, 100,000 sessions,” Boughanmi said. “And then basically what happens is that the computer, instead of learning one song, one outcome, skip or listen, the computer learns there is a context of five songs that are skipped or not.”

The immediacy of taking in new information and adjusting accordingly allows for companies to earn the trust of new customers, especially in instances where more data isn’t available.

“There are instances where we can only collect limited sets of data, then this will offer a solution to that,” he said. “Of course, it’s not going to be perfect, but it will be a very good solution compared to what other techniques would do.”

The research focused on music, but Boughanmi said that he sees vast potential for meta-learning technology across industries.

“This kind of technology could be useful for all sorts of things… just think about the impact when it comes to medical treatments, for example,” he said. “The math is the same. It’s the application that changes.”