To Engage Video Viewers, Make More Personal Recommendations

Platforms like TikTok could boost viewing times by grouping users to better match their preferences

Based on the research of Junyu Cao and Yan Leng

In the nine years since TikTok debuted, it’s helped transform the way people view and absorb information, along with other short-form video platforms such as Instagram and Snapchat. Every month, TikTok alone has nearly 1.6 billion active users globally.

But despite their maturity, such platforms are falling short in a basic function: recommending the right video to the right viewer at the right time. So say Junyu Cao and Yan Leng, both assistant professors of information, risk, and operations management at Texas McCombs.

At NetEase Cloud Music, a popular music streaming service in China, they found 70% of videos got no views on any given day, while 87% of users did not interact with any of the videos they saw.

Using 2019 data from NetEase, the researchers devised a new approach to learning users’ preferences for short-form videos, with the goal of improving recommendations. Testing it against existing algorithms, they found the new approach could substantially boost viewing time.

Why do current recommendation methods perform so poorly? The researchers cite a couple of factors.

Limited data on individuals. Platforms often see sparse feedback for any given user and video. With few interactions per person, existing methods find it hard to fully learn the individual viewer’s interests.

Missing the future. Viewers’ interests can change quickly based on news, memes that bubble up, holidays and seasons, and many other factors. It’s hard to predict what video topics might be popular in the future.

“Content is evolving really fast in those platforms,” Cao says. “Future content could be very different from the current content.”

That uncertainty presents a new challenge for short-form video platforms, Leng adds. “Short-form video is very different from a movie or a product on Amazon, where what platform is recommending is relatively stable.”

Predicting the Unpredictable

The researchers tackled both problems together. Their algorithm, they reasoned, might be able to base some predictions on seasonal trends, such as Christmas holidays, and new products/movies.

They could also better group individuals, using an approach the researchers call an adaptive acquisition tree (AAT).

The tree can split viewers into specific groups, based on their hometowns, age, and other demographic information. It can also group videos into different categories, such as dance, comedy, or education.

Based on such groupings, the algorithm recommends videos to users. Then, it gathers information on what they liked and didn’t like, based on their viewing time.

Using that feedback, the algorithm can split the users into smaller groups. As it learns their preferences over time, it can serve them more personalized and high-quality video recommendations.

“We simultaneously do this split and information acquisition, and so we are able to learn user preferences with greater granularity across different users,” Cao says.

Running simulations on the data, the researchers compared their AAT with existing recommendation algorithms. Their approach increased viewing time by 30% to 100%.

The AAT model need not be limited to video and social media platforms, the researchers say. The same process of algorithmic grouping, recommendations, and feedback could be applied to search engines, advertising, and even content moderation.

The two are already working to improve algorithms for online labor platforms, such as Mechanical Turk, an Amazon marketplace that lets companies recruit workers for specific tasks.

“We want to design algorithms that show workers the tasks that best match their expertise and experience,” Leng says.

“Ultimately, it’s the same goal: acquiring data to better understand the people on the platform — whether they’re workers on Mechanical Turk or consumers on a short-form video platform.”

“Adaptive Data Acquisition for Personalized Recommendations With Optimality Guarantees on Short-Form Video Platforms” is published in Management Science.

Story by Omar Gallaga