Written by Josh Yeh
HONG KONG (Reuters) – The content recommendation algorithm behind online short video platform TikTok has been cut short after the U.S. ordered Chinese owner ByteDance to sell the app's U.S. assets or face a nationwide ban. is once again attracting attention.
Here's how it works and why it's garnering more discussion than the technology used by rivals like Meta's Instagram, Google's YouTube, and Snapchat.
algorithm
These algorithms are considered core to ByteDance's entire operation, and ByteDance would rather shut down the app than sell it, Reuters reported, citing sources.
China revised its export laws in 2020, giving it approval rights to export algorithms and source code, further complicating any efforts to sell apps.
Academics and former company staff say TikTok's global success is not just due to its algorithm, but also to the way it works with the short video format.
It's also an app
Before TikTok, many believed that technology that connects users' social connections was the secret to a successful social media app, given the popularity of Meta's Facebook and Instagram.
But TikTok showed that algorithms based on understanding users' interests can become more powerful. TikTok executives, including CEO Shou Zi Chew, said that rather than building the algorithm on a “social graph” like Meta, the algorithm is based on “interest signals.”
Katarina Goanta, an associate professor at Utrecht University, said competitors have similar interest-based algorithms, but TikTok can significantly increase the effectiveness of its algorithms with the short video format.
“Their recommender system is very common. But what really sets TikTok apart as an app is the design and content,” she said.
The short video format allows TikTok's algorithm to be more dynamic and even track changes in user preferences and interests over time, including what users prefer at certain times of the day. can also be tracked in detail.
Rapid data collection
Additionally, Jason Huang, TikTok's former head of gaming, said the short video format allows TikTok to learn user preferences faster.
“Because it's a bite-sized format and short videos, we can collect data about user preferences much faster than on YouTube, where the average video is probably less than 10 minutes long,” he said. On average, data about users is collected every 10 minutes, but every few seconds. ”
TikTok was also positioned from the beginning as an app built for mobile devices, giving it an advantage over rival platforms that had to adapt their interfaces from computer screens.
TikTok's early entry into the short video market also gave it a significant first-mover advantage. Instagram didn't launch Reels until 2020, and YouTube launched Shorts in 2021, but both lag behind TikTok in years of data and product development experience.
allow exploration
TikTok also regularly recommends content that is out of user interest, and company executives have repeatedly stated that this is essential to the TikTok user experience.
A study published last month by U.S. and German researchers examined data from 347 TikTok users and five automated bots and found that TikTok's algorithm “targets users' interests by 30% to 50% of recommended videos.” It turned out that it was being used.
“This finding suggests that TikTok's algorithm uses a large number of ,” the researchers wrote. In a paper titled “TikTok and the Art of Personalization.”
Mobilize users into groups
Ari Reitman, a professor at Carnegie Mellon University, said another effective tactic used by TikTok is to encourage users to form groups publicly through hashtags.
By encouraging users to form public groups, TikTok can more effectively learn about users' behaviors, interests, alignments and ideologies, he said.
Reitman said that if TikTok were to be banned in the United States, American tech giants certainly have the ability to recreate TikTok in their products, but not to recreate the user culture enabled by TikTok. may be a bigger challenge, he said.
China's advantage
TikTok's recommendation algorithm is also largely adapted from its Chinese sister app Douyin, which was released in 2016. Although ByteDance often emphasizes that TikTok and Douyin are separate apps, a source with direct knowledge of the matter said the algorithms for the two remain similar. .
Second, Douyin's AI is greatly enhanced by the company's ability to take advantage of low labor costs in China, where it has hired many content annotators to painstakingly tag all content and users on the platform. Ta.
“Around 2018-2019, Douyin worked on tagging all users. So they started tagging all video clips manually. And they started tagging users based on the videos they watched. “We started adding tags,” said Yikai Li, a manager at advertising agency Nativex. He is also a former director of ByteDance. “They then applied this tactic to TikTok as well.”
Hiring annotators to tag data is now common and an important practice for AI companies, but ByteDance was an early adopter of this strategy.
“It's a lot of work to organize these tags. It's very time-consuming,” he said. “So Chinese companies have an advantage in this respect. They can afford to hire far more people. The cost is lower. It's cheaper than North American companies.”
(Reporting by Josh Ye; Additional reporting by Krystal Hu in San Francisco; Editing by Brenda Goh and Jacqueline Wong)
