A few days ago TikTok made public some information on how its recommendation algorithm works, that is the system that proposes new videos to users as they scroll the app screen in the “For you” section, where public videos are shown. TikTok explained how the algorithm works with a post on his company blog and in person to some American journalists, who were also able to analyze and test the code.
Recommendation algorithms are very important for a social network, because they determine the quality of engagement, i.e. the time and activity that each user dedicates to the app. The TikTok algorithm, then, is famous in the world of technology for the accuracy of the recommendations it can give. Eugene Wei, an American entrepreneur and analyst, wrote on his personal blog at the beginning of August: “Before TikTok, I would have said that YouTube had the best algorithm in the video industry, but compared to that of TikTok, the algorithm of YouTube seems rudimentary”.
TikTok's algorithm, according to TikTok, works like this. The first time a person accesses the social network, the system asks him to select some categories of interest, for example “animals” and “travel”, on the basis of which to start the recommendations. At that point, the algorithm shows the user eight initial videos, and then another eight depending on how the user behaves with the first ones. To understand if a video is liked or not, the algorithm uses a series of criteria, including interactions (likes, comments, followers), clicks on hashtags, duration of viewing, the use of certain filters. The algorithm also takes into account other data, such as the user's language, his country of origin, the type of telephone used.
– Read also: TikTok CEO Kevin Mayer resigned
Based on the information gathered, the algorithm begins to show the user videos similar to the ones they have already liked. To understand if the new videos correspond to the user's tastes, use parameters such as the written content of the posts, the hashtags, the sounds and the songs.
As Axios wrote, when it began to get an idea of what the user likes, the algorithm inserts it into one or more clusters, that is, into categories of interest. A user can be identified as belonging to a cluster of lovers of basketball, opera or horse riding. At the same time, all videos are categorized within clusters, and according to Axios the algorithm creates associations based on the “proximity” between user clusters and content clusters. This means that users classified as lovers of horseback riding will be able to see both videos of the “horse riding” cluster and videos that other users in their own cluster have enjoyed. TikTok says it uses artificial intelligence (machine learning) to make this association between video and user but doesn't explain precisely how. How TikTok's artificial intelligence works is its main trade secret, and the company doesn't want to reveal it.
TikTok also has systems to understand if the user does not like a video and to avoid boredom: the algorithm tries not to show two videos in a row with the same song or created by the same person, and from time to time tries to submit to the user videos that do not exactly correspond to his tastes, to make him discover new things. Despite this, TikTok recognizes that its algorithm can lead to the creation of “filter bubbles”, that is, a recommendation system that continuously shows videos that are always homogeneous in terms of themes and contents and always adhering to user preferences.
TikTok announced in July that it would make some features of its algorithm public, to demonstrate transparency and accountability. Michael Beckerman, vice president of TikTok with responsibility for public relations in the United States, told Axios that TikTok is “a company born two years ago that has to deal with the expectations of a ten-year-old company”, which it must support ” a lot of skepticism towards platforms, about how they moderate content and how their algorithms work “.
– Read also: Microsoft wants to remove all the carbon dioxide it has produced since 1975
TikTok's algorithm is also the reason selling the US side of the company is complicated. In early August, US President Donald Trump signed an executive order that TikTok must sell all of its local business to a US company to avoid being banned to users in the US. Trump's motivation is that TikTok is owned by ByteDance, a Chinese company based in Beijing, and therefore constitutes a national security threat. Trump gave TikTok until mid-November (initially the deadline was September 20, but was then extended) to find a buyer, but the negotiations, first with Microsoft and then with Oracle, were complicated by the fact that ByteDance he does not want to give up the formula of his algorithm. Furthermore, in early September, the Chinese government approved new restrictions that prevent local companies from exporting certain technologies, including those relating to artificial intelligence.
The only agreement hypothesized so far, anticipated by the Wall Street Journal but not yet officially announced, provides for Oracle to become TikTok's technology partner without ByteDance having to sell its American business. The American and Chinese governments must give their approval.
