TikTok told how his algorithm recommendations

Service short videos TikTok revealed information about how the algorithm for its recommendations.

In a post published on the company's blog also provides tips for personalization tape, so that users can avoid irrelevant recommendations.

  • The signals used in selecting the recommendations

TikTok recommendations algorithm is based on the involvement of factors. The way users interact with the application, including the publication of comments and subscribe to the accounts, affects the selection of recommendations.

For example, if a user is signed only for accounts with videos of animals and "laykaet" / comments on these videos, the TikTok will show him more animals.

In this case, the signals of involvement - is only one part of the equation. TikTok also takes into account information about the video (signature sounds and hashtags), the user's device and settings in their account.

Meanwhile, the last two factors receive less weight compared with other signals, because users are not actively express them as preferences.

Also it has a value of the degree of involvement. If the user closes the video without having watched it till the end, then TikTok is a fairly strong indicator of interest.

  • problem areas

Algorithm recommendations TikTok is appreciated by users and allows the service as long as possible to keep them in the application. However, in this system has its weak points.

So, when optimizing for the purpose of personalization and relevance there is a risk of formation of "bubble filter" when a user begins to see more and more monotonous video. And developers are working on this issue.

TikTok also tries not to show malicious content. Those videos which depict a medical procedure, or consumption of "regulated" products, such as alcohol, are not allowed in the recommendations, as they may be perceived as "shocking" if they show a wider audience - in particular, children.

For more information about the algorithm TikTok recommendations can be found in the company's blog .