T-Tok short video recommendation mechanism

12-01 2022

First of all, since we chose short video delivery, we need to understand the recommendation system of T-Tok short video, which mainly includes three parts: user portrait, content portrait, matching between user and content, and video recommendation ordering.

user portrait


After the account is registered, the system will collect data sets based on the user's basic attributes, such as gender, age, interests and hobbies, and preliminarily define relevant labels for the account. Many people don't understand the difference between age options when registering an account, if you have registered an account under the age of 18+ and an account under the age of 18, you will find that the account under the age of 18 has many functional limitations. As a delivery account, we need short video tags, not user viewing tags, and only short video tags will accurately distribute our videos to people marked with such user tags.

content portrait


The system will analyze the characteristics according to the hierarchical classification, keywords, solid words, etc. of the videos we post, and label all kinds of content with relevant labels, according to the official reviewers of D sound, manual marking their monthly assessment should be 90% correct, that is, if you happen to meet that 10%, it will become metaphysics. When reviewing content, you will look at the quality of video content, divided into useless, weak useful, strong useful, for example, you sent a food tutorial, maybe you have text on it, but you don't say it, he will think that you are useless, there is text but not detailed enough, that is, weak and useful, each step tells how many grams of seasoning in detail, will think that this content is strongly useful.

User and content matching:

With user tags and content tag thickness, the system will match the user's favorite content in the content pool according to the user portrait and content portrait and then display it.


The system has to face hundreds of millions of users and content, while also considering that the user's preferences will continue to change, in order to make the selected content closer to the user's desire and more in line with the user's liking, the system will sort the content. The system will sort according to your completion rate, likes, comments, retweets, and followers, and videos of similar products will be recommended according to ranking. This is important!

Short video weight 100% = user portrait 5% + content portrait 25% + user and content matching 10% + sorting 60%

T-Tok for short videos, like China, has a double review mechanism, divided into machine review and manual review.

machine audit


Detect and identify video images and keywords through artificial intelligence technology. Review whether the title violates the rules, and if there is a suspected machine interception, prompt manual attention. Extract the pictures in the video, match the existing works in the large database to reduce the weight, and make low-traffic recommendations or downgrade recommendations if the content is repeated.

manual review:

Selected suspected illegal works for re-review for machine review. Manually conduct detailed review one by one, and if the violation is determined, punishments such as deleting videos, reducing rights, and banning accounts will be imposed according to the offending account.

focuses on reviewing video titles, cover screenshots, and video keyframes. Why do video works never pass the review, all video works published on the site have to go through a double review process. If the video work is not reviewed by the machine, it will be intercepted by the machine, enter the manual review, and then change due to manpower problems and the amount of submissions , review delays and audit backlogs occur. When a video work is under review, if no violation notice is received, it does not mean that the stream is restricted or the video content is problematic, and it can be reviewed for review. After the video work passes the review normally, it will not affect the subsequent data traffic, and the key is the content quality.

The platform recommendation mechanism is that after the video has been reviewed, it will be recommended by the system, the recommendation is decentralized, traffic pool planning, artificial intelligence distribution system, the platform will provide a traffic pool for each work, whether you are large or small, as long as the video work has been reviewed, it will be recommended, can the video explode in the later stage,

depends on the data performance of this traffic pool (number of plays, likes, comments, retweets).

recommendations are divided into three stages: Basic recommendation: newly released video works, traffic distribution is mainly based on proximity and attention, and then intelligently distributed to a small number of users with user tags and content tags, depending on the feedback of users, if the proportion of users likes is larger, it is possible to enter the second round of superimposed recommendations.

Overlay recommendation


newly released videos will go through the previous round of intelligent distribution, according to the weight of the account distribution of about 100-500 views, such as the number of retweets up to 10 (for example), the algorithm will judge the popular content, automatically weight the content, superimpose recommendations to you 1000, retweets up to 100 (for example), the algorithm continues to superimpose recommendations

to 10000, the number of forwards reaches 1000 (for example), and then the recommendation is superimposed to 10w, and then the push is accumulated. When the forwarding volume reaches a certain magnitude, it is based on the mechanism of combining big data algorithms and manual operations.

time effect:

How does the T-Tok platform judge the quality of a video? When a new video work is released, the first round of basic recommendations will assign you 300-500 online users according to the size of your account, and judge your video quality according to their data feedback. What data feedback video quality? Four core indicators: like rate, comment rate, completion rate, and retweet rate.

The recommended mechanism for T-Tok is: Accounts that log in more than 10-20 times a month and post videos 10 times are users that the platform must find a way to retain; Login 5-10 times a month, release 0-10 video accounts, for the core conversion of users, the long-term development of the platform needs such users to maintain, the main group of people who receive advertisements; Accounts that log in 0-20 times a month and publish videos 0-5 times are content providers and have unstable traffic; Log in 0-10 times a month, do not send short videos, the platform will mainly convey the value of the product, and strive to cultivate user habits, which is why when you register a new account, you rarely brush ads.

Based on this, in order to improve the account level, each account must log in to the platform at least 20 times a month and publish 10 pieces of content. Based on this, it can be calculated that when the account is stable, a video is played 500, the more videos are sent, the wider the traffic and the higher the conversion rate.

       This article is published by the columnist, and the copyright belongs to the original author. The article is the author's personal opinion, does not represent the position of the scholar, please contact the original author for reprinting. If you have any questions, please contact sales@shusheng.com