Short video products have been booming in the last two years. Baidu’s short video brand – Good Video has an average usage time of 70 minutes, and the playback volume of short video has exceeded 3 billion.

The full text is 2433 words, and the expected reading time is 8 minutes.

As of June 2020, there were more than 900 million online audiovisual users in China, and 820 million short video users, who spent nearly two hours a day watching short videos on average, according to the Report on the Development of Online Audiovisual in China 2020. At the same time, Baidu’s short video brand – Qiaowei, the average usage time is 70 minutes, short video has been played more than 3 billion, the massive content of qiaowei video is also a very important supplier of search.

For recommendation products, content is the cornerstone. So what is the content system of a good video? First, there are millions of content creators constantly producing content in the current system of good-looking video. This includes 100,000 + creators of premium content; At the same time, we also continue to create and export homemade dramas, such as “Getting Better”, “Your Life is Better” and so on.

1. What problems does the recommendation system solve?

There are three roles in the recommendation platform or information flow products: user, creator and advertiser, and the experience problems of user and creator to be solved by the recommendation system:

1) Client: optimize the experience to meet the highly personalized consumer needs of thousands of users. In Baidu, we have rich user portraits and understanding. Meanwhile, in App, we can perceive the use environment and further capture users’ interest based on signals such as user interest expression.

2) Creators: let high-quality content get more distribution, retain creators, and realize the survival of the fittest. In addition, with macro-control at the ecological level, high-quality creators can be guaranteed to input more content continuously on this platform to the greatest extent.

Second, the overall picture of the video recommendation system

As shown in the figure below, after the creator uploads the content to the platform, it will enter a unified forward arrangement process; When the user opens the APP request, the system will recall all relevant content (display, implicit, etc.), with the purpose of “Make Everything Happens”, and then select the user’s favorite content through the three funnel of coarse arrangement, fine arrangement and fusion, and finally present the final content through mechanism regulation.

3. User interest characterization in the interactive form of video products

After the user opens the APP, the recommended video will automatically play, just like the TV at home. If they like it, they will continue to stay on the page, and if they don’t like it, they will cross to the next line. This is a new interaction pattern in the APP, but also brings challenges to recommendation: how to depict the user’s portrait? How to determine whether users like it or not?

We can draw lessons from the traditional point-and-select interaction mode, that is, users first decide whether they are interested, and then click on the video if they are interested. Satisfied will continue to watch and play; If not, the user may opt out. Under the new interaction form of auto-play, we define the behavior of “hurting” when we don’t like it and “satisfying” when we watch a video for a long time or watch it completely. The recommendation system will use these three signals (harm, duration, end of play) to describe the user’s interest.

In addition to depicting users’ interests through playing behaviors, attention, likes and favorites are all signals of users expressing their preferences. When designing a recommendation system, understand the product first, then design it. For the interaction behavior of short video recommendation system, there are the following signals, which are classified into the four quadrants below and introduced into the recommendation system.

Four, the application of multi – objective sorting

1) Multi-objective modeling

Multi-objective modeling developed from the basic shared-bottom DNN modeling to MMOE, and finally adopted population-based MMOE to model N recommendation system goals. When it was really realized, experts with low, medium and high activity levels were trained separately to make joint decisions to prevent high-activity population samples from leading the whole model. To ensure the accuracy of the system.

2) Multi-objective fusion sequencing

After N multiple targets are estimated by the model, these multiple targets need to be fused together. The common method is simple polynomial fusion, which is relatively basic and simple, effective and easy to introduce experience values. However, the disadvantages are also obvious, requiring frequent adjustments and not adaptive. Now the deepES method is developed, that is, personalized fusion of scenes. Each time, multiple groups of parameters are obtained by disturbing the internal parameters of the model, and then the optimal parameters are selected according to the Reward designed. Its features introduce various states such as device type, state and refresh rhythm, and the logical diagram is as follows:

5. Recommendation system for long-term benefit targets

The recommendation goals discussed above are all sorted according to the consumption of the videos currently being recommended, but we move the perspective to a longer time series. The user’s consumption is divided into past, present and future stages. As for the content of the past, it can be regarded as the sample of model training to depict the characteristics of user interest. The current recommendation goal is: the content that may be consumed in the future is essentially the “continuation” of current interest; Current interest can be regarded as the “stimulation” of future interest, and the Value of future consumption content can be attributed to the current video, that is, the long-term Value of the future.

Through the way below design LTV, for example: users see a talk show video today value defined as where V0, assuming that the future users will also see such video V2,.vn, then we can be the V2,.vn length and consumption value due to the current where V0, specific design is divided into 2 steps: 1, find relevant content, design attenuation factor; 2. Use model to fit LTV.

6. Multi-objective discussion of video recommendation system

1) Are multiple goals of equal value?

For example, does a user like a video as much as they like it, or does the first like of a video have the same value as the 10th?

2) Is the current goal group-optimal?

Is there synergistic value when the video is played? If the current video is not the most valuable to the user, but the most valuable to the whole, can I recommend it?

3) Can you design a goal/system/model to capture retention?

Multi-purpose design, as well as future revenue design, is all about retention. Can you build retention models directly?

All of the above are worthy of students’ in-depth thinking, we can communicate and discuss together if we have the opportunity.

Recommended reading:

Baidu credibility certification in Taiwan architecture analysis

Graph database in Baidu Chinese application

Learn Flutter development in 5 minutes

———- END ———-

Baidu said Geek

Baidu official technology public number online!

Technical dry goods, industry information, online salon, industry conference

Recruitment information · Internal push information · technical books · Baidu surrounding

Welcome to your attention