Video retention and adoption is an important distribution strategy for video platforms. Different from ordinary recommendation algorithms, video support is usually required to ensure certain exposure of the specified video material and minimize the negative impact on the effect of the recommendation system for the consideration of business and painting style. In this paper, we mainly introduce how the intelligent support system developed by IQiyi based on the improved Budget Pacing algorithm is to achieve the maximum magnitude and the most efficient support effect under the premise of ensuring the influence on the user consumption index is controllable.
01
background
For iQiyi and other online video service platforms, in addition to the ordinary recommendation algorithm, due to various considerations, it is often required to increase the exposure of designated new hot videos.
In specific application scenarios, iQiyi and instant short video streams, super fans, e-commerce videos and all kinds of operating material videos of TV series and variety shows all need to use support to ensure the maximum amount of playback.
Figure 1 maintenance video in the feed stream
Exposure resources of feed stream are limited, and there is competition among contents. Increasing exposure of supporting videos will inevitably squeeze exposure resources of videos produced by recommendation algorithm, which will have a negative impact on recommendation system.
On the other hand, there are also differences between supported videos. Some videos consume too fast and even complete the guaranteed amount target within the first few minutes, leaving the competition early and failing to reach more matched users, thus reducing consumption indicators. Other videos were not distributed effectively. This is a typical maintenance problem.
Therefore, we designed an eco-intelligent support system to directly support and maintain the quantity at the video level, and customize the support according to the quantity objectives of different videos and the real consumption situation. In addition, all videos that maintain the quantity compete with each other for constant display resources to reduce the impact on consumption. In addition, the Budget Pacing mechanism similar to the calculation of advertising field is used to allocate the guaranteed amount of display targets in accordance with the actual flow curve in the guaranteed amount cycle, and dynamically adjust the degree of guaranteed amount according to the specific consumption situation, so that the video is evenly distributed in the guaranteed amount cycle to complete the guaranteed amount target.
In addition, the eco-intelligent support system uses the way of promoting the position of video in the fine arrangement to enhance the weight of video, so as to ensure the certainty of the effect of supporting and preserving quantity.
Figure 2: Flow structure diagram of eco-intelligent support system
02
Problem analysis
In our opinion, the problem of video maintenance and budget control in the field of computing advertising has a similar place: the core problem of budget control is how to smoothly consume the budget of advertisers and help advertisers optimize the conversion effect. The core problem of video maintenance is how to smoothly consume the target exposure of video, and ensure that the impact on the overall consumption is reduced as much as possible. Therefore, the following two goals should be achieved for video maintenance:
(1) Uniform distribution
The video delivery speed is controlled by video target exposure, current exposure and exposure curve.
(2) Improve video consumption indicators
Ensure that the impact of supporting video on overall consumption is as low as possible.
Budget control schemes for calculating advertising are currently divided into two broad categories: Probabilistic Throttling and Bid modification control the frequency of advertising bidding through a probability to achieve the purpose of controlling budget consumption speed. Bid modification, on the other hand, modiates bidding to control the speed of budget consumption.
Figure 4 (a) : Probabilistic
throttling
Figure 4 (b) : the Bid modification
For the video cultivation and maintenance scenario, the target exposure amount of the video is the “budget”, so the “bidding” of each exposure is a constant value, that is, exposure times, and cannot be modified by Bid modification. Therefore, we used the Probabilistic Throttling as the basic framework to design the intelligent support system.
03
The framework design
First of all, the system should ensure that the distribution trend of the video is consistent with the overall market exposure trend. We divided the day into 288 time slices in 5 minutes and calculated the overall market exposure trend in a day based on historical data.
Figure 4(a) : Time slice – flow curve
Figure 4(b) : Time slice – cumulative flow curve
It should be noted that in the original Probabilistic throttling scheme for calculating advertising, advertising materials had separate display resources, so the advertising materials only competed with each other through probability, and the winner could occupy the display space reserved for advertising.
But the video quality fostering scenario, not only required to ensure the competition between the video opportunity to foster cultivate quality video also and ordinary video competition, so in addition to the use of probability, we introduced the fine line position power mechanism, to ensure the needs to foster the video, according to the target exposure, promotion video directly on the ranking position of fine line, It makes the supporting video more dominant in the competition with ordinary video, so as to achieve the purpose of expanding exposure.
Therefore, we need to calculate the relationship curve between the ranking position and exposure amount according to the historical data. It should be noted here that each time slice has an independent position-exposure curve due to the different exposure efficiency of different time slices (e.g. the exposure efficiency of evening peak is greater than that of early morning).
Figure 5: Position-exposure curve
The specific algorithm is as follows:
For any supported video, the whole day’s expected guaranteed exposure is denoted as, and a day is split into time slice set; for time slice, represents the cumulative expected exposure at the end of the time slice; then, according to the cumulative exposure curve described in FIG. 4 (b), it can be split into the cumulative exposure set of each time slice, where. At the same time for time slice, it also represents the cumulative real exposure at the end of the time slice.
For the next coming time slice, the expected exposure in this time slice, according to Figure 5, can meet the maximum position of exposure requirements, set as.
Similarly, we can also obtain a time slice that has just ended, and the real exposure in this time slice can also be obtained according to Figure 5. The maximum position of the corresponding exposure can be set as.
For the time slice, is the mean of the finishing position achieved by the supporting amount preserving video for its expected exposure, while for the time slice, is the mean of the actual finishing position achieved by the supporting amount preserving video. Therefore, in the time slice, the position we need to improve in each finishing result is:
σ
Guarantee σ indicates that the weight reduction is not performed for videos that are excessively distributed.
In addition, since the corresponding relationship between the precise placement position and the exposure amount in Figure 5 is discrete, the corresponding exposure amount (set as) will generally be greater than the requirement. Therefore, in order to prevent the over-growth of the supporting amount, the probability of the supporting amount video participating in the competition in the time slice is agreed as follows:
Finally, the recommendation engine can adjust the placement of any video in any time slice according to σ and, thus achieving the effect of maintaining quantity.
04
Effect of online
** Daily support amount: ** Daily support exposure of ecological intelligent support system reaches 100 million level.
** Exposure completion rate: ** The old version of the guaranteed exposure completion rate (actual exposure/exposure target >80%) is less than 5%, the ecological intelligent support system guaranteed completion degree is greatly improved, with the engraving end to reach 65-70%, baseline end to reach 50-60%, the guaranteed completion rate increases about 20 times.
** Impact on the system: ** realizes intelligent speed control for supporting and maintaining video distribution, reducing the influence of per capita playback time on the system from 3.5% to 2.5%, and the influence of per capita video display from 2% to 0.15%.
** Business Achievements: ** assisted in the distribution of high quality content such as “Biu Bi Sayo” and encyclopedia and unpacking. This ensures the exposure of high-quality videos and creators, which is of great significance to improve the platform stickiness of high-quality authors.
05
Summary and Prospect
The above are some of our recent work in the field of ecological support and conservation. The practice has proved that the ecological intelligent support system based on Budget Pacing can indeed be of great help to the task of video support and conservation. We will further optimize from the following perspectives:
(1) Order optimization of the support system and related support recall to further reduce the impact on overall consumption on the basis of ensuring the completion of support;
(2) The current support system is similar to the way of advertising bidding consumption, which cannot guarantee that the target exposure of the video can be completed. Therefore, we should consider increasing the guarantee delivery mechanism to ensure the amount of video exposure.
References:
1. Agarwal D, Ghosh S, Wei K, et al. Budget pacing for targeted online advertisements at linkedin[C]//Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 2014: 1613-1619.
2.Xu J, Lee K, Li W, et al. Smart pacing for effective online ad campaign optimization[C]//Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2015: 2217-2226.
Did you see the heart?
Iqiyi Business Division has multiple positions waiting for talents
Including technology, product, operations, commercialization and growth
Interested students please send your resume to
Talk directly to the boss
Ps. Background reply “Recruiting 01”
You can get specific job postings.
More job information in IQiyi
Click “Read the article” to get it!
Maybe you’d like to see more
How to improve link target consistency? Iqiyi short video recommendation coarse layout model optimization process
The evolution of multi-interest recall technology in iQiyi short video recommendation Technology
Script evaluation, intelligent casting, online film review, interactive master IMF, etc., iQiyi intelligent production is becoming a “new productivity” to improve the quality of content, and will also become a key step to promote the industrialization of film and television. # Intelligent production # Intelligent casting # Art Club # Online Film Review # IQiyi World Congress # Film and Television Industrialization # Black technology