The key technology of content traffic management
One: Business Background
The volume maintenance strategy is a very important delivery strategy for video content. New hot video content needs to increase its own exposure resources to maximize the amount of playback, but the overall resources of each scene (home page, channel page, etc.) are limited and the daily exposure resources of each drawer are limited, so the exposure resources of each content Allocation has a race problem. In addition, different scenarios are independent of each other, and each scenario optimizes efficiency and experience according to its own goals, but traffic coordination between scenarios cannot be achieved by optimizing a single scenario.
Assigning exposure to content involves modeling issues about exposure and clicks, and predicting the future clicks of content. Content exposure, clicks, and playback constitute a complex nonlinear chaotic system, which depends not only on the quality of the content itself, but also on the content update time, update strategy, and user click habits. The traditional statistical forecasting model cannot explain the various disturbance factors of the external environment and the chaotic characteristics of the system, that is, it cannot describe the essence of the system from the mechanism. In response to this problem, we firstly analyzed the historical exposure click logs of new hot content, and established a new hot content exposure sensitive model using ordinary differential equations, namely the pv-click-ctr model (P2C model for short). Based on the P2C model, combined with the exposure resource constraints of each scene and drawer, a multi-objective optimization framework and algorithm under exposure resource constraints are given.
Two: Content Exposure Sensitivity Model
Normally, the click PV (click) increases with the increase of the exposure PV, that is, high exposure leads to high clicks. However, the number of content consumers is limited, and repeated exposure of a single content to the same consumer will not bring more clicks. This click "saturation" phenomenon can be observed from the content's historical exposure click logs. Inspired by this phenomenon, based on the historical data characteristics of content exposure PV and click PV, we established an ordinary differential equation (Ordinary Differential Equation, ODE) model that can describe the change trend of content clicks with exposure, that is, pv-click-ctr ( P2C) model, the overall structure is shown in Figure 3.
Due to the limitations of its own factors and the external environment, a content has a maximum or saturation value of image.png in terms of clicks. When given an exposure image.png, there exists a unique hit image.png and saturation image.png. For a hit image.png, the saturation image.png is defined as the ratio of the gap between the current hit and the saturation value to the saturation value, that is
For any content, as the pv increases, the click saturation decreases, and the ratio of the click increment brought by the unit pv (referred to as click increment) to the current click shows a downward trend. That is to say, there is a positive correlation between click increment and saturation, which can be expressed by the following formula:
Among them, image.png is the positive correlation coefficient. According to formula (2), the ordinary differential equation model in which click increases with pv can be obtained.
Integrating the two ends of equation (3) after separating the variables, we can get
Among them, image.png and image.png are the initial pv and click respectively.
For the parameters image.png and image.png in formula (4), the least square method can be used for fitting. Here, it is first necessary to filter and preprocess the historical pv and click data and parameters.
(a) Sample point filtering principle. Select the largest incremental subsequence in the daily historical pv and click data sequences respectively.
(b) Parameter preprocessing. Since the order of magnitude of the click saturation value image.png is usually large, and the order of magnitude of the correlation coefficient image.png is usually small, in order to avoid the phenomenon of "big numbers eat small numbers", the data conversion of these two parameters is performed separately, namely: image. png
(c) Sample point preprocessing. In order to prevent the least square method from falling into local optimum when fitting parameters, data transformation is performed on the historical samples (click value y, pv value x) respectively, namely: image.png, image.png. After the parameter fitting process, a single content pv-click function relationship can be obtained. Further, pv-click-ctr prediction can be performed. Here, the numerical solution method of finite difference can be used for prediction, and the data points can also be substituted into formula (4) for prediction.
Three: Guaranteed Model & Algorithm
Based on the P2C model established in the previous section, the task of this section is to give approximately optimal exposure for each content when the exposure resources of each scene and drawer are limited. The overall program flow is as follows:
First, based on the ordinary differential equation (ODE) model predicted by pv-click-ctr, for each content in the content pool, two parameters in the ODE are fitted by least squares: click saturation value image.png and click with pv intrinsic growth rate image.png. Thus, the pv-click function relation of each content is given.
Second, based on the given optimization objectives and constraints, a multi-objective nonlinear optimization model for pv allocation can be established. Before abstracting the business problem into a mathematical model, it is necessary to explain the symbols in the model, as follows.
The optimization objectives of the above model include two: maximization of multi-scene vv, and minimum variance of content pool content ctr. It should be noted that the minimum ctr variance here is a formal description of exposure fairness, which is used to balance "overexposure" and "underexposure". Constraints represent the exposure PV constraints of scenes, drawers, pits, and content, respectively. Since we use numerical methods to solve the objective function, the above optimization model cannot be solved by traditional gradient-based algorithms. The evolutionary algorithm provides a solution, and the genetic algorithm (GA) is chosen here to solve it. It should be noted that the fitness value function calculation in GA adopts the P2C model.
Four: Experimental results
We select a number of new and popular content, and respectively give the prediction effect of the P2C model and the offline effect of the guaranteed model. The evaluation metrics here are Root Mean Square Error (RMSE) and Absolute Percent Error (APE). The P2C model and the smoothed ctr method [1] are used to predict the click volume of new hot content respectively. From the table, it can be seen that the P2C model can effectively predict clicks, and outperforms the smooth ctr method in terms of RMSE.
In the online experiment part, we set up a bucket experiment. The benchmark bucket adopts a manual strategy to maintain volume; the experimental bucket adopts the strategy proposed in this paper. During the experiment, we pay attention to and compare the daily delivery effects of the benchmark bucket and the experimental bucket (CTR variance, the overall CTR of the strategy on the scene, etc.). The 30-day and 7-week volume maintenance effect data are given below. Compared with the manual strategy results, it is found that the volume maintenance strategy has different degrees of improvement in terms of CTR variance and the overall CTR of the scene. In particular, in terms of CTR variance, the effect of the quantity preservation strategy is very obvious, with an average relative improvement of +50%.
V: Summary & Outlook
The content maintenance strategy aims to solve the contradiction between limited traffic resources and excessive demand, and provide an optimized exposure suggestion for each content, so that the exposure resources of each scene can generate greater value. This paper proposes a resource-constrained model and algorithm framework for the multi-scenario VV maintenance requirements of new hot content. This framework consists of two stages: prediction and optimization. We conducted offline tests and bucketing experiments in some scenarios, and the experimental results reflect the feasibility and effectiveness of the strategy in this paper. In the future, there are many aspects that need to be continuously explored and improved, such as PUV volume maintenance, volume maintenance cold start issues, etc.
The volume maintenance strategy is a very important delivery strategy for video content. New hot video content needs to increase its own exposure resources to maximize the amount of playback, but the overall resources of each scene (home page, channel page, etc.) are limited and the daily exposure resources of each drawer are limited, so the exposure resources of each content Allocation has a race problem. In addition, different scenarios are independent of each other, and each scenario optimizes efficiency and experience according to its own goals, but traffic coordination between scenarios cannot be achieved by optimizing a single scenario.
Assigning exposure to content involves modeling issues about exposure and clicks, and predicting the future clicks of content. Content exposure, clicks, and playback constitute a complex nonlinear chaotic system, which depends not only on the quality of the content itself, but also on the content update time, update strategy, and user click habits. The traditional statistical forecasting model cannot explain the various disturbance factors of the external environment and the chaotic characteristics of the system, that is, it cannot describe the essence of the system from the mechanism. In response to this problem, we firstly analyzed the historical exposure click logs of new hot content, and established a new hot content exposure sensitive model using ordinary differential equations, namely the pv-click-ctr model (P2C model for short). Based on the P2C model, combined with the exposure resource constraints of each scene and drawer, a multi-objective optimization framework and algorithm under exposure resource constraints are given.
Two: Content Exposure Sensitivity Model
Normally, the click PV (click) increases with the increase of the exposure PV, that is, high exposure leads to high clicks. However, the number of content consumers is limited, and repeated exposure of a single content to the same consumer will not bring more clicks. This click "saturation" phenomenon can be observed from the content's historical exposure click logs. Inspired by this phenomenon, based on the historical data characteristics of content exposure PV and click PV, we established an ordinary differential equation (Ordinary Differential Equation, ODE) model that can describe the change trend of content clicks with exposure, that is, pv-click-ctr ( P2C) model, the overall structure is shown in Figure 3.
Due to the limitations of its own factors and the external environment, a content has a maximum or saturation value of image.png in terms of clicks. When given an exposure image.png, there exists a unique hit image.png and saturation image.png. For a hit image.png, the saturation image.png is defined as the ratio of the gap between the current hit and the saturation value to the saturation value, that is
For any content, as the pv increases, the click saturation decreases, and the ratio of the click increment brought by the unit pv (referred to as click increment) to the current click shows a downward trend. That is to say, there is a positive correlation between click increment and saturation, which can be expressed by the following formula:
Among them, image.png is the positive correlation coefficient. According to formula (2), the ordinary differential equation model in which click increases with pv can be obtained.
Integrating the two ends of equation (3) after separating the variables, we can get
Among them, image.png and image.png are the initial pv and click respectively.
For the parameters image.png and image.png in formula (4), the least square method can be used for fitting. Here, it is first necessary to filter and preprocess the historical pv and click data and parameters.
(a) Sample point filtering principle. Select the largest incremental subsequence in the daily historical pv and click data sequences respectively.
(b) Parameter preprocessing. Since the order of magnitude of the click saturation value image.png is usually large, and the order of magnitude of the correlation coefficient image.png is usually small, in order to avoid the phenomenon of "big numbers eat small numbers", the data conversion of these two parameters is performed separately, namely: image. png
(c) Sample point preprocessing. In order to prevent the least square method from falling into local optimum when fitting parameters, data transformation is performed on the historical samples (click value y, pv value x) respectively, namely: image.png, image.png. After the parameter fitting process, a single content pv-click function relationship can be obtained. Further, pv-click-ctr prediction can be performed. Here, the numerical solution method of finite difference can be used for prediction, and the data points can also be substituted into formula (4) for prediction.
Three: Guaranteed Model & Algorithm
Based on the P2C model established in the previous section, the task of this section is to give approximately optimal exposure for each content when the exposure resources of each scene and drawer are limited. The overall program flow is as follows:
First, based on the ordinary differential equation (ODE) model predicted by pv-click-ctr, for each content in the content pool, two parameters in the ODE are fitted by least squares: click saturation value image.png and click with pv intrinsic growth rate image.png. Thus, the pv-click function relation of each content is given.
Second, based on the given optimization objectives and constraints, a multi-objective nonlinear optimization model for pv allocation can be established. Before abstracting the business problem into a mathematical model, it is necessary to explain the symbols in the model, as follows.
The optimization objectives of the above model include two: maximization of multi-scene vv, and minimum variance of content pool content ctr. It should be noted that the minimum ctr variance here is a formal description of exposure fairness, which is used to balance "overexposure" and "underexposure". Constraints represent the exposure PV constraints of scenes, drawers, pits, and content, respectively. Since we use numerical methods to solve the objective function, the above optimization model cannot be solved by traditional gradient-based algorithms. The evolutionary algorithm provides a solution, and the genetic algorithm (GA) is chosen here to solve it. It should be noted that the fitness value function calculation in GA adopts the P2C model.
Four: Experimental results
We select a number of new and popular content, and respectively give the prediction effect of the P2C model and the offline effect of the guaranteed model. The evaluation metrics here are Root Mean Square Error (RMSE) and Absolute Percent Error (APE). The P2C model and the smoothed ctr method [1] are used to predict the click volume of new hot content respectively. From the table, it can be seen that the P2C model can effectively predict clicks, and outperforms the smooth ctr method in terms of RMSE.
In the online experiment part, we set up a bucket experiment. The benchmark bucket adopts a manual strategy to maintain volume; the experimental bucket adopts the strategy proposed in this paper. During the experiment, we pay attention to and compare the daily delivery effects of the benchmark bucket and the experimental bucket (CTR variance, the overall CTR of the strategy on the scene, etc.). The 30-day and 7-week volume maintenance effect data are given below. Compared with the manual strategy results, it is found that the volume maintenance strategy has different degrees of improvement in terms of CTR variance and the overall CTR of the scene. In particular, in terms of CTR variance, the effect of the quantity preservation strategy is very obvious, with an average relative improvement of +50%.
V: Summary & Outlook
The content maintenance strategy aims to solve the contradiction between limited traffic resources and excessive demand, and provide an optimized exposure suggestion for each content, so that the exposure resources of each scene can generate greater value. This paper proposes a resource-constrained model and algorithm framework for the multi-scenario VV maintenance requirements of new hot content. This framework consists of two stages: prediction and optimization. We conducted offline tests and bucketing experiments in some scenarios, and the experimental results reflect the feasibility and effectiveness of the strategy in this paper. In the future, there are many aspects that need to be continuously explored and improved, such as PUV volume maintenance, volume maintenance cold start issues, etc.
Related Articles
-
A detailed explanation of Hadoop core architecture HDFS
Knowledge Base Team
-
What Does IOT Mean
Knowledge Base Team
-
6 Optional Technologies for Data Storage
Knowledge Base Team
-
What Is Blockchain Technology
Knowledge Base Team
Explore More Special Offers
-
Short Message Service(SMS) & Mail Service
50,000 email package starts as low as USD 1.99, 120 short messages start at only USD 1.00