Video broadcasting method and device, equipment and storage medium

Abstract

The embodiment of the invention provides a video broadcasting method, device and equipment and a storage medium, which can be applied to various scenes such as artificial intelligence, cloud technology, smart traffic and auxiliary driving, and the method comprises the following steps: obtaining a plurality of different candidate video sets according to a plurality of pieces of video information to be broadcasted on a video platform; from at least one preset recommendation feature dimension, feature extraction is carried out on video information contained in the multiple candidate video sets, and recommendation feature sets corresponding to the multiple candidate video sets are obtained; according to the embodiment of the invention, on the basis of the preset weight value of each recommendation feature, the respective estimated retention rate of the plurality of recommendation feature sets is obtained, and the recommended video set is selected from the plurality of candidate video sets on the basis of the obtained estimated retention rates, so that the retention rate of the video platform is improved by broadcasting the recommended video set.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a video scheduling method, device, equipment and storage medium. Background Art

[0002] With the continuous progress and innovation of digital media entertainment, more and more people choose to watch online video content as their first choice for daily entertainment. In order to better meet the viewing needs of people, each video platform will adjust the video content played in a specific time period, that is, adjust the scheduled video content to attract more people to continue to visit the video platform and improve the retention rate of the video platform.

[0003] Under the related technology, it is usually based on periodic time factors (such as seasons, holidays, etc.). In different time periods, according to the historically scheduled video sets and their corresponding retention rate results, matching video sets are selected for video scheduling to improve the retention rate of the video platform.

[0004] However, the video sets scheduled in this way only consider the influence of general time factors and rely too much on specific time periods and specific video sets in historical records. As a result, the video sets used for scheduling lack flexibility and foresight to adapt to future changes, and thus fail to bring the expected retention rate to the video platform. There is even a problem of being unable to schedule video sets similar to those in the past, which lacks business guidance and practicability.

[0005] Furthermore, this method is prone to repeated scheduling and closed scheduling, which is not conducive to expanding the scheduled video content, resulting in insufficient novelty and diversity in the scheduled video sets, unable to meet the video viewing needs of the target audience, and resulting in a lower retention rate of the video platform.

[0006] For example, suppose that in the past year, video platform X used video set A for scheduling during the summer vacation and video set B for scheduling during holidays, which brought a retention rate of up to 60% for video platform X. Therefore, video platform X plans to give priority to video sets similar to video set A during the summer vacation and video sets similar to video set B during holidays in the new year, hoping to produce a retention rate close to or even exceeding 60%.

[0007] However, if video platform X relies too much on historical successful experience, it may overfit the historical data of last year and ignore the actual situation of this year, resulting in similar video sets failing to achieve the expected retention rate results. Even in the new year, video platform X may not have videos to be scheduled that are similar to the video content contained in video sets A or B, making it impossible to use similar video sets for video scheduling based on existing technologies.

[0008] Furthermore, video platform X uses a video set similar to that of the past year to schedule videos, which may cause the subject to repeatedly see similar video content, reducing the diversity of video content on the video platform and the subject's video viewing experience, resulting in a decrease in the retention rate of the video platform. Summary of the invention

[0009] The embodiments of the present application provide a video scheduling method and device for improving the retention rate of a video platform.

[0010] In a first aspect, an embodiment of the present application provides a video scheduling method, the method comprising:

[0011] Based on the multiple video information to be broadcasted, a plurality of different candidate video sets are obtained; wherein each candidate video set is used to indicate: at least one of the video information to be broadcasted in a preset time period;

[0012] Based on at least one preset recommendation feature dimension, feature extraction is performed on at least one video information contained in each of the multiple candidate video sets to obtain recommendation feature sets corresponding to each of the multiple candidate video sets; wherein each recommendation feature dimension represents: a video attribute that causes a deviation in the estimated retention rate of the video platform;

[0013] Based on the preset weight value of each recommended feature, the estimated retention rates corresponding to each of the plurality of recommended feature sets are obtained respectively; the preset weight value is determined during the model training process based on the preset weight constraint condition; the weight constraint condition indicates that the degree of influence of each recommended feature on the estimated retention rate meets the preset scheduling scenario condition;

[0014] Based on the obtained estimated retention rates, a recommended video set is selected from the multiple candidate video sets, and the recommended video set is scheduled for broadcast.

[0015] In a second aspect, an embodiment of the present application provides a video scheduling device, including:

[0016] An acquisition unit, configured to obtain a plurality of different candidate video sets based on a plurality of video information to be broadcasted; wherein each candidate video set is used to indicate: at least one of the video information to be broadcasted in a preset time period;

[0017] The extraction unit is used to extract features of at least one video information contained in each of the plurality of candidate video sets based on at least one preset recommendation feature dimension, so as to obtain recommendation feature sets corresponding to each of the plurality of candidate video sets; wherein each recommendation feature dimension represents: a video attribute that causes a deviation in the estimated retention rate of the video platform;

[0018] An estimation unit, for obtaining the estimated retention rates corresponding to each of the plurality of recommended feature sets based on a preset weight value of each recommended feature; the preset weight value is determined based on a preset weight constraint during model training; the weight constraint indicates that the degree of influence of each recommended feature on the estimated retention rate complies with a preset scheduling scenario condition;

[0019] The broadcast scheduling unit is used to select a recommended video set from the multiple candidate video sets based on the obtained estimated retention rates, and schedule the recommended video set for broadcast.

[0020] Optionally, the extraction unit is specifically used to:

[0021] For each candidate video set in the multiple candidate video sets, perform the following operations respectively:

[0022] Based on at least one preset recommendation feature dimension, feature extraction is performed on each video information contained in a candidate video set to obtain corresponding independent recommendation features, each independent recommendation feature represents: an independent impact of a corresponding video information on the estimated retention rate;

[0023] Perform feature cross extraction on every two independent recommendation features to obtain a corresponding cross recommendation feature set, where each cross recommendation feature represents: the superimposed impact of the corresponding two video information on the estimated retention rate;

[0024] Based on the obtained independent recommendation features and the cross-recommendation feature set, a recommendation feature set corresponding to the candidate video set is obtained.

[0025] Optionally, the extraction unit is specifically used to:

[0026] For each piece of video information, perform the following operations respectively:

[0027] For a piece of video information, based on a plurality of preset recommendation feature dimensions, feature extraction is performed on the video information to obtain a plurality of recommendation sub-features corresponding to the video information, where the plurality of recommendation sub-features correspond one to one to the plurality of recommendation feature dimensions;

[0028] Perform feature fusion processing on the multiple recommendation sub-features to obtain an independent recommendation feature corresponding to the one video information.

[0029] Optionally, the estimation unit is specifically used for:

[0030] For the multiple recommended feature sets obtained, perform the following operations respectively:

[0031] For a recommendation feature set, based on a first preset weight corresponding to each independent recommendation feature and a second preset weight corresponding to each cross recommendation feature, weighted fusion is performed on the recommendation feature set to obtain a corresponding recommendation fusion feature;

[0032] Feature mapping is performed on the obtained recommended fusion features to obtain an estimated retention rate corresponding to the recommended feature set.

[0033] Optionally, the scheduling unit is specifically used to:

[0034] Based on a preset retention threshold, selecting a target estimated retention rate from the multiple estimated retention rates, and using a target candidate video set associated with the target estimated retention rate as the recommended video set;

[0035] Based on a preset video quantity condition, a target video set subset is screened out from the multiple candidate video sets, and based on a preset retention rate threshold, the recommended video set is selected from the target video set subset; the video quantity condition indicates that each target video set contains a preset amount of video information;

[0036] Based on preset video attribute conditions, a target video set subset is screened out from the multiple candidate video sets, and based on a preset retention rate threshold, the recommended video set is selected from the target video set subset; the video attribute conditions indicate that each target video set contains video information of preset video attributes.

[0037] Optionally, the weight constraint condition includes a feature constraint condition and a dimension constraint condition;

[0038] The feature constraint condition indicates that: each recommended feature has a positive impact on the estimated retention rate;

[0039] The dimension constraint condition indicates that each recommended sub-feature under each recommended feature dimension has a different degree of influence on the estimated retention rate.

[0040] Optionally, the feature constraint condition includes a single constraint condition and a cross constraint condition;

[0041] The single constraint condition indicates that each independent recommendation feature has a positive impact on the estimated retention rate, and the independent recommendation feature indicates the individual impact of a video information on the estimated retention rate;

[0042] The cross constraint condition indicates that each cross recommendation feature has a positive impact on the estimated retention rate, and the cross recommendation feature indicates the superimposed impact of two video information on the estimated retention rate.

[0043] Optionally, the estimation unit is specifically used for:

[0044] Through the trained retention estimation model, perform the following operations on multiple recommended feature sets:

[0045] For a recommended feature set, based on the preset weight value, weighted fusion is performed on the recommended feature set to obtain corresponding recommended fusion features;

[0046] Performing feature mapping on the recommended fusion features to obtain an estimated retention rate of the recommended feature set;

[0047] The retention prediction model is obtained by training a plurality of sample data, and each sample data includes: a sample video set, and a sample retention rate corresponding to the sample video set.

[0048] Optionally, the video scheduling device further includes a training unit, which is used to:

[0049] The retention prediction model to be trained is iteratively trained based on the sample data to obtain the trained retention prediction model; wherein each iterative training process includes the following operations:

[0050] Based on the selected sample video set, obtain the corresponding estimated retention rate;

[0051] Obtaining a target loss value based on the sample retention rate corresponding to the sample video set and the estimated retention rate;

[0052] Based on the weight constraint condition, the target loss value is used to adjust the weight parameters of the retention estimation model.

[0053] In a third aspect, an embodiment of the present application provides a computer device, comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes any one of the video scheduling methods in the first aspect.

[0054] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which includes a computer program. When the computer program is run on a computer device, the computer program is used to enable the computer device to execute any one of the video scheduling methods in the first aspect above.

[0055] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program stored in a computer-readable storage medium; when a processor of a computer device reads the computer program from the computer-readable storage medium, the processor executes the computer program, so that the computer device executes any one of the video scheduling methods in the above-mentioned first aspect.

[0056] The beneficial effects of this application are as follows:

[0057] In the embodiment of the present application, based on the multiple video information to be scheduled for broadcast on the video platform, multiple different candidate video sets are obtained to indicate one or more video information to be broadcast on the video platform in a preset time period. In this way, compared with the use of a fixed historical video set, the present application expands the video set used for video broadcasting, avoids repeated broadcasting and closed broadcasting caused by over-reliance on historical video sets for video broadcasting, improves the novelty and diversity of the video content broadcast by the video platform, meets the video viewing needs of the changing objects, and thus improves the retention rate of the video platform.

[0058] Furthermore, the present application extracts the recommended features corresponding to each candidate video set from at least one preset recommended feature dimension for each candidate video set obtained, thereby obtaining a corresponding recommended feature set. Each recommended feature dimension represents a video attribute that causes a deviation in the estimated retention rate of the video platform, thereby more comprehensively considering the impact of video information of different video attributes on the retention rate, so that the estimated retention rate generated by using the candidate video set for video scheduling is closer to the actual video scheduling scenario, reducing the estimated error of the retention rate, and improving the accuracy of the estimated retention rate, thereby improving the accuracy of the subsequent selection of recommended video sets based on the estimated retention rate, so that the recommended video set is used for video scheduling, which can more accurately produce the expected retention rate results for the video platform. In addition, the present application introduces weight constraints in the training process of the retention estimation model to limit the range of values ​​of the preset weight values ​​of each recommended feature, ensuring that the weight value of each recommended feature can more truly reflect the degree of influence of the corresponding video attribute on the retention rate in the actual scheduling scenario, thereby further improving the estimated accuracy of the retention rate. Moreover, through the preset weight constraints, the trained retention estimation model can be more adapted to the actual video scheduling scenario, thereby improving the generalization ability and robustness of the model.

[0059] Furthermore, the present application selects a recommended video set from a large number of candidate video sets by obtaining the estimated retention rate of each candidate video set, so as to schedule the video content corresponding to the recommended video set in a specific time period, thereby attracting more objects to continue to visit the video platform and improve the retention rate of the video platform. In addition, the present application can flexibly select the desired recommended video set based on the estimated retention rate of each candidate video set according to the business objectives of different video scheduling, so that the video platform can better adapt to the ever-changing user needs and market trends.

[0060] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0062] Figure 1 A schematic diagram of an application scenario provided in an embodiment of the present application;

[0063] Figure 2 A schematic diagram of a process for a video platform to obtain a recommended video set provided in an embodiment of the present application;

[0064] Figure 3 A schematic diagram of the overall process of training the retention estimation model provided in the embodiment of the present application;

[0065] Figure 4 A schematic diagram of a flow chart of iterative training of a retention estimation model provided in an embodiment of the present application;

[0066] Figure 5 A schematic diagram of the process of iterative training of another retention estimation model provided in an embodiment of the present application;

[0067] Figure 6 A schematic diagram of performance parameters of a retention estimation model provided in an embodiment of the present application;

[0068] Figure 7 A flowchart of a video scheduling method provided in an embodiment of the present application;

[0069] Figure 8 A schematic diagram of a process for generating a candidate video set provided in an embodiment of the present application;

[0070] Fig. 9 A schematic diagram of various video attributes of video information provided in an embodiment of the present application;

[0071] Fig.10 A schematic diagram of feature coding of a recommended sub-feature provided in an embodiment of the present application;

[0072] Fig.11 A schematic diagram of a feature fusion process provided in an embodiment of the present application;

[0073] Fig.12 A schematic diagram of a feature cross-processing process provided in an embodiment of the present application;

[0074] Fig.13A schematic diagram of a result of estimating retention rate provided in an embodiment of the present application;

[0075] Fig.14 A schematic diagram of a video scheduling device provided in an embodiment of the present application;

[0076] Fig.15 A schematic diagram of the structure of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the drawings in the embodiment of the present application. Obviously, the described embodiment is only a part of the embodiment of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be arbitrarily combined with each other. In addition, although the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in an order different from that here.

[0078] The terms "first" and "second" in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the term "comprising" and any of their variations are intended to cover non-exclusive protection. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. "Multiple" in the present application can mean at least two, for example, can be two, three or more, and the embodiments of the present application are not limited.

[0079] The term "and / or" in the embodiments of the present application is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0080] It is understandable that in the following specific implementations of the present application, data related to retention rate and the like are involved. When the various embodiments of the present application are applied to specific products or technologies, relevant licenses or consents need to be obtained, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions. For example, when it is necessary to obtain retention rate data related to objects on a video platform, relevant volunteers can be recruited and relevant agreements for volunteer authorization data can be signed, and then the data of these volunteers can be used for implementation; or, by implementing within the authorized organization, the following implementation methods can be implemented using the data of members within the organization to make relevant recommendations to internal members; or, the relevant data used in the specific implementation are all simulated data, such as simulated data generated in a virtual scene.

[0081] To facilitate understanding of the technical solutions provided in the embodiments of the present application, some key terms used in the embodiments of the present application are explained here:

[0082] Linear regression: A statistical analysis method used to build a linear model to describe the relationship between independent variables (features) and dependent variables (result vectors). Linear regression can determine the coefficients of the model by minimizing the squared error between the predicted value and the actual value to predict and explain the changes in the dependent variable. In this application, a linear regression can be used to build a retention estimation model, where the independent variables are the recommended features corresponding to each candidate video set, and the target variable is the estimated retention rate corresponding to each candidate video set.

[0083] Lasso Regression (Lasso Regression): A variant of linear regression used to select and regularize features when building a model. By adding the Lasso Regularization term (L1 regularization) to the loss function to penalize the complexity of the model, some features in the model coefficients become zero, thereby generating sparse coefficients, that is, shrinking some coefficients to zero, making Lasso regression advantageous in high-dimensional data sets and achieving the purpose of feature selection.

[0084] L1 regularization: A regularization technique used in linear regression and other machine learning models. It limits the complexity of the model by adding an L1 penalty term to the model's loss function, thereby preventing overfitting and forcing the model to select the most important features. L1 regularization achieves feature selection by sparsifying the coefficients (or weights), i.e. shrinking some coefficients to zero.

[0085] Least squares method: An optimization technique commonly used in fitting mathematical models and regression analysis to estimate model parameters to minimize the sum of squared errors between the model's predicted values ​​and actual observed values. The least squares method can be used to obtain the optimal parameter values ​​to minimize the loss function (the sum of squared errors), thereby fitting the data and finding the best fitting line or curve. In this application, the least squares method can be used to fit the parameters of the retention rate prediction model so as to minimize the error between the model's predicted values ​​and the actual retention rate. Through the least squares method, the best linear relationship or curve can be found to describe the relationship between the retention rate and various features.

[0086] Quadratic Programming: A method for solving a class of mathematical optimization problems in which the objective function and constraints are quadratic functions. Quadratic programming problems usually involve finding a vector that minimizes or maximizes the objective function under the constraints. In this application, the Lasso regression problem is transformed into a standard quadratic programming problem through quadratic programming to optimize the regularization term in the retention estimation model. By performing quadratic programming on the objective function, the optimal model parameter settings can be found to minimize the retention rate prediction error while considering the regularization term.

[0087] Constraints: In mathematical modeling, constraints are a set of constraints used to limit the solution space of an optimization problem. Constraints can be equations or inequalities that are used to restrict the solution of an optimization problem to meet specific requirements.

[0088] Cross-validation: A model evaluation and selection technique that comprehensively evaluates model performance by splitting the dataset into multiple different subsets of training and test sets and training and testing the model multiple times.

[0089] Sigmoid function: An S-shaped curve function, also known as the Logistic function. It can be used to map input values ​​to a range between 0 and 1. It is often used to map the output of a linear model to a probability value, indicating the probability that a sample belongs to a certain class. In this application, the Sigmoid function can be used to map the output of a linear regression model to a probability value of the retention rate to help determine whether a user will be retained.

[0090] Video attributes: characteristics used to describe and classify video content, such as video type, video rating, video playback status, etc.

[0091] Video playback status: also known as network broadcast status. Video content can be divided into exclusive broadcast and non-exclusive broadcast content according to the network broadcast status. Exclusive broadcast means that the broadcast rights of the video are monopolized by a video platform, which has exclusive resources and can only be broadcast on this video platform, and the audience can only watch it on this video platform. Non-exclusive broadcast means that multiple video platforms have the broadcast rights of the video content.

[0092] Video scheduling: refers to online video platforms or TV broadcasting platforms determining when and what video content to play in order to meet audience needs, improve retention rates or achieve other business goals.

[0093] Retention rate: refers to the degree to which the subjects of a video platform continue to visit, use or watch the video platform. The higher the retention rate, the longer the subjects will use the video platform, watch more content, and may bring more advertising benefits or membership subscriptions to the video platform.

[0094] Artificial Intelligence (AI): It is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making. Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level technology and software-level technology. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0095] Machine Learning (ML): It is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithmic complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning by teaching. In an embodiment of the present application, a deep learning technology can be used to train a retention prediction model, and then the retention prediction model is used to obtain and select recommended video sets from each candidate video set based on the recommended feature sets of each candidate video set, and the corresponding video schedule is performed based on the estimated retention rates corresponding to each candidate video set.

[0096] The following is a brief introduction to the design concept of the embodiment of the present application:

[0097] Retention rate is a key indicator to measure the user stickiness and attractiveness of a video platform. When formulating operations and product portfolios, one of the most important issues that video platforms pay attention to is how to effectively schedule various video content, that is, by scheduling video collections containing specific video content to attract more users to continue to visit the video platform and improve the retention rate of the video platform.

[0098] In related technologies, it is usually based on periodic time factors (such as seasons, holidays, etc.), and in different time periods, according to the historical video sets recorded in the historical records and their corresponding retention rate results, to select matching video sets, so as to schedule the video content contained in the video set in a specific time period to improve the retention rate of the video platform. However, the video sets selected in this way only consider the influence of general time factors, and are too dependent on the specific time period and specific video sets in the historical records, resulting in the lack of flexibility and foresight in the video sets used for scheduling to adapt to future changes, thereby failing to bring the expected retention rate, and even the problem of being unable to select video sets similar to the past, lacking business guidance significance and practicability. In addition, this method is prone to repeated scheduling and closed scheduling, which is not conducive to expanding the scheduled video content, resulting in insufficient novelty and diversity of the scheduled video sets, and failing to meet the video viewing needs of the target, resulting in a lower retention rate of the video platform.

[0099] In view of the above problems, the embodiment of the present application provides a video scheduling method, which obtains multiple different candidate video sets based on multiple video information to be scheduled on the video platform, and is used to indicate one or more video information to be played on the video platform in a preset time period. In this way, compared with the use of a fixed historical video set, the present application expands the video set used for scheduling, avoids repeated scheduling and closed scheduling caused by over-reliance on historical video sets, improves the novelty and diversity of the scheduled video sets, meets the video viewing needs of the changing objects, and thus improves the retention rate of the video platform.

[0100] Furthermore, the present application extracts the recommended features corresponding to each candidate video set from at least one preset recommended feature dimension for each candidate video set obtained, thereby obtaining a corresponding recommended feature set. Each recommended feature dimension represents a video attribute that causes a deviation in the estimated retention rate of the video platform, thereby more comprehensively considering the impact of video information of different video attributes on the retention rate, making the estimated retention rate of the candidate video set closer to the actual video scheduling scenario, reducing the estimated error of the retention rate, and improving the accuracy of the estimated retention rate, thereby improving the accuracy of the subsequent selection of recommended video sets based on the estimated retention rate, so that the use of recommended video sets for video scheduling can bring the expected retention rate results to the video platform. Furthermore, the present application introduces weight constraints in the model training process to limit the range of values ​​of the preset weight values ​​of each recommended feature, ensuring that the weight value of each recommended feature can more realistically reflect the degree of influence of different video attributes on the retention rate in the actual scheduling scenario, and further improving the estimated accuracy of the retention rate. Moreover, through the preset weight constraints, the trained model is more adaptable to the actual video scheduling scenarios, which improves the generalization ability and robustness of the model.

[0101] Furthermore, the present application selects a recommended video set from a large number of candidate video sets by obtaining the estimated retention rate of each video set, so as to schedule the video content corresponding to the recommended video set in a specific time period, thereby attracting more objects to continue to visit the video platform and improve the retention rate of the video platform. In addition, the present application can flexibly select a recommended video set that meets the business objectives based on the estimated retention rate of each candidate video set according to the business objectives of different video scheduling, so that the video platform can schedule videos through the recommended video set to meet the ever-changing user needs and market trends.

[0102] In order to further improve the estimated accuracy of the retention rate, and thus improve the retention rate of the video platform, this application takes into account that the joint scheduling of multiple videos may have a superimposed chain effect, which will affect the user's viewing decision, thereby affecting the retention rate of the video platform. Therefore, when the candidate video set contains multiple video information to be played, this application extracts independent recommendation features for each video information separately to accurately obtain the impact of each video information on the estimated retention rate. Cross-recommendation features are extracted for every two independent recommendation features to obtain the superimposed impact of two different video information on the estimated retention rate, thereby improving the accuracy of retention rate estimation for the candidate video set composed of multiple video information.

[0103] In order to further improve the estimated accuracy of retention rate, and thus improve the retention rate of the video platform, this application also extracts recommended sub-features for a video information from multiple recommended feature dimensions, so as to obtain different video attributes of each video information in different dimensions. In this way, the recommended sub-features reflect the multi-faceted characteristics of the video information, which helps to better understand the characteristics of each video information, thereby improving the accuracy of estimating the retention rate based on the recommended features. In addition, after obtaining multiple recommended sub-features, this application further fuses the recommended sub-features, integrating the feature information on different dimensions into a comprehensive independent recommended feature, so as to reduce the redundancy of the features and improve the expressive power of the features. The independent recommended features obtained by fusion can more comprehensively reflect the comprehensive characteristics of the video, further improving the accuracy of estimating the retention rate based on the recommended features, thereby accurately improving the retention rate of the video platform.

[0104] In order to further improve the estimated accuracy of the retention rate, thereby improving the retention rate of the video platform, the embodiment of the present application controls the influence of each recommended feature on the estimated retention rate through a preset weight value to conform to the actual video scheduling scenario, so that according to the importance and contribution of different features, the information of each recommended feature is integrated into a recommended fusion feature, and the influence of key features is highlighted, and the influence of secondary features is reduced, so as to more accurately obtain the key factors affecting the retention rate, making the estimated results of the retention rate more interpretable, so as to better meet the actual business needs of the video platform. Further, through the feature mapping process, the relationship between the recommended feature set and the estimated retention rate is modeled, and the multi-dimensional recommended features are mapped to a single output, so as to more accurately obtain the estimated retention rate results by capturing the relationship and nonlinear effects between the features. At the same time, both feature fusion and feature mapping can be performed using a trained retention estimation model, so as to accurately predict the estimated retention rates of different candidate video sets, so that the video platform can more accurately estimate the retention effects that different candidate video sets can bring, and provide an important basis for business decision-making.

[0105] In order to further improve the generalization ability and interpretability of the retention prediction model, the embodiment of the present application continuously uses different sample video sets and retention rate data to optimize the model during the iterative training process of the retention prediction model, which not only improves the accuracy of the estimated retention rate output by the model, but also improves the generalization ability of the retention prediction model to deal with different video sets, so that it can better adapt to different actual scheduling scenarios. In addition, by presetting weight constraints, the influence of each recommended feature on the estimated retention rate during the parameter adjustment process of the model is limited to the actual video scheduling scenario, so that the trained retention prediction model is more in line with actual business cognition, more reasonable and interpretable, so as to better meet the business needs in the actual scheduling scenario.

[0106] In order to further improve the generalization ability and interpretability of the retention prediction model, the embodiment of the present application sets weight constraints such as feature constraints and dimension constraints during the training process of the retention prediction model to limit the range of preset weight values ​​of the recommended features and improve the rationality and interpretability of the retention prediction model. Among them, the feature constraints are used to limit each recommended feature to have a positive impact on the estimated retention rate, and further set single constraints and cross constraints to ensure that both independent recommended features and cross-recommended features have a positive impact on the estimated retention rate, reduce the overfitting risk of the retention prediction model, and improve the generalization ability of the retention prediction model. On the other hand, by setting dimension constraints, the different recommended sub-features under each recommended feature dimension are limited to have different degrees of influence on the retention rate, so that the retention prediction model can be weighted according to the contribution and importance of different sub-features, so as to more accurately obtain the difference in the influence of different sub-features on the retention rate under each feature dimension, and further improve the generalization ability of the model.

[0107] In order to further improve the retention rate of the video platform, the embodiment of the present application selects the target estimated retention rate according to the preset retention rate threshold, thereby ensuring that the recommended video set can bring the expected retention rate, helping the video platform to more efficiently manage and control the user retention rate and achieve the business goal of video scheduling. At the same time, based on the preset retention rate threshold, the recommended video set is selected in combination with the video quantity condition and the video attribute condition. According to different business scenarios and the interests and preferences of different target audiences, a recommended video set containing video information of specific video attributes or a specific number of videos can be selected to perform more personalized video scheduling, meet the diverse video content needs of different objects, increase user satisfaction, and thus improve the retention rate of the video platform.

[0108] In the embodiment of the present application, part of the retention estimation model involves AI and machine learning technology, and is designed based on image recognition technology and ML in artificial intelligence.

[0109] Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence.

[0110] Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. Artificial intelligence technology mainly includes computer vision technology, natural language processing technology, and machine learning / deep learning. With the research and progress of artificial intelligence technology, artificial intelligence has been studied and applied in many fields, such as common smart homes, smart customer service, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, robots, smart medical care, etc. It is believed that with the development of technology, artificial intelligence will be applied in more fields and play an increasingly important role.

[0111] Machine learning is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Compared with data mining, which seeks mutual characteristics between big data, machine learning focuses more on algorithm design, allowing computers to automatically "learn" rules from data and use the rules to predict unknown data.

[0112] Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and other technologies. Reinforcement learning (RL), also known as reinforcement learning, evaluation learning or enhanced learning, is one of the paradigms and methodologies of machine learning. It is used to describe and solve the problem of how an agent can maximize its rewards or achieve specific goals through learning combinations in the process of interacting with the environment.

[0113] After introducing the design ideas of the embodiments of the present application, the following briefly introduces the application scenarios to which the technical solutions of the embodiments of the present application can be applied. It should be noted that the application scenarios introduced below are only used to illustrate the embodiments of the present application and are not limited. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0114] The solution provided in the embodiment of the present application can be applied to any scenario where a video platform schedules video content, for example, a video on demand platform schedules specific video content for viewing by objects accessing the platform. Figure 1 As shown, it is a schematic diagram of an application scenario of an embodiment of the present application. The application scenario diagram includes any terminal device 101 among multiple terminal devices and any server 102 among multiple servers.

[0115] In the embodiment of the present application, the terminal device 101 includes but is not limited to mobile phones, tablet computers (PADs), laptop computers, desktop computers, smart TVs, smart car devices, smart voice interaction devices, smart home appliances, and smart wearable devices. The terminal device 101 may be installed with a client related to the video platform service, which may be software, a web page, a small program, etc., which is not specifically limited in the embodiment of the present application.

[0116] Server 102 is a backend server corresponding to software or web pages, applets, etc., or a backend server specifically used for scheduling videos for clients, which is not specifically limited in this application. Server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms. Server 102 may include one or more processors 1021, a memory 1022, and an I / O interface 1023 for interacting with a terminal. In addition, server 102 may also be configured with a database 1024, which may be used to store video data information, retention rate information, and trained model parameters. Among them, the memory 1022 of the server 102 can also store program instructions of the video scheduling method provided in the embodiment of the present application. When these program instructions are executed by the processor 1021, they can be used to implement the steps of the video scheduling method provided in the embodiment of the present application to schedule the recommended video set.

[0117] The terminal device 101 and the server 102 may be directly or indirectly connected to each other through one or more networks 103. The network 103 may be a wired network or a wireless network, for example, a mobile cellular network or a wireless fidelity (Wireless-Fidelity, WIFI) network, or other possible networks, which are not limited in the embodiment of the present invention.

[0118] The video scheduling method in the embodiment of the present application can be executed by a computer device, which can be a terminal device 101 or a server 102, that is, the method can be executed by the terminal device 101 or the server 102 alone, or can be executed by the terminal device 101 and the server 102 together.

[0119] Specifically, when the terminal device 101 and the server 102 jointly perform video scheduling, when the object opens the client through the terminal device 101, the client needs to display the video information that needs to be scheduled in the time period, that is, the client needs to initiate a video scheduling request to the server 102 in advance. Based on the scheduling request, the server 102 obtains multiple video information to be scheduled in the time period, and determines multiple candidate video sets to be selected for the scheduling request, and estimates the retention rate of each recommended feature set by extracting the recommended feature sets of the multiple candidate video sets. Based on the estimated retention rate results of each candidate video set, a recommended video set is selected from the multiple candidate video sets, and a return result is generated based on the recommended video set, which is returned to the client of the terminal device 110, so that the terminal device 101 schedules the corresponding video content based on the recommended video set.

[0120] It should be noted that, in the embodiment of the present application, the number of terminal devices 101 may be one or more, and similarly, the number of servers 102 may be one or more, that is, for Figure 1 What is shown is just an example. In fact, the number and communication mode of terminal devices and servers are not limited and are not specifically limited in the embodiments of the present application.

[0121] In actual application, server 102 and the client's background server can also be different servers. In this case, server 102 can be a dedicated video scheduling server. When the object triggers a scheduling request, the client's background server can request the recommended video set to be scheduled from server 102. After server 102 determines the recommended video set, it pushes it to the client through the client's background server.

[0122] In a possible implementation, the video scheduling method of the embodiment of the present application can be applied to a video platform to schedule specific video content to obtain a higher retention rate of video scheduling scenarios. For example, a video streaming platform has multiple video contents to be scheduled, including movies, TV series, variety shows, animations and other different video contents. The platform hopes to schedule one or more of these videos in a specific time period to attract as many objects as possible to visit the platform and improve the retention rate of the video platform.

[0123] Specifically, Figure 2As shown, take the example that a video platform has two video contents to be broadcast, including video a and video b, and needs to formulate a video broadcast plan for the next week. Based on the video broadcast method of the embodiment of the present application, the video platform can generate three different candidate video sets according to the two videos to be broadcast, including scheduling video a alone, scheduling video b alone, and scheduling video a and video b. And according to different recommended feature dimensions such as preset video type and video level, feature extraction is performed on the video information contained in each of the three candidate video sets to obtain the recommended feature sets corresponding to each of the three candidate video sets. At the same time, the video platform can also pre-collect different video sets used in its history, as well as the retention rate brought by each historical video set to obtain sample data, and use the sample data to iteratively train the model to obtain a trained retention estimation model. Therefore, the trained retention estimation model is used to calculate the estimated retention rates corresponding to the three candidate video scheduling feature sets, and based on the estimated retention rates obtained, a recommended video set that meets its scheduling goals is selected from the three candidate video sets, and the recommended video set is scheduled for broadcast in the next week. For example, the candidate video set with the highest estimated retention rate is selected for broadcast, so as to maximize the retention rate of the video platform in the next week.

[0124] In an embodiment of the present application, a retention estimation model can be used to implement the process of obtaining the above-mentioned estimated retention rate. Before the retention estimation model is put into use, it is necessary to perform model training in advance so that the retention estimation model converges. Next, in order to facilitate the description of the application process, the training process of the retention estimation model is first introduced.

[0125] Please refer to Figure 3 , which is a schematic diagram of the training process of the retention prediction model provided in the embodiment of the present application, and the process can be Figure 1 The process may be executed by the server 102 or the terminal device 101, or may be executed jointly by the server 102 and the terminal device 101. Here, the iterative training process of the retention estimation model in the embodiment of the present application is mainly described in detail by taking the execution by the server 102 as an example.

[0126] Step 301: Obtain multiple sample data for model training.

[0127] In the embodiment of the present application, each sample data includes at least: a sample video set, and a sample retention rate corresponding to the sample video set. The video set represents one or more video contents played on the video platform in a specific time period, and the retention rate represents how many objects are attracted by the video content scheduled by the video platform to continue to choose the video platform. By using a large amount of sample data as supervisory data for model training, the model can learn the correlation between the correct video set and the retention rate, so that the retention rates of different video sets can be more accurately estimated, and the generalization ability of the retention estimation model to deal with different video sets can be improved.

[0128] In one possible implementation, the historical video sets used by the video platform in different historical time periods and their corresponding retention rate data can be obtained through the system logs, database records, object data, etc. of the video platform, so as to obtain the sample video sets and sample retention rates used for model training.

[0129] Specifically, the system log of the video platform may include: the video content scheduled to be broadcast by the video platform in various historical time periods, basic information of objects on the video platform, platform access of objects, video viewing and other operations. The embodiment of the present application can process the above video data and object data, such as filtering, cleaning and splicing, and abstract the original data into a data expression containing complete information according to a set standardized format, thereby processing it into sample data that can be used for training the retention estimation model.

[0130] For example, based on the system log of the video platform, obtain the different video content scheduled for daily broadcast on the video platform in the past month, as well as the information of objects that visit the video platform daily. And further determine whether there are objects that have watched the scheduled videos yesterday (i.e., retained objects) among the objects that visit the video platform daily, and count the number of retained objects every day. Finally, the daily retention rate of the video platform is calculated based on the total number of objects that visited the video platform that day and the number of retained objects. Combined with the video content scheduled for daily broadcast on the video platform and the calculated daily retention rate, sample data for training the retention estimation model can be obtained. Among them, the calculation method of the retention rate is as follows:

[0131] Retention rate = (number of retained objects / total number of objects accessing the video platform) * 100%

[0132] Step 302: Based on the sample data, the retention estimation model to be trained is iteratively trained to obtain a trained retention estimation model.

[0133] In the embodiment of the present application, after the training samples are obtained, model training can be performed. Specifically, through various historical video sets of a certain video platform and their corresponding historical retention rates, the implicit video interest preferences and behavioral habits of the objects of the video platform are mined, and the mined multiple behavioral outputs are added to the retention estimation model learning, thereby predicting the retention rates of different video sets.

[0134] In a possible implementation, the Lasso regression method can be used to preliminarily construct a retention prediction model to be trained. The retention prediction model is as follows:

[0135]

[0136] Among them, y represents the target variable of the retention estimation model, that is, the estimated retention rate corresponding to the video set that the retention estimation model needs to predict.

[0137] x j represents the jth independent variable of the retention prediction model, that is, the jth recommended feature representing the video information of different video sets in the embodiment of the present application. j Both are related to the corresponding β j Multiply them together to determine their contribution to retention rate y.

[0138] β j Represents the coefficient of the independent variable (i.e., feature). Each β j Represents the corresponding feature x j The impact on the target variable y, that is, the impact of different recommended features on the estimated retention rate. j When it is a positive number, it means feature x j The increase of is positively correlated with the increase of retention rate y; j When it is a negative number, it means that the feature x j The increase in is negatively correlated with the increase in retention rate y.

[0139] β 0 Represents the intercept term of the retention estimation model, also known as the bias term. 0 Represents when all independent variables (i.e., recommended features x j ) is equal to zero, the expected value of the retention rate y, that is, the baseline value of the retention rate of the video platform when the input video set is zero.

[0140] ∈ represents the error term of the retention estimation model, that is, the random noise or error that the model cannot explain. It reflects the impact of other factors that the model cannot capture on the retention rate y. Usually, ∈ is set to be a random variable that follows a normal distribution.

[0141] In summary, the retention prediction model is used to express the estimated retention rate (y) and the recommended features (x) contained in different video sets. j ). The intercept term (β 0 ) represents the basic retention rate of the video platform when there is no influence of any recommended features. The coefficient of each feature (β j ) indicates the influence of the feature on the retention rate, that is, the weight value corresponding to the recommended feature, and the positive and negative signs indicate the direction of the influence. The goal of the retention estimation model is to find the best β j The coefficient combination that best explains the observed retention data.

[0142] Specifically, the objective function of the retention estimation model is as follows:

[0143]

[0144] Among them, X represents the feature matrix of m*n structure, where m represents the number of sample data and n represents the number of recommended features. That is, the feature matrix X contains the feature information of all sample data, each row represents a sample, and each column represents a feature.

[0145] y represents the result vector of m*1 structure, which contains the estimated retention rate corresponding to each sample data.

[0146] β is the coefficient vector of the retention prediction model, that is, the weight or coefficient of each recommended feature in the retention prediction model. β is a vector containing multiple elements, each element β j Corresponding to a recommended feature in the retention estimation model, it determines the impact of the recommended feature on the estimated retention rate.

[0147] ||β|| 1 is the L1 regularization term, also known as the L1 norm of the coefficient vector β. represents the sum of the absolute values ​​of the coefficient vector β. The complexity of the model is penalized by the L1 norm of β to make the coefficient vector β sparse. The L1 regularization term can make the retained estimation model more inclined to select a few important features, while reducing the coefficients of irrelevant or unimportant features to zero, thereby improving the generalization ability and interpretability of the model.

[0148] RSS is the residual sum of squares of the objective function, which is used to measure the fit of the retention prediction model. In statistics and machine learning, the residual is the difference between the observed value and the model predicted value. By minimizing the residual sum of squares, the retention prediction model can fit the observed data as much as possible, that is, find the coefficient that best matches the actual data.

[0149] Represents the normalization factor used to match the scale of the RSS term with the scale of the L1 regularization term to ensure that they are appropriately weighted in the objective function.

[0150] λ is a regularization parameter, also known as a penalty parameter, which controls the strength of regularization and thus the complexity of the retention estimation model. A larger λ value will shrink more coefficients to zero, making the coefficients sparse. By penalizing unimportant features with λ, the model can be prevented from overfitting.

[0151] In summary, by adjusting the value of the coefficient vector β, the influence of each recommended feature in the retention prediction model can be controlled, thereby affecting the degree of fit and complexity of the model. The regularization parameter λ controls the degree of penalty for β, thereby adjusting the selection of features and the complexity of the model. In the process of fitting the retention prediction model by the least squares method and other methods, by solving the optimal coefficient vector β, a model that fits the retention rate well and has a certain sparsity (some coefficients are zero) can be obtained to better achieve the estimation of the retention rate.

[0152] In one possible implementation, due to various realistic factors in actual video scheduling scenarios, such as different video attributes causing different degrees of deviation in the estimated retention rate of the video platform, in order to make the coefficient vector β of each recommended feature more truly reflect the impact of the video attributes on the retention rate in the actual scenario, the embodiment of the present application converts the Lasso regression into a quadratic programming problem with constraints, and restricts the coefficient β of the retention estimation model by converting the objective function and setting weight constraints, such as restricting the L1 norm of the coefficient β or restricting the value range of the coefficient β, so as to obtain a retention estimation model that is more in line with actual business cognition, and the estimated retention rate predicted by the retention estimation model is closer to the actual situation of the actual video scheduling scenario.

[0153] Specifically, the standard form of the quadratic programming problem is as follows:

[0154]

[0155] subject to G*x≤h

[0156] Ax=b

[0157] x represents the variable to be solved for the quadratic programming problem. In the embodiment of the present application, x may correspond to the coefficient vector β to be solved in the Lasso regression. And XT represents the transpose of the matrix X. When X is an m*n matrix, XT is the corresponding n*m matrix to achieve the correct dimension matching in the matrix multiplication.

[0158] P represents an n*n semi-positive definite matrix, which represents the quadratic coefficient matrix of the quadratic objective function. In Lasso regression, P corresponds to the product of the transpose of the feature matrix X and X, which is used to capture the relationship between features. The existence of P makes the problem a quadratic programming problem.

[0159] q represents a column vector of length n, which represents the coefficient of the first-order term in the objective function. In Lasso regression, q can correspond to the product of the transpose of the feature matrix X and the result vector y. Similarly, qT represents the transpose of the vector q. The transpose of the vector itself does not change the dimension of the vector, but turns the column vector into a row vector.

[0160] G represents an m*n matrix, which represents the coefficient matrix of the inequality constraints. In Lasso regression, G can be used to represent the weight constraints on the coefficient vector β.

[0161] h represents a column vector of length m, which indicates the right-hand side value of the inequality constraint. It can be used together with the matrix G to define a set of inequality constraints.

[0162] A represents a p*n matrix, which represents the coefficient matrix of the equality constraints. In Lasso regression, there are usually no equality constraints, so A may be empty.

[0163] b represents a column vector of length p, which represents the right side value of the equality constraint. Together with the matrix A, it defines a set of equality constraints.

[0164] In summary, the standard form of quadratic programming is used to solve optimization problems involving linear and quadratic objective functions, inequality constraints, and equality constraints. In Lasso regression, the original problem can be transformed into a quadratic programming problem by selecting appropriate parameters such as P, q, G, h, A, and b to meet the optimization requirements of the retained prediction model with constraints, so that the retained prediction model can comprehensively consider the trade-off between fitting data and regularization to obtain the optimal coefficient vector β.

[0165] Furthermore, the conversion process of converting Lasso regression into a quadratic programming problem in the embodiment of the present application is as follows:

[0166] First, the first term in the objective function of the original retention estimation model, the residual sum of squares RSS, is decomposed into:

[0167] RSS=(yX*β).T*(yX*β)

[0168] =yT*y-2*β.T*XT*y+β.T*XT*X*β

[0169] Among them, yT*y is a constant and will not affect the minimization of the objective function. Therefore, the first term RSS of the objective function can be converted to:

[0170]

[0171] Among them, Q = XT * X is a symmetric matrix, which is obtained by multiplying the transpose of the feature matrix X by the feature matrix X. Its function is to capture the correlation and influence between different features. In the objective function, it represents the interaction between recommended features, helping the retention estimation model to determine which recommended features have a significant impact on the retention rate, as well as the correlation between features.

[0172] p = -XT*y is a column vector of length n, which is obtained by multiplying the result vector y by the transpose of the feature matrix X. In the retention prediction model, each component of p represents the degree of association between the result vector y and the corresponding recommended feature, so that the model can be fitted according to the observed data.

[0173] Furthermore, in order to add the L1 regularization term to the objective function of the quadratic programming, it can be written in the form of an absolute value to penalize each coefficient vector. The penalty coefficient is the hyperparameter λ, which is used to control the strength of regularization. Adding the coefficient of the L1 regularization term to the coefficient matrix of the first-order term can limit the value range of each coefficient to a certain extent. Therefore, the second term λ||β|| in the objective function of the original retained estimation model can be 1 , that is, the L1 regularization term is decomposed into:

[0174] λ||β|| 1 =sum(abs(β))=1 T *abs(β)=[1,1,…,1]*max(β,-β)

[0175] Among them, λ represents a hyperparameter, i.e., a regularization parameter or penalty coefficient, which is used to control the strength of L1 regularization. The larger the value of λ, the stronger the penalty, and the coefficient β of the model will be more likely to be compressed to zero, thereby reducing the complexity of the model. The fit and regularization of the model can be balanced by adjusting λ. In machine learning, cross-validation and other techniques are usually used to select a suitable λ value.

[0176] abs(β) represents the absolute value vector of β, which indicates the absolute value of each coefficient vector β. By measuring the size of each coefficient vector β, the contribution of each recommended feature to the retention rate is determined. The regularization term will penalize the above absolute value, thereby affecting the value of the coefficient vector β.

[0177] [1,1,…,1] represents a unit vector of length n (i.e., the number of features), which is multiplied by the absolute value vector of β to add the absolute value of each coefficient to obtain the value of the L1 regularization term. The 1 here is only used to weigh the effects of each coefficient so that each coefficient is subject to the same regularization constraint.

[0178] max(β, -β) represents the operation of taking the absolute value of each component in the coefficient vector β to ensure that each component is positive. This is because L1 regularization is used to encourage coefficient sparsity, even if some coefficients are compressed to zero, thereby achieving the effect of feature selection. Taking the absolute value is to ensure that the penalty is applied to the absolute value of the coefficient.

[0179] In summary, the second term of the objective function, λ*|(|β|)|_1, is used to regularize the coefficient vector β in the retained prediction model. The size of this penalty term is controlled by λ, and by constraining the absolute value of the coefficient, the retained prediction model can determine a smaller set of coefficient vectors, thereby simplifying the model and improving generalization performance.

[0180] Therefore, according to the above process, the objective function of Lasso regression can be transformed into:

[0181]

[0182] Among them, since yT*y is a constant, it will not affect the minimization of the objective function. Therefore, the objective function of Lasso regression can be simplified to the standard quadratic programming form:

[0183]

[0184] Among them, β represents the coefficient vector to be solved, that is, the weight value of each recommended feature in the retention estimation model. The purpose of this objective function is to solve the coefficient vector β corresponding to each recommended feature by minimizing the quadratic programming problem. The fit of the data is considered by minimizing the residual sum of squares, and the size and sparsity of the coefficients are limited by L1 regularization to prevent overfitting of the retention estimation model and select important features, thereby improving the generalization performance of the model.

[0185] X represents a feature matrix of m*n structure, where m represents the number of sample data and n represents the number of recommended features. That is, the feature matrix X contains the feature information of all sample data, each row represents a sample, and each column represents a feature.

[0186] X T Represents the transposed matrix of the feature matrix X, which is used to convert the columns of the feature matrix X into rows and the rows into columns, so as to calculate the matrix multiplication so as to consider the relationship between the features and the coefficients in the objective function. T *X represents the product of the feature matrix X and its own transposed matrix, which is called the covariance matrix of the feature matrix. It is used to describe the correlation and importance between the recommended features and optimize the objective function.

[0187] y represents the result vector of m*1 structure, which contains the estimated retention rate corresponding to each sample data.

[0188] (X T *y) T Represents the transposed matrix X T The product of the feature matrix and the result vector y is transposed again to represent the relationship between the feature matrix and the result vector.

[0189] 1 T represents a unit vector of length n, all elements of which are 1, used to penalize the L1 regularization term.

[0190] Specifically, refer to Figure 4 When the server executes step 302, it will specifically perform the following operations in each round of iterative training:

[0191] Step 3021: Adopt the retention estimation model used in this round, and obtain the corresponding estimated retention rate based on the selected sample video set.

[0192] In the embodiment of the present application, the prediction process of estimating the retention rate during training is the same as the prediction process actually applied in the video scheduling process, so the process will be introduced in detail in the subsequent application process and will not be elaborated here.

[0193] Step 3022: Obtain a target loss value based on the sample retention rate and the estimated retention rate corresponding to the sample video set.

[0194] In an embodiment of the present application, as described above, the predicted estimated retention rate can be compared with the corresponding sample retention rate, and then the loss value of the retention estimation model can be determined based on the comparison result, and then it can be determined whether the retention estimation model meets the convergence condition based on the loss value.

[0195] Among them, any possible loss function can be selected to calculate the loss value, and the embodiments of the present application do not limit this.

[0196] Step 3023: Determine whether the retention estimation model meets the convergence condition; if not, execute step 3024; if yes, execute step 3025.

[0197] Combination Figure 5 As shown, in an embodiment of the present application, to determine whether the retention estimation model meets the convergence conditions, it is possible to determine whether the target loss value of the retention estimation model converges to a preset target value. If the target loss value is less than or equal to the preset target value, it indicates that the difference between the estimated retention rate and the sample retention rate is small enough, and thus the accuracy of the retention estimation model meets the requirements. Therefore, it can be determined that the retention estimation model converges. Otherwise, the retention estimation model does not converge.

[0198] Alternatively, when the change in the model loss value obtained after N consecutive trainings is less than or equal to the preset target value, it is determined that the model loss value has converged to the preset target value, indicating that the retained estimation model meets the convergence condition; otherwise, it means that the retained estimation model does not meet the convergence condition. Alternatively, when the number of model training iterations meets the preset value, the model can also be considered to have converged.

[0199] Step 3024: Based on the weight constraint, the target loss value is used to adjust the parameters of the retention estimation model.

[0200] In the embodiment of the present application, if the model loss value does not converge, the weight parameters of the model are adjusted using the loss value according to the adjustment direction indicated by the preset weight constraint condition, so that the weight parameter corresponding to each recommended feature (i.e., the degree of its influence on the estimated retention rate) conforms to the actual situation of the video scheduling scenario. After adjusting the model parameters, return to step 3021 to continue the next round of training process.

[0201] In a possible implementation, the weight parameter adjustment of the model can adopt an optimization algorithm to minimize the target loss value. In each iteration, the model updates the weight parameter to reduce the loss value. The direction and size of the weight parameter adjustment are limited by the weight constraint.

[0202] Specifically, the optimization algorithm includes but is not limited to the gradient descent method, the quasi-Newton method, etc. Take the gradient descent method as an example. The model initializes the weight parameters, which can be initialized randomly or in other ways. In the first round of iterations, for the initialized weight parameters, the model calculates the corresponding target loss value, such as the mean square error or other suitable loss function value. And calculate the gradient of the loss function with respect to the weight parameters. The gradient is the rate of change of the target loss value with respect to the parameter, indicating in which direction the current model should update the weight parameters to reduce the loss. According to the calculated gradient and the preset weight constraints, the model updates the weight parameters, and the amplitude and direction of the update are affected by the gradient and the constraints. The above steps will be repeated iteratively many times until the model reaches the convergence condition. In this way, the model gradually learns the weight parameters that minimize the loss.

[0203] In a possible implementation, in combination with the video scheduling scenario used in the present application, the preset weight constraints may include feature constraints and dimension constraints. Among them, the feature constraints represent that each recommended feature should have a positive impact on the estimated retention rate; the dimension constraints represent that the recommended sub-features of different feature dimensions have different degrees of influence on the estimated retention rate. During the iterative training process, when the weight parameter of a recommended feature is less than zero, the model needs to adjust it to a positive number to meet the feature constraints, and the model needs to balance the weights of different recommended sub-features under each feature dimension to meet the dimension constraints.

[0204] Specifically, the weight constraints can be converted into mathematical expressions by constructing a parameter matrix and applied to model optimization to ensure that the model parameter adjustment will be carried out with the weight constraints taken into account to meet the specific needs of the business objectives. Therefore, the inequality constraint matrix corresponding to the weight constraints can be expressed as:

[0205]

[0206] Among them, the non-negative constraint β≥0 represents the feature constraint condition, and each element in the restriction coefficient vector β cannot be less than 0, that is, each recommended feature should have a positive impact on the estimated retention rate.

[0207] β a -β b ≥0 represents a dimensional constraint, where β a is the recommended sub-feature a under a specific feature dimension, β b Represents the recommended sub-feature b under the same feature dimension. The impact of the recommended sub-feature a on the estimated retention rate is greater than or equal to the impact of the recommended sub-feature b on the estimated retention rate.

[0208] For example, β a represents the recommended sub-feature corresponding to the S+ level under the video level dimension, β b Represents the recommended sub-feature corresponding to the S level under the video level dimension, and the impact of S+ level video content on the estimated retention rate is greater than or equal to that of S level video content.

[0209] In a possible implementation, due to the actual video scheduling scenario, the joint scheduling of multiple videos may have a superimposed chain effect, that is, multiple videos are scheduled together in a specific time period, which will affect the viewing decision of the object, thereby affecting the retention rate of the video platform. For example, when an S-level exclusive variety show and an S+-level exclusive TV series are scheduled at the same time, the retention rate of the video platform is improved more than when the two are scheduled separately, that is, when the two are played at the same time, more objects will be attracted and the retention of the objects will be increased. Therefore, the embodiment of the present application further divides the feature constraints into single constraints and cross constraints. The single constraint restricts each independent recommended feature to have a positive impact on the estimated retention rate, that is, it ensures that each video information has an impact on the estimated retention rate alone; the cross constraint restricts each cross-recommended feature to have a positive impact on the estimated retention rate, that is, it ensures that the superimposed impact of multiple video information on the estimated retention rate is a positive impact, reducing the overfitting risk of the retention estimation model and improving the generalization ability of the retention estimation model.

[0210] Specifically, a single constraint can be expressed as:

[0211]

[0212] Among them, i represents the index of coefficient β, that is, each coefficient β in the retention estimation model i must be greater than or equal to zero. That is, each independent recommendation feature should have a positive impact on the estimated retention rate. i Less than zero, that is, the corresponding independent recommendation feature has a negative impact on the estimated retention rate, and the model needs to adjust it to a non-negative value to meet this single constraint.

[0213] The cross constraint can be expressed as:

[0214] β i +β ij *x j ≥0 formax(x j ), min (x j )

[0215] Among them, β i Represents the weight parameter corresponding to the recommended feature i in the retention estimation model, that is, the impact of the recommended feature i on the estimated retention rate.

[0216] β ij Represents the cross-effect between feature i and feature j in the retention estimation model, that is, the impact of the interaction between feature i and feature j on the estimated retention rate.

[0217] x j represents another recommended feature j that is different from feature i. In this cross constraint, x j The value of will change according to the change of feature j, so as to ensure the interaction between feature i and any other feature j, that is, the product β ij *x j , no matter how the feature j changes, it is always greater than or equal to zero.

[0218] The cross constraints restrict the interaction between feature i and another feature j to have a positive impact on the estimated retention rate, ensuring that the model considers the interaction between different features to better fit the retention rate.

[0219] Step 3025: If the model converges, the training process ends and a trained retention estimation model is obtained.

[0220] In an embodiment of the present application, if it is determined that the loss value converges, the retention estimation model used in this round is used as the trained retention estimation model.

[0221] In one possible implementation, cross-validation can be used to evaluate the generalization performance of the retention prediction model to select appropriate parameter settings for the retention prediction model and ensure that the trained retention prediction model has good predictive power for new video sets. Cross-validation can mitigate the adverse effects of uneven or random division of sample data sets, provide more reliable performance estimates, and thus select appropriate model parameters to avoid model overfitting or underfitting problems.

[0222] Specifically, the pre-collected sample data set is divided into K non-overlapping subsets, each containing the same proportion of sample data. The retention estimation model is trained for K iterations. In each iteration, one of the subsets is selected as the test set, and the remaining K-1 subsets are used as the training set. The retention estimation model is trained on the training set, and the performance is evaluated on the test set. The performance measurement results of K iterations are averaged to obtain the average performance of the model on different data subsets, so as to eliminate the influence of specific data subsets and more comprehensively reflect the performance of the retention estimation model.

[0223] For example, at each iteration, various performance metrics can be used to evaluate the performance of the model, such as mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R-squared, R 2 ), root mean square error (RMSE), mean absolute percentage error (MAPE) or prediction accuracy of retention rate, etc. The specific selection of the metric depends on the actual business needs and is not specifically limited in the embodiments of the present application.

[0224] like Figure 6 The figure shows a performance parameter diagram of a retention estimation model provided in an embodiment of the present application, wherein MSE represents the expected value of the square of the difference between the predicted value and the actual value, and is used to measure the average deviation between the estimated retention rate predicted by the retention estimation model and the actual retention rate.

[0225] MAE represents the average of the absolute differences between the predicted value and the actual value, and is used to measure the average absolute deviation between the estimated retention rate predicted by the retention estimation model and the actual retention rate.

[0226] R 2 Represents the degree to which the model explains the variability of the data, that is, the correlation between the estimated retention rate predicted by the retention estimation model and the actual retention rate. 2 The value range is between 0 and 1. The closer it is to 1, the better the retention estimation model can explain the data.

[0227] RMSE is the square root of MSE and is used to measure the average deviation between the estimated retention rate predicted by the retention estimation model and the actual retention rate, and is expressed in the same units as the actual value.

[0228] MAPE represents the average of the percentage errors between the predicted values ​​and the actual values, and it is used to measure the percentage error degree of the retained estimation model.

[0229] The above metrics are used to evaluate the performance of the retention estimation model. MSE, MAE, RMSE, and MAPE are used to measure the difference between the model prediction and the actual value. The smaller the difference, the better the prediction performance of the retention estimation model. 2 This measures the degree of fit of the retention estimation model to the data. The closer it is to 1, the better the fit performance of the retention estimation model. The above metrics can help evaluate the accuracy and effectiveness of the model, thereby guiding decision-making and improving the model.

[0230] In one possible implementation, during the cross-validation process, different parameter settings (such as weight parameters and regularization parameters for each recommended feature) can be tried to evaluate the performance of the retention estimation model under different parameters. The parameter setting with the best performance is selected as the final model configuration. When the retention estimation model determines the final parameter configuration, the entire data set (not just the cross-validated subset) is used to perform a final evaluation of the model. If the model performs well on the entire data set, consider using it as a trained retention estimation model and deploying it in actual applications.

[0231] See also Figure 7 As shown, it is a flow chart of the video scheduling method provided in the embodiment of the present application. Here, an example is given with a server as the execution subject. The specific implementation process of the method is as follows:

[0232] Step 701: Based on information of multiple videos to be broadcasted, obtain multiple different candidate video sets.

[0233] In the embodiment of the present application, scheduling represents the process of organizing and arranging the playback of different video content by the video platform. In scenarios such as live programs and linear channels (such as TV stations), video platforms usually apply a scheduling mode to automatically play one or more video contents according to a certain video set. Therefore, when a video platform has multiple video information to be scheduled, it is necessary to decide which one or several video information to play in a specific scheduling time period, that is, to decide how to schedule the video to maximize the attraction of the target, improve the target satisfaction, or achieve specific business goals.

[0234] For example, Figure 8As shown, the video platform has 4 video contents to be broadcast, including TV series a, movie b, and variety show c. For the above video contents to be broadcast, the video platform can obtain 7 different candidate video sets without limiting the number of videos played during the broadcast time period, including three candidate video sets for separately broadcasting TV series a, movie b, and variety show c, three candidate video sets for jointly broadcasting any two of the videos, and one candidate video set for jointly broadcasting three videos. The video platform needs to decide which candidate video set to use for video broadcasting on its client on a daily or weekly basis in the future to attract more objects to continue to visit its platform. It is worth noting that the candidate video sets can be flexibly selected and combined according to the business needs of the video platform and the preferences of the objects, such as limiting the broadcast time period to only play two video information, etc., and the embodiments of the present application do not limit this.

[0235] Step 702: Based on at least one preset recommendation feature dimension, feature extraction is performed on at least one video information contained in each of the multiple candidate video sets to obtain recommendation feature sets corresponding to each of the multiple candidate video sets.

[0236] In the embodiment of the present application, one or more recommendation feature dimensions are set in advance according to the business requirements and video characteristics of the video platform, and are used to extract the recommendation features corresponding to each video information in the candidate video set. Each recommendation feature dimension may correspond to a video attribute that causes a deviation in the estimated retention rate of the video platform, and the recommendation features of the video information extracted based on the recommendation feature dimension represent the degree of influence of the video content of the video attribute on the estimated retention rate.

[0237] In a possible implementation, Fig. 9 As shown, each video information may contain multiple video attributes, such as video type, video level, video playback status and other attributes, which correspond to the recommended feature dimensions such as video type dimension, video level dimension, video status dimension and so on:

[0238] (1) Video type: also known as video category, it represents the category to which the video content belongs, which can include TV series, movies, variety shows, animation and other types, and is used to describe the style and content type of the video. For example, a TV series is usually a series of episodes, while a movie is a complete video story.

[0239] (2) Video grade: Video grade represents the rating of video content, which can reflect the production quality and production cost of the video content. Video grades can be divided into multiple levels, such as S+, S, A, etc., where S+ represents the highest grade and A represents a relatively low grade. Different video platforms may have different rating systems, which usually include video quality standards, video content costs, etc. A higher video grade may mean that the video content has better shooting equipment and special effects production, higher audio quality, more professional actors and production team, higher investment budget, etc.

[0240] (3) Video playback status: represents the playback mode of video content on the video platform, which may include exclusive broadcast, non-exclusive broadcast, paid viewing, free viewing, etc. Exclusive broadcast means that the video content is only played on this video platform, while non-exclusive broadcast means that the video content can be played on multiple video platforms. Paid viewing means that the object needs to pay to watch, while free viewing means that the object can watch without paying.

[0241] In a possible implementation, the video information used to extract features may include, but is not limited to: video description text information of the video content, video cover, screenshots, and other video image information.

[0242] Specifically, for the video description text information, the recommended features of the video information can be extracted through feature encoding. Any encoder structure that can perform text encoding can be used, and they are not listed one by one here. For example, a recurrent neural network (RNN) or a transformer can be used for implementation. RNN processes sequence data by recursively updating hidden states, mapping each word in the text to a vector representation of a fixed length. The transformer directly calculates the context representation of each word through the attention mechanism, and then averages or merges the context representations of all words to obtain a text vector representation. Alternatively, natural language processing technology can be used for text analysis, and the recommended features corresponding to the video information can be extracted by extracting keywords, sentiment analysis, and the like. For video image information, computer vision technology can be used for image analysis to extract the recommended features corresponding to the video information. Alternatively, feature encoding is performed on the video image information to extract recommended features corresponding to the video information, and any encoder structure capable of video encoding can be used, for example, a convolutional neural network (CNN) can be used. CNN can gradually extract low-level and high-level features of the image through multi-layer convolution and pooling operations. Common CNN structures include VGG, ResNet, Inception, etc., which are not listed one by one here.

[0243] In a possible implementation, a video attribute may include multiple specific video attribute values, for example, video types include four types: TV series, movies, variety shows, and animations; video levels include three levels: S+, S, and A; and video playback states include two playback states: exclusive and non-exclusive. Therefore, under the recommendation feature dimension corresponding to the video attribute, there may also be recommendation sub-features corresponding to different attribute values ​​of the video attribute.

[0244] Specifically, Fig.10 As shown, by using feature encoding techniques such as one-hot encoding, each recommended sub-feature under each recommended feature dimension can be mapped to a binary vector to better achieve the modeling of the retention estimation model. For example, under the video type dimension, there are four recommended sub-features representing four different video types, including TV series, movies, variety shows, animation, and documentaries, which can be encoded as feature vectors such as [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1] respectively. Under the video level dimension, different video levels can be mapped to different numerical features. For example, S+ level can be mapped to [3], S level can be mapped to [2], and A level can be mapped to [1], and these numerical features can be used as the corresponding recommended sub-features under the video level dimension. Under the video status dimension, binary features can be used to represent different video playback states. For example, exclusive broadcast can be represented as [1] and non-exclusive broadcast can be represented as [0].

[0245] In a possible implementation, Fig.11 As shown, for a video information, feature extraction is performed from multiple preset recommendation feature dimensions respectively to obtain the corresponding recommendation sub-features of the video information under each recommendation feature dimension. Feature fusion processing is performed on the obtained multiple recommendation sub-features to obtain independent recommendation features of the video information.

[0246] Specifically, take video information with a video rating of S+, exclusively broadcast on a video platform, and a video type of a TV series as an example of video information to be extracted as features, and based on the preset three recommendation feature dimensions of video type, video rating, and video status, extract the recommended sub-features under the corresponding dimensions of the video information to represent the attribute values ​​corresponding to the video information under different video attributes. For example, the recommended sub-feature [1,0,0,0] represents that the video information is a TV series type, the recommended sub-feature [3] represents that the video information is S+ rated, and the recommended sub-feature [1] represents that the video information is exclusively broadcast. The above-mentioned recommended sub-features are spliced ​​and fused to obtain the independent recommended feature code [1,0,0,0,3,1] corresponding to the independent recommended feature {TV series, S+, exclusively broadcast} of the video information.

[0247] In one possible implementation, the present application takes into account that the joint scheduling of multiple videos may have a superimposed chain effect, which will mutually influence the user's viewing decision and thus affect the retention rate of the video platform. When extracting features from the multiple video information contained in the candidate video set to obtain the independent recommendation features corresponding to each video information, feature cross-extraction will also be performed on each two independent recommendation features obtained to obtain a corresponding cross-recommendation feature set, where each cross-recommendation feature represents the superimposed impact of the corresponding two video information on the estimated retention rate. The independent recommendation features obtained and the cross-recommendation feature set constitute the recommendation feature set corresponding to the candidate video set.

[0248] Specifically, refer to Fig.12 As shown, taking the candidate video set containing two video information, video information a1 and video information a2, as an example, according to the preset recommendation feature dimension, the independent recommendation feature x1 corresponding to video information a1 and the independent recommendation feature x2 corresponding to video information a2 are extracted. Considering that there may be a superposition chain effect when two video information are broadcasted together, the independent recommendation feature x and the independent recommendation feature x2 of the two are cross-extracted to obtain the corresponding cross feature x12, and finally the recommendation feature set {x1, x2, x12} corresponding to the candidate video set is obtained. Similarly, when the candidate video set contains three video information, video information a1, video information a2 and video information a3, according to the preset recommendation feature dimension, the independent recommendation features x1, x2, x3 corresponding to each video information are extracted respectively, and the features of each two independent recommendation features obtained are cross-extracted to obtain the corresponding cross recommendation feature set, including {x12, x13, x23}, and finally the recommendation feature set {x1, x2, x12, x13, x23} corresponding to the candidate video set is obtained. The feature extraction process for the candidate video set of 4 or more video information is the same as the principle of the above process, and will not be described one by one here.

[0249] Step 703: Based on the preset weight value of each recommended feature, obtain the estimated retention rate corresponding to each candidate feature set.

[0250] In the embodiment of the present application, the preset weight value of each recommended feature represents the influence of the recommended feature on the estimated retention rate, that is, the higher the preset weight value, the greater the possibility that the video attribute corresponding to the recommended feature attracts the object to retain, which may bring a higher retention rate to the video platform. Therefore, when the recommended feature set corresponding to the candidate video set is determined, the preset weight value corresponding to each recommended feature in the recommended feature set can be used to predict the estimated retention rate that the candidate video set may bring to the video platform.

[0251] In one possible implementation, the preset weight value of each recommended feature can be obtained based on the retention estimation model that has been trained in the aforementioned process. According to the aforementioned step 3024, during the training process of the retention estimation model, the present application uses the target loss value to iteratively adjust the weight parameters of the retention estimation model, so that in the trained model, the weight parameter of each recommended feature (that is, the preset weight value corresponding to the recommended feature) satisfies the preset weight constraint condition (that is, the degree of influence of each recommended feature on the estimated retention rate is consistent with the actual video scheduling scenario), and the specific weight parameter adjustment process will not be repeated here.

[0252] Specifically, the recommended feature set corresponding to the candidate video set includes n recommended features {x 1 ,x 2 ,...,x n} as an example, according to the trained retention estimation model, it can be known that the preset weight value corresponding to each recommended feature is β 1 ,β 2 ,...,β n , the estimated retention rate is

[0253] y=β 0 +β 1 *x 1 +β 2 *x 2 +…+β n *x n

[0254] y represents the quantitative result of the estimated retention rate, that is, the retention rate that the candidate video set may bring to the video platform, β 0 Represents the intercept term of the retention estimation model, that is, when the recommended feature x is equal to zero, the expected value of the estimated retention rate y, that is, when the recommended feature contained in the input video set is zero, the baseline value of the retention rate of the video platform. The weight value β of each recommended feature determines the contribution of the recommended feature to the estimated retention rate. The larger the weight value, the greater the impact of the corresponding recommended feature on the estimated retention rate. Integrating the multidimensional information of multiple recommended features into a single numerical result can be used to predict the retention rate indicator.

[0255] In one possible implementation, taking into account the possible overlapping chain effects of the joint scheduling of multiple videos, during the training process of the retention estimation model, the independent recommendation features corresponding to one video information and the cross-recommendation features corresponding to two video information are learned separately, and the first preset weight corresponding to each independent recommendation feature and the second preset weight corresponding to each cross-recommendation feature are obtained. Therefore, for the recommendation feature set including independent recommendation features and cross-recommendation features, the recommendation feature set can be weightedly fused through the first preset weight corresponding to each independent recommendation feature and the second preset weight corresponding to each cross-recommendation feature to obtain the corresponding recommendation fusion feature. The obtained recommendation fusion feature is then feature mapped to obtain the estimated retention rate corresponding to the candidate video set.

[0256] Specifically, taking the candidate video set containing video a1 and video a2 as an example, the recommended feature set corresponding to the candidate video set is {x1, x2, x12}. If the first preset weight values ​​corresponding to the independent recommended feature x1 and the independent recommended feature x2 are β 1 and β 2 , the second preset weight value of the cross-recommendation feature x12 is β 12 , then through feature weighted fusion, the recommended fusion features are obtained as follows:

[0257] R=β 1 *x+β 2 *x 2 +β 12 *x 12

[0258] Among them, the recommended fusion feature R represents the linear combination of each recommended feature and its corresponding weight value.

[0259] In a possible implementation, in the embodiment of the present application, feature mapping processing is performed on the recommended fusion features, and a linear mapping algorithm, a logistic regression algorithm, a deep neural network method, etc. may be used. The number of mapping processing times may be one, two, three or more times, etc., and there is no limitation on the specific method and number of times.

[0260] Specifically, taking the linear mapping algorithm as an example, the recommended fusion features can be mapped into a numerical range, such as a continuous numerical value between (0, 1), through a linear mapping function to represent the estimated retention rate. The specific mapping range can be adjusted according to specific business needs, and this application does not make specific limitations. It is worth noting that in the embodiment of the present application, the trained retention estimation model has determined the parameters of the linear mapping function, that is, through supervised learning of the training data, the retention estimation model can map the recommended fusion features to the corresponding estimated retention rate value.

[0261] For example, the S-shaped function (Sigmoid function) can be used to map the recommended fusion features to the range of (0,1). The specific mapping function is as follows:

[0262]

[0263] Among them, R represents the recommended fusion feature, and f(R) represents the estimated retention rate corresponding to the recommended fusion feature after feature mapping. When R approaches positive infinity, f(R) approaches 1, and when R approaches negative infinity, f(R) approaches 0.

[0264] Step 704: based on the obtained estimated retention rates, a recommended video set is selected from the plurality of candidate video sets.

[0265] In an embodiment of the present application, after obtaining the estimated retention rate corresponding to each candidate recommendation feature, a recommended video set will be selected from multiple candidate video sets based on the size of each estimated retention rate and a specific video set selection method for the video platform to schedule videos to achieve specific business goals.

[0266] Specifically, the video platform has 4 video information to be broadcast, namely, A-level non-exclusive TV series, S+-level exclusive TV series, S-level non-exclusive movie and S-level exclusive variety show, and plans to broadcast at most two video information in a specific broadcast time period. Fig.13 As shown in FIG. 1 , for the 4 to-be-scheduled video information of the video platform, multiple candidate video sets can be generated, and the estimated retention rate corresponding to each candidate video set can be obtained. ( Fig.13 Not all candidate video sets are shown. Fig.13 From the sorting diagram of the estimated retention rate results shown, it can be seen that the co-broadcasting of A-level non-exclusive TV series and S-level non-exclusive movies has a superimposed chain effect, and the co-broadcasting of S+-level exclusive TV series and S-level exclusive variety shows has a superimposed chain effect. In the above two video superposition methods, the improvement of the retention rate by scheduling any one video alone is small, but scheduling the two video information at the same time will have a superimposed effect on the improvement of the retention rate. In addition, the superimposed effect is also proportional to the number of videos scheduled. For example, in the same scheduling time period, if more video content of S-level exclusive variety shows is scheduled, the video content corresponding to the S+-level exclusive TV series can bring a higher retention rate to the video platform.

[0267] In a possible implementation, in order to improve the retention rate of the video platform, a certain retention rate threshold can be pre-set according to the business needs of the video platform, and a target estimated retention rate greater than the preset retention rate threshold can be selected from multiple estimated retention rates, and the target candidate video set associated with the target estimated retention rate can be used as a recommended video set. This ensures that the selected recommended video set can bring the expected retention rate to the video platform, helping the video platform to more efficiently manage and control the retention of objects and achieve business goals.

[0268] In a possible implementation, based on a preset retention rate threshold, the target video set can also be selected in combination with video quantity conditions and video attribute conditions, that is, based on different business scenarios and timing business needs such as the interests and preferences of different target audiences, a recommended video set containing video information of specific video attributes or a specific number of videos is selected to provide more personalized video content recommendations, meet the diverse video content needs of different objects, increase user satisfaction, and thus improve the retention rate of the video platform.

[0269] Specifically, by setting the retention rate threshold to 60%, candidate video sets with an estimated retention rate exceeding 60% can be selected. At the same time, in combination with the actual business objectives of video scheduling, certain video attribute conditions can be set to limit the video attributes of the selected candidate video sets, such as scheduling or not scheduling S+ level video content, or setting certain video quantity conditions to limit the number of videos included in the selected video set, such as scheduling no more than 2 video contents, so as to screen out some candidate video sets that meet the above conditions, thereby controlling the scheduling cost of the video platform while making the estimated retention rate exceed the preset retention rate threshold.

[0270] See also Fig.14 Based on the same inventive concept, the embodiment of the present application further provides a video scheduling device 140, which includes:

[0271] The acquisition unit 1401 is used to obtain a plurality of different candidate video sets based on a plurality of video information to be broadcasted; wherein each candidate video set is used to indicate: at least one video information to be broadcasted in a preset time period.

[0272] The extraction unit 1402 is used to extract features of at least one video information contained in each of the multiple candidate video sets based on at least one preset recommendation feature dimension, so as to obtain recommendation feature sets corresponding to each of the multiple candidate video sets; wherein each recommendation feature dimension represents: a video attribute that causes a deviation in the estimated retention rate of the video platform.

[0273] The estimation unit 1403 is used to obtain the estimated retention rates corresponding to each of the multiple recommended feature sets based on the preset weight value of each recommended feature; the preset weight value is determined during the model training process based on the preset weight constraint condition; the weight constraint condition characterizes the degree of influence of each recommended feature on the estimated retention rate in accordance with the preset scheduling scenario condition.

[0274] The broadcast scheduling unit 1404 is used to select a recommended video set from multiple candidate video sets based on the obtained estimated retention rates, and schedule the recommended video set for broadcast.

[0275] Optionally, the extraction unit 1402 is specifically configured to:

[0276] For each of the multiple candidate video sets, perform the following operations respectively:

[0277] Based on at least one preset recommendation feature dimension, feature extraction is performed on each video information contained in a candidate video set to obtain corresponding independent recommendation features, each independent recommendation feature represents: an independent impact of a corresponding video information on the estimated retention rate;

[0278] Perform feature cross extraction on every two independent recommendation features to obtain a corresponding cross recommendation feature set, where each cross recommendation feature represents: the superimposed impact of the corresponding two video information on the estimated retention rate;

[0279] Based on the obtained independent recommendation features and the cross-recommendation feature set, a recommendation feature set corresponding to the candidate video set is obtained.

[0280] Optionally, the extraction unit 1402 is specifically configured to:

[0281] For each video information, perform the following operations respectively:

[0282] For a piece of video information, based on a plurality of preset recommendation feature dimensions, feature extraction is performed on the video information to obtain a plurality of recommendation sub-features corresponding to the video information, where the plurality of recommendation sub-features correspond one to one to the plurality of recommendation feature dimensions;

[0283] Perform feature fusion processing on multiple recommendation sub-features to obtain an independent recommendation feature corresponding to the video information.

[0284] Optionally, the estimation unit 1403 is specifically configured to:

[0285] For the multiple recommended feature sets obtained, perform the following operations respectively:

[0286] For a recommendation feature set, based on a first preset weight corresponding to each independent recommendation feature and a second preset weight corresponding to each cross recommendation feature, the recommendation feature set is weightedly fused to obtain a corresponding recommendation fusion feature;

[0287] Perform feature mapping on the obtained recommended fusion features to obtain the estimated retention rate corresponding to the recommended feature set.

[0288] Optionally, the scheduling unit 1404 is specifically used for:

[0289] Based on a preset retention threshold, a target estimated retention rate is selected from multiple estimated retention rates, and a target candidate video set associated with the target estimated retention rate is used as a recommended video set;

[0290] Based on a preset video quantity condition, a target video set subset is screened out from multiple candidate video sets, and based on a preset retention rate threshold, a recommended video set is selected from the target video set subset; the video quantity condition characterizes that: each target video set contains a preset amount of video information;

[0291] Based on preset video attribute conditions, a target video set subset is screened out from multiple candidate video sets, and based on a preset retention rate threshold, a recommended video set is selected from the target video set subset; video attribute condition representation: each target video set contains video information of preset video attributes.

[0292] Optionally, the weight constraint condition includes a feature constraint condition and a dimension constraint condition;

[0293] Characterization of feature constraints: Each recommended feature has a positive impact on the estimated retention rate;

[0294] Representation of dimension constraints: Each recommended sub-feature under each recommended feature dimension has a different degree of influence on the estimated retention rate.

[0295] Optionally, the feature constraint condition includes a single constraint condition and a cross constraint condition;

[0296] Single constraint condition representation: Each independent recommendation feature has a positive impact on the estimated retention rate. The independent recommendation feature represents the individual impact of a video information on the estimated retention rate.

[0297] Cross-constraint condition representation: Each cross-recommendation feature has a positive impact on the estimated retention rate. The cross-recommendation feature represents the superimposed impact of two video information on the estimated retention rate.

[0298] Optionally, the estimation unit 1403 is specifically configured to:

[0299] Through the trained retention estimation model, perform the following operations on multiple recommended feature sets:

[0300] For a recommended feature set, based on a preset weight value, the recommended feature set is weightedly fused to obtain a corresponding recommended fusion feature;

[0301] Perform feature mapping on the recommended fusion features to obtain the estimated retention rate of the recommended feature set;

[0302] The retention estimation model is obtained by training multiple sample data, each of which includes: a sample video set and a sample retention rate corresponding to the sample video set.

[0303] Optionally, the video scheduling device further includes a training unit 1405, which is used to:

[0304] The retention estimation model to be trained is iteratively trained based on the sample data to obtain a trained retention estimation model; wherein each iterative training process includes the following operations:

[0305] Based on the selected sample video set, obtain the corresponding estimated retention rate;

[0306] Based on the sample retention rate and estimated retention rate corresponding to the sample video set, a target loss value is obtained;

[0307] Based on the weight constraints, the target loss value is used to adjust the weight parameters of the retention estimation model.

[0308] The device can be used to execute the video scheduling method provided in each embodiment of the present application. Therefore, the functions that can be implemented by each functional module of the device can be referred to the description of the aforementioned embodiments and will not be elaborated herein.

[0309] See also Fig.15 Based on the same technical concept, the present application also provides a computer device. In one embodiment, the computer device can be Figure 1 The server or terminal device shown in FIG. Fig.15 As shown, it includes a memory 1501 , a communication module 1503 and one or more processors 1502 .

[0310] The memory 1501 is used to store computer programs executed by the processor 1502. The memory 1501 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and programs required for running the instant messaging function, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0311] The memory 1501 may be a volatile memory, such as a random-access memory (RAM); the memory 1501 may also be a non-volatile memory, such as a read-only memory, a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); or the memory 1501 may be any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 1501 may be a combination of the above memories.

[0312] The processor 1502 may include one or more central processing units (CPU) or a digital processing unit, etc. The processor 1502 is used to implement the above-mentioned video scheduling method when calling the computer program stored in the memory 1501 .

[0313] The communication module 1503 is used to communicate with terminal devices or other servers.

[0314] The specific connection medium between the memory 1501, the communication module 1503 and the processor 1502 is not limited in the embodiment of the present application. Fig.15 In the embodiment, the memory 1501 and the processor 1502 are connected via a bus 1504. The bus 1504 is connected to the processor 1502 via a bus 1504. Fig.15 The connections between the other components are described with bold lines, which are only for illustration and are not intended to be limiting. The bus 1504 can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Fig.15 The diagram shows that only one thick line is used, but this does not mean that there is only one bus or only one type of bus.

[0315] The memory 1501 stores a computer storage medium, and the computer storage medium stores computer executable instructions. The computer executable instructions are used to implement the video scheduling method of the embodiment of the present application. The processor 1502 is used to execute the video scheduling method of each of the above embodiments.

[0316] Based on the same inventive concept, an embodiment of the present application also provides a storage medium, which stores a computer program. When the computer program runs on a computer, the computer executes the steps of the video scheduling method according to various exemplary embodiments of the present application described above in this specification.

[0317] In some possible implementations, various aspects of the video scheduling method provided by the present application may also be implemented in the form of a computer program product, which includes a computer program. When the program product is run on a computer device, the computer program is used to enable the computer device to execute the steps of the video scheduling method according to various exemplary embodiments of the present application described above in this specification. For example, the computer device may execute the steps of each embodiment.

[0318] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0319] The program product of the embodiment of the present application may adopt a portable compact disk read-only memory (CD-ROM) and include a computer program, and can be run on a computer device. However, the program product of the present application is not limited thereto. In the present application, the readable storage medium may be any tangible medium containing or storing a program, and the computer program included therein may be used by or in combination with a command execution system, apparatus or device.

[0320] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, wherein a readable computer program is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A readable signal medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with a command execution system, apparatus, or device.

[0321] The computer program embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0322] Computer programs for performing the operations of the present application may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" language or similar programming languages.

[0323] It should be noted that, although several units or subunits of the device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided into multiple units to be embodied.

[0324] In addition, although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0325] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0326] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0327] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A video scheduling method, characterized in that: The method comprises: Based on the multiple video information to be broadcasted, a plurality of different candidate video sets are obtained; wherein each candidate video set is used to indicate: at least one of the video information to be broadcasted in a preset time period; Based on at least one preset recommendation feature dimension, feature extraction is performed on at least one video information contained in each of the multiple candidate video sets to obtain recommendation feature sets corresponding to each of the multiple candidate video sets; wherein each recommendation feature dimension represents: a video attribute that causes a deviation in the estimated retention rate of the video platform; Based on the preset weight value of each recommended feature, the estimated retention rates corresponding to each of the plurality of recommended feature sets are obtained respectively; the preset weight value is determined during the model training process based on the preset weight constraint condition; the weight constraint condition indicates that the degree of influence of each recommended feature on the estimated retention rate meets the preset scheduling scenario condition; Based on the obtained estimated retention rates, a recommended video set is selected from the multiple candidate video sets, and the recommended video set is scheduled for broadcast.

2. The method according to claim 1, characterized in that The candidate video set includes multiple video information; The extracting features of at least one video information contained in each of the plurality of candidate video sets based on at least one preset recommendation feature dimension to obtain recommendation feature sets corresponding to each of the plurality of candidate video sets includes: For each candidate video set in the multiple candidate video sets, perform the following operations respectively: Based on at least one preset recommendation feature dimension, feature extraction is performed on each video information contained in a candidate video set to obtain corresponding independent recommendation features, each independent recommendation feature represents: an independent impact of a corresponding video information on the estimated retention rate; Perform feature cross extraction on every two independent recommendation features to obtain a corresponding cross recommendation feature set, where each cross recommendation feature represents: the superimposed impact of the corresponding two video information on the estimated retention rate; Based on the obtained independent recommendation features and the cross-recommendation feature set, a recommendation feature set corresponding to the candidate video set is obtained.

3. The method according to claim 2, characterized in that The extracting features of each video information contained in a candidate video set based on at least one preset recommendation feature dimension to obtain corresponding independent recommendation features includes: For each piece of video information, perform the following operations respectively: For a piece of video information, based on a plurality of preset recommendation feature dimensions, feature extraction is performed on the video information to obtain a plurality of recommendation sub-features corresponding to the video information, where the plurality of recommendation sub-features correspond one to one to the plurality of recommendation feature dimensions; Perform feature fusion processing on the multiple recommendation sub-features to obtain an independent recommendation feature corresponding to the one video information.

4. The method according to claim 2, characterized in that The step of obtaining the estimated retention rates corresponding to the plurality of recommendation feature sets based on the preset weight value of each recommendation feature includes: For the multiple recommended feature sets obtained, perform the following operations respectively: For a recommendation feature set, based on a first preset weight corresponding to each independent recommendation feature and a second preset weight corresponding to each cross recommendation feature, weighted fusion is performed on the recommendation feature set to obtain a corresponding recommendation fusion feature; Feature mapping is performed on the obtained recommended fusion features to obtain an estimated retention rate corresponding to the recommended feature set.

5. The method according to any one of claims 1 to 4, characterized in that: The weight constraint conditions include feature constraint conditions and dimension constraint conditions; The feature constraint condition indicates that: each recommended feature has a positive impact on the estimated retention rate; The dimension constraint condition indicates that each recommended sub-feature under each recommended feature dimension has a different degree of influence on the estimated retention rate.

6. The method according to claim 5, characterized in that The characteristic constraint conditions include single constraint conditions and cross constraint conditions; The single constraint condition indicates that each independent recommendation feature has a positive impact on the estimated retention rate, and the independent recommendation feature indicates the individual impact of a video information on the estimated retention rate; The cross constraint condition indicates that each cross recommendation feature has a positive impact on the estimated retention rate, and the cross recommendation feature indicates the superimposed impact of two video information on the estimated retention rate.

7. The method according to any one of claims 1 to 4, characterized in that: The selecting a recommended video set from the plurality of candidate video sets based on the obtained plurality of estimated retention rates includes any one of the following: Based on a preset retention threshold, selecting a target estimated retention rate from the multiple estimated retention rates, and using a target candidate video set associated with the target estimated retention rate as the recommended video set; Based on a preset video quantity condition, a target video set subset is screened out from the multiple candidate video sets, and based on a preset retention rate threshold, the recommended video set is selected from the target video set subset; the video quantity condition indicates that each target video set contains a preset amount of video information; Based on a preset video attribute condition, a target video set subset is screened out from the multiple candidate video sets, and based on a preset retention rate threshold, the recommended video set is selected from the target video set subset; The video attribute condition indicates that each target video set contains video information of preset video attributes.

8. The method according to any one of claims 1 to 4, characterized in that: The step of obtaining the estimated retention rates corresponding to the plurality of recommendation feature sets based on the preset weight value of each recommendation feature includes: Through the trained retention estimation model, perform the following operations on multiple recommended feature sets: For a recommended feature set, based on the preset weight value, weighted fusion is performed on the recommended feature set to obtain corresponding recommended fusion features; Performing feature mapping on the recommended fusion features to obtain an estimated retention rate of the recommended feature set; The retention prediction model is obtained by training a plurality of sample data, and each sample data includes: a sample video set, and a sample retention rate corresponding to the sample video set.

9. The method according to claim 8, characterized in that The training process of the retention estimation model includes: The retention prediction model to be trained is iteratively trained based on the sample data to obtain the trained retention prediction model; wherein each iterative training process includes the following operations: Based on the selected sample video set, obtain the corresponding estimated retention rate; Obtaining a target loss value based on the sample retention rate corresponding to the sample video set and the estimated retention rate; Based on the weight constraint condition, the target loss value is used to adjust the weight parameters of the retention estimation model.

10. A video scheduling device, characterized in that: The device comprises: An acquisition unit, configured to obtain a plurality of different candidate video sets based on a plurality of video information to be broadcasted; wherein each candidate video set is used to indicate: at least one of the video information to be broadcasted in a preset time period; The extraction unit is used to extract features of at least one video information contained in each of the plurality of candidate video sets based on at least one preset recommendation feature dimension, so as to obtain recommendation feature sets corresponding to each of the plurality of candidate video sets; wherein each recommendation feature dimension represents: a video attribute that causes a deviation in the estimated retention rate of the video platform; An estimation unit, for obtaining the estimated retention rates corresponding to each of the plurality of recommended feature sets based on a preset weight value of each recommended feature; the preset weight value is determined based on a preset weight constraint during model training; the weight constraint indicates that the degree of influence of each recommended feature on the estimated retention rate complies with a preset scheduling scenario condition; The broadcast scheduling unit is used to select a recommended video set from the multiple candidate video sets based on the obtained estimated retention rates, and schedule the recommended video set for broadcast.

11. A computer device, characterized in that: include: Memory for storing computer programs; A processor, configured to call a computer program stored in the memory, and execute the method according to any one of claims 1 to 9 according to the obtained program.

12. A computer-readable non-volatile storage medium, characterized in that: The computer-readable non-volatile storage medium stores a program, and when the program is run on a computer, the computer is enabled to implement the method according to any one of claims 1 to 9.

13. A computer program product, characterized in that The method comprises a computer program, wherein the computer program is stored in a computer-readable storage medium; when a processor of a computer device reads the computer program from the computer-readable storage medium, the processor executes the computer program, so that the computer device executes the method according to any one of claims 1 to 9.

Images