Prediction method and device of content potential, equipment and storage medium

Through multimodal fusion characteristics and the target prediction model of multiple classification prediction networks, the problem of the diversity of distribution quality points of information flow products such as short videos is solved, the efficiency and efficiency of content potential prediction is improved, and the model is enhanced Generalization and generalization ability.

CN119089373APending Publication Date: 2024-12-06TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310665467.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-06-06
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In the prior art, the distribution quality points of information flow products such as short videos vary, resulting in the need to train a large number of potential prediction models, which increases the difficulty of data sample collection and low model training efficiency, and reduces the potential prediction efficiency of information flow products. and distribution efficiency.

Method used

Using a target prediction model based on multimodal fusion features, training through multiple classification prediction networks, obtain the potential prediction results of the content to be estimated at multiple interaction levels, and use the correlation between different tasks to be reduced modeling. Cost and improve the generalization effect of the model.

Benefits of technology

It realizes unified modeling of multiple interaction levels, reduces modeling costs, improves potential estimate efficiency and distribution efficiency, and enhances the universality and generalization capabilities of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089373A_ABST
    Figure CN119089373A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of information, in particular to a content potential estimation method and device, equipment and a storage medium, and the method comprises the steps: obtaining a corresponding multi-modal fusion feature based on a basic feature and an attribute feature of a to-be-estimated content; the basic feature at least represents a content category to which the to-be-estimated content belongs, and the attribute feature represents a multi-modal content feature of the to-be-estimated content; and adopting a plurality of classification estimation networks contained in the target estimation model to respectively obtain potential estimation results of the to-be-estimated content in a plurality of interaction levels based on the multi-modal fusion features. Therefore, unified modeling is carried out on the to-be-estimated content in the dimensions of multiple interaction levels, the modeling cost is reduced, the generalization effect of a single model is improved by using the correlation between different tasks, and meanwhile, the universality of the target estimation model is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0002] With the rapid development of the Internet, the upload volume of information flow products such as short videos has increased exponentially. For content consumers, information flow products are suitable for the screens of mobile terminals, and they can continuously obtain fresh content by sliding their fingers up and down like water; and for content producers, information flow products are a very good product form for obtaining value benefits.

[0003] However, with the development and growth of various content creation organizations such as self-media, while bringing massive information flow products to the market, the quality of their content has also become uneven. In order to allow content consumers to obtain more content of interest and at the same time allow content producers to obtain more value benefits, it is necessary to estimate the distribution quality points such as the likes potential, comment potential and attention potential corresponding to information flow products such as short videos, so as to determine the distribution strategy of information flow products.

[0004] Under the relevant technology, since the distribution quality points of information flow products such as short videos are different, in order to specifically estimate the different potentials of information flow products such as short videos, it is usually necessary to train corresponding potential estimation models for different distribution quality points. In this way, it is difficult to collect data samples, resulting in a huge workload for building a model alone and low efficiency of model training.

[0005] For example, in order to determine whether a short video has the potential to be liked by content consumers, a like rate prediction model 1 is trained; in order to determine whether a short video has the potential to be commented on by content consumers, a comment rate prediction model 2 is trained; in order to determine whether the content producer corresponding to the short video has the potential to be followed by content consumers, an attention rate prediction model 3 is trained, and so on.

[0006] Therefore, when there are many categories of distribution quality points of information flow products, a large number of potential estimation models need to be trained. Moreover, since a single potential estimation model can only estimate one category of distribution quality points, multiple potential estimation models corresponding to different distribution quality points are needed to complete the estimation of each distribution quality point, which reduces the potential estimation efficiency of information flow products and further reduces the distribution efficiency of information flow products. Summary of the invention

[0007] The present application provides a method for estimating content potential and a related device, which are used to improve the efficiency of potential estimation for each content distribution quality point during the content distribution process.

[0008] In a first aspect, an embodiment of the present application provides a method for estimating content potential, the method comprising:

[0009] Based on the basic features and attribute features of the content to be estimated, corresponding multimodal fusion features are obtained; wherein the basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated;

[0010] The target estimation model includes multiple classification estimation networks, and based on multimodal fusion features, the potential estimation results of the content to be estimated at multiple interaction levels are obtained; wherein, one classification estimation network corresponds to a potential estimation task at an interaction level, and one classification estimation network outputs a potential estimation result at an interaction level, and the number of layers of each classification estimation network is positively correlated with the complexity of the corresponding potential estimation task. The target estimation model is obtained based on the fitting training of the loss functions corresponding to each of the multiple classification estimation networks.

[0011] In a second aspect, an embodiment of the present application further provides a device for estimating content potential, the device comprising:

[0012] The first processing module is used to obtain corresponding multimodal fusion features based on basic features and attribute features of the content to be estimated; wherein the basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated;

[0013] The second processing module is used to adopt multiple classification prediction networks included in the target estimation model to obtain potential estimation results of the content to be estimated at multiple interaction levels based on multimodal fusion features; wherein, one classification estimation network corresponds to a potential estimation task at an interaction level, and one classification estimation network outputs a potential estimation result at an interaction level, and the number of layers of each classification estimation network is positively correlated with the complexity of the corresponding potential estimation task, and the target estimation model is obtained based on the fitting training of the loss functions corresponding to each of the multiple classification estimation networks.

[0014] In a possible implementation, after respectively obtaining the potential estimation results of the content to be estimated at multiple interaction levels, the second processing module is further configured to:

[0015] Obtain the business objectives to be optimized for the content to be estimated, as well as the historical interaction data of multiple candidate distribution objects at multiple interaction levels;

[0016] Based on the potential estimation results of the target interaction level corresponding to the business goal to be optimized and the obtained historical interaction data, the interaction evaluation values ​​of multiple candidate objects at the target interaction level for the content to be estimated are obtained;

[0017] Based on the interaction evaluation value, at least one target distribution object of the content to be estimated is determined.

[0018] In a possible implementation, when the target estimation model is obtained based on the loss functions corresponding to the multiple classification estimation networks, the second processing module is used to:

[0019] Based on multiple historical contents whose cumulative exposure values ​​are greater than a preset exposure threshold, a training sample set is constructed, and interaction logs of each consumer object associated with each sample content in the training sample set are obtained;

[0020] Based on each interaction log, determine the sample interaction labels of each sample content at multiple interaction levels;

[0021] Using the training sample set and the interaction labels of each sample, as well as the corresponding loss functions of multiple classification estimation networks, the estimation model to be trained is trained for multiple rounds of iterative training to obtain the target estimation model.

[0022] In a possible implementation, the estimation model includes at least a first feature network, a second feature network, and a plurality of classification estimation networks;

[0023] During one round of iterative training of the estimation model, the first processing module and the second processing module perform the following operations:

[0024] Extracting a sample basic feature of the sample content based on the first feature network, and extracting a sample attribute feature of the sample content based on the second feature network;

[0025] Based on the basic features and attribute features of the samples, the multimodal fusion features of the samples are obtained, and multiple classification estimation networks are used to obtain the sample potential estimation results of the sample content at multiple interaction levels based on the multimodal fusion features of the samples;

[0026] Based on the loss functions corresponding to multiple classification estimation networks, the loss values ​​between the sample potential estimation results at multiple interaction levels and the corresponding sample interaction labels are obtained, and the parameters of the estimation model are adjusted based on the obtained loss values.

[0027] In one possible implementation, the number of layers of each classification estimation network is obtained in the following way:

[0028] For each potential estimation task, perform the following operations:

[0029] Determine a task relevance category set for a task identifier corresponding to the first potential estimation task; wherein the first potential estimation task is any one of the potential estimation tasks, and the task relevance category corresponding to the first potential estimation task represents: a degree of relevance between multiple potential estimation tasks in a potential estimation task set to which the first potential estimation task belongs;

[0030] Determining the complexity level corresponding to the first potential estimation task based on the task relevance category and the correspondence between the preset task relevance category and the complexity level;

[0031] Based on the complexity corresponding to the first potential estimation task, the number of layers of the classification estimation network corresponding to the potential estimation task is set.

[0032] In a possible implementation, based on each interaction log, sample interaction tags of each sample content at multiple interaction levels are determined, and the first processing module is used to:

[0033] For each interaction level, perform the following operations on each sample content:

[0034] Based on the interaction logs of each consumer object associated with a sample content, the interaction rate of a sample content at the current interaction level is obtained;

[0035] Obtain the average interaction rate of all historical content at the current interaction level under the content category to which a sample interactive content belongs;

[0036] If the interaction rate corresponding to a sample content at the current interaction level is greater than the average interaction rate, the sample interaction label of the sample content is determined to be a positive sample, otherwise, the sample interaction label of the sample content is determined to be a negative sample.

[0037] In a possible implementation, based on each interaction log, sample interaction tags of each sample content at multiple interaction levels are determined, and the first processing module is used to:

[0038] For each interaction level, perform the following operations on each sample content:

[0039] Based on the interaction logs of each consumer object associated with a sample content, the interaction rate of a sample content at the current interaction level is obtained;

[0040] Obtain the historical interaction rates of all similar sample contents under the content category to which a sample content belongs, and divide the similar sample contents into multiple interaction levels according to the preset division rules;

[0041] The target interaction level to which the interaction rate corresponding to a sample content at the current interaction level belongs is used as a sample interaction label of the sample content at the current interaction level.

[0042] In a possible implementation, after adjusting the parameters of the estimation model based on the obtained loss values, the first processing module and the second processing module are further used to:

[0043] If multiple classification estimation networks meet the convergence conditions, the training is stopped and the target estimation model is obtained;

[0044] If there is a target network that does not meet the convergence conditions among multiple classification estimation networks, then based on the training sample set and the interaction labels of each sample, and the loss function corresponding to the target network, when the estimation model to be trained is subsequently iteratively trained, the parameters of the target network are adjusted, and the parameters of the first feature network, the second feature network, and the converged network that meets the convergence conditions among the multiple classification estimation networks are kept unchanged.

[0045] In a possible implementation, the classification prediction network is composed of one or more multi-layer perceptrons (MLPs).

[0046] In a third aspect, an embodiment of the present application provides a computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any one of the first aspects when executing the computer program.

[0047] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any method described in the first aspect.

[0048] In a fifth aspect, an embodiment of the present application provides a computer program product, which, when called by a computer, enables the computer to execute the method described in the first aspect.

[0049] The beneficial effects of this application are as follows:

[0050] In the embodiment of the present application, a method, device, equipment and storage medium for estimating content potential are proposed. The server obtains corresponding multimodal fusion features based on basic features and attribute features of the content to be estimated; the basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated; multiple classification estimation networks included in the target estimation model are used to obtain potential estimation results of the content to be estimated at multiple interaction levels based on the multimodal fusion features;

[0051] In different classification prediction networks, different element weights are set for multimodal fusion features, and multi-task classification can be performed on the same multimodal fusion feature to achieve potential estimation of different interactive levels of the content to be estimated. In this way, unified modeling is performed for the dimensions of the content to be estimated at multiple interactive levels, which reduces the cost of modeling, and also utilizes the correlation between different tasks to increase the generalization effect of a single model, while improving the versatility of the target estimation model.

[0052] 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

[0053] Figure 1 A schematic diagram of an application scenario of an embodiment of the present application;

[0054] Figure 2A A schematic diagram of the architecture of a target estimation model provided in an embodiment of the present application;

[0055] Figure 2B This is a structural diagram of a first characteristic network and a second characteristic network in an embodiment of the present application;

[0056] Figure 3 A flow chart of a training method for a target estimation model provided in an embodiment of the present application;

[0057] Figure 4 A user interface diagram of a content platform in an embodiment of the present application;

[0058] Figure 5 A first flow chart for determining a sample interaction tag in an embodiment of the present application;

[0059] Figure 6 A second flow chart for determining sample interaction tags in an embodiment of the present application;

[0060] Figure 7 This is a schematic diagram of an interactive gear division in an embodiment of the present application;

[0061] Figure 8 A detailed flowchart of the training estimation model in the embodiment of the present application;

[0062] Fig. 9 This is a schematic diagram of the division of vermicelli gears in an embodiment of the present application;

[0063] Fig.10 A multi-layer perceptron structure diagram provided in an embodiment of the present application;

[0064] Fig.11 A flow chart of a method for setting the number of layers of each classification estimation network in an embodiment of the present application;

[0065] Fig.12 This is a flow chart of a method for estimating content potential in an embodiment of the present application;

[0066] Fig.13 A schematic diagram of a scenario for performing potential estimation on content to be estimated in an embodiment of the present application;

[0067] Fig.14 A schematic diagram of a process for obtaining a potential estimation result based on a classification estimation network in an embodiment of the present application;

[0068] Fig.15 This is a flow chart of a content recommendation method in an embodiment of the present application;

[0069] Fig.16 A structural diagram of a recommendation system provided in an embodiment of the present application;

[0070] Fig.17 It is a structural schematic diagram of a device for estimating content potential in an embodiment of the present application;

[0071] Fig.18 A schematic diagram of a hardware structure of a computing device in an embodiment of the present application;

[0072] Fig.19 A schematic diagram of the hardware composition structure of another computing device in an embodiment of the present application. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the technical solution of the present application, rather than all of the embodiments. Based on the embodiments recorded in the application documents, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the technical solution of the present application.

[0074] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0075] Some of the terms used in the embodiments of the present application are explained below to facilitate understanding by those skilled in the art.

[0076] Multi-Channel Network (MCN): is a product form of a multi-channel network that unites PGC content and brings stable income to content creators through reasonable traffic distribution.

[0077] Professional Generated Content (PGC): For example, video websites, etc. It is used to refer to content personalization, diversified perspectives, virtual social relationships, etc.

[0078] User Generated Content (UGC) is the original content created by users and displayed or provided to other users through the Internet platform.

[0079] Content: refers to media information in the form of text, images, videos, sounds, animations, etc.; for example, videos, short videos, graphics, etc.

[0080] Attribute feature: It is feature information that integrates at least one of the audio information, text information, and visual information of the content; and the attribute feature is a vectorized Embedding feature.

[0081] One-hot encoding: One-hot encoding, also known as one-bit effective encoding, is a method of using an N-bit state register to encode N states. Each state has its own independent register bit, and at any time, only one of them is valid.

[0082] Machine Learning (ML): It is a multi-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. 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.

[0083] Deep learning: Deep learning is a form of machine learning that originates from the study of artificial neural networks. It consists of multiple input, output, and hidden layers. Each layer contains units that convert input data into information for the next layer to use for specific prediction tasks. Thanks to this structure, the machine can learn through its own data processing.

[0084] Multi-Layer Perceptron (MLP): is a directional structured artificial neural network that maps a set of input vectors to a set of output vectors. MLP can be viewed as a directed graph consisting of multiple node layers, each layer is fully connected to the next layer, and except for the input node, each node is a neuron (or processing unit) with a nonlinear activation function.

[0085] The following is a brief introduction to the design concept of the embodiment of the present application:

[0086] With the development and growth of various content creation organizations such as self-media, more and more information flow products are pouring into the market. The essence of information flow products is content distribution, that is, accurately distributing and pushing content according to the preferences of content consumers. Before distributing content, it is necessary to first stratify the quality of the content itself, so as to effectively distribute high-quality content and reduce unnecessary traffic waste.

[0087] One of the evaluation criteria for content quality is distribution quality, and the distribution quality points included in distribution quality mainly include: attention potential, sharing potential, comment potential, like potential, collection potential, etc.

[0088] Under the relevant technology, since the distribution quality points of information flow products such as short videos are different, in order to specifically estimate the different potentials of information flow products such as short videos, it is usually necessary to train corresponding potential estimation models for different distribution quality points. In this way, it is difficult to collect data samples, resulting in a huge workload for building a model alone and low efficiency of model training.

[0089] However, the differences in the features of different model inputs are not large. For example, the potential for likes and collections are often related to the information contained in the content itself, and the potential for attention is related to the identity of the content creator. High-quality content creators also have certain commonalities, and high-quality content creators are more likely to create better quality content. Therefore, there is a certain correlation between different tasks. Unified modeling of content at various distribution quality points can reduce the cost of modeling, increase the generalization effect of a single model, and be more in line with actual needs.

[0090] In view of this, in the embodiments of the present application, a method, apparatus, device and storage medium for estimating the potential of content are proposed.

[0091] In an embodiment of the present application, a target estimation model is provided, which includes a first feature network, a second feature network and multiple classification estimation networks, the first feature network and the second feature network are respectively used to extract and process basic features and attribute features of content to be estimated, and obtain corresponding multimodal fusion features through the basic features and attribute features; wherein the basic features at least characterize the content category to which the content to be estimated belongs, and the attribute features characterize the multimodal content characteristics of the content to be estimated; the multiple classification estimation networks are used to obtain potential estimation results of the content to be estimated at multiple interaction levels based on the multimodal fusion features.

[0092] In different classification prediction networks, different element weights are set for multimodal fusion features, so that multi-task classification can be performed on the same multimodal fusion features to achieve potential prediction of different contents to be estimated at the interaction level.

[0093] In this way, unified modeling is performed on the dimensions of the content to be estimated at multiple interaction levels, which reduces the cost of modeling, utilizes the correlation between different tasks, increases the generalization effect of a single model, and improves the versatility of the target estimation model.

[0094] In the embodiments of the present application, the estimation model and the target estimation model involve artificial intelligence (AI) and machine learning technology, and are designed based on speech technology, natural language processing technology and machine learning (ML) in artificial intelligence.

[0095] 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.

[0096] 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.

[0097] 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.

[0098] 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 strategies during its interaction with the environment.

[0099] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In addition, the embodiments of the present application and the features in the embodiments may be combined with each other if there is no conflict.

[0100] See also Figure 1 As shown, it is a schematic diagram of an application scenario of an embodiment of the present application.

[0101] The application scenario includes a terminal device 110 and a server 120 , and the terminal device 110 and the server 120 can communicate with each other through a communication network.

[0102] In an optional implementation, the communication network may be a wired network or a wireless network. Therefore, the terminal device 110 and the server 120 may be directly or indirectly connected via a wired or wireless communication. For example, the terminal device 110 may be indirectly connected to the server 120 via a wireless access point, or the terminal device 110 may be directly connected to the server 120 via the Internet, and this application does not limit this.

[0103] In the embodiment of the present application, the terminal device 110 includes but is not limited to mobile phones, tablet computers, laptop computers, desktop computers, e-book readers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals and other devices; various clients may be installed on the terminal device, and the client may be an application (such as a browser, game software, etc.), or a web page, a small program, etc.;

[0104] The server 120 is a background server corresponding to the client installed in the terminal device 110. The server 120 may 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 (CDNs), and big data and artificial intelligence platforms.

[0105] It should be noted that the content potential estimation method in the embodiment of the present application can be deployed in a computing device, which can be a server or a terminal device, wherein the server can be Figure 1 The server 120 shown in FIG. 1 may also be Figure 1 Other servers other than the server 120 shown in FIG. 1 ; the terminal device may be Figure 1 The terminal device 110 shown in FIG. 1 may also be Figure 1Other terminal devices other than the terminal device 110 shown in FIG. That is, the method can be executed by a server or a terminal device alone, or by a server and a terminal device together.

[0106] When the terminal device executes alone, for example, the terminal device may obtain the content to be estimated, and perform potential estimation on the content to be estimated based on the trained target estimation model, thereby obtaining potential estimation results of the content to be estimated at multiple interaction levels.

[0107] When the server executes alone, for example, the server may obtain the content to be estimated, and perform potential estimation on the content to be estimated based on the trained target estimation model, thereby obtaining potential estimation results of the content to be estimated at multiple interaction levels.

[0108] When the server and the terminal device execute together, for example, the terminal device can obtain the content to be estimated; then, the terminal device transmits the obtained content to be estimated to the server, and the server estimates the potential of the content to be estimated based on the trained target estimation model, and obtains the potential estimation results of the content to be estimated at multiple interaction levels.

[0109] It should be noted that, in the following, the example of execution by a server alone is mainly used for illustration, and no specific limitation is made here.

[0110] Figure 1 The figure is only an example. In fact, the number of terminal devices 110 and servers 120 is not limited and is not specifically limited in the embodiments of the present application.

[0111] In the embodiment of the present application, when there are multiple servers 120, the multiple servers 120 can form a blockchain, and the servers 120 are nodes on the blockchain; as disclosed in the embodiment of the present application, the method for estimating the content potential, the processing method of the potential estimation of the target estimation model involved, etc. can be saved on the blockchain.

[0112] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In addition, the embodiments and features in the embodiments of the present application may be combined with each other if there is no conflict.

[0113] In an embodiment of the present application, the server performs potential estimation on the content to be estimated based on the target estimation model. In order to ensure the accuracy of the potential estimation results, before performing potential estimation on the content to be estimated, the estimation model to be trained is trained based on a training sample set composed of historical content to obtain a target estimation model.

[0114] First, the architecture of the target estimation model is briefly introduced.

[0115] For example, see Figure 2A As shown, it is a schematic diagram of the architecture of a target estimation model provided in an embodiment of the present application, and the target estimation model includes a first feature network, a second feature network and multiple classification estimation networks.

[0116] The first feature network is mainly used to process some discrete features in the form of one-hot encoding, explore the co-occurrence in historical data, and improve the memory capacity of the model; the second feature network is mainly used to process embedded continuous features, explore new feature combinations that have hardly appeared in historical data based on migration correlation, and improve the generalization ability of the model; the weight parameters corresponding to multiple classification prediction networks are different, and they can output different types of potential prediction results after calculating the same input features.

[0117] See also Figure 2B As shown, it is an architecture diagram of a first feature network and a second feature network in an embodiment of the present application. The architecture refers to the Deep&Wide model architecture and is divided into two parts: the Wide part on the left and the Deep part on the right, wherein the Wide side is composed of a single-layer neural network, and an ordinary generalized linear model can be adopted, such as a logistic regression model (Logistic Regression, LR), and its input features are in one-hot form; the Deep side is composed of a multi-layer neural network, and the input features are embedded in the first layer and sent to the second layer for forward propagation.

[0118] It is worth noting that in the embodiments of the present application, Figure 2B The output results corresponding to the model architecture shown are not used as the corresponding potential estimation results. Instead, the features processed by Wide and Deep are fused and sent to multiple classification estimation networks to obtain different types of potential estimation results.

[0119] In an embodiment of the present application, the content processed by the target estimation model is video content. For the video content to be estimated, the target estimation model can output potential estimation results at multiple interaction levels, including but not limited to: video view count (Video View, VV) estimation, completion estimation, fast scroll rate estimation, sharing rate estimation, like rate estimation, attention rate estimation and comment rate estimation.

[0120] Then, the training process of the estimation model is introduced in detail.

[0121] In the embodiment of the present application, Figure 2A and Figure 2B The target estimation model shown in FIG. 1 provides a training method for a target estimation model. Figure 3As shown, it is a flow chart of a training method for a target estimation model provided in an embodiment of the present application. Figure 3 , and explain in detail the specific operations performed:

[0122] Step S301: construct a training sample set based on a plurality of historical contents whose cumulative exposure values ​​are greater than a preset exposure threshold, and obtain an interaction log of each consumer object associated with each sample content in the training sample set.

[0123] Specifically, in the embodiment of the present application, when constructing a training sample set, in order to ensure that the influence results of the sample content are credible, it is necessary to screen out video content with a cumulative exposure value greater than 300 (the exposure threshold in the embodiment of the present application is 300) as sample content, and in order to ensure that the training sample set has a sufficient scale, data is usually accumulated within the time window of the validity period of the video content (such as 3 months).

[0124] When users consume video content, corresponding interaction logs are generated to determine the degree of interaction when users consume video content. The data format is usually (t, uid i ,vid v , exp i,v ,clk i,v , comment i,v , share i,v , great i,v , collect i,v ), where t is the timestamp of behavior reporting, uid i For user i, vid v is the video v, exp i,v is the number of times video v is exposed to user i, clk i,v is the number of times user i clicks to play video v, comment i,v is the number of comments made by user i on video v, share i,v is the number of times user i shares video v, great i,v is the number of likes given by user i to video v, collect i,v is the number of times user i has collected video v.

[0125] After obtaining the interaction logs associated with all sample contents, group all users’ interaction logs according to the content categories (first-level categories) to which the sample contents belong, aggregate the interaction data such as exposure times, clicks, comments, shares, likes, favorites, and swipes, and obtain the proportion of all interaction data in the number of times the corresponding sample contents are played, i.e., the posterior data of VV, completion rate, swipe rate, share rate, like rate, attention rate, and comment rate. According to the content category to which the sample contents belong, count the relevant posterior data corresponding to all sample contents under the category.

[0126] For example, see Figure 4 As shown, it is a user interface diagram of a content platform in an embodiment of the present application. In this interface, the video content includes video content 1 and video content 2, as well as four interactive behavior records of collection, comment, like and share obtained by the users of video content 1 and video content 2 respectively (the number of exposures and the number of clicks to play are displayed in the user interface). The record is obtained through the statistics of the interactive logs generated when each user consumes the video content.

[0127] Furthermore, according to the interaction logs of all users in the "Martial Arts" category, the posterior data corresponding to all video contents in this category are obtained, including the VV, completion rate, fast scroll rate, sharing rate, like rate, attention rate and comment rate corresponding to video content 1 and video content 2 in the "Martial Arts" category.

[0128] Optionally, when obtaining sample content, you can obtain sample content for the same content platform, or you can use a web crawler to crawl corresponding content and corresponding interaction logs from various content platforms, and filter data according to the above rules as supplementary samples, so that the distribution of sample content is more extensive and the constructed model has better generalization characteristics.

[0129] Step S302: Based on each interaction log, determine the sample interaction tags of each sample content at multiple interaction levels.

[0130] Specifically, in the embodiments of the present application, when determining the sample interaction labels of each sample content at multiple interaction levels, a classification method, a regression method, or a combination of a classification method and a regression method may be used, and the present application does not impose any restrictions on this.

[0131] See also Figure 5 As shown, it is a first flow chart for determining sample interaction tags in an embodiment of the present application. Figure 5 , the specific steps of determining sample interaction labels by classification are described in detail:

[0132] For each interaction level, perform the following operations on each sample content:

[0133] Step S3021-A: Based on the interaction logs of each consumer object associated with a sample content, obtain the interaction rate of a sample content at the current interaction level.

[0134] For example, assuming that the current interaction level is likes, based on the method described in step S301, through the interaction logs of each user associated with video content 1, the like rate of video content 1 is obtained to be 65.2%.

[0135] Step S3022-A: Obtain the average interaction rate of all historical contents at the current interaction level under the content category to which a sample interactive content belongs.

[0136] If the content category to which video content 1 belongs is "martial arts", then the like rates of all historical content under the "martial arts" category are obtained, and the average like rate of each historical content is calculated, where the historical content here can be all "martial arts" video content that has been published on the content platform, or it can be "martial arts" video content that has been published and meets the aforementioned exposure rate greater than the preset exposure threshold, or it can be the "martial arts" video content that has been initially filtered by other filtering methods for the historical content that has been published, such as: filtering out video content with a total duration of less than 3 seconds; filtering out video content with a single playback duration of less than 10% of the total duration, etc. This application does not impose any restrictions on this.

[0137] Step S3023-A: If the interaction rate corresponding to a sample content at the current interaction level is greater than the average interaction rate, the sample interaction label of the sample content is determined to be a positive sample, otherwise, the sample interaction label of the sample content is determined to be a negative sample.

[0138] Assuming that the average like rate of the "martial arts" video content that has been released is 51.3%, and the like rate of video content 1 is 65.2%, which is greater than the average like rate of similar video content, then video content 1 is a positive sample in terms of likes.

[0139] Optionally, at the likes level, after determining that the sample interaction label of video content 1 is a positive sample, the same method can be used to determine the sample interaction labels of video content 1 at other interaction levels. That is to say, in addition to being a positive sample at the likes level, video content 1 can also be used as a positive sample or negative sample at the attention level, collection level or other interaction levels. This application does not impose any restrictions on this.

[0140] On the other hand, see Figure 6 As shown, it is a second flow chart for determining sample interaction tags in the embodiment of the present application. Figure 6 , the specific steps of determining sample interaction labels by regression are described in detail:

[0141] For each interaction level, perform the following operations on each sample content:

[0142] Step S3021-B: Based on the interaction logs of each consumer object associated with a sample content, the interaction rate of a sample content at the current interaction level is obtained.

[0143] Similarly, assuming that the current interaction level is comment, based on the method described in step S301, through the interaction logs of each user associated with the video content 2, the comment rate of the video content 2 is obtained to be 73.1%.

[0144] Step S3022-B: Obtain the historical interaction rates of all similar sample contents under the content category to which a sample content belongs, and divide the similar sample contents into multiple interaction levels according to a preset division rule.

[0145] Specifically, if the content category to which video content 2 belongs is "martial arts", then the historical comment rates of all similar sample contents under the "martial arts" category are obtained, and the similar sample contents are sorted according to the size of each historical comment rate, and each similar sample content is divided into multiple interaction levels.

[0146] For example, see Figure 7 As shown, it is a schematic diagram of an interactive gear division in an embodiment of the present application. Sample content with a comment rate greater than 90% is classified as the first gear, sample content with a comment rate in the range of 75% to 90% is classified as the second gear, and sample content with a comment rate in the range of 55% to 75% is classified as the third gear…

[0147] Step S3023-B: taking a target interaction level to which an interaction rate corresponding to a sample content at the current interaction level belongs as a sample interaction label of the sample content at the current interaction level.

[0148] In combination with step S3021-B and step S3022-B, since the comment rate of video content 2 is 73.1%, the sample interaction label of video content 2 at the comment level is the third tier.

[0149] Optionally, all sample contents may be equally divided into two parts, one part using a classification method to determine sample interaction labels of the sample contents, and the other part using a regression method to determine sample interaction labels of the sample contents.

[0150] Step S303: using the training sample set and the interaction labels of each sample, as well as the loss functions corresponding to each of the multiple classification estimation networks, the estimation model to be trained is subjected to multiple rounds of iterative training to obtain a target estimation model.

[0151] Specifically, in the embodiment of the present application, the first feature network and the second feature network together form a Deep&Wide model architecture, and multiple classification estimation networks form a multi-task classification regression layer. Figure 8 As shown, it is a detailed flow chart of the training prediction model in the embodiment of the present application. Figure 8 , and explain the specific execution steps in detail:

[0152] During one round of iterative training of the estimated model, the following operations are performed:

[0153] Step S3031: extracting a sample basic feature of the sample content based on the first feature network, and extracting a sample attribute feature of the sample content based on the second feature network.

[0154] In an embodiment of the present application, the first feature network is the Wide part in the Deep&Wide model architecture, which is composed of an artificial neural network and a neural network transformer based on a self-attention mechanism, and performs feature extraction and processing on the basic information of the content.

[0155] The basic information of the content mainly includes: the account information of the content creator, the content category to which the content belongs, and meta-information related to the content publishing.

[0156] The account information of content creators mainly includes account level, account fan level and account statistics.

[0157] Among them, account level refers to the experience growth level of the content creator's account from registration or certification to the present, which is usually divided into levels 1-5, and level 5 accounts are usually some authoritative media accounts.

[0158] Account fan tiers are usually divided into multiple tiers according to the number of fans, reflecting the popularity of the account. For example, see Fig. 9 As shown in FIG. 1 , it is a schematic diagram of fan level division in an embodiment of the present application. For an account with less than 1,000 fans, its account fan level is level 0, and for an account with a fan level between 1,000 and 10,000 fans, its account fan level is level 1…

[0159] Account statistics mainly include account dimension enabled content, such as the number of comments obtained in the past day, the number of comments obtained historically, the number of postings in the past day, the historical number of postings, the number of shares obtained in the past day, the historical number of shares obtained, the number of exposures obtained in the past day, the historical exposures obtained, the number of follows in the past day, the historical number of follows, the number of followers obtained in the past day, the historical number of followers obtained, etc.

[0160] The content categories to which the content belongs are divided into primary, secondary and tertiary categories. For example, the primary categories are: movies, food, games, etc.; the secondary categories of the primary category "movies" are: movie clips, movie commentary, etc.; the secondary categories of the secondary category "movie clips" are: starring star A, directed by director B, etc.

[0161] Content publishing related meta-information mainly includes: content publishing time, content size, duration and other information.

[0162] In addition, the basic information of the content may also include the account nickname of the content creator, etc., which is not limited in this application.

[0163] The first feature network obtains the sample basic features in one-hot encoding form from the above basic information, and performs a series of operations on the sample basic features in combination with various initial parameters in the first feature network to obtain a first output result.

[0164] In an embodiment of the present application, the second feature network is the Deep part in the Deep&Wide model architecture, which is composed of a cross-modal transformer and performs feature extraction and processing on multimodal content characteristics consisting of visual modality, audio modality and text modality of the content.

[0165] (1) Visual modality

[0166] In an embodiment of the present application, the video image and time information are input into the Video Swin Transformer network to extract the visual features of the video content, and the two dimensions of the video image and time information are input at the same time to make the determined visual features more accurate.

[0167] (2) Audio mode

[0168] The pre-trained wavlm-base-plus is used as a feature extractor to realize the conversion of audio features into vectors. The wavlm-base-plus is trained on 94,000 hours of unsupervised English data and has achieved good results on multiple speech-related datasets, making the determined audio features more accurate.

[0169] (3) Text Mode

[0170] The text modality includes content title, content tag and content category. In the embodiment of the present application, a pre-trained BERT model based on large-scale text prediction of information flow is used to extract the semantic features of the text content.

[0171] The second feature network obtains corresponding features from the visual modality, audio modality and text modality respectively, cross-fused and mapped to the same vectorized embedding space to obtain sample attribute features, and combined with the initial parameters in the second feature network, performs a series of operations on the basic features to obtain the second output result.

[0172] Step S3032: Obtain sample multimodal fusion features based on sample basic features and sample attribute features, and adopt multiple classification estimation networks to obtain sample potential estimation results of sample content at multiple interaction levels based on sample multimodal fusion features.

[0173] Specifically, based on the first feature network and the second feature network, the sample basic features and the sample attribute features are processed respectively to obtain the first output result and the second output result are feature fused to obtain the sample multimodal fusion features, and the sample multimodal fusion features are input into the classification and regression layer, and seven classification prediction networks are used to obtain the sample potential prediction results of the sample content at seven interaction levels based on the sample multimodal fusion features.

[0174] Among them, the seven classification prediction networks are composed of one or more multi-layer perceptrons (MLPs), and the parameters of the MLPs in different classification prediction networks are different. The relevant parameters in the MLP in each classification prediction network only serve the corresponding tasks and have no effect on other tasks.

[0175] See also Fig.10 As shown in FIG. 1 , it is a multi-layer perceptron structure diagram provided in an embodiment of the present application, wherein the bottom layer is the input layer, the middle layer is the hidden layer, and the last layer is the output layer. If the feature vector of the MLP input is an n-bit vector, then the input layer has n neuron nodes ( Fig.10 3 neuron nodes are drawn as an example in the figure), the hidden layer is fully connected to the input layer, all neuron nodes have the functions of input, output and data storage, and each connection between adjacent layers contains different weights, indicating the importance of the corresponding neuron node in the previous layer, which is generally saved in matrix form.

[0176] In the embodiment of the present application, the input feature of MLP is the multimodal fusion feature of the sample. Assuming that the input of the input layer is represented by vector X (vector X contains vector elements x1, x2, x3), the output of the hidden layer is S=f(W1X+b1), wherein W1 is the weight of the hidden layer, b1 is the bias parameter of the hidden layer, and function f is the activation function, and the sigmoid function or the tanh function is usually selected.

[0177] The output layer can be regarded as a logistic regression of a classification task, namely softmax regression, so the output of the output layer is y=softmax(W2S+b2), where S represents the output vector of the hidden layer (vector S contains vector elements s1, s2, s3).

[0178] In the above calculation process, W1 and W2 are parameters that control the importance of the input signal, while the bias parameters b1 and b2 are parameters that adjust the ease with which the neuron nodes are activated. Normally, the neuron nodes will be activated only when the weighted sum of the input signals exceeds 1. Sometimes, the parameters W1, W2, b1 and b2 are collectively referred to as the weight parameters of MLP.

[0179] In an embodiment of the present application, based on the output results of the seven classification prediction networks, the sample VV estimation, sample completion estimation, sample quick scroll rate estimation, sample sharing rate estimation, sample like rate estimation, sample attention rate estimation and sample comment rate estimation results corresponding to the current sample content are obtained.

[0180] Step S3033: Based on the loss functions corresponding to the multiple classification estimation networks, the loss values ​​between the sample potential estimation results at the multiple interaction levels and the corresponding sample interaction labels are obtained, and based on the obtained loss values, the parameters of the estimation model are adjusted.

[0181] Taking the VV estimation task as an example, the mean squared error (MSE) combined with the kernel density estimation (KDE) is used as the loss function of the VV estimation task. Among them, KDE is mainly used to solve the VV long tail problem and increase the loss weight of high VV samples.

[0182] Assume that the estimation model outputs a VV estimation result of sample content i as Y', and the sample interaction label corresponding to sample content i is Y, then the difference between the VV estimation result and the sample interaction label is:

[0183]

[0184] in, Used to calculate the proximity between Y' and the sample interaction label Y.

[0185] Specifically, in the embodiment of the present application, the underlying feature extraction backbone network part of the estimation model, namely the first feature network on the Wide side and the second feature network on the Deep side, share their structure and parameters.

[0186] Therefore, after adjusting the parameters of the estimation model once, if multiple classification estimation networks all meet the convergence conditions, the training is stopped and the target estimation model is obtained; if there is a target network that does not meet the convergence conditions among the multiple classification estimation networks, then based on the training sample set and the interaction labels of each sample, and the loss function corresponding to the target network, when the estimation model to be trained is subsequently iterated, the parameters of the target network are adjusted, and the parameters of the first feature network, the second feature network, and the converged network that meets the convergence conditions among the multiple classification estimation networks are kept unchanged.

[0187] For example, if the network of the sharing rate estimation task has reached a convergence state, but the rest of the network has not yet converged, in the next parameter adjustment process, in order not to affect the original convergence state of the network of the sharing rate estimation task, the parameters of the MLP part corresponding to the sharing part and the sharing rate estimation task can be frozen, and only the parameters of other classification estimation tasks themselves can be adjusted until the network corresponding to each classification estimation task has completed convergence. In practical applications, since each classification task is related, once the network of one classification task converges, the parameters of other parts only need a small amount of fine-tuning.

[0188] It should be noted that in the architecture of the above estimation model, the number of layers of each classification estimation network, that is, the number of MLP layers contained in each classification estimation network, is positively correlated with the complexity of each potential estimation task.

[0189] For details, see Fig.11 As shown, it is a flow chart of the method for setting the number of layers of each classification estimation network in the embodiment of the present application. Fig.11 , and explain the specific execution steps in detail:

[0190] For each potential estimation task, perform the following operations:

[0191] Step S1101: Determine the task relevance category set for the task identifier corresponding to the first potential estimation task.

[0192] The first potential estimation task is any one of the potential estimation tasks, and the task relevance category corresponding to the first potential estimation task represents: the degree of relevance between multiple potential estimation tasks in the potential estimation task set to which the first potential estimation task belongs.

[0193] Specifically, in an embodiment of the present application, when designing the structure of each classification estimation network in the estimation model, each potential estimation task can be divided into multiple potential estimation task sets according to the degree of correlation between the potential estimation tasks. The so-called degree of correlation refers to the overlap of the feature sources relied on during the execution of each potential estimation task. The higher the overlap of the feature sources relied on by any two potential estimation tasks, the higher their degree of correlation.

[0194] In some embodiments, when determining the degree of correlation between two potential estimation tasks, the degree of correlation between the sources of relevant features of the two potential estimation tasks can be calculated first, and the degree of correlation between the two potential estimation tasks can be determined according to the overlap interval where the overlap is located. In some embodiments, the potential estimation task set can be divided according to the correlation interval where the degree of correlation is located, and the potential estimation tasks with the degree of correlation located in the same interval are divided into one potential estimation task set. Among them, the overlap interval and the correlation interval can be set according to experience, and can also be determined according to multiple experiments, which are not limited here. In some embodiments, multiple potential estimation task sets can also be obtained by clustering according to the degree of correlation between the potential estimation tasks.

[0195] For example, the feature sources of some potential estimation tasks are mainly prior features, such as multimedia information conveyed by the content itself, content category, author characteristics, etc. The correlation levels of such potential estimation tasks, such as playback time estimation and completion rate estimation, are in the same interval and belong to the first potential estimation task set; the feature sources of some potential estimation tasks are mainly a posteriori features, such as user interaction data such as likes, shares, collections, and attention with content. The correlation levels of such potential estimation tasks, such as like rate estimation, share rate estimation, collection rate estimation, and attention rate estimation, are in the same interval and belong to the second potential estimation task set.

[0196] After different potential estimation task sets are divided based on the above method, the number of the potential estimation task set to which each potential estimation task belongs may be used as the task identifier corresponding to each potential estimation task.

[0197] Since there are many potential estimation tasks required in practical applications, when dividing the potential estimation task set, the degree of correlation can also be subdivided to obtain more potential estimation task sets, and this application does not impose any restrictions on this.

[0198] Step S1102: Based on the complexity corresponding to the task relevance category, the number of layers of the classification estimation network corresponding to the first potential estimation task is set.

[0199] Generally speaking, the fewer potential estimation tasks are included in the same potential estimation task set, the more reference information, that is, the more influencing factors, are required to execute the corresponding potential estimation tasks, and therefore the greater its complexity. Therefore, in the embodiment of the present application, the number of potential estimation tasks in the first potential estimation task set is less than that in the second potential estimation task set, and therefore the complexity of each potential estimation task in the first potential estimation task set is higher than the complexity of each potential estimation task in the second potential estimation task set.

[0200] In the embodiment of the present application, the complexity level corresponding to the task relevance category may be determined based on a preset correspondence between the task relevance category and the complexity level, or may be directly set manually based on experience, or obtained through repeated experiments.

[0201] Since the underlying parameters of each potential estimation task in the embodiment of the present application are shared, subsequent training mainly depends on the representation range of the upper-layer classification estimation network parameters. Therefore, for tasks with greater complexity, the classification estimation network has more variable layers. It is understandable that potential estimation tasks such as playback duration estimation and completion rate estimation have more influencing factors, so a larger adjustable parameter range and space are required, so the corresponding classification estimation network has more MLP layers, where the specific number of layers is the balance value found by relevant technicians after multiple experiments between the minimum amount of computation and the optimal reasoning performance.

[0202] In the embodiment of the present application, when all networks in the estimation model have converged, the training is stopped to obtain the target estimation model, and the target estimation model is used to perform the corresponding potential estimation task on the estimated content. The following is a detailed description of performing the potential estimation task through the target estimation model.

[0203] See also Fig.12 As shown, it is a flow chart of a method for estimating content potential in an embodiment of the present application, comprising the following steps:

[0204] Step S1201: Based on the basic features and attribute features of the content to be estimated, corresponding multimodal fusion features are obtained.

[0205] The basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated.

[0206] For details, see Fig.13 As shown, it is a schematic diagram of a scenario for potential estimation of the content to be estimated in an embodiment of the present application. When the target estimation model obtains the multimodal fusion feature, the following operations are performed:

[0207] On the one hand, the target estimation model obtains the account information, content category, and content publishing-related meta-information corresponding to the content to be estimated, inputs the Wide part on the left, extracts discrete features in the form of one-hot encoding, and obtains the first output result after processing by MLP and the neural network transformer based on the self-attention mechanism.

[0208] On the other hand, the target estimation model obtains the visual features of the content to be estimated based on the visual information, obtains the audio features of the content to be estimated based on the audio information, and obtains the text features of the content to be estimated based on the text information through the Deep part on the right, and cross-integrates the visual features, audio features and text features through a cross-modal transformer network to obtain the second output result.

[0209] Exemplarily, the image and time information of the content to be estimated are input into the Video Swin Transformer network to obtain the visual features of the content to be estimated, and the multimedia image and time information dimensions are input at the same time to make the determined visual features more accurate.

[0210] The audio information is input into the wavlm-base-plus network to obtain audio features. The wavlm-base-plus is trained on 94,000 hours of unsupervised English data and has achieved good results on multiple speech-related data sets, making the determined audio features more accurate.

[0211] The text information is input into the pre-trained BERT network based on large-scale text prediction of information flow to obtain the text features of the multimedia to be processed.

[0212] Finally, the first output result and the second output result are fused to obtain the multimodal fusion features of the content to be estimated.

[0213] Step S1202: using multiple classification estimation networks included in the target estimation model, based on multimodal fusion features, respectively obtain potential estimation results of the content to be estimated at multiple interaction levels.

[0214] Among them, one classification prediction network corresponds to a potential prediction task at the interaction level, and one classification prediction network outputs a potential prediction result at the interaction level. The number of layers of each classification prediction network is positively correlated with the complexity of the corresponding potential prediction task. The target prediction model is obtained based on the fitting training of the loss functions corresponding to multiple classification prediction networks.

[0215] For details, see Fig.14 As shown, it is a schematic diagram of the process of obtaining potential estimation results based on the classification estimation network in an embodiment of the present application.

[0216] Among them, in multiple classification estimation networks, the following operations are performed respectively:

[0217] Step S1202-1: Based on the element weights of each feature element included in the multimodal fusion feature in a classification estimation network, the amount of information of the characterization potential estimation reference basis corresponding to each feature element is obtained.

[0218] Among them, the element weights in the classification prediction network are the connection coefficients W between adjacent layers. Each connection coefficient W is associated with a bias parameter b. W is used to control the importance of the input signal, and b is used to adjust the ease with which the neuron nodes are activated. In the stage of training the prediction model based on the training sample set to obtain the target prediction model, the element weights in all classification prediction networks have been adjusted to constant values, which can achieve correct classification of known samples.

[0219] Combination Fig.10 As shown in the figure, it is assumed that the classification prediction network is a one-layer MLP network. Each neuron node in the input layer represents a feature element in the multimodal fusion feature. The input layer is only used to transmit data to the hidden layer. The hidden layer and the output layer will calculate the input features to obtain the corresponding output results.

[0220] Assuming that a feature element contained in the multimodal fusion feature is x, when the input layer transmits the feature element x to the first neuron node of the hidden layer, the weighted sum W1x+b1 is obtained through the connection coefficient W1 and the bias parameter b1 between the input layer and the hidden layer, which can measure the amount of information corresponding to the feature element x as a reference for estimating the representation potential.

[0221] Step S1202-2: Based on at least one target feature element whose information amount is greater than a preset information threshold, a potential estimation result of an interaction level corresponding to a classification estimation network is obtained.

[0222] In combination with step S1202-1, only when a feature element contains sufficient information will its corresponding neuron node be activated and transmit the corresponding information to the next neuron node in the output layer, where the amount of information contained in the feature element can characterize the influencing factors when performing the corresponding potential estimation task.

[0223] For example, when the value of W1x+b1 is greater than 1, the neuron node corresponding to x can be activated, and the activation method is f(W1x+b1), where f is the activation function, generally using the sigmoid function or the tanh function to perform nonlinear mapping on the data, so that the MLP has the ability to fit any continuous function.

[0224] Furthermore, when the neuron node corresponding to x is activated, the hidden layer continues to transmit a1=f(W1x+b1) to the output layer, and combines the connection coefficient W2 and the bias parameter b2 between the hidden layer and the output layer to obtain a2=f(W2x+b2).

[0225] It is worth noting that a2 here is only the calculation result obtained by transmitting the feature element x to the output layer. The calculation results of the other feature elements are not specified in the above example. In addition, there can be multiple hidden layers corresponding to MLP. Therefore, the calculation result of the output layer is the sum of all feature elements after weighted summation and activation operation.

[0226] Furthermore, the final output result of the classification prediction network is the softmax regression of the output layer calculation result. Since the weight parameters corresponding to each classification prediction network are different, the output categories of each MLP are different. Fig.12 As shown, in the embodiment of the present application, MLP1 outputs the VV estimation result corresponding to the content to be estimated, MLP2 outputs the completion estimation result corresponding to the content to be estimated, MLP3 outputs the fast scroll rate estimation result corresponding to the content to be estimated, MLP4 outputs the sharing rate estimation result corresponding to the content to be estimated, MLP5 outputs the like rate estimation result corresponding to the content to be estimated, MLP6 outputs the attention rate estimation result corresponding to the content to be estimated, and MLP7 outputs the comment rate estimation result corresponding to the content to be estimated.

[0227] It should be noted that in this application, the VV estimation task and the completion estimation task are more complex, so their corresponding MLP1 and MLP2 are both composed of two layers of MLP, and the remaining potential estimation tasks are less complex, and their corresponding MLP2, MLP3, MLP4, MLP5, MLP6 and MLP7 are all composed of one layer of MLP.

[0228] Optionally, since during the target estimation model training stage, the sample interaction labels of the sample content are divided in two different ways: classification and regression, the target estimation model also has two forms of expression for the potential estimation results of the content to be estimated. For example, the sharing rate estimation result is 57.95%, or the sharing rate estimation result is the third tier. It can also be a combination of the two forms, and this application does not impose any restrictions on this.

[0229] In an embodiment of the present application, after obtaining the potential estimation result of the content to be estimated, content recommendation can be performed based on the corresponding potential estimation result.

[0230] See also Fig.15 As shown, it is a flow chart of a content recommendation method in an embodiment of the present application, comprising the following steps:

[0231] Step S1501: Obtain the business goal to be optimized of the content to be estimated, and the historical interaction data of multiple candidate distribution objects at multiple interaction levels.

[0232] Assuming that the content to be estimated is video content M, and its corresponding business goal to be optimized is to increase the number of followers, that is, to increase the number of fans of the creator of video content M, then in the process of distributing video content M, you can combine the attention preferences reflected in the historical interaction data of all users on the target content platform to choose which target users to distribute video content M to.

[0233] Step S1502: Based on the potential estimation result of the target interaction level corresponding to the business goal to be optimized and the obtained historical interaction data, the interaction evaluation values ​​of multiple candidate objects at the target interaction level for the content to be estimated are obtained.

[0234] Assume that the estimated result of the attention rate of video content M is 64.34%. Based on the estimated result of the attention rate, the predicted score of the increase in fans of video content M is calculated: score 涨粉预测 =α*norm(wilson(64.34%))+β. At the same time, based on the user's historical interaction data, the attention preference score of each user is calculated: Among them, α and β are hyperparameters, and the empirical values ​​are generally α=0.5 and β=0.1. The number of flowing attention behaviors represents the average number of attention behaviors generated when the user consumes multiple video contents within a set time period.

[0235] At the same time, based on experience, determine the category score socre 关注品类 =1.2. Generally speaking, if a user has followed content creators in the same category as video content M, the higher the score of the category they follow.

[0236] Based on the fan increase prediction score and each user's attention preference score, the interaction evaluation value of each user for the video content M is obtained: score = pow (score 涨粉预测 , socre 关注偏好 )*socre 关注品类 .

[0237] In the above calculation process, the specific implementation methods corresponding to the function norm(), the function wilson(), and the function pow() are all existing technologies and will not be elaborated in this application.

[0238] Step S1503: Based on the interaction evaluation value, determine at least one target distribution object of the content to be estimated.

[0239] Based on the interaction evaluation value of each user for the video content M, at least one target user to whom the video content M is distributed is determined.

[0240] Based on steps S1501 to S1503, personalized user-content recommendation is achieved based on the target estimation model for the attention rate estimation results of the content to be estimated, which can promote the fan growth rate of content creators in the content platform, improve the publishing rate of high-quality content creators, and at the same time increase the supply of content in the content distribution pool, reduce the cost of introducing content, and improve user stickiness and retention rate.

[0241] In addition, after obtaining the potential estimation results of the content to be estimated, it is also possible to formulate plans such as cold start optimization of the content and operational push based on the corresponding potential estimation results.

[0242] For example, traffic allocation and optimization are performed based on the completion estimation results of the content to be estimated. If the completion estimation results indicate that the viewing time of the user consuming the content to be estimated accounts for a small proportion of the total content time, the content will not be personalized recommended and distributed, thereby reducing unnecessary traffic waste and improving cold start efficiency.

[0243] For example, based on the estimated results of the like rate and sharing rate of the content to be estimated, high-quality potential content can be screened, the distribution weight of high-quality content can be increased, the review priority of low-quality content can be lowered, and the cost of manual screening and the resource consumption of the recommendation system can be reduced.

[0244] Based on the above embodiments, the present application also provides a recommendation system for content recommendation based on the potential estimation results of the content, see Fig.16 As shown, it is a structural diagram of a recommendation system provided in an embodiment of the present application, and the system includes:

[0245] Content production end 1600, uplink and downlink content interface service unit 1601, dispatch center service unit 1602, content storage service unit 1603, manual review system 1604, content database 1605, target estimation model training unit 1606, target estimation model storage unit 1607, content consumption end 1608, wherein:

[0246] The content production end has more than 1,600 content publishers’ terminals. Content publishers produce and edit content on the content production end. They are mainly used by: PGC organizations or UGC individuals, and content producers of MCN organizations, who provide content through mobile terminals or back-end interface API systems. These are the main sources of content distribution;

[0247] The uplink and downlink content interface service unit 1601 is mainly used to: communicate directly with the content production end, write the content submitted by the front end, usually the title, publisher, summary, cover image, and release time of the multimedia content, such as file size, bit rate, resolution, title, release time, author and other information into the content database; synchronize the published submitted content to the scheduling center service unit for subsequent content processing and circulation;

[0248] The scheduling center service unit 1602 is mainly used to: be responsible for the entire scheduling process of content flow, receive the content stored in the content database and the content storage service unit through the uplink and downlink content interface servers, and then obtain the meta information of the content from the content database; schedule the manual review system and the machine processing system, and control the scheduling order and priority;

[0249] The content storage service unit 1603 is mainly used to: store content entity information other than content meta-information, such as content source files and content cover image source files; communicate with the content consumption terminal, and the terminal directly downloads the corresponding file from the content storage server through the content access address for playback and display;

[0250] The manual review system 1604 is mainly used for: being a carrier of manual service capabilities, mainly for reviewing content that cannot be determined by the filtering machine, and also for secondary confirmation of the classification label of the content; during the manual review process, the information in the content database will be read, and the results and status of the manual review will also be transmitted back to the content database;

[0251] The content database 1605 is mainly used to store the meta information of the content published by all content producers, including but not limited to: file size, cover image link, bit rate, file format, title, publishing time, author, multimedia file size, multimedia format, whether it is original or first release, and the category of the content in the manual review process (including primary, secondary, tertiary categories and tag information);

[0252] The target estimation model training unit 1606 is mainly used to: collect sample content and auxiliary information according to the model training method provided in the embodiment of the present application, and perform necessary cleaning, filtering and enhancement processing on the content; then, based on the collected sample content, based on the pre-training model structure and training method provided in the embodiment of the present application, complete the training of the target estimation model;

[0253] The target estimation model storage unit 1607 is mainly used to store the trained target estimation model, and in the content distribution process, perform corresponding potential estimation for the content, and implement personalized content recommendation, cold start optimization, operation push and other solutions in combination with actual scenarios;

[0254] The content consumption end 1608 is the terminal of the content consumer. The content consumer watches the content through the content consumption end. It is mainly used to: communicate with the upstream and downstream content interface service units to obtain the content index information of the accessed content, and then communicate with the content storage service unit to obtain the corresponding content including the recommended content and the subscribed content.

[0255] Based on the same technical concept, see Fig.17 As shown, the embodiment of the present application further provides a content potential estimation device 1700, which includes:

[0256] The first processing module 1701 is used to obtain corresponding multimodal fusion features based on basic features and attribute features of the content to be estimated; wherein the basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated;

[0257] The second processing module 1702 is used to adopt multiple classification estimation networks included in the target estimation model to obtain potential estimation results of the content to be estimated at multiple interaction levels based on multimodal fusion features; wherein, one classification estimation network corresponds to a potential estimation task at an interaction level, and one classification estimation network outputs a potential estimation result at an interaction level, and the number of layers of each classification estimation network is positively correlated with the complexity of the corresponding potential estimation task, and the target estimation model is obtained based on the fitting training of the loss functions corresponding to each of the multiple classification estimation networks.

[0258] In a possible implementation, after respectively obtaining the potential estimation results of the content to be estimated at multiple interaction levels, the second processing module 1702 is further configured to:

[0259] Obtain the business objectives to be optimized for the content to be estimated, as well as the historical interaction data of multiple candidate distribution objects at multiple interaction levels;

[0260] Based on the potential estimation results of the target interaction level corresponding to the business goal to be optimized and the obtained historical interaction data, the interaction evaluation values ​​of multiple candidate objects at the target interaction level for the content to be estimated are obtained;

[0261] Based on the interaction evaluation value, at least one target distribution object of the content to be estimated is determined.

[0262] In a possible implementation, when the target estimation model is obtained based on the loss functions corresponding to the multiple classification estimation networks, the second processing module 1702 is used to:

[0263] Based on multiple historical contents whose cumulative exposure values ​​are greater than a preset exposure threshold, a training sample set is constructed, and interaction logs of each consumer object associated with each sample content in the training sample set are obtained;

[0264] Based on each interaction log, determine the sample interaction labels of each sample content at multiple interaction levels;

[0265] Using the training sample set and the interaction labels of each sample, as well as the corresponding loss functions of multiple classification estimation networks, the estimation model to be trained is trained for multiple rounds of iterative training to obtain the target estimation model.

[0266] In a possible implementation, the estimation model includes at least a first feature network, a second feature network, and a plurality of classification estimation networks;

[0267] During a round of iterative training of the estimation model, the first processing module 1701 and the second processing module 1702 perform the following operations:

[0268] Extracting a sample basic feature of the sample content based on the first feature network, and extracting a sample attribute feature of the sample content based on the second feature network;

[0269] Based on the basic features and attribute features of the samples, the multimodal fusion features of the samples are obtained, and multiple classification estimation networks are used to obtain the sample potential estimation results of the sample content at multiple interaction levels based on the multimodal fusion features of the samples;

[0270] Based on the loss functions corresponding to multiple classification estimation networks, the loss values ​​between the sample potential estimation results at multiple interaction levels and the corresponding sample interaction labels are obtained, and the parameters of the estimation model are adjusted based on the obtained loss values.

[0271] In one possible implementation, the number of layers of each classification estimation network is obtained in the following way:

[0272] For each potential estimation task, perform the following operations:

[0273] Determine a task relevance category set for a task identifier corresponding to the first potential estimation task; wherein the first potential estimation task is any one of the potential estimation tasks, and the task relevance category corresponding to the first potential estimation task represents: a degree of relevance between multiple potential estimation tasks in a potential estimation task set to which the first potential estimation task belongs;

[0274] Determining the complexity level corresponding to the first potential estimation task based on the task relevance category and the correspondence between the preset task relevance category and the complexity level;

[0275] Based on the complexity corresponding to the first potential estimation task, the number of layers of the classification estimation network corresponding to the potential estimation task is set.

[0276] In a possible implementation, based on each interaction log, sample interaction tags of each sample content at multiple interaction levels are determined, and the first processing module 1701 is used to:

[0277] For each interaction level, perform the following operations on each sample content:

[0278] Based on the interaction logs of each consumer object associated with a sample content, the interaction rate of a sample content at the current interaction level is obtained;

[0279] Obtain the average interaction rate of all historical content at the current interaction level under the content category to which a sample interactive content belongs;

[0280] If the interaction rate corresponding to a sample content at the current interaction level is greater than the average interaction rate, the sample interaction label of the sample content is determined to be a positive sample, otherwise, the sample interaction label of the sample content is determined to be a negative sample.

[0281] In a possible implementation, based on each interaction log, sample interaction tags of each sample content at multiple interaction levels are determined, and the first processing module 1701 is used to:

[0282] For each interaction level, perform the following operations on each sample content:

[0283] Based on the interaction logs of each consumer object associated with a sample content, the interaction rate of a sample content at the current interaction level is obtained;

[0284] Obtain the historical interaction rates of all similar sample contents under the content category to which a sample content belongs, and divide the similar sample contents into multiple interaction levels according to the preset division rules;

[0285] The target interaction level to which the interaction rate corresponding to a sample content at the current interaction level belongs is used as a sample interaction label of the sample content at the current interaction level.

[0286] In a possible implementation, after adjusting the parameters of the estimation model based on the obtained loss values, the first processing module 1701 and the second processing module 1702 are further used to:

[0287] If multiple classification estimation networks meet the convergence conditions, the training is stopped and the target estimation model is obtained;

[0288] If there is a target network that does not meet the convergence conditions among multiple classification estimation networks, then based on the training sample set and the interaction labels of each sample, and the loss function corresponding to the target network, when the estimation model to be trained is subsequently iteratively trained, the parameters of the target network are adjusted, and the parameters of the first feature network, the second feature network, and the converged network that meets the convergence conditions among the multiple classification estimation networks are kept unchanged.

[0289] In a possible implementation, the classification prediction network is composed of one or more multi-layer perceptrons (MLPs).

[0290] After introducing the method and apparatus for estimating content potential according to an exemplary embodiment of the present application, next, a computing device according to another exemplary embodiment of the present application is introduced.

[0291] Those skilled in the art will appreciate that various aspects of the present application may be implemented as a system, method or program product. Therefore, various aspects of the present application may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to as "circuit", "module" or "system" herein.

[0292] Based on the same inventive concept as the above method embodiment, the present application embodiment further provides a computing device 1800. In one embodiment, the computing device may be a server, such as Figure 1 In this embodiment, the structure of the computing device is as follows: Fig.18 As shown, it may include at least a memory 1801 , a communication module 1803 , and at least one processor 1802 .

[0293] The memory 1801 is used to store computer programs executed by the processor 1802. The memory 1801 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.

[0294] The memory 1801 may be a volatile memory, such as a random-access memory (RAM); the memory 1801 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 1801 may be any other medium that can be used to carry or store a desired computer program in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 1801 may be a combination of the above memories.

[0295] The processor 1802 may include one or more central processing units (CPU) or a digital processing unit, etc. The processor 1802 is configured to implement the above-mentioned method for estimating the content potential when calling the computer program stored in the memory 1801 .

[0296] The communication module 1803 is used to communicate with terminal devices and other servers.

[0297] The specific connection medium between the memory 1801, the communication module 1803 and the processor 1802 is not limited in the embodiment of the present application. Fig.18 In the embodiment, the memory 1801 and the processor 1802 are connected via a bus 1804. The bus 1804 is connected to the processor 1802 via a bus 1804. Fig.18 The connections between the other components are only for illustration and are not intended to be limiting. The bus 1804 can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Fig.18 The diagram shows that only one thick line is used, but this does not mean that there is only one bus or one type of bus.

[0298] The memory 1801 stores a computer storage medium, and the computer storage medium stores computer executable instructions, and the computer executable instructions are used to implement the content potential estimation method of the embodiment of the present application. The processor 1802 is used to execute the above-mentioned content potential estimation method.

[0299] In another embodiment, the computing device may also be other computing devices, such as Figure 1 In this embodiment, the structure of the computing device can be as follows: Fig.19As shown, it includes: a communication component 1910, a memory 1920, a display unit 1930, a camera 1940, a sensor 1950, an audio circuit 1960, a Bluetooth module 1970, a processor 1980 and other components.

[0300] The communication component 1910 is used to communicate with the server. In some embodiments, a wireless fidelity (WiFi) module may be included. The WiFi module belongs to a short-range wireless transmission technology. The electronic device can help the object to send and receive information through the WiFi module.

[0301] The memory 1920 can be used to store software programs and data. The processor 1980 executes various functions and data processing of the terminal device 110 by running the software programs or data stored in the memory 1920. The memory 1920 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device. The memory 1920 stores an operating system that enables the terminal device 110 to run. In the present application, the memory 1920 can store an operating system and various application programs, and can also store a computer program for executing the method for estimating the content potential of the embodiment of the present application.

[0302] The display unit 1930 may also be used to display information input by the object or information provided to the object and a graphical user interface (GUI) of various menus of the terminal device 110. Specifically, the display unit 1930 may include a display screen 1932 disposed on the front of the terminal device 110. The display screen 1932 may be configured in the form of a liquid crystal display, a light emitting diode, etc. The display unit 1930 may be used to display the game simulation game interface, etc. in the embodiment of the present application.

[0303] The display unit 1930 can also be used to receive input digital or character information, and generate signal input related to the object setting and function control of the terminal device 110. Specifically, the display unit 1930 may include a touch screen 1931 set on the front of the terminal device 110, which can collect touch operations of objects on or near it, such as clicking a button, dragging a scroll box, etc.

[0304] The touch screen 1931 can be covered on the display screen 1932, or the touch screen 1931 and the display screen 1932 can be integrated to realize the input and output functions of the terminal device 110. The integrated touch screen can be referred to as a touch display screen. In this application, the display unit 1930 can display the application and the corresponding operation steps.

[0305] The camera 1940 can be used to capture static images, and the subject can publish the images taken by the camera 1940 through the application. The camera 1940 can be one or more. The object generates an optical image through the lens and projects it onto the photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the optical signal into an electrical signal, and then transmits the electrical signal to the processor 1980 to convert it into a digital image signal.

[0306] The terminal device may further include at least one sensor 1950, such as an acceleration sensor 1951, a distance sensor 1952, a fingerprint sensor 1953, and a temperature sensor 1954. The terminal device may also be configured with other sensors such as a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, a light sensor, and a motion sensor.

[0307] The audio circuit 1960, the speaker 1961, and the microphone 1962 can provide an audio interface between the object and the terminal device 110. The audio circuit 1960 can transmit the electrical signal converted from the received audio data to the speaker 1961, which is converted into a sound signal for output. The terminal device 110 can also be configured with a volume button for adjusting the volume of the sound signal. On the other hand, the microphone 1962 converts the collected sound signal into an electrical signal, which is received by the audio circuit 1960 and converted into audio data, and then the audio data is output to the communication component 1910 to be sent to, for example, another terminal device, or the audio data is output to the memory 1920 for further processing.

[0308] The Bluetooth module 1920 is used to exchange information with other Bluetooth devices having Bluetooth modules through the Bluetooth protocol. For example, the terminal device can establish a Bluetooth connection with a wearable electronic device (such as a smart watch) that also has a Bluetooth module through the Bluetooth module 1920 to exchange data.

[0309] The processor 1980 is the control center of the terminal device. It uses various interfaces and lines to connect various parts of the entire terminal. It executes various functions of the terminal device and processes data by running or executing software programs stored in the memory 1920 and calling data stored in the memory 1920. In some embodiments, the processor 1980 may include one or more processing units; the processor 1980 may also integrate an application processor and a baseband processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the baseband processor mainly processes wireless communications. It is understandable that the above-mentioned baseband processor may not be integrated into the processor 1980. In the present application, the processor 1980 can run the operating system, application programs, user interface display and touch response, as well as the method for estimating the content potential of the embodiment of the present application. In addition, the processor 1980 is coupled to the display unit 1930.

[0310] In addition, it should be noted that in the specific implementation of this application, object data related to the estimation of content potential is involved. When the above embodiments of this application are applied to specific products or technologies, it is necessary to obtain the object's permission or consent, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0311] In some possible implementations, various aspects of the method for estimating content potential provided by the present application may also be implemented in the form of a program product, which includes a computer program. When the program product is run on a computing device, the computer program is used to enable the computing device to execute the steps of the method for estimating content potential according to various exemplary implementations of the present application described above in this specification.

[0312] 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.

[0313] 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 an electronic device. However, the program product of the present application is not limited thereto, and in this document, a readable storage medium may be any tangible medium containing or storing a program, which can be used by or in combination with a command execution system, apparatus, or device.

[0314] 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.

[0315] 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.

[0316] The computer program 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" or similar programming languages. The computer program may be executed entirely on the user computing device, partially on the user computing device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on the remote computing device. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0317] 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.

[0318] 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.

[0319] 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 method for estimating content potential, It is characterized in that include: Based on the basic features and attribute features of the content to be estimated, corresponding multimodal fusion features are obtained; wherein the basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated; The target estimation model comprises multiple classification estimation networks included in the target estimation model, and based on the multimodal fusion features, the potential estimation results of the content to be estimated at multiple interaction levels are obtained; wherein, one classification estimation network corresponds to a potential estimation task at an interaction level, and one classification estimation network outputs a potential estimation result at an interaction level, and the number of layers of each classification estimation network is positively correlated with the complexity of the corresponding potential estimation task, and the target estimation model is obtained based on the fitting training of the loss functions corresponding to each of the multiple classification estimation networks.

2. The method according to claim 1, It is characterized in that After obtaining the potential estimation results of the content to be estimated at multiple interaction levels, the method further includes: Acquire the business goal to be optimized of the content to be estimated, and the historical interaction data of each of the plurality of candidate distribution objects at the plurality of interaction levels; Based on the potential estimation result of the target interaction level corresponding to the business goal to be optimized and the obtained historical interaction data, obtaining the interaction evaluation values ​​of the multiple candidate objects at the target interaction level for the content to be estimated; Based on the interaction evaluation value, at least one target distribution object of the content to be estimated is determined.

3. The method according to claim 1, It is characterized in that The target estimation model is obtained by training the loss functions corresponding to the multiple classification estimation networks, including: Based on a plurality of historical contents whose cumulative exposure values ​​are greater than a preset exposure threshold, a training sample set is constructed, and interaction logs of each consumer object associated with each sample content in the training sample set are obtained; Based on the interaction logs, determine sample interaction tags of the sample contents at the multiple interaction levels; The training sample set and the interaction labels of each sample, as well as the loss functions corresponding to each of the multiple classification estimation networks, are used to perform multiple rounds of iterative training on the estimation model to be trained to obtain the target estimation model.

4. The method according to claim 3, It is characterized in that The estimation model at least includes a first feature network, a second feature network, and a plurality of classification estimation networks; During one round of iterative training of the estimation model, the following operations are performed: Extracting a sample basic feature of the sample content based on the first feature network, and extracting a sample attribute feature of the sample content based on the second feature network; Obtaining sample multimodal fusion features based on the sample basic features and the sample attribute features, and using the multiple classification estimation networks to obtain sample potential estimation results of the sample content at multiple interaction levels based on the sample multimodal fusion features; Based on the loss functions corresponding to each of the multiple classification estimation networks, the loss values ​​between the sample potential estimation results at the multiple interaction levels and the corresponding sample interaction labels are obtained, and based on the obtained loss values, the parameters of the estimation model are adjusted.

5. The method according to any one of claims 1 to 4, It is characterized in that The number of layers of each classification estimation network is obtained in the following way: For each potential estimation task, perform the following operations respectively: Determine a task relevance category set corresponding to a task identifier of a first potential estimation task; wherein the first potential estimation task is any one of the potential estimation tasks, and the task relevance category corresponding to the first potential estimation task represents: a degree of relevance between multiple potential estimation tasks in a potential estimation task set to which the first potential estimation task belongs; The number of layers of the classification estimation network corresponding to the first potential estimation task is set based on the complexity corresponding to the task relevance category.

6. The method according to claim 3, It is characterized in that The determining, based on the interaction logs, sample interaction tags of the sample contents at the multiple interaction levels includes: For each interaction level, the following operations are performed on each sample content: Based on the interaction logs of each consumer object associated with a sample content, obtaining the interaction rate of the sample content at the current interaction level; Obtain the average interaction rate of all historical content at the current interaction level under the content category to which a sample content belongs; If the interaction rate corresponding to the sample content at the current interaction level is greater than the average interaction rate, the sample interaction label of the sample content is determined to be a positive sample; otherwise, the sample interaction label of the sample content is determined to be a negative sample.

7. The method according to claim 3, It is characterized in that The determining, based on the interaction logs, sample interaction tags of the sample contents at the multiple interaction levels includes: For each interaction level, the following operations are performed on each sample content: Based on the interaction logs of each consumer object associated with a sample content, obtaining the interaction rate of the sample content at the current interaction level; Obtaining historical interaction rates of all similar sample contents under the content category to which the sample content belongs, and dividing the similar sample contents into a plurality of interaction levels according to a preset division rule; The target interaction level to which the interaction rate corresponding to the sample content at the current interaction level belongs is used as the sample interaction label of the sample content at the current interaction level.

8. The method according to claim 4, It is characterized in that After adjusting the parameters of the estimation model based on the obtained loss values, the method further includes: If the multiple classification prediction networks all meet the convergence condition, the training is stopped and the target prediction model is obtained; If there is a target network that does not meet the convergence condition among the multiple classification prediction networks, then based on the training sample set and the interaction labels of each sample, and the loss function corresponding to the target network, when the estimation model to be trained is subsequently iteratively trained, the parameters of the target network are adjusted, and the parameters of the first feature network, the second feature network, and the converged network that meets the convergence condition among the multiple classification prediction networks are kept unchanged.

9. The method according to any one of claims 1 to 4, It is characterized in that The classification prediction network is composed of one or more multi-layer perceptrons.

10. A device for estimating content potential, It is characterized in that include: The first processing module is used to obtain corresponding multimodal fusion features based on basic features and attribute features of the content to be estimated; wherein the basic features at least represent the content category to which the content to be estimated belongs, and the attribute features represent the multimodal content characteristics of the content to be estimated; The second processing module is used to adopt the multiple classification prediction networks included in the target estimation model to obtain potential estimation results of the content to be estimated at multiple interaction levels based on the multimodal fusion features; wherein, one classification estimation network corresponds to a potential estimation task at an interaction level, and one classification estimation network outputs a potential estimation result at an interaction level, and the number of layers of each classification estimation network is positively correlated with the complexity of the corresponding potential estimation task, and the target estimation model is obtained based on the fitting training of the loss functions corresponding to each of the multiple classification estimation networks.

11. A computing device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the computer program, the method according to any one of claims 1 to 9 is implemented.

12. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

13. A computer program product, It is characterized in that When the computer program product is called by a computer, the computer executes the method according to any one of claims 1 to 9.