Information processing method and device, storage medium and computer device

Abstract

The application discloses an information processing method, comprising: obtaining push information, object information and scene information; inputting the push information, the object information and the scene information into a target prediction model to output a predicted push effect of target push content in different scenes; the target prediction model is obtained by performing multi-task learning on a preset prediction network based on sample fusion features; the sample fusion features are obtained by performing vector fusion on sample gating features based on sample general features and sample scene features after gate adjustment; and weight parameters of a prediction layer in the prediction network are determined based on sample rearrangement features and global parameters. The application applies artificial intelligence technology, performs representation learning of scene features and replaces global parameters of a subnetwork based on scene adaptive parameters, so that the model can capture difference information and similar information existing between different scenes, thereby improving the accuracy and stability of the target prediction model in predicting push effects in multiple scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and more particularly, to an information processing method and device, a storage medium and a computer device. BACKGROUND

[0002] An advertisement system is a comprehensive framework for online advertisement placement and management on the Internet based on computational advertising technology. The advertisement system can allocate traffic based on the predicted click-through rate (CTR) and click value rate (CVR) of the to-be-exposed advertisements, so as to determine the optimal value advertisements for online placement.

[0003] Traditional click-through rate and conversion rate prediction is usually based on large-scale logistic regression calculation and factor decomposition based on machine learning. For example, second-order feature cross is automatically performed to capture advertisement interaction information. Currently, the prediction model of the click-through rate and the conversion rate is mainly based on the joint modeling of shared embedding and multilayer perception based on deep learning, which can embed a large amount of information.

[0004] This modeling method converts information into a dense vector of the same dimension through embedding representation, introduces to-be-fitted parameters for the model, and increases the representation space. However, the accuracy of the model in predicting the click-through rate and the conversion rate in different advertisement placement scenarios is uneven, thereby affecting the accuracy and stability of the model in predicting the click-through rate and the conversion rate in multiple scenarios. SUMMARY

[0005] The embodiments of the present application provide an information processing method and device, a storage medium and a computer device. The accuracy of the model in predicting the click-through rate and the conversion rate in different scenarios can be unified, and thus the accuracy and stability of the model in predicting the click-through rate and the conversion rate in multiple scenarios can be improved.

[0006] In one aspect, the embodiments of the present application provide an information processing method, which comprises: obtaining push information, object information and scene information; inputting the push information, the object information and the scene information into a target prediction model to output a predicted push effect of target push content in different scenes, the predicted push effect comprising at least one of a predicted click-through rate or a predicted conversion rate; wherein the target prediction model is obtained by performing multi-task learning on a preset prediction network based on sample fusion features, the sample fusion features are obtained by vector fusion of sample general features based on sample object features and sample push features and sample gating features after gating adjustment of sample scene features, and the weight parameters of a prediction layer in the prediction network are determined based on sample rearrangement features after vector conversion of the sample gating features and global parameters.

[0007] On the other hand, this application embodiment also provides an information processing device, which includes: an information acquisition module for acquiring push information, object information, and scene information; and an effect prediction module for inputting the push information, object information, and scene information into a target prediction model and outputting the predicted push effect of the target push content in different scenarios, wherein the predicted push effect includes at least one of predicted click-through rate or predicted conversion rate; wherein the target prediction model is obtained by multi-task learning of a preset prediction network based on sample fusion features in different scenarios, the sample fusion features are sample general features concatenated based on sample object features and sample push features, and are obtained by vector fusion with sample gating features after gating adjustment of sample scene features, and the weight parameters of the prediction layer in the prediction network are determined by sample rearrangement features after vector transformation based on sample gating features and global parameters.

[0008] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the above-described information processing method is executed when the computer program is run by a processor.

[0009] On the other hand, embodiments of this application also provide a computer device, which includes a processor and a memory. The memory stores a computer program, which, when called by the processor, executes the information processing method described above.

[0010] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program stored in a storage medium; a processor of a computer device reads the computer program from the storage medium and executes the computer program, causing the computer device to perform the steps in the above-described computing engine determination method.

[0011] This application provides an information processing method that can acquire push information, object information, and scene information. Further, the push information, object information, and scene information are input into a target prediction model, which outputs the predicted push effect of the target push content in different scenarios. The predicted push effect includes at least one of predicted click-through rate or predicted conversion rate. The target prediction model is obtained by multi-task learning of a preset prediction network based on sample fusion features for different scenarios. The sample fusion features are obtained by vector fusion of a general sample feature (concatenated from sample object features and sample push features) with a gating feature (sample scene features after gating adjustment). The weight parameters of the prediction layer in the prediction network are determined by the sample rearranged features (sample gating features after vector transformation) and global parameters.

[0012] Therefore, by fusing sample scene features into sample general features, the prediction network can capture differentiated information between different scenes during representation learning, thus avoiding inconsistent prediction accuracy of the target prediction model across multiple scenarios due to differences in different scenes. Simultaneously, the weight parameters of the prediction layer in the prediction network are determined based on the sample gating features after vector transformation and global parameters. Furthermore, multi-task learning of the pre-defined prediction network across multiple scenarios is performed based on the sample fusion features, ensuring consistent convergence of the prediction network in different scenarios. This improves the accuracy and stability of the target prediction model in predicting push effects across different scenarios. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A schematic diagram of a system framework provided in an embodiment of this application is shown.

[0015] Figure 2 A flowchart illustrating an information processing method provided in an embodiment of this application is shown.

[0016] Figure 3 This paper illustrates a system architecture diagram for advertising delivery provided in an embodiment of this application.

[0017] Figure 4 A flowchart illustrating another information processing method provided in an embodiment of this application is shown.

[0018] Figure 5 The diagram illustrates an application scenario of an information processing method provided in an embodiment of this application.

[0019] Figure 6 This illustration shows a sub-network architecture diagram of a prediction network provided in an embodiment of this application.

[0020] Figure 7 A network architecture diagram of a prediction network provided in an embodiment of this application is shown.

[0021] Figure 8 This is a block diagram of an information processing device provided in an embodiment of this application.

[0022] Figure 9 This is a block diagram of a computer device provided in an embodiment of this application.

[0023] Figure 10 This is a block diagram of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0024] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0025] In some processes described in the specification, claims, and the foregoing drawings, multiple steps are included in a specific order. However, it should be clearly understood that these steps may be performed in any order or in parallel. The step numbers are merely used to distinguish different steps and do not represent any execution order. Furthermore, descriptions such as "first" and "second" in this document are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0026] To enable those skilled in the art to better understand the solutions of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0027] It should be noted that, in the specific implementation of this application, the push information, object information and scene information and other related data involved, when applied to the specific products or technologies of this application embodiment, require user permission or consent, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and subsequent data use and processing activities shall be carried out within the scope of the laws, regulations and authorization of the personal information subject.

[0028] In the online advertising field, advertising systems, as an effective advertising delivery tool, optimize conversion costs. They can intelligently and dynamically adjust bids based on the click-through rate (CTR) and conversion rate of individual traffic, helping advertisers effectively control conversion costs and improve advertising efficiency. During the advertising process, well-established CTR and conversion rate prediction models can be used to calculate the CTR and conversion rate of candidate ads.

[0029] Furthermore, the effective cost per mille (ECPM) of candidate ads is calculated based on click-through rate (CTR) and conversion rate (CTR), and ads with higher placement value are selected for placement based on their ECPM scores. In existing technologies, CTR prediction models and conversion rate prediction models are mainly based on deep neural networks (DNNs).

[0030] However, such models typically have exclusive network parameters in upper-layer sub-networks and shared embedding representations in lower-layer sub-networks. Because exclusive network parameters lead to inconsistent model convergence, while shared embedding representations cause a seesaw effect in model prediction—that is, the accuracy of model predictions varies significantly across different scenarios—the predictive performance of the model is unstable. To address these issues, the inventors, through research, proposed the information processing method provided in the embodiments of this application.

[0031] The system architecture of the information processing method involved in this application will be introduced below.

[0032] like Figure 1 As shown, the information processing method provided in this application embodiment can be applied in system 300, and the data acquisition device 310 is used to acquire training data. For the information processing method in this application embodiment, the training data can be training samples used for model training. These training samples include sample push information, sample object information, sample scene information, and push effect labels. The type of training sample can be at least one of exposure samples or click samples, and the push effect labels can be manually labeled. After acquiring the training data, the data acquisition device 310 can store the training data in database 320. The training device 330 can train the target model 301 based on the training data maintained in database 320.

[0033] Specifically, the training device 330 can train a preset neural network based on the input training data until the preset neural network meets preset conditions, thus obtaining the trained target model 301. The preset conditions may include: the total loss value of the target loss is less than a preset value, the total loss value of the target loss no longer changes, or the number of training iterations reaches a preset number, etc. This target model 301 can be used to implement the information processing method in the embodiments of this application. Specifically, the target model 301 in the embodiments of this application can be a deep neural network model, such as a combination of a feature representation layer and a multilayer perceptron (MLP), a Transformer, etc., which is not limited here.

[0034] It should be noted that in practical application scenarios, the training data maintained in database 320 may not all come from data acquisition device 310; it may also be received from other devices. For example, client device 360 ​​may also act as a data acquisition end, using the acquired data as new training data and storing it in database 320. Furthermore, training device 330 may not necessarily train the preset neural network entirely based on the training data maintained in database 320; it may also train the preset neural network based on training data obtained from the cloud or other devices. The above description should not be construed as limiting the embodiments of this application.

[0035] The target model 301 trained by the training device 330 can be applied to different systems or devices, such as... Figure 1 The execution device 340 shown can be a terminal, such as a mobile terminal, tablet computer, or laptop computer, or a server or cloud service, but is not limited thereto. The execution device 340 can be used to interact with external devices. For example, a user (hereinafter referred to as the target object) can use a client device 360 ​​to send input data to the execution device 340 via a network.

[0036] In this embodiment of the application, the input data may include a request message sent by the client device 360. During the preprocessing of the request message by the execution device 340, or during the calculation and other related processing performed by the execution module 341 of the execution device 340, the execution device 340 may call data, programs, etc. in the data storage system 350 for corresponding calculation processing, and store the processing results and instructions obtained from the calculation processing into the data storage system 350.

[0037] Finally, the execution device 340 can return the processing result, that is, the set of target entity pairs generated based on the target model 301, to the client device 360 ​​via the network, so that the target object can query the processing result on the client device 360. It is worth noting that the training device 330 can generate corresponding target models 301 based on different training data for different targets or different tasks. The corresponding target model 301 can be used to achieve the above-mentioned targets or complete the above-mentioned tasks, thereby providing the required results to the target object.

[0038] For example, Figure 1The system 300 shown can be a Client-Server (C / S) system architecture, the execution device 340 can be a cloud server deployed by a service provider, and the client device 360 ​​can be a laptop used by the target object. For example, the target object can use social software on the laptop, and if the target object's information is authorized for advertising, the laptop can send a request message to the cloud server via the network when the target object browses the web.

[0039] Furthermore, upon receiving a request message, the cloud server utilizes the target model 301, also known as the target prediction model, to predict the click-through rate (CTR) and conversion rate of the candidate push content for the target object in different scenarios, based on object information, push information of the candidate push content, and scenario information. Subsequently, the cloud server calculates the exposure price of the candidate push content based on the predicted CTR and conversion rate, selects push content with higher delivery value based on the exposure price, and sends the selected push content to the push content exposure position on the webpage for exposure (delivery).

[0040] It is worth noting that, Figure 1 This is merely a schematic diagram of a system architecture provided in this application embodiment. The system architecture and application scenarios described in this application embodiment are for the purpose of more clearly illustrating the technical solutions of this application embodiment and do not constitute a limitation on the technical solutions provided in this application embodiment. For example, Figure 1 The data storage system 350 is an external memory relative to the execution device 340. In other cases, the data storage system 350 can also be placed within the execution device 340. As those skilled in the art will recognize, with the evolution of system architecture and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to solving similar technical problems.

[0041] The information processing method in this application relates to Artificial Intelligence (AI) technology. AI technology utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce new intelligent machines capable of reacting in a manner similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0042] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0043] Deep learning (DL) is a sub-problem of machine learning, whose main purpose is to automatically learn effective feature representations from data. Through multi-layered feature transformation, raw data is transformed into higher-level, more abstract representations. These learned representations can replace manually designed features, thus avoiding "feature engineering." The abstract representations are then input into a prediction function to obtain the final result. For example, in the embodiments of this application, the target prediction model can learn representations from input push information, object information, and scene information to obtain corresponding features.

[0044] Please see Figure 2 , Figure 2 A flowchart illustrating an embodiment of the information processing method provided in this application is shown. In a specific embodiment, the information processing method is applied to, for example... Figure 8 The information processing device 500 and the computer equipment 600 equipped with the information processing device 500 are shown. Figure 9 ).

[0045] The following will use a computer device as an example to illustrate the specific process of this embodiment. It is understood that the computer device used in this embodiment can be a server or a terminal, etc. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, blockchain, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The information processing method may specifically include the following steps:

[0046] Step 110: Obtain push information, object information, and scene information.

[0047] Considering the need to extend the data and model capabilities of the main business scenarios to more smaller business scenarios as advertising business scenarios expand, in order to improve the resource utilization of algorithm iteration and the return on investment of advertising, a multi-scenario modeling approach for click-through rate (CTR) or conversion rate (CTR) prediction is proposed. That is, for the target audience, the CTR or CTR of the target content, i.e., the target advertisement, is predicted in different scenarios.

[0048] Contextual information refers to information about the advertising scenario, which can include two dimensions: advertising time and ad placement. Advertising time can be the current moment the ad is being placed or a time period set according to business needs, such as peak or off-peak seasons for advertising. Ad placement can be a traffic indicator or a page placement indicator, etc.

[0049] Traffic identifiers can be used to indicate the specific application in which the advertisement is placed, such as a social network or open application platform. Placement identifiers are used to indicate the form of advertisement placement, such as a splash screen ad in an application or a landscape ad in a video stream within a video player. Optionally, ad duration and ad placement can be set according to the specific business rules of the advertising operation, and are not limited here.

[0050] Push notifications refer to information related to candidate ads, which may include promotional materials, creative content, and ad identifiers. For example, promotional materials could be the product or message being promoted, creative content could be images or text related to the promotional material, and ad identifiers could be the advertiser's name and identifier. Candidate ads refer to ads for which click-through rate or conversion rate prediction is performed. Target audience information refers to the basic information (e.g., identity identifier) ​​and historical behavior information (e.g., records of ad clicks) of the target audience.

[0051] As one implementation method, in response to the prediction request of the target object, the object information of the target object can be obtained, and the push information of each candidate push content and the scene information of different scenarios can be obtained from the candidate push set.

[0052] For example, when the target object is browsing a webpage, a prediction request can be generated. Furthermore, in response to the prediction request, the target object's object information can be obtained from the service log, and the push information of each candidate advertisement can be obtained from the candidate advertisement set in the advertisement database. Furthermore, the time when the target object browses the webpage and the advertisement position on the webpage can be obtained.

[0053] It should be noted that when obtaining scene information and target audience information, it is necessary to comply with legal and regulatory requirements and adhere to the principles of legality, legitimacy, necessity, and good faith, collecting only the information necessary for ad placement. For example, before ad placement, a pop-up window can be displayed to the target audience to show them a personalized ad management interface, allowing them to determine whether they need to enable personalized push / ad placement. If the target audience enables it, personalized push / ad placement can then be performed.

[0054] Step 120: Input the push information, target information, and scene information into the target prediction model, and output the predicted push effect of the target push content in different scenes.

[0055] When modeling multiple scenarios for click-through rate or conversion rate prediction tasks, existing technologies train models directly based on shared features for different scenarios to improve the learning efficiency of prediction tasks for each scenario. However, this approach ignores the specificity of different scenarios. For example, the target audience's preferences for web page ads and streaming media ads in short videos may differ, and their exposure formats are also different. When training models for multiple scenarios based on shared features, the trained models may exhibit a seesaw phenomenon in different prediction tasks. For example, the learning efficiency may be high for some prediction tasks, while the learning efficiency may be low for others.

[0056] Furthermore, existing technologies employ model training based on shared features within independent sub-networks for prediction tasks. This approach ignores the similarities or correlations between different scenarios. For instance, vertical and horizontal ads in short videos are both streaming media ads. Since each ad placement scenario corresponds to an independent sub-network, although each sub-network learns based on shared features, the complete isolation of the sub-network parameters leads to inconsistent sample distributions across different scenarios during training, resulting in inconsistent convergence levels of the model in different scenarios. These shortcomings cause the trained model to exhibit inaccurate and unstable prediction performance when predicting ad effects across multiple scenarios. Therefore, this application, considering the specificity and similarity between different scenarios, creatively proposes a target prediction model to calculate the predicted push effect in different scenarios.

[0057] The predicted push effect includes at least one of predicted click-through rate or predicted conversion rate. The target prediction model is obtained by multi-task learning of a pre-defined prediction network based on sample fusion features for different scenarios. The sample fusion features are obtained by vector fusion of sample general features (concatenated from sample object features and sample push features) with sample gating features (gated to adjust sample scenario features). The weight parameters of the prediction layer in the prediction network are determined by the sample rearranged features (vector-transformed from sample gating features) and global parameters.

[0058] In this way, when training the prediction network, feature information from different scenarios can be fused into the general features of the samples, enabling the prediction network to learn personalized representations for different scenarios and thus avoiding the "seesaw effect." Simultaneously, multi-task learning is performed based on the sample rearranged features (derived from sample gating features through vector transformation) and the prediction layer weight parameters determined by global parameters, ensuring consistent model convergence across different scenarios. This improves the accuracy and stability of the target prediction model in predicting push notification effects across multiple scenarios.

[0059] In a typical advertising scenario, when a target audience sees an ad displayed by the advertising system, it's called an impression. Further, there's a certain probability that the target audience will click on the displayed ad to enter the product details page. If satisfied, the target audience will purchase the product; this is called a conversion. A conversion is an action that "may" occur only after a click.

[0060] Therefore, to ensure diversity in the prediction of push notification effects, the training samples used to train the prediction network can include exposure samples and click samples. Exposure samples include positive samples that were exposed and clicked, and negative samples that were exposed but not clicked. Click samples include positive samples that were clicked and converted, and negative samples that were clicked but not converted. This application can train target prediction models that perform different tasks based on different types of training samples.

[0061] For example, depending on the needs of the advertising business, a target prediction model can be trained based on exposure samples to predict click-through rate, or based on click samples to predict conversion rate, or based on both exposure and click samples to predict conversion rate and target prediction model. This makes the prediction of advertising effectiveness more diverse.

[0062] As one implementation method, push information, target information, and scenario information can be input into the target prediction model to predict the push effect. For the target target, the predicted click-through rate and predicted conversion rate of each candidate push content in different scenarios can be predicted.

[0063] Please see Figure 3 ,Figure 3 This diagram illustrates a system architecture for ad delivery, such as... Figure 3 As shown, object information can be obtained from the database that stores service logs, and scene information and candidate ad sets (i.e., candidate push sets) can also be obtained online. The candidate ad set stores multiple candidate ads (i.e., candidate push content) obtained through ad coarse ranking.

[0064] Furthermore, advertising information (i.e., push notifications), target information, and scene information can be input into the target prediction model to predict advertising effectiveness (i.e., push notification effectiveness). Specifically, the target prediction model can perform representation learning on advertising information, target information, and scene information separately to obtain corresponding advertising features, target features, and scene features, where each feature can be a vector in the embedding space. The advertising features and target features are then concatenated to obtain a general feature. The target prediction model can also perform gating adjustments on the scene features to obtain gated features, and then the general feature and the gated features are multiplied positionally to obtain the fused feature.

[0065] Furthermore, the fused features can be passed to the corresponding sub-networks (i.e., prediction layers) of the target prediction model for each scenario to predict the prediction effect. Thus, the prediction layer for each scenario can output the corresponding advertising prediction effect {CTR}. d1 CVR d1 CTR d2 CTR d3 ,…CTR d* Furthermore, it is possible to obtain bidding adjustment parameters and the target conversion bid for each candidate ad. These bidding adjustment parameters can be pre-set by business personnel according to advertising business rules, for example, to adjust the exposure ratio of different ads participating in the bidding. The target conversion bid refers to the advertiser's expected cost per conversion. For example, if the advertiser hopes to spend 30 yuan on advertising to complete a sale, then the target conversion bid is 30 yuan.

[0066] Furthermore, based on predicted click-through rate, predicted conversion rate, bidding adjustment parameters, and the target conversion bid for each candidate ad, the exposure price for each candidate ad in the corresponding scenario can be determined. For example, the exposure price can be the revenue per thousand impressions (ECPM) score, and the calculation formula can be as follows:

[0067] ECPM t＝ CVR t ×CTR t ×θ×CPA t

[0068] Among them, CVR t CTR t and CPAt Let θ represent the conversion rate, click-through rate, and target conversion bid of the t-th candidate ad, respectively, and θ be the bidding adjustment parameter. Further, the ECPM score for each candidate ad can be calculated. Based on the ECPM scores, the ads are sorted in descending order. Then, a fine-grained ranking is performed on the multiple candidate ads obtained from the initial ranking. According to the advertising business's delivery requirements, the top-k candidate ads corresponding to the highest ECPM scores are selected as target ads (i.e., target content). Finally, the target ads are placed in the corresponding scenarios, such as on the webpage of a social media platform that the target audience is currently browsing.

[0069] In this embodiment, push information, object information, and scene information can be obtained. Further, the push information, object information, and scene information are input into a target prediction model, which outputs the predicted push effect of the target push content in different scenarios. The predicted push effect includes at least one of predicted click-through rate or predicted conversion rate. The target prediction model is obtained by multi-task learning of a preset prediction network based on sample fusion features for different scenarios. The sample fusion features are obtained by vector fusion of a general sample feature (concatenated from sample object features and sample push features) with a gating-adjusted sample gating feature (sample scene features). The weight parameters of the prediction layer in the prediction network are determined by the sample rearranged features (sample gating features transformed by vector) and global parameters.

[0070] In this way, by fusing sample scene features into sample general features, the personalized representation of the network under different scenarios is enhanced, thereby avoiding inconsistent prediction accuracy of click-through rate and conversion rate in the trained target prediction model. Based on the sample gating features transformed into vector-based sample rearranged features and global parameters, the weight parameters of the prediction layer in the prediction network are determined. Furthermore, based on the sample fusion features, the pre-defined prediction network is subjected to multi-task learning for different scenarios, making the model's parameter learning space more sufficient and ensuring consistent model convergence across different scenarios. This improves the prediction performance of the target prediction model.

[0071] Based on the methods described in the above embodiments, the following examples will provide further detailed explanations.

[0072] The information processing method provided in this embodiment specifically relates to the technology of artificial intelligence. The following description will take the integration of an information processing device into a computer device as an example, and will focus on… Figure 4 The process shown combines Figure 4 The application scenarios shown are described in detail. This computer device can be a server or a terminal device, etc. Please refer to [link / reference]. Figure 5 , Figure 5 This application illustrates another information processing method provided by an embodiment of the present application. In a specific embodiment, this information processing method can be applied to, for example... Figure 5The scenario shown is where the advertisement is placed.

[0073] The advertising service provider offers a service platform, which includes a cloud training server 410 and a cloud execution server 430. The cloud training server 410 can be used to train a target prediction model for advertising delivery, and the cloud execution server 430 is used to deploy the target prediction model trained on the cloud training server 410. Based on advertising information (i.e., push information), object information, and scene information, the target prediction model predicts the predicted click-through rate and predicted conversion rate of each candidate advertisement in different scenes for the target object.

[0074] Furthermore, the cloud execution server 430 can be used to determine the predicted click-through rate, predicted conversion rate, bidding adjustment parameters, and corresponding target conversion bid for each candidate ad in different scenarios. This determines the exposure price for each candidate ad (i.e., candidate push content) in different scenarios, and then determines the target ad (i.e., target push content) based on the exposure price, and exposes the target ad on the client. For example, the client can be a smartphone 420, and the cloud execution server 430 can deliver the target ad to the social media software 421 installed on the smartphone 420 for display.

[0075] It should be noted that, Figure 5 This is merely one application scenario provided by an embodiment of this application. The application scenario described in this embodiment is for the purpose of more clearly illustrating the technical solution of this application and does not constitute a limitation on the technical solution provided by this embodiment. For example, in other cases, the smartphone 420 can also deploy a target prediction model, thereby completing the entire process of ad delivery directly on the smartphone 420. Those skilled in the art will understand that with the evolution of system architecture and the emergence of new application scenarios, the technical solution provided by this embodiment is also applicable to solving similar technical problems. The information processing method may specifically include the following steps:

[0076] Step 210: The computer device acquires training samples.

[0077] The training samples include sample push information, sample object information, sample scenario information, and push effect tags. Training samples can include at least one of exposure samples and click samples. Considering the differences in business needs for content push in different scenarios, targeted prediction models for specific prediction tasks can be trained based on different types of training samples, thereby improving the predictive diversity of push effects.

[0078] Each type of training sample corresponds to positive and negative samples. Specifically, the training samples can be labeled using the positive and negative feedback of the target object's click behavior and conversion behavior after exposure, thus obtaining the push effect label corresponding to each training sample.

[0079] For example, for click-through rate prediction, model training can be based on the exposure hypothesis, where the target audience's click behavior represents positive feedback, indicating that the advertisement corresponding to the training sample is attractive to the target audience, and the push effect label corresponding to the training sample can be marked as 1. Conversely, exposure without clicks represents negative feedback, and the push effect label can be marked as 0.

[0080] For conversion rate prediction, model training can be based on clicks. If the target audience clicks and then converts, it represents positive feedback, indicating that the advertisement corresponding to the training sample is attractive to the target audience and meets their deeper needs. The push effect label for this training sample can be marked as 1. Conversely, if there is no conversion after clicking, it represents negative feedback, and the push effect label can be marked as 0.

[0081] Considering the large scale of advertising operations in real-world scenarios, resulting in a large amount of scenario information and complex relationships between different scenario information, this application proposes a method for selecting sample scenario information from multiple candidate scenario information. By selecting a suitable number of scenario information that is strongly correlated with the model's modeling objective as sample scenario information, the representation capability of the model can be improved, resulting in a feature subset that is both effective and efficient—that is, the features corresponding to the sample scenario information. This, in turn, enhances the predictive performance of the trained target prediction model.

[0082] As one implementation, the cloud training server 410 can acquire multiple candidate scene information, which may include at least advertising time and traffic identifiers. Based on the candidate features of each candidate scene information, a filtering parameter corresponding to each candidate scene information is calculated, and the filtering parameter satisfies the Bernoulli distribution of the hyperparameters corresponding to the filtering parameter.

[0083] Furthermore, the cloud training server 410 can determine a preset number of sample scene information from multiple candidate scene information based on each filtering parameter, and determine training samples based on sample push information, sample object information, and sample scene information. For example, given a candidate scene information set S = {f1, f2, ..., f...} M}, where f j It is the j-th candidate scene information in S, and M is the number of candidate scene information.

[0084] To select sample information that is both effective and efficient, a corresponding filtering parameter Z can be set for each candidate scene information.j ∈{0,1}, to represent that candidate scene information is discarded (Z j =0) or reserved (Z) j =1). Candidate scene information f can be obtained based on representation learning. j The corresponding feature vector v j and the feature vector v j With the filtering parameter Z j Multiply, and filter parameter Z j The Bernoulli distribution that satisfies the hyperparameters corresponding to the screening parameters is shown in the following formula:

[0085] Z j ~Bern(σ) j )

[0086] Among them, hyperparameter σ j It is candidate scene information f j The prior probabilities that can be preserved can be modeled as a feature complexity function c. j The formula is shown below:

[0087] σ j =H(c j )=1-τ(c j )

[0088] c j =G(o j ,e j ,n j )

[0089] Where τ(·) can be the Sigmoid function. j σ j and c j The modeling method, through feature complexity c j Can measure candidate scene information f j The storage and computational complexity, for example, computational complexity O. j The dimension e of the feature vector j and the number of eigenvector values ​​n j Therefore, candidate scene information with higher complexity will have a lower prior probability of being retained.

[0090] Furthermore, the cloud training server 410 can determine m sample scene information from the candidate scene information set S based on each filtering parameter, and determine training samples based on sample push information, sample object information, and sample scene information, thereby constructing a training set D = {(x i ,y i )|i=1,2,…,N}, where x i Let y represent the i-th training sample. iIndicates training sample x i Corresponding push effect tags. Training sample x i This can include sample scene information. (cd* represents the *th scene).

[0091] It is worth noting that the information processing method provided in this application embodiment includes training a preset prediction network. This training process can be performed based on an acquired training set D. Subsequently, whenever click-through rate and conversion rate predictions need to be performed, the trained target prediction model can be used to directly calculate the values, without having to repeatedly train based on the training set D each time click-through rate and conversion rate predictions are performed. For example, the cloud training server 410 can obtain the training set D from a database storing the training set.

[0092] Step 220: The computer device inputs the sample push information, sample object information and sample scene information into the prediction network for feature extraction, and obtains the corresponding sample push features, sample object features and sample scene features.

[0093] Please see Figure 6 , Figure 6 A sub-network architecture diagram of a prediction network is shown. (For example...) Figure 6 As shown, the prediction network's sub-networks can include an embedding layer and a gating layer. The computer device can input the sample push information, sample object information, and sample scene information of the training samples into the prediction network. The embedding layer of the prediction network performs feature extraction to obtain the corresponding sample push features, sample object features, and sample scene features. Furthermore, the gating layer can be used to adjust the gating of the sample scene features.

[0094] For example, the cloud training server 410 can provide training samples, which may include sample push information x. ad Sample object information x user and sample scene information from three different scenarios {x cd1 ,x cd2 ,x cd3 The input is fed into the prediction network. Then, through the embedding layer representation learning of the prediction network, the corresponding sample push feature `emb` is obtained. ad Sample object features (emb) user and sample scene features emb context .

[0095] Step 230: The computer device uses a prediction network to predict the push effect based on sample push features, sample object features, and sample scene features, and obtains the sample push effect of the training samples for different scenes.

[0096] Since target prediction models can be trained to perform specific prediction tasks based on different types of training samples, the predicted push effect can include at least one of predicted click-through rate or predicted conversion rate. In this embodiment, predicted click-through rate and predicted conversion rate are used as examples to illustrate the predicted push effect.

[0097] As one implementation method, the computer device predicts the push effect based on sample push features, sample object features, and sample scene features using a prediction network, and obtains the sample push effect of training samples corresponding to different scenes. The steps may include:

[0098] (1) The computer device uses a prediction network to concatenate the sample object features and sample push features into vectors to obtain the sample general features.

[0099] For example, the cloud training server 410 obtains the sample push feature emb ad and sample object features emb user At this time, the two can be concatenated to obtain the common feature emb of the sample. com For example, sample push feature emb. ad = (0.585, 0.168, 2.785), sample object feature emb user = (4.565, 1.136, 0.113). Push feature emb to the sample. ad and sample object features emb user By concatenating the vectors, we can obtain the general feature emb of the samples. com =(0.585,0.168,2.785,4.565,1.136,0.113)

[0100] (2) The computer equipment performs gating adjustment on the sample scene features to obtain the sample gating features corresponding to the sample scene features.

[0101] Gating refers to inputting the obtained sample scene features into a gating layer to extract significant latent information from the sample scene features. In this embodiment, the gating layer may include two fully connected layers: a first fully connected layer and a second fully connected layer.

[0102] For example, the cloud training server 410 can obtain intermediate features of the sample by performing a first nonlinear calculation on the sample scene features based on a first activation function through a first fully connected layer. Furthermore, the cloud training server 410 can obtain gated features of the sample by performing a second nonlinear calculation on the intermediate features of the sample based on a second activation function and a scaling parameter through a second fully connected layer.

[0103] For example, the cloud training server 410 can generate sample scene features (emb). context The output is fed to the first fully connected layer, and a first nonlinear calculation is performed on the sample scene features based on the first activation function to obtain the intermediate sample feature emb. mid and the intermediate features of the samples, emb mid The input is fed into the second fully connected layer, where a second nonlinear computation is performed on the intermediate features of the sample based on the second activation function and scaling parameters to obtain the sample gated feature emb. gate The calculation formula can be as follows:

[0104] emb mid =Relu(emb context ×w1+b1)

[0105] emb gate =λ×Sigmoid(emb mid ×w2+b2)

[0106] In this context, the ReLU function is the first activation function, and the Sigmoid function is the second activation function. `w` represents the weight parameter, and `b` is the penalty term. `λ` is the scaling parameter (λ is usually 2), used to scale the output values ​​of all Sigmoid functions to ensure that `emb` is optimized. gate This is more in line with the physical meaning.

[0107] In this way, by using gating to extract significant latent information from the features of the sample scene, the prediction network can more efficiently and implicitly capture high-order cross features, thereby improving prediction performance.

[0108] (3) The computer equipment performs a positional multiplication operation on the sample gating features and the sample general features to obtain the sample fusion features.

[0109] Given that existing technologies use a learning method that directly shares features across different scenarios, this method ignores the specificity of different scenarios, leading to a seesaw effect in model predictions. This is because samples from all scenarios participate in representation learning, thus ignoring the different preferences of target objects in different scenarios.

[0110] To this end, this application proposes to superimpose the influence of scene features when the prediction network performs representation learning based on samples, that is, to perform element-wise product operation, so that the feature representation learned by the network can perceive the target object's preference information in different scenes.

[0111] For example, the cloud training server 410 can perform positional multiplication on the sample gating features and sample scene features to obtain sample fusion features.

[0112] (4) The computer device predicts the push effect based on the sample fusion feature and sample gating feature, and obtains the sample push effect corresponding to each scene in the training sample.

[0113] Existing technologies employ a learning approach based on shared features within independent sub-networks, neglecting the similarities or connections between different scenarios. This leads to inconsistent model convergence and consequently, poor predictive performance. To address this, this application considers the specificity and similarity between different scenarios. It proposes that when the prediction network learns based on sample fusion features, it starts with scene features, using a common prediction layer as a base and overlaying adaptive parameters relevant to the scene. This approach considers both the specificity and similarity of different scenarios.

[0114] Please see Figure 7 , Figure 7 A network architecture diagram of a predictive network is shown. For example... Figure 7 As shown, the prediction network can include a lower-level sub-network embedding layer for representation learning to obtain sample push features, sample object features, and sample scene features. This embedding layer is used to perform a positional product operation between the general sample features (concatenated from the sample object features and sample push features) and the gated sample scene features (after gating adjustment). The prediction network can also include a prediction layer for each scene, with each scene corresponding to an MPL (Multi-Level Model). d* The weight parameters are the sample rearranged features based on the sample gating features after vector transformation and the common MPL. common The global parameters are determined.

[0115] In some embodiments, the step of the computer device predicting the push effect based on sample fusion features and sample gating features to obtain the sample push effect corresponding to each scenario in the training samples may include:

[0116] (4.1) Perform vector segmentation on the gating features of the samples to obtain the sample segmentation features corresponding to each scene.

[0117] To enhance the model's parameter space, the sample gating features can be vector-segmented based on the desired model fit scenario. For example, during model training, the sample scenario information {x} of the training samples... cd1 ,x cd2 ,x cd3 The information corresponds to three scenarios, which allows the sample gating features to be segmented into sample segmentation features corresponding to each scenario.

[0118] For example, in conjunction with reference Figure 6 The cloud training server 410 can perform gating feature extraction on samples (emb). gate Perform vector segmentation to obtain sample segmentation features corresponding to each scene. Sample segmentation features and sample segmentation features

[0119] (4.2) Perform matrix rearrangement on each sample segmentation feature to obtain the sample rearrangement feature corresponding to each sample segmentation feature.

[0120] To avoid ignoring the similarities between different scenarios and causing inconsistent model convergence, this application proposes to use a common prediction layer as a base and superimpose adaptive parameters related to the scenario to determine the private prediction layer corresponding to each scenario.

[0121] Specifically, the global parameters of the common prediction layer can be multiplied in pairs with the adaptive parameters corresponding to each scene. To achieve this pairwise multiplication, the sample segmentation features corresponding to each scene can be reshaped to obtain the corresponding rearranged sample features.

[0122] For example, in conjunction with reference Figure 6 The cloud training server 410 can segment features for samples corresponding to each scene. Sample segmentation features and sample segmentation features Perform matrix rearrangement to obtain the corresponding sample rearrangement features. Sample rearrangement features and sample rearrangement features

[0123] For example, during the calculation process, the global parameters of the public prediction layer are in the format of an a×b matrix, while the sample segmentation features... The format is a 1×k matrix (vector). At this point, the sample segmentation features in the 1×k format can be used. Through matrix rearrangement, the sample rearrangement features are transformed into a×b format.

[0124] (4.3) Perform positional multiplication of each sample rearranged feature with the global parameters of the prediction layer in the prediction network to obtain the weight parameters of the prediction layer corresponding to each scenario.

[0125] This application proposes to perform positional multiplication of the global parameters of the public prediction layer with the adaptive parameters corresponding to each scenario, where the adaptive parameters are the sample rearrangement features. The prediction layer refers to a sub-network in the prediction network that performs feature cross-cutting of arbitrary order, thereby capturing the interaction information between the target object and the pushed content (e.g., advertisements) in different scenarios, resulting in better performance in click-through rate and conversion rate prediction.

[0126] In the embodiments of this application, the prediction layer can be a third-order multilayer perceptron. Optionally, the prediction layer can also be a mixed logistic regression network (Large Scale Piece-wise Linear Model) or a factorization machine (FM), and there is no limitation herein.

[0127] Please see Figure 6 In this application, the sample segmentation features corresponding to each scene are used as adaptive parameters and respectively multiplied with the global parameters of the common multilayer perceptron. For example, different scenes may include a first sample scene, a second sample scene, and a third sample scene.

[0128] Specifically, the sample rearrangement features corresponding to the first sample scenario are multiplied in pairs with the global parameters of the prediction layer in the prediction network to obtain the first weight parameters of the prediction layer corresponding to the first sample scenario. The sample rearrangement features corresponding to the second sample scenario are multiplied in pairs with the global parameters of the prediction layer in the prediction network to obtain the second weight parameters of the prediction layer corresponding to the second sample scenario. The sample rearrangement features corresponding to the third sample scenario are multiplied in pairs with the global parameters of the prediction layer in the prediction network to obtain the third weight parameters of the prediction layer corresponding to the third sample scenario.

[0129] For example, a public multilayer perceptron (MPL) common The global parameter F is used to perform sample reordering features with the sample reordering features corresponding to each scenario (domain), for example, This leads to the creation of a private multilayer perceptron (MPL) for each scenario. d* The weight parameters.

[0130] (4.4) Based on the weight parameters of the prediction layer and the sample fusion features corresponding to each scenario, the push effect is predicted to obtain the sample push effect corresponding to each sample scenario of the training sample.

[0131] In some embodiments, when the private multilayer perceptron (MPL) corresponding to each scene is obtained... d* When setting the weight parameters, the sample fusion features can be output to each private multilayer perceptron (MPL). d* This yields the sample push effect corresponding to each sample scenario of the training samples. In this embodiment, a target prediction model capable of simultaneously predicting click-through rate (CTR) and conversion rate (CTR) is specifically trained based on different types of training samples. The sample push effect may include the first sample CTR. d1 Second sample conversion rate (CVR) d2 and the third sample conversion rate (CVR) d3 .

[0132] For example, in conjunction with reference Figure 6 The cloud training server 410 can predict the push effect based on the first weight parameters of the prediction layer corresponding to the first sample scenario and the sample fusion features, and obtain the first sample click-through rate (CTR) of the first sample scenario corresponding to the training samples. d1 Furthermore, based on the second weight parameters of the prediction layer corresponding to the second sample scenario and the sample fusion features, the push effect can be predicted to obtain the second sample conversion rate (CVR) corresponding to the second sample scenario of the training samples. d2 Based on the third weight parameters of the prediction layer corresponding to the third sample scenario and the sample fusion features, the push effect is predicted to obtain the third sample conversion rate (CVR) of the training samples corresponding to the third sample scenario. d3 .

[0133] Step 240: The computer device performs binary cross-entropy calculation based on the sample push effect and the push effect label to determine the target loss corresponding to the prediction network.

[0134] Considering that the prediction network needs to perform different click-through rate or conversion rate prediction subtasks for each scenario, this application constructs the target loss of the prediction network based on the loss of each prediction subtask.

[0135] As one implementation method, the cloud training server 410 can determine the first loss L1(x) based on the first sample click rate and the first sample click label. i ,y i Based on the conversion rate and conversion label of the second sample, the second loss L2(x) is determined. i ,y i Based on the conversion rate and conversion label of the third sample, the third loss L3(x) is determined. i ,y i ).

[0136] Furthermore, the cloud training server 410 can adjust the first loss L1(x) based on the training loss L1(x). i ,y i ), second loss L2(x i ,y i ) and the third loss L3(x i ,y i Determine the target loss L corresponding to the prediction network. For example, the target loss Loss calculation formula can be as follows:

[0137]

[0138] Wherein, L1(x i ,y i L2(x) i,y i ) and L3(x i ,y i Cross-entropy loss can be used for binary classification. and These are the weight parameters of the scene-specific private multilayer perceptron.

[0139] Step 250: The computer device iteratively trains the prediction network for different scenarios based on the target loss until the prediction network meets the preset conditions, thus obtaining the target prediction model.

[0140] As one implementation method, the cloud training server 410 can iteratively train the prediction network using the target loss for multi-task learning in different scenarios until the prediction network meets preset conditions, thus obtaining the target prediction model. These preset conditions may include: the total target loss value being less than a preset value, the total target loss value no longer changing, or the number of training iterations reaching a preset number, etc. Optionally, an optimizer can be used to optimize the target loss, setting the learning rate, batch size, and epoch based on experimental experience.

[0141] After the target prediction model is trained on the cloud training server 410, it can be deployed on the cloud execution server 430 to provide advertising delivery services.

[0142] Step 260: The computer device responds to the prediction request of the target object and obtains push information, object information and scene information.

[0143] For example, when a target object browses the interface of social software 421 on a smartphone 420, the smartphone 420 can generate a prediction request. Furthermore, the smartphone 420 can send the prediction request to the cloud execution server 430. In response to the prediction request, the cloud execution server 430 can obtain the target object's object information from the service log and obtain the advertising information of each candidate advertisement from the candidate advertisement set in the advertising database. Furthermore, it can obtain the time when the target object browses the interface of social software 421 and the advertisement position in the interface.

[0144] Step 270: The computer device uses the target prediction model to predict the push effect based on the characteristics of the predicted object, the predicted push features, and the predicted scenario features, and obtains the predicted click-through rate and predicted conversion rate of each candidate push content in different scenarios.

[0145] For example, the cloud execution server 430 can use a target prediction model to predict the advertising effect based on the characteristics of the prediction object, the characteristics of the prediction push, and the characteristics of the prediction scenario. In this way, it can obtain the predicted click-through rate and the predicted conversion rate of each candidate advertisement in different scenarios.

[0146] Step 280: The computer device determines the exposure price of each candidate push content in different scenarios based on the predicted click-through rate, predicted conversion rate, bidding control parameters, and the preset pricing corresponding to each candidate push content.

[0147] For example, the cloud execution server 430 can obtain bidding control parameters and the preset price corresponding to each candidate advertisement. Further, for example, the cloud execution server 430 can determine the exposure price of each candidate advertisement in different scenarios based on the predicted click-through rate, predicted conversion rate, bidding control parameters and the preset price corresponding to each candidate advertisement.

[0148] Step 290: Determine the target push content based on the exposure price of each candidate push content in different scenarios.

[0149] For example, the cloud execution server 430 can sort each exposure price in descending order and determine a preset number of candidate ads corresponding to the previous exposure prices as target ads. Further, the cloud execution server 430 can determine multiple target ads and then display these multiple target ads on the interface of the social software 421. For example, please refer to... Figure 5 As the target object swipes up on the interface, each target advertisement can be displayed sequentially.

[0150] In this embodiment, training samples can be obtained, including sample push information, sample object information, sample scene information, and push effect labels. The sample push information, sample object information, and sample scene information are input into a prediction network for feature extraction to obtain corresponding sample push features, sample object features, and sample scene features. Then, the push effect is predicted based on the sample push features, sample object features, and sample scene features through the prediction network to obtain the sample push effect of the training samples for different scenarios. The predicted push effect includes at least one of predicted click-through rate or predicted conversion rate.

[0151] Furthermore, based on the sample push effect and the push effect label, binary classification cross-entropy calculation is performed to determine the target loss corresponding to the prediction network. Then, based on the target loss, the prediction network is iteratively trained using multi-task learning in different scenarios until the prediction network meets preset conditions, thus obtaining the target prediction model. Next, push information, object information, and scene information are acquired and input into the target prediction model to output the predicted push effect of the target push content in different scenarios.

[0152] Therefore, by fusing sample scene features into sample general features, the personalized representation of the network under different scenarios is enhanced, thus avoiding inconsistent prediction accuracy of click-through rate and conversion rate in the trained target prediction model. Based on the sample gating features transformed into vector-based sample rearranged features and global parameters, the weight parameters of the prediction layer in the prediction network are determined. Furthermore, based on the sample fusion features, the pre-defined prediction network is subjected to multi-task learning under different scenarios, making the model's parameter learning space more sufficient and ensuring consistent model convergence across different scenarios. This improves the accuracy and stability of the target prediction model in predicting push effects under different scenarios.

[0153] Please see Figure 8 This document illustrates a structural block diagram of an information processing device 500 provided in an embodiment of this application. The information processing device 500 includes: an information acquisition module 510, used to acquire push information, object information, and scene information; and an effect prediction module 520, used to input the push information, object information, and scene information into a target prediction model, and output the predicted push effect of the target push content in different scenarios. The predicted push effect includes at least one of predicted click-through rate or predicted conversion rate. The target prediction model is obtained by multi-task learning of a preset prediction network based on sample fusion features for different scenarios. The sample fusion features are obtained by vector fusion of a general sample feature (concatenated from sample object features and sample push features) with a gating feature (sample scene features after gating adjustment). The weight parameters of the prediction layer in the prediction network are determined by the sample rearranged features (sample gating features after vector transformation) and global parameters.

[0154] In some embodiments, the information processing device 500 may further include: a sample acquisition module, a feature extraction module, an effect prediction module, a loss determination module, and a network training module.

[0155] The system comprises the following modules: a sample acquisition module for acquiring training samples, including sample push information, sample object information, sample scene information, and push effect labels; a feature extraction module for inputting the sample push information, sample object information, and sample scene information into the prediction network for feature extraction, resulting in corresponding sample push features, sample object features, and sample scene features; an effect prediction module for predicting push effects based on the sample push features, sample object features, and sample scene features through the prediction network, obtaining the sample push effects corresponding to different scenarios for the training samples, including at least one of predicted click-through rate or predicted conversion rate; a loss determination module for performing binary cross-entropy calculation based on the sample push effects and push effect labels to determine the target loss corresponding to the prediction network; and a network training module for iteratively training the prediction network for multi-task learning in different scenarios based on the target loss until the prediction network meets preset conditions, thus obtaining the target prediction model.

[0156] In some embodiments, the effect prediction module may include a splicing unit, a gating unit, a fusion unit, and a prediction unit. The splicing unit is used to perform vector splicing on sample object features and sample push features through a prediction network to obtain sample general features; the gating unit is used to perform gating adjustment on sample scene features to obtain sample gating features corresponding to the sample scene features; the fusion unit is used to perform positional multiplication on the sample gating features and sample general features to obtain sample fusion features; and the prediction unit is used to predict the push effect based on the sample fusion features and sample gating features to obtain the sample push effect corresponding to each scene in the training samples.

[0157] In some embodiments, the gating unit may be specifically used to: perform a first nonlinear calculation on the sample scene features based on a first activation function through a first fully connected layer to obtain sample intermediate features; and perform a second nonlinear calculation on the sample intermediate features based on a second activation function and a scaling parameter through a second fully connected layer to obtain sample gating features.

[0158] In some embodiments, the prediction unit may include: a segmentation unit, a rearrangement unit, a product unit, and a prediction unit. The segmentation unit performs vector segmentation on the sample gating features to obtain sample segmentation features corresponding to each scene; the rearrangement unit performs matrix rearrangement on each sample segmentation feature to obtain sample rearrangement features corresponding to each sample segmentation feature; the product unit performs positional multiplication of each sample rearrangement feature with the global parameters of the prediction layer in the prediction network to obtain the weight parameters of the prediction layer corresponding to each scene; and the prediction unit predicts the push effect based on the weight parameters of the prediction layer corresponding to each scene and the sample fusion features to obtain the sample push effect corresponding to each sample scene of the training samples.

[0159] In some embodiments, the scenario may include a first sample scenario, a second sample scenario, and a third sample scenario. The product sub-unit may be specifically used to: perform a positional product operation on the sample rearrangement features corresponding to the first sample scenario and the global parameters of the prediction layer in the prediction network to obtain the first weight parameters of the prediction layer corresponding to the first sample scenario; perform a positional product operation on the sample rearrangement features corresponding to the second sample scenario and the global parameters of the prediction layer in the prediction network to obtain the second weight parameters of the prediction layer corresponding to the second sample scenario; and perform a positional product operation on the sample rearrangement features corresponding to the third sample scenario and the global parameters of the prediction layer in the prediction network to obtain the third weight parameters of the prediction layer corresponding to the third sample scenario.

[0160] In some embodiments, the sample push effect may include a first sample click-through rate, a second sample conversion rate, and a third sample conversion rate. The prediction subunit may be specifically used to: predict the push effect based on the weight parameters of the prediction layer and the sample fusion features to obtain the sample push effect corresponding to each scenario of the training sample, including: predicting the push effect based on the first weight parameters of the prediction layer and the sample fusion features corresponding to the first sample scenario to obtain the first sample click-through rate corresponding to the first sample scenario of the training sample; predicting the push effect based on the second weight parameters of the prediction layer and the sample fusion features corresponding to the second sample scenario to obtain the second sample conversion rate corresponding to the second sample scenario of the training sample; and predicting the push effect based on the third weight parameters of the prediction layer and the sample fusion features corresponding to the third sample scenario to obtain the third sample conversion rate corresponding to the third sample scenario of the training sample.

[0161] In some embodiments, the sample push effect includes a first sample click-through rate, a second sample conversion rate, and a third sample conversion rate, and the push effect labels include a first sample click label, a second sample conversion label, and a third sample conversion label. The loss determination module can be specifically used to: determine a first loss based on the first sample click-through rate and the first sample click label; determine a second loss based on the second sample conversion rate and the second sample conversion label; determine a third loss based on the third sample conversion rate and the third sample conversion label; and determine the target loss corresponding to the prediction network based on the first loss, the second loss, and the third loss.

[0162] In some embodiments, the sample acquisition module may be specifically used to: acquire multiple candidate scene information; calculate the filtering parameters corresponding to each candidate scene information based on the candidate features of each candidate scene information, wherein the filtering parameters satisfy the Bernoulli distribution of the hyperparameters corresponding to the filtering parameters; determine a preset number of sample scenes from multiple candidate scenes according to each filtering parameter; and determine training samples based on sample push information, sample object information, and sample scene information corresponding to the sample scenes.

[0163] In some embodiments, the information acquisition module 510 may be specifically used to: in response to a prediction request for a target object, acquire object information of the target object; acquire push information of each candidate push content from the candidate push set; and acquire scene information of different scenarios corresponding to the prediction request.

[0164] In some embodiments, the effect prediction module 520 may be specifically used to: predict the push effect based on the characteristics of the predicted object, the predicted push features, and the predicted scenario features through the target prediction model, and obtain the predicted click-through rate and predicted conversion rate of each candidate push content in different scenarios.

[0165] In some embodiments, the information processing device 500 may further include: a price acquisition module, a price determination module, and a push determination module. The price acquisition module is used to acquire bidding control parameters and the target conversion bid corresponding to each candidate push content; the price determination module is used to determine the exposure price corresponding to each candidate push content in different scenarios based on the predicted click-through rate, predicted conversion rate, bidding control parameters, and the target conversion bid corresponding to each candidate push content; the push determination module is used to sort each exposure price in descending order and determine a preset number of candidate push contents corresponding to the preceding exposure prices as target push contents, so as to expose the target push contents.

[0166] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0167] In the several embodiments provided in this application, the coupling between modules can be electrical, mechanical, or other forms of coupling.

[0168] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0169] The solution provided in this application can acquire push information, object information, and scene information. Further, it inputs these information into a target prediction model and outputs the predicted push effect of the target content in different scenarios. The predicted push effect includes at least one of predicted click-through rate or predicted conversion rate. The target prediction model is obtained by multi-task learning of a pre-defined prediction network based on sample fusion features for different scenarios. The sample fusion features are obtained by vector fusion of a general sample feature (concatenated from sample object features and sample push features) with a gating feature (sample scene features after gating adjustment). The weight parameters of the prediction layer in the prediction network are determined by the sample rearranged features (based on vector transformation of the sample gating features) and global parameters.

[0170] Therefore, by fusing sample scene features into sample general features, the personalized representation of the network under different scenarios is enhanced, thus avoiding inconsistent prediction accuracy of click-through rate and conversion rate in the trained target prediction model. Based on the sample gating features transformed into vector-based sample rearranged features and global parameters, the weight parameters of the prediction layer in the prediction network are determined. Furthermore, based on the sample fusion features, the pre-defined prediction network is subjected to multi-task learning under different scenarios, making the model's parameter learning space more sufficient and ensuring consistent model convergence across different scenarios. This improves the prediction performance of the target prediction model.

[0171] like Figure 9 As shown in the figure, this application embodiment also provides a computer device 600, which includes a processor 610, a memory 620, a power supply 630, and an input unit 640. The memory 620 stores a computer program, and when the computer program is called by the processor 610, it can execute the various method steps provided in the above embodiments. Those skilled in the art will understand that the structure of the computer device shown in the figures does not constitute a limitation on the computer device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0172] The processor 610 may include one or more processing cores. The processor 610 connects to various parts of the entire battery management system using various interfaces and lines. It executes instructions, programs, instruction sets, or program sets stored in the memory 620, calls data stored in the memory 620, performs various functions and processes data within the battery management system, and performs various functions and processes data within the computer device, thereby providing overall control of the computer device. Optionally, the processor 610 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 610 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 610 and may be implemented separately using a communication chip.

[0173] The memory 620 may include random access memory (RAM) or read-only memory (ROM). The memory 620 can be used to store instructions, programs, instruction sets, or program assemblies. The memory 620 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described above. The data storage area may also store data created during the use of the computer device (such as phonebook and audio / video data). Accordingly, the memory 620 may also include a memory controller to provide the processor 610 with access to the memory 620.

[0174] The power supply 630 can be logically connected to the processor 610 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 630 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0175] The input unit 640 can be used to receive input digital or character information, and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0176] Although not shown, the computer device 600 may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 610 in the computer device loads the executable files corresponding to the processes of one or more computer programs into the memory 620 according to the following instructions, and the processor 610 runs the data stored in the memory 620, such as telephone books and audio and video data, thereby implementing the various method steps provided in the foregoing embodiments.

[0177] like Figure 10 As shown, this application embodiment also provides a computer-readable storage medium 700, which stores a computer program 710. The computer program 710 can be invoked by a processor to execute various method steps provided in this application embodiment.

[0178] Computer-readable storage media can be electronic storage devices such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, computer-readable storage media includes non-volatile computer-readable storage medium. Computer-readable storage medium 700 has storage space for computer programs that perform any of the method steps in the above embodiments. These computer programs can be read from or written to one or more computer program products. The computer programs can be compressed in an appropriate form.

[0179] According to one aspect of this application, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the various method steps provided in the above embodiments.

[0180] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. An information processing method, characterized in that, The method includes: Obtain push notifications, recipient information, and context information; The push information, the object information and the scene information are input into the target prediction model, and the predicted push effect of the target push content in different scenes is output. The predicted push effect includes at least one of the predicted click-through rate or the predicted conversion rate. The target prediction model is obtained by multi-task learning of a preset prediction network based on sample fusion features in different scenarios. The sample fusion features are sample general features concatenated based on sample object features and sample push features, and obtained by vector fusion with sample gating features after gating adjustment of sample scene features. The weight parameters of the prediction layer in the prediction network are determined based on the sample rearranged features after vector transformation of the sample gating features and global parameters.

2. The method according to claim 1, characterized in that, The target prediction model is trained through the following steps: Obtain training samples, which include sample push information, sample object information, sample scene information, and push effect tags; The sample push information, the sample object information, and the sample scene information are input into the prediction network for feature extraction to obtain the corresponding sample push features, sample object features, and sample scene features. The prediction network is used to predict the push effect based on the sample push features, the sample object features and the sample scene features, so as to obtain the sample push effect of the training sample corresponding to different scenes. The sample push effect includes at least one of the predicted click-through rate or the predicted conversion rate. Based on the sample push effect and the push effect label, perform binary classification cross-entropy calculation to determine the target loss corresponding to the prediction network; Based on the target loss, the prediction network is iteratively trained using multi-task learning in different scenarios until the prediction network meets the preset conditions, thus obtaining the target prediction model.

3. The method according to claim 2, characterized in that, The step of predicting the push effect based on the sample push features, the sample object features, and the sample scene features through the prediction network to obtain the sample push effect of the training samples corresponding to different scenes includes: The sample object features and the sample push features are vector-concatenated using the prediction network to obtain the sample general features. Gating the sample scene features is performed to obtain the sample gating features corresponding to the sample scene features; Perform a positional product operation on the sample gating features and the sample general features to obtain the sample fusion features; Based on the sample fusion features and the sample gating features, the push effect is predicted to obtain the sample push effect corresponding to each scenario in the training samples.

4. The method according to claim 3, characterized in that, The step of gating the sample scene features to obtain the sample gating features corresponding to the sample scene features includes: Through the first fully connected layer, the sample scene features are subjected to a first nonlinear calculation based on the first activation function to obtain the sample intermediate features; Through the second fully connected layer, a second nonlinear calculation is performed on the intermediate features of the sample based on the second activation function and scaling parameters to obtain the sample gating features.

5. The method according to claim 3 or 4, characterized in that, The step of predicting the push effect based on the sample fusion features and the sample gating features to obtain the sample push effect corresponding to each scene in the training samples includes: The sample gating features are vector-segmented to obtain sample segmentation features corresponding to each scene; Perform matrix rearrangement on each of the sample segmentation features to obtain the sample rearrangement features corresponding to each of the sample segmentation features; Each of the sample rearranged features is multiplied in place with the global parameters of the prediction layer in the prediction network to obtain the weight parameters of the prediction layer for each scene. Based on the weight parameters of the prediction layer corresponding to each scenario and the sample fusion features, the push effect is predicted to obtain the sample push effect corresponding to each scenario of the training samples.

6. The method according to claim 5, characterized in that, The scenarios include a first sample scenario, a second sample scenario, and a third sample scenario. The step of performing a positional multiplication operation between the rearranged features of each sample and the global parameters of the prediction layer in the prediction network to obtain the weight parameters of the prediction layer corresponding to each scenario includes: The sample rearrangement features corresponding to the first sample scenario are multiplied in place with the global parameters of the prediction layer in the prediction network to obtain the first weight parameters of the prediction layer corresponding to the first sample scenario. The sample rearrangement features corresponding to the second sample scenario are multiplied in place with the global parameters of the prediction layer in the prediction network to obtain the second weight parameters of the prediction layer corresponding to the second sample scenario. The sample rearrangement features corresponding to the third sample scenario are multiplied in tandem with the global parameters of the prediction layer in the prediction network to obtain the third weight parameters of the prediction layer corresponding to the third sample scenario.

7. The method according to claim 6, characterized in that, The sample push effect includes the click-through rate of the first sample, the conversion rate of the second sample, and the conversion rate of the third sample. The push effect prediction is performed based on the weight parameters of the prediction layer corresponding to each scenario and the sample fusion features to obtain the sample push effect for each scenario of the training samples, including: Based on the first weight parameters of the prediction layer corresponding to the first sample scenario and the sample fusion features, the push effect is predicted to obtain the first sample click rate corresponding to the first sample scenario of the training sample. Based on the second weight parameters of the prediction layer corresponding to the second sample scenario and the sample fusion features, the push effect is predicted to obtain the second sample conversion rate corresponding to the second sample scenario of the training sample. Based on the third weight parameters of the prediction layer corresponding to the third sample scenario and the sample fusion features, the push effect is predicted to obtain the third sample conversion rate corresponding to the third sample scenario of the training samples.

8. The method according to claim 7, characterized in that, The push effect labels include a first sample click label, a second sample conversion label, and a third sample conversion label. The step of performing binary classification cross-entropy calculation based on the sample push effect and the push effect labels to determine the target loss corresponding to the prediction network includes: Based on the click-through rate and the click tags of the first sample, determine the first loss; The second loss is determined based on the conversion rate of the second sample and the conversion label of the second sample; The third loss is determined based on the conversion rate and conversion label of the third sample; The target loss corresponding to the prediction network is determined based on the first loss, the second loss, and the third loss.

9. The method according to any one of claims 2 to 4, characterized in that, The acquisition of training samples includes: Obtain information on multiple candidate scenarios; Based on the candidate features of each candidate scene information, the filtering parameters corresponding to each candidate scene information are calculated, and the filtering parameters satisfy the Bernoulli distribution of the hyperparameters corresponding to the filtering parameters. Based on each of the filtering parameters, a preset number of sample scenarios are determined from multiple candidate scenarios; Training samples are determined based on sample push information, sample object information, and sample scene information corresponding to the sample scene.

10. The method according to claim 1, characterized in that, The acquisition of push information, object information, and scene information includes: In response to a prediction request for a target object, obtain the object information of the target object; Retrieve push information for each candidate push content from the candidate push set; Obtain scene information for different scenarios corresponding to the prediction request.

11. The method according to claim 10, characterized in that, The step of inputting the push information, the object information, and the scene information into the target prediction model and outputting the predicted push effect of the target push content in different scenes includes: Using the target prediction model, the push effect is predicted based on the predicted object features, predicted push features, and predicted scenario features, to obtain the predicted click-through rate and predicted conversion rate of each candidate push content in different scenarios.

12. The method according to claim 11, characterized in that, The method further includes: Obtain the bidding control parameters and the target conversion bid corresponding to each candidate push content; Based on the predicted click-through rate, the predicted conversion rate, the bidding control parameters, and the target conversion bid corresponding to each candidate push content, the exposure price corresponding to each candidate push content in different scenarios is determined. Each exposure price is sorted in descending order, and a preset number of candidate push contents corresponding to the preceding exposure prices are determined as target push contents, so that the target push contents can be exposed.

13. An information processing device, characterized in that, The device includes: The information acquisition module is used to acquire push information, object information, and scene information; The effect prediction module is used to input the push information, the object information and the scene information into the target prediction model, and output the predicted push effect of the target push content in different scenes. The predicted push effect includes at least one of the predicted click-through rate or the predicted conversion rate. The target prediction model is obtained by multi-task learning of a preset prediction network based on sample fusion features in different scenarios. The sample fusion features are sample general features concatenated based on sample object features and sample push features, and obtained by vector fusion with sample gating features after gating adjustment of sample scene features. The weight parameters of the prediction layer in the prediction network are determined based on the sample rearranged features after vector transformation of the sample gating features and global parameters.

14. The apparatus according to claim 13, characterized in that, The information processing device further includes: The sample acquisition module is used to acquire training samples, which include sample push information, sample object information, sample scene information, and push effect tags. The feature extraction module is used to input the sample push information, the sample object information and the sample scene information into the prediction network for feature extraction, so as to obtain the corresponding sample push features, sample object features and sample scene features; The effect prediction module is used to predict the push effect based on the sample push features, the sample object features and the sample scene features through the prediction network, so as to obtain the sample push effect of the training sample corresponding to different scenes. The sample push effect includes at least one of the predicted click-through rate or the predicted conversion rate. The loss determination module is used to perform binary cross-entropy calculation based on the sample push effect and the push effect label to determine the target loss corresponding to the prediction network. The network training module is used to iteratively train the prediction network for multi-task learning in different scenarios based on the target loss until the prediction network meets the preset conditions to obtain the target prediction model.

15. The apparatus according to claim 14, characterized in that, The effect prediction module includes: The concatenation unit is used to perform vector concatenation of the sample object features and the sample push features through the prediction network to obtain the sample general features; A gating unit is used to perform gating adjustment on the sample scene features to obtain sample gating features corresponding to the sample scene features; The fusion unit is used to perform a positional multiplication operation on the sample gating features and the sample general features to obtain sample fusion features; The prediction unit is used to predict the push effect based on the sample fusion features and the sample gating features, so as to obtain the sample push effect corresponding to each scene in the training samples.

16. The apparatus according to claim 15, characterized in that, The gating unit is also used for: Through the first fully connected layer, the sample scene features are subjected to a first nonlinear calculation based on the first activation function to obtain the sample intermediate features; Through the second fully connected layer, a second nonlinear calculation is performed on the intermediate features of the sample based on the second activation function and scaling parameters to obtain the sample gating features.

17. The apparatus according to claim 15 or 16, characterized in that, The prediction unit includes: The molecular segmentation unit is used to perform vector segmentation on the sample gating features to obtain the sample segmentation features corresponding to each scene; The rearrangement subunit is used to perform matrix rearrangement on each of the sample segmentation features to obtain the sample rearrangement features corresponding to each of the sample segmentation features; The product sub-unit is used to perform a positional product operation between each of the sample rearranged features and the global parameters of the prediction layer in the prediction network to obtain the weight parameters of the prediction layer corresponding to each scene. The prediction subunit is used to predict the push effect based on the weight parameters of the prediction layer corresponding to each scenario and the sample fusion features, so as to obtain the sample push effect of each scenario corresponding to the training samples.

18. The apparatus according to claim 17, characterized in that, The scenarios include a first sample scenario, a second sample scenario, and a third sample scenario, and the product sub-unit is configured as follows: The sample rearrangement features corresponding to the first sample scenario are multiplied in place with the global parameters of the prediction layer in the prediction network to obtain the first weight parameters of the prediction layer corresponding to the first sample scenario. The sample rearrangement features corresponding to the second sample scenario are multiplied in place with the global parameters of the prediction layer in the prediction network to obtain the second weight parameters of the prediction layer corresponding to the second sample scenario. The sample rearrangement features corresponding to the third sample scenario are multiplied in tandem with the global parameters of the prediction layer in the prediction network to obtain the third weight parameters of the prediction layer corresponding to the third sample scenario.

19. The apparatus according to claim 18, characterized in that, The sample push effect includes the first sample click-through rate, the second sample conversion rate, and the third sample conversion rate. The prediction subunit is configured as follows: Based on the first weight parameters of the prediction layer corresponding to the first sample scenario and the sample fusion features, the push effect is predicted to obtain the first sample click rate corresponding to the first sample scenario of the training sample. Based on the second weight parameters of the prediction layer corresponding to the second sample scenario and the sample fusion features, the push effect is predicted to obtain the second sample conversion rate corresponding to the second sample scenario of the training sample. Based on the third weight parameters of the prediction layer corresponding to the third sample scenario and the sample fusion features, the push effect is predicted to obtain the third sample conversion rate corresponding to the third sample scenario of the training samples.

20. The apparatus according to claim 19, characterized in that, The push effect tags include a first sample click tag, a second sample conversion tag, and a third sample conversion tag. The loss determination module is configured as follows: Based on the click-through rate and the click tags of the first sample, determine the first loss; The second loss is determined based on the conversion rate of the second sample and the conversion label of the second sample; The third loss is determined based on the conversion rate and conversion label of the third sample; The target loss corresponding to the prediction network is determined based on the first loss, the second loss, and the third loss.

21. The apparatus according to any one of claims 14 to 16, characterized in that, The sample acquisition module is configured as follows: Obtain information on multiple candidate scenarios; Based on the candidate features of each candidate scene information, the filtering parameters corresponding to each candidate scene information are calculated, and the filtering parameters satisfy the Bernoulli distribution of the hyperparameters corresponding to the filtering parameters. Based on each of the filtering parameters, a preset number of sample scenarios are determined from multiple candidate scenarios; Training samples are determined based on sample push information, sample object information, and sample scene information corresponding to the sample scene.

22. The apparatus according to claim 13, characterized in that, The information acquisition module is configured as follows: In response to a prediction request for a target object, obtain the object information of the target object; Retrieve push information for each candidate push content from the candidate push set; Obtain scene information for different scenarios corresponding to the prediction request.

23. The apparatus according to claim 22, characterized in that, The effect prediction module is configured as follows: Using the target prediction model, the push effect is predicted based on the predicted object features, predicted push features, and predicted scenario features, to obtain the predicted click-through rate and predicted conversion rate of each candidate push content in different scenarios.

24. The apparatus according to claim 23, characterized in that, The information processing device further includes: The price acquisition module is used to acquire bidding control parameters and the target conversion bid corresponding to each candidate push content; The pricing module is used to determine the exposure price of each candidate push content in different scenarios based on the predicted click-through rate, the predicted conversion rate, the bidding control parameters, and the target conversion bid corresponding to each candidate push content. The push determination module is used to sort each of the exposure prices in descending order, determine a preset number of candidate push contents corresponding to the previous exposure prices as target push contents, so as to expose the target push contents.

25. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be invoked by a processor to perform the method as described in any one of claims 1 to 12.

26. A computer device, characterized in that, include: Memory; One or more processors are coupled to the memory; One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, the one or more computer programs being configured to perform the method as described in any one of claims 1 to 12.

27. A computer program product, characterized in that, The computer program product includes a computer program stored in a storage medium; a processor of a computer device reads the computer program from the storage medium and executes the computer program, causing the computer device to perform the method as described in any one of claims 1 to 12.

Images