Data processing method and device, computer equipment, storage medium and program product

Abstract

The application provides a data processing method and device, computer equipment, a storage medium and a program product. Business data of a business object is obtained, the business data including object features of the business object, material features of material in a first media resource, and style features of N media styles; first correlation analysis is performed on the object features and the material features to predict a first interaction degree between the business object and the first media resource; second correlation analysis is performed on the object features, the material features, and the style features of the N media styles to predict N second interaction degrees between the business object and the first media resource; interaction degree fusion processing is performed on the first interaction degree and the N second interaction degrees to obtain N interest degrees of the business object to the first media resource, and a recommended media style for recommending the first media resource is determined from the N media styles according to the N interest degrees. The application can accurately predict the interest degree of a business object to a media resource.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a data processing method, a data processing apparatus, a computer device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] Current recommendation systems can predict a business user's interest in media resources by predicting their interest in media materials within those materials. However, practical experience has shown that this interest does not accurately reflect the user's overall interest in the media resources. A user's interest in media resources may be influenced by other factors. Considering only media materials as a single influencing factor when recommending media resources leads to low posterior click-through rates and conversion rates. Therefore, current recommendation systems cannot accurately predict a business user's interest in media resources. Summary of the Invention

[0003] This application provides a data processing method, apparatus, computer equipment, storage medium, and program product that can accurately predict the interest of business users in media resources.

[0004] On one hand, embodiments of this application provide a data processing method, which includes:

[0005] Obtain business data of the business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1.

[0006] Perform a first correlation analysis on object features and material features to predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object will interact with the first media resource.

[0007] A second correlation analysis is performed on the object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree is used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style.

[0008] The first interaction degree and N second interaction degrees are fused to obtain N interest degrees of the business object for the first media resource. Based on the N interest degrees, the recommended media style for recommending the first media resource is determined from N media styles. The N interest degrees correspond one-to-one with the N media styles.

[0009] Accordingly, embodiments of this application provide a data processing apparatus, which includes:

[0010] The acquisition unit is used to acquire business data of the business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1.

[0011] The processing unit is used to perform a first correlation analysis on the object features and material features, and predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object will interact with the first media resource.

[0012] The processing unit is also used to perform a second correlation analysis on the object features, material features and style features of N media styles, and predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree is used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style.

[0013] The processing unit is also used to perform interaction degree fusion processing on the first interaction degree and N second interaction degrees to obtain N interest degrees of the business object to the first media resource, and to determine the recommended media style for recommending the first media resource from N media styles based on the N interest degrees, with each of the N interest degrees corresponding to one of the N media styles.

[0014] In one implementation, the processing unit, when performing a first correlation analysis on object features and material features to predict the first degree of interaction between the business object and the first media resource, specifically executes the following steps:

[0015] The object features and material features are divided into first features and second features. The complexity of the second feature is higher than that of the first feature.

[0016] Perform a first feature crossover process on the first feature to obtain the first crossover feature;

[0017] The second feature is subjected to a second feature crossover process to obtain a second crossover feature; wherein the first feature crossover process is different from the second feature crossover process.

[0018] The first and second cross features are fused to obtain the fused features.

[0019] Based on the integration characteristics, predict the first degree of interaction between the business object and the first media resource.

[0020] In one implementation, the processing unit is further configured to perform the following steps:

[0021] Based on the first cross feature, the second cross feature is subjected to feature enhancement processing to obtain the enhanced feature;

[0022] The processing unit is used to perform feature fusion processing on the first cross feature and the second cross feature to obtain the fused feature. Specifically, it performs the following steps:

[0023] The first cross feature and the enhanced feature are fused together to obtain the fused feature.

[0024] In one implementation, the processing unit, when performing feature enhancement processing on the second cross feature based on the first cross feature to obtain the enhanced feature, specifically performs the following steps:

[0025] The first cross feature is subjected to a first feature transformation process to obtain the first transformed feature;

[0026] The second cross feature is subjected to a second feature transformation process to obtain the second transformed feature; wherein the first feature transformation process is different from the second feature transformation process.

[0027] The enhanced features are obtained by performing feature multiplication between the first transformed features and the second transformed features.

[0028] In one implementation, the processing unit, used to perform a second correlation analysis on object features, material features, and style features of N media styles, and to predict N second interaction degrees between the business object and the first media resource, specifically performs the following steps:

[0029] Obtain the integrated features obtained by integrating object features and material features;

[0030] The style features of N media styles are subjected to feature cross processing to obtain the cross features of N media styles;

[0031] The integrated features are then fused with the cross features of the N media styles to obtain the fused features corresponding to the N media styles.

[0032] Based on the fusion characteristics corresponding to N media styles, predict N second interaction degrees between the business object and the first media resource.

[0033] In one implementation, a prediction model is deployed in the data processing device, and N second interaction degrees are predicted by the style prediction module in the prediction model. The style prediction module includes a style processing structure shared by various media styles.

[0034] The processing unit is used to perform feature cross-processing on the style features of N media styles respectively. When obtaining the cross features of N media styles, it specifically performs the following steps:

[0035] The style processing structure is copied N-1 times to obtain N-1 copied style processing structures;

[0036] The N style processing structures are called to perform feature cross-processing on the style features of the N media styles, resulting in the cross-features of the N media styles.

[0037] In one implementation, a prediction model is deployed in the data processing device. N second interaction degrees are predicted by the style prediction module in the prediction model. The style prediction module includes P style processing structures, each of which corresponds one-to-one with a P media style. The N media styles are the media styles that match the business object among the P media styles, where P is an integer greater than or equal to N.

[0038] The processing unit is used to perform feature cross-processing on the style features of N media styles respectively. When obtaining the cross features of N media styles, it specifically performs the following steps:

[0039] Call the N style processing structures corresponding to the N media styles from the P style processing structures, and perform feature cross processing on the style features of the N media styles respectively to obtain the cross features of the N media styles.

[0040] In one implementation, obtaining the unit is also used to perform the following steps:

[0041] Obtain recommendation features related to the business clients of the business object;

[0042] The processing unit is used to perform a first correlation analysis on object features and material features, and to predict the first interaction degree between the business object and the first media resource. Specifically, it is used to perform the following steps: perform a first correlation analysis on object features, material features and recommendation features, and predict the first interaction degree between the business object and the first media resource.

[0043] The processing unit is used to perform a second correlation analysis on object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. Specifically, it performs the following steps: performing a second correlation analysis on object features, material features, recommendation features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource.

[0044] In one implementation, the first media resource is any one of the media resources in the set to be recommended; the processing unit is further configured to perform the following steps:

[0045] Based on the interest level of the business users towards each media resource in the media resource set under the corresponding recommended media style, the media resources are sorted.

[0046] Determine the media resources of interest to recommend to business users from the ranking results;

[0047] Recommend media resources of interest to business users based on the recommended media styles corresponding to the media resources of interest.

[0048] In one implementation, the processing unit is further configured to perform the following steps:

[0049] Select M media styles from N media styles that match the media material in the first media resource, where M is a positive integer less than or equal to N;

[0050] The processing unit, used to perform a second correlation analysis on object features, material features, and style features of N media styles, and to predict the N second interaction degrees between the business object and the first media resource, specifically executes the following steps:

[0051] A second correlation analysis is performed on the object characteristics, material characteristics, and style characteristics of M media styles to predict N second interaction degrees between the business object and the first media resource.

[0052] In one implementation, the processing unit, when filtering M media styles from N media styles that match the media material in the first media resource, specifically performs the following steps:

[0053] Obtain style files for N media styles;

[0054] Perform material type identification processing on the style files of N media styles respectively to obtain the first material type information corresponding to the N media styles. The first material type information is used to indicate the material types that are allowed to be presented in the corresponding media style.

[0055] The media materials in the first media resource are processed to identify the material type, and the second material type information is used to indicate the material type contained in the media materials in the first media resource.

[0056] Among N media styles, the media style that matches the first material type information with the second material type information is determined as the M media styles that match the media material of the first media resource.

[0057] In one implementation, the data processing device deploys a prediction model, which includes a material prediction module and a style prediction module; the acquisition unit is further configured to perform the following steps:

[0058] Obtain sample data for training the prediction model. The sample data includes sample object features of sample business objects, sample material features of media materials in sample media resources, and style features of target media styles. Among them, sample media resources are media resources recommended to sample business objects, and target media styles are media styles used to recommend sample media resources.

[0059] The processing unit is also used to perform the following steps:

[0060] The material prediction module is invoked to perform a first correlation analysis on the characteristics of the sample object and the characteristics of the sample material, and to predict the first sample interaction degree between the sample business object and the sample media resource.

[0061] The style prediction module is invoked to perform a second correlation analysis on the style features of the sample object, sample material, and target media style, and to predict the second sample interaction degree between the sample business object and the sample media resource under the target media style.

[0062] The interaction degree of the first sample and the interaction degree of the second sample are fused to obtain the sample interest degree of the sample business object on the sample media resource under the target media style, and the prediction model is trained based on the sample interest degree.

[0063] In one implementation, the processing unit, when training the prediction model based on sample interest, specifically performs the following steps:

[0064] The style prediction module includes a style processing structure shared by P types of media styles, and the style processing structure is trained based on the sample interest.

[0065] Alternatively, the style prediction module includes style processing structures corresponding to each of the P media styles, and the style processing structures corresponding to the target media style are trained based on the sample interest.

[0066] Among them, the target media style is the media style used to recommend sample media resources from among the P types of media styles.

[0067] Accordingly, embodiments of this application provide a computer device, which includes:

[0068] A processor is a tool for implementing computer programs.

[0069] A computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the above-described data processing method.

[0070] Accordingly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when read and executed by a processor of a computer device, causes the computer device to perform the aforementioned data processing method.

[0071] Accordingly, this application provides a computer program product 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 data processing method described above.

[0072] In this embodiment, business data of the business object can be obtained. This business data may include the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1. Therefore, this embodiment, when predicting the business object's interest in the media resource, considers not only the media material but also the media style, taking into account multiple factors that may affect the interest level. A first correlation analysis can be performed on the object characteristics and material characteristics to predict the first interaction degree between the business object and the first media resource. The first interaction degree can be represented by the probability of the business object interacting with the first media resource. Thus, by analyzing the correlation between the object characteristics and material characteristics, the probability of the business object interacting with the first media resource can be predicted under the influencing factor of media material. A second correlation analysis can be performed on object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles, and any one of the second interaction degrees can be used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style. It can be seen that by analyzing the correlation between object features, material features, and style features, the probability of the business object interacting with the first media resource can be predicted under the two influencing factors of media material and media style. The first interaction degree and N second interaction degrees can be fused to obtain N interest degrees of the business object for the first media resource. Based on the N interest degrees, the recommended media style for recommending the first media resource is determined from N media styles. The N interest degrees correspond one-to-one with the N media styles. It can be seen that the embodiments of this application not only consider the influence of media materials on interest degrees, but also the influence of media materials combined with media styles on interest degrees. This can more accurately predict the business object's interest degree for media resources, which is conducive to determining the media style for recommending media resources based on interest degrees from multiple media styles. This can improve the recommendation effect of media resources. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0074] Figure 1 This is a schematic diagram illustrating a presentation format of a media resource provided in an embodiment of this application;

[0075] Figure 2 This is a schematic diagram of a media style provided in an embodiment of this application;

[0076] Figure 3 This is a schematic diagram of the architecture of a recommendation system provided in an embodiment of this application;

[0077] Figure 4 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of this application;

[0078] Figure 5 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0079] Figure 6 This is a schematic diagram illustrating the integration of style features into a feature sequence, as provided in an embodiment of this application.

[0080] Figure 7 This is a schematic diagram of the prediction logic of a prediction model provided in an embodiment of this application;

[0081] Figure 8 This is a schematic diagram of an interaction degree fusion processing provided in an embodiment of this application;

[0082] Figure 9 This is a flowchart illustrating another data processing method provided in an embodiment of this application;

[0083] Figure 10 This is a schematic diagram of the framework of a prediction model provided in an embodiment of this application;

[0084] Figure 11 This is a schematic diagram of the structure of a prediction model provided in an embodiment of this application;

[0085] Figure 12 This is a flowchart illustrating another data processing method provided in an embodiment of this application;

[0086] Figure 13 This is a schematic diagram of model training provided in an embodiment of this application;

[0087] Figure 14This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0088] Figure 15 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0089] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0090] To better understand the technical solutions provided in the embodiments of this application, the technical terms used in the embodiments of this application will be introduced first:

[0091] I. Media Resources and Media Materials:

[0092] Media resources refer to information resources recommended to business users through media platforms. Business users refer to the target audience of these media resources, and the media platform refers to the business client of the business users. In other words, media resources are information resources recommended to business users through their business clients. Media materials refer to the content of media resources, usually provided by the media resource provider. The types of media materials can include at least one of the following: text, images, video, and audio, etc. For example, an advertisement is a media resource, and the advertising materials within an advertisement can be called creative materials. Creative materials are the advertising content provided by the advertiser.

[0093] II. Media Formats:

[0094] Media styles refer to the presentation format of media resources. Media materials within a media resource can be combined with different types of media styles to form media resources presented in various formats. A media style file is a file used to encapsulate the media style. A media style can include two elements: templates and components. Correspondingly, a media style file can include template files and component files. Template files encapsulate templates, and component files encapsulate components. Templates describe the layout of the style, specifically the layout structure between areas used to fill different media materials (e.g., text, images, titles, or components). For example, a template can describe the layout of areas for titles, images, and components, such as top-middle-bottom or left-middle-right layouts. Components are controls that can be used to interact with media resources, such as click components, shake components, twist components, red envelope rain components, top card components, or bottom card components.

[0095] Figure 1 An exemplary presentation of a media resource is shown. In this presentation, the components include a Topcard component and a click component. The areas for filling the Topcard component, the areas for filling the image material, the areas for filling the title and description material, and the areas for filling the click component are arranged in a top-middle-bottom layout. The topmost area is for filling the Topcard component, the middle area is for filling the image material and the area for filling the title and description material, and the bottommost area is for filling the click component.

[0096] A media style pool is a database that stores style files for media styles. It can contain various media styles derived from combinations of different templates and components. For example, if a media style includes a region for filling a component, X templates and Y components can be combined to create X*Y media styles, where X and Y are both integers greater than 1. Similarly, if a media style includes two regions for filling components, X templates and Y components can be combined to create X*Y*Y media styles, where X and Y are both integers greater than 1. Figure 2 As shown, the second template is paired with two different components (e.g., Figure 2 Combining component 1 and component 2 in the code can produce two different media styles.

[0097] III. Recommendation System:

[0098] A recommender system is a system used to identify media resources that a business user might be interested in and recommend those resources to that user. The main task of a recommender system is to predict the degree of interest a business user has in media resources. For example... Figure 3As shown, a recommendation system can include coarse-ranking and fine-ranking devices. The coarse-ranking device can deploy a coarse-ranking model, and the fine-ranking device can deploy a fine-ranking model. The coarse-ranking model is typically a lightweight (here, lightweight specifically means simple structure and few parameters) AI model, while the fine-ranking model is typically a complex (here, complex specifically means complex structure and many parameters) AI model. The fine-ranking model is used in the recommendation system to predict the interest of business users in media resources. The coarse-ranking device can use the coarse-ranking model to filter a portion of media resources from the large number of recalled resources to enter the fine-ranking stage. The fine-ranking device can use the fine-ranking model to predict the interest of business users in each filtered media resource, and rank the filtered media resources according to the interest. Then, it can select the Top K media resources (i.e., the media resources ranked in the top K positions, where K is a positive integer) from the ranking results to recommend to the business users. The filtering operation of the coarse-ranking model can reduce the computational burden on the fine-ranking model.

[0099] Based on the above-described technical terms, this application proposes a data processing method. This method addresses the following issues in recommendation systems: different presentation formats of media resources significantly impact the interactive behavior of business users, particularly in the large differences between posterior click-through rates and conversion rates. When the fine-ranking model lacks awareness of media styles, significant granular prediction biases occur at the media style level. To address this issue, this data processing method proposes a scheme that considers media styles during the fine-ranking process to predict the interest of business users in media resources. Specifically, this method achieves granular prediction of media styles during the fine-ranking process, considering the impact of media styles on interest, thus enabling more accurate prediction of business users' interest in media resources, achieving personalized media style recommendations, and optimizing the personalized recommendation effect of media resources.

[0100] Furthermore, this data processing method can encapsulate the object features of the business object, the material features of the media material, and the style features of various media styles using feature sequences. This allows the fine-mapping model to make a single request based on the feature sequence, and the fine-mapping model can predict in parallel the business object's interest in media resources presented using different media styles. This enables efficient prediction of the business object's interest in media resources.

[0101] The data processing system involved in the embodiments of this application is described below.

[0102] like Figure 4As shown, the data processing system may include a business client 401 and a recommendation system 402. The recommendation system 402 may include a coarse ranking device 4021 and a fine ranking device 4022. This application embodiment does not limit the connection method between the business client 401 and the recommendation system 402. The business client 401 and the recommendation system 402 can establish a direct communication connection via wired communication, or they can establish an indirect communication connection via wireless communication. Similarly, this application embodiment does not limit the connection method between the coarse ranking device 4021 and the fine ranking device 4022. The coarse ranking device 4021 and the fine ranking device 4022 can establish a direct communication connection via wired communication, or they can establish an indirect communication connection via wireless communication.

[0103] The business client 401 of the business object may have a need to recommend media resources; for example, there may be a need to recommend media resources during the startup process of the business client 401, or when certain conditions are triggered during the operation of the business client 401, or when switching back to the business client 401 from other clients; which media resources the business client 401 recommends to the business object, and according to what media style the business client 401 recommends media resources, are determined by the recommendation system 402.

[0104] This application embodiment does not limit the service type of the service client 401. Service type refers to the type of service functions that the service client 401 can provide. The service client 401 can be any type of service client. For example, the service client 401 can be an audio / video client (which can provide functions related to audio / video playback), a game client (which can provide functions related to game playing), or an e-commerce client (which can provide functions related to online goods transactions), and so on. This application embodiment does not limit the operating type of the service client 401. Operating type refers to the way the service client 401 operates on the terminal device. For example, the service client 401 can be an application, a mini-program, or a web (World Wide Web) application, and so on.

[0105] For example, if the media resource is an advertisement and the business client 401 is an e-commerce application, the e-commerce application can recommend the advertisement to the business object when the business object starts the e-commerce application or switches from other applications to the e-commerce application. This recommended advertisement can be called a splash screen advertisement. If the media resource is an advertisement and the business client 401 is a game mini-program, the game mini-program can recommend the advertisement to the business object when the current game level ends. The business object can unlock the next game level by watching the recommended advertisement for a period of time. This recommended advertisement can be called a rewarded advertisement. If the media resource is an advertisement and the business client 401 is an audio / video web application, the audio / video web application can recommend the advertisement to the business object when the business object pauses video playback. This recommended advertisement can be called an interstitial advertisement.

[0106] In the data processing system consisting of business client 401 and recommendation system 402, the data processing flow may include:

[0107] ①The business client 401 can send recommendation requests to the recommendation system 402.

[0108] The recommendation system can respond to recommendation requests with a 402 error and execute the following steps ②-⑥:

[0109] ② The coarse-ranking device 4021 in the recommendation system 402 can call the coarse-ranking model to filter out a portion of candidate media resources from the recalled media resources (e.g., Figure 4 The candidate media resources (1-10) in the media style pool are selected and sent to the fine-ranking device 4022 in the recommendation system 402; the coarse-ranking device 4021 can also select a portion of candidate media styles from the media style pool (e.g., Figure 4 The candidate media styles 1-3 are selected and sent to the fine-ranking device 4022 in the recommendation system 402.

[0110] ③ For any candidate media resource, the fine-mapping device 4022 can predict the business object's interest in the candidate media resource across multiple candidate media styles based on the object characteristics of the business object, the material characteristics of the media material in the candidate media resource, and the style characteristics of the three candidate media styles; the fine-mapping device 4022 can then determine the recommended media style of the candidate media resource from among the multiple candidate media styles based on the business object's interest in the candidate media resource across multiple candidate media styles. Taking any candidate media resource as an example... Figure 4 Taking candidate media resource 1 as an example, the fine-sorting device 4022 can predict the interest of business objects in candidate media resource 1 under three candidate media styles (i.e., candidate media style 1 to candidate media style 3), and can determine the candidate media style 3 with the highest interest as the recommended media style of candidate media resource 1.

[0111] ④ For multiple candidate media resources, the ranking device 4022 can sort the candidate media resources according to the business object's interest in the corresponding recommended media style, and determine the Top K media resources as the media resources of interest in the ranking results. Figure 4 For example, the fine sorting device 4022 can sort candidate media resources 1 to 10 based on their respective interest levels, and determine the top 3 (i.e., Top 3) candidate media resources 5, 3 and 8 as the interest media resources of the business object.

[0112] ⑤ The 4022 fine-formatting device can combine media materials from interest-based media resources with recommended media styles to obtain interest-based media resources presented according to the recommended media styles. Figure 4 For example, the recommended media style for candidate media resource 5 is candidate media style 1, the recommended media style for candidate media resource 3 is candidate media style 2, and the recommended media style for candidate media resource 8 is candidate media style 3. The fine-formatting device 4022 can combine the media materials in candidate media resource 5 with candidate media style 1 to obtain candidate media resource 5 presented according to candidate media style 1, can combine the media materials in candidate media resource 3 with candidate media style 2 to obtain candidate media resource 3 presented according to candidate media style 2, and can combine the media materials in candidate media resource 8 with candidate media style 3 to obtain candidate media resource 8 presented according to candidate media style 3.

[0113] It is worth noting that, in Figure 4In the illustrated embodiment, the process of combining media materials in the interest media resource with the recommended media style of the interest media resource is performed by the fine-sorting device 4022. Optionally, the process of combining media materials in the interest media resource with the recommended media style of the interest media resource can also be performed by other devices in the recommendation system 402. Other devices refer to devices in the recommendation system 402 other than the coarse-sorting device 4021 and the fine-sorting device 4022. Other devices can combine the media materials in the interest media resource with the recommended media style of the interest media resource to obtain the interest media resource presented according to the recommended media style. Optionally, the process of combining media materials in the interest media resource with the recommended media style of the interest media resource can also be performed by the business client 401. The recommendation system 402 can send the style files of the media materials in the interest media resource and the recommended media style of the interest media resource to the business client 401. The business client 401 can combine the media materials in the interest media resource with the style file of the recommended media style to obtain the interest media resource presented according to the recommended media style.

[0114] ⑥ The recommendation system 402 can send media resources of interest presented according to the recommended media style to the business client 401. Figure 4 For example, the recommendation system 402 can send candidate media resource 5 presented according to candidate media style 1, candidate media resource 3 presented according to candidate media style 2, and candidate media resource 8 presented according to candidate media style 3 to the business client 401.

[0115] ⑦ When there is a recommendation need, the business client 401 can recommend media resources of interest presented in the recommended media style to the business object.

[0116] It is evident that the data processing system considers the impact of media style on interest during the fine-tuning process. By predicting the granularity of media style during the fine-tuning process, it can more accurately predict the interest of business objects in media resources under different media styles, thereby improving the accuracy of fine-tuning.

[0117] exist Figure 4In the data processing system shown, the business client 401 can run on a terminal device, and both the coarse sorting device 4021 and the fine sorting device 4022 can be servers. The terminal devices mentioned in this embodiment can include, but are not limited to, any of the following: smartphones, tablets, laptops, desktop computers, smartwatches, smart home appliances, smart vehicle terminals, and aircraft. The servers mentioned in this embodiment can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers 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 (Content Delivery Network), and big data and artificial intelligence platforms, etc.

[0118] Figure 4 The data processing system shown in the embodiments is intended to more clearly illustrate the technical solutions of the embodiments of this application, and does not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0119] In the embodiments of this application, the relevant data collection and processing (e.g., the collection and processing of object characteristics of business objects) should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject (or have a legal basis), and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0120] The data processing method provided in the embodiments of this application will be described in detail below.

[0121] This application provides a data processing method, which includes constructing a feature sequence and processing the prediction results (i.e., interest level). This data processing method can be executed by a computer device, such as a ranking device deployed with a prediction model (i.e., the ranking model mentioned above). Figure 5 As shown, the data processing method may include, but is not limited to, the following steps S501-S505:

[0122] S501, Obtain business data of the business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1.

[0123] In step S501, the first media resource can be any media resource in the set of media resources to be recommended. The set of media resources can be obtained by the coarse-ranking model from the recalled media resources; that is, the media resources in the set are the candidate media resources mentioned above. The N media styles can be obtained by the coarse-ranking model from the media style pool; that is, the media styles among the N media styles are the candidate media styles mentioned above. Each media resource in the set of media resources corresponds to N media styles.

[0124] In step S501, the object characteristics of a business object refer to information used to describe the object's features. These characteristics may include object attribute characteristics, object device characteristics, and object behavior characteristics. Object attribute characteristics describe the basic attributes of the business object. These characteristics typically remain unchanged over a period of time and are generally simple, making feature learning relatively easy. For example, object attribute characteristics may include the object's identifier (e.g., nickname, account, avatar, and description), age, gender, and interest tags. Object device characteristics describe the characteristics of the terminal device used by the business object (i.e., the terminal device running the business client). These characteristics also typically remain unchanged and are generally simple, making feature learning relatively easy. For example, object device characteristics may include the type of terminal device and the type of operating system running on the terminal device. Object behavior characteristics refer to information used to describe the historical interaction behavior of business objects in business clients. Object behavior characteristics change continuously over time, and are usually quite complex, making feature learning of object behavior characteristics difficult. For example, the object behavior characteristics of a business object may include media resources that the business object clicked on in the past 3 days, media resources that the business object clicked and converted in the past 3 days, videos that the business object watched in the past 3 days, and products that the business object browsed in the past 3 days.

[0125] The material characteristics of media materials in media resources can include the media material itself, as well as information describing the characteristics of the media material, such as the material identifier, the material description information, and the material title. The style characteristics of a media style refer to the information describing the style characteristics of the media style. Style characteristics can include the template identifier (i.e., template ID) sub-characteristic used in the media style, the layout sub-characteristic used in the media style (e.g., the layout sub-characteristic describes that the various areas in the media style are arranged in a top-middle-bottom layout), and the material type sub-characteristic allowed to be filled in the media style (e.g., the material type sub-characteristic describes that the media style allows filling with materials of types such as images, text, and components, and can also describe the specific component types).

[0126] The business data of a business object can be represented as a feature sequence. A single request is used to ask the ranking model to predict the business object's interest in the first media resource across N media styles. In other words, the business data of a business object is a feature sequence composed of the object features of the business object, the material features of the media material in the first media resource, and the style features of the N media styles. Formula 1 below illustrates the feature sequence form of the business data of the business object:

[0127] [u1 a1…(t1 t2… t N )]→[p1 p2… p N ] Formula 1

[0128] In Formula 1 above, u1 represents the object characteristics of the business object, a1… represents the material characteristics of the media material in the first media resource, and t1t2…t… N Represents the style features of N media styles, p1p2…p N This represents the N degrees of interest of a business object towards the first media resource under N media styles. It can be seen that the business object's business data encapsulates the object features of the business object, the material features of the media material in the first media resource, and the style features of the N media styles in the form of a feature sequence. Furthermore, the N media styles are encapsulated in an ordered manner within the feature sequence and passed to the fine-ranking model as style features. This allows the fine-ranking model to obtain the style features of the N media styles. Thus, by requesting the fine-ranking model only once through the feature sequence, the fine-ranking model can predict the business object's degree of interest towards the first media resource under N media styles, improving prediction efficiency.

[0129] Compared to traditional multimedia style prediction schemes, which are shown in Formula 2 below:

[0130]

[0131] Formula 2 above describes a traditional multimedia style prediction scheme, where N media styles are encapsulated into N requests. For example, [u1a1…t1] is encapsulated into one request, [u1a1…t1] is encapsulated into one request. N The N requests are encapsulated into another request, and these N requests need to be made to the ranking model N times sequentially. In other words, the ranking model can only predict the business object's interest in the first media resource under N media styles by making N requests, resulting in low prediction efficiency. Compared to traditional multimedia style prediction schemes, the feature sequence in this embodiment can improve prediction efficiency.

[0132] More specifically, the style features of a media style are multi-valued features. Multi-valued features mean that the style features of a media style include multiple sub-features. As described above, the style features of a media style can include the template identifier sub-feature used in the media style, the layout sub-feature used in the media style, and the material type sub-feature allowed to be filled in the media style. Thus, the multi-valued features of N media styles can be encapsulated in a feature sequence.

[0133] by Figure 6 For example, each media resource in the fine-sorting queue (i.e., the queue composed of individual media resources in the media resource set) corresponds to 3 media styles (e.g., Figure 6 (Media styles 1-3 in the text). Taking media resource 2 as an example, the feature sequence of media resource 2 can include style features of three media styles; the style feature of media style 1 is a multi-valued feature, which can include template identifiers (e.g., Figure 6 The template style 1) corresponds to the sub-features and layout structure (e.g., Figure 6 The layout structure 1) corresponds to the sub-features, and the material type (e.g., Figure 6 The sub-features corresponding to material type 1) in the example; similarly, the style features of media style 2 are multi-valued features, which may include template identifiers (e.g., Figure 6 The template style 2) corresponds to the sub-features and layout structure (e.g., Figure 6 The layout structure 2) corresponds to the sub-features, and the material type (e.g., Figure 6 The sub-features corresponding to material type 2) in the media style 3 are multi-valued features, which may include template identifiers (e.g., Figure 6 The template style 3) corresponds to the sub-features and layout structure (e.g., Figure 6 The layout structure 3) corresponds to the sub-features, and the material type (e.g., Figure 6 Sub-features corresponding to material type 3 in the data.

[0134] Before introducing steps S502-S505 in the embodiments of this application, the general idea of ​​the interest prediction model (i.e., the fine-grained ranking model) in the embodiments of this application will be introduced first. When the prediction model obtains the style features of N media styles from the feature sequence and performs style-granular prediction, considering the cost constraints of the prediction model, it is necessary to control the computational amount of the prediction model under multiple media styles. Considering that the object features of the business object and the material features of the media resources corresponding to each media style are the same in a request, therefore, as Figure 7 As shown, the prediction model can be split into two parts: a material prediction module (Main Tower) that predicts only once and a style prediction module (Style Tower) that predicts multiple times. In other words, the prediction model can include a material prediction module and a style prediction module.

[0135] The Main Tower module takes as input the object features of the business object and the material features of the media materials in the media resources. Its output is the probability at the material granularity level (i.e., the probability that the business object will interact with the media resources). The Style Tower module takes as input the style features of multiple media styles and estimates the probability at the style granularity level of multiple media styles in parallel during a single forward computation (i.e., the probability that the business object will interact with the media resources under multiple media styles). The prediction model can then fuse these two probabilities to obtain the probability at the material + style granularity level (i.e., the business object's interest in the media resources under multiple media styles). As can be seen, by splitting the prediction model into two parts, the probability at the material granularity level of a media resource is predicted only once. If the prediction model were not split into two parts, it would require multiple predictions; for example, N media styles would require N predictions. This reduces the prediction workload and improves prediction efficiency.

[0136] The prediction process will be described in detail below, in conjunction with steps S502-S505.

[0137] S502, perform a first correlation analysis on the object features and material features to predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object will interact with the first media resource.

[0138] Step S502 is performed by the material prediction module in the business model. The first correlation analysis refers to the first feature interaction learning of object features and material features. The first feature interaction learning refers to learning the relationship and mutual influence between object features and material features, so as to predict the first interaction degree between the business object and the first media resource. The first interaction degree can be used to indicate the probability of the business object interacting with the first media resource (i.e., the probability at the material granularity).

[0139] In detail, the first correlation analysis process may include: performing feature segmentation on object features and material features to obtain first features and second features; performing first feature cross processing on the first features to obtain first cross features; performing second feature cross processing on the second features to obtain second cross features; performing feature fusion processing on the first cross features and second cross features to obtain fused features; and predicting the first interaction degree between the business object and the first media resource based on the fused features.

[0140] The following technical points ①-③ are involved in the above-mentioned first correlation analysis:

[0141] ① The complexity of the second feature is higher than that of the first feature. This complexity can be used to measure the learning difficulty of a feature; features with low learning difficulty have low complexity, meaning they easily capture explicit linear relationships, such as the explicit linear relationship between the age of a business object and the type of media material. Features with high learning difficulty have high complexity, meaning they are difficult to capture implicit non-linear relationships, such as the implicit non-linear relationship between the access time of a business object in the business client and the type of media material in the business object's historical behavior.

[0142] For example, the sub-features in the first feature obtained by feature segmentation processing may include object attribute features and object device features in the object features, as well as all material features, and the sub-features in the second feature may include all object features and all material features.

[0143] ② The first feature cross-processing differs from the second feature cross-processing. This difference manifests in two aspects: First, the order of feature cross-processing differs; the second feature cross-processing corresponds to a higher-order feature cross-processing than the first. In other words, the second feature cross-processing is a higher-order feature cross-processing, while the first is a lower-order one. Second, the functions of feature cross-processing differ. The first feature cross-processing can be used to learn the explicit linear relationships between the sub-features of the first feature, while the second feature cross-processing can be used to learn the implicit non-linear relationships between the sub-features of the second feature. It's easy to see that the first and second feature cross-processing can interactively learn object features and material features from different dimensions. The first feature cross-processing captures the explicit linear relationships between object features and material features, while the second feature cross-processing captures the implicit non-linear relationships, which helps improve the accuracy of the first correlation analysis.

[0144] ③ Feature fusion processing refers to superimposing the first cross feature and the second cross feature.

[0145] S503, perform a second correlation analysis on the object features, material features and style features of N media styles, predict N second interaction degrees between the business object and the first media resource, the N second interaction degrees correspond one-to-one with the N media styles, and any second interaction degree is used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style.

[0146] Step S503 is executed by the style prediction module in the business model. The second correlation analysis refers to the second feature interaction learning of the object features, material features and style features of N media styles. The second feature interaction learning refers to learning the relationship and mutual influence between the style features of N media styles and the object features and material features, so as to predict N second interaction degrees between the business object and the first media resource under N media styles. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree can be used to indicate the probability (i.e., the probability at the style granularity) of the business object interacting with the first media resource presented using the corresponding media style.

[0147] In detail, the second correlation analysis process may include: obtaining integrated features obtained by integrating object features and material features; performing feature cross processing on the style features of N media styles to obtain cross features of N media styles; performing feature fusion processing on the integrated features and the cross features of N media styles to obtain fused features corresponding to N media styles; and predicting N second interaction degrees between the business object and the first media resource based on the fused features corresponding to N media styles.

[0148] The following technical points ①-② are involved in the above-mentioned second correlation analysis:

[0149] ① Feature integration processing refers to integrating the feature learning results of the material prediction module on object features and material features. Specifically, it refers to integrating the first cross feature and the second cross feature in step S502 to obtain N integrated features corresponding to N media styles. It can be seen that although the second correlation analysis process of the style prediction module introduces object features and material features, it does not repeatedly calculate these features. Instead, it directly introduces the calculation results of object features and material features from the first correlation analysis process of the material prediction module. This saves time spent on repeatedly calculating object features and material features, improving the efficiency of the second correlation analysis and thus improving the prediction efficiency of interest.

[0150] ② Feature fusion processing refers to superimposing the cross features of N media styles with the corresponding integrated features in the N integrated features to obtain the fused features corresponding to the N media styles.

[0151] S504, perform interaction degree fusion processing on the first interaction degree and N second interaction degrees to obtain N interest degrees of the business object on the first media resource, and the N interest degrees correspond one-to-one with N media styles.

[0152] In step S504, the interaction degree fusion processing refers to the addition or multiplication of interaction degrees. That is, the first interaction degree and N second interaction degrees can be added or multiplied to obtain N interest degrees of the business object towards the first media resource. These N interest degrees correspond one-to-one with N media styles. Interest degrees can be used to represent at least one of CTR (Click-Through-Rate) and CVR (Conversion Rate).

[0153] For example, such as Figure 8 As shown, N media styles are 3 media styles (for example, 3 media styles are...). Figure 8 Taking media styles 1-3 as an example, the first interactivity output by the material prediction module can be copied three times, and then added to the second interactivity corresponding to the three media styles output by the style prediction module to obtain the interest scores corresponding to the three media styles. The first interactivity is added to or multiplied by the second interactivity corresponding to media style 1 to obtain the interest score of 0.1 for media style 1, the first interactivity is added to or multiplied by the second interactivity corresponding to media style 2 to obtain the interest score of 0.4 for media style 2, and the first interactivity is added to or multiplied by the second interactivity corresponding to media style 3 to obtain the interest score of 0.8 for media style 3.

[0154] S505, determine the recommended media style from N media styles based on N interest levels to recommend the first media resource.

[0155] In step S505, the method for determining the recommended media style may include: determining the media style corresponding to the highest interest among N interest levels as the recommended media style for recommending the first series of media resources. Continuing with... Figure 8 For example, the interest score for media style 1 is 0.1, the interest score for media style 2 is 0.4, and the interest score for media style 3 is 0.8. Media style 3 has the highest interest score, so media style 3 can be identified as the recommended media style for recommending the first media resource.

[0156] For each media resource in the media resource set, the process of determining the recommended media style for the first media resource can be referenced to determine the recommended media resource corresponding to each media resource in the media resource set. Based on this, the media resources can be sorted according to the business object's interest in each media resource in the media resource set under the corresponding recommended media style; the recommended media resources can be determined from the sorting results; and the recommended media resources can be recommended to the business object according to the recommended media style corresponding to the media resources of interest. The sorting can be based on descending order of interest, and the media resources of interest refer to the Top K (i.e., the first K positions) media resources in the sorting results.

[0157] For example, a media resource set includes media resource 1, media resource 2, and media resource 3. The recommended media style for media resource 1 is media style 2, the recommended media style for media resource 2 is media style 3, and the recommended media style for media resource 3 is media style 1. The business object's interest in media resource 1 under media style 2 is 0.8, the business object's interest in media resource 2 under media style 3 is 0.9, and the business object's interest in media resource 3 under media style 1 is 0.7. The ranking result is media resource 2, media resource 1, and media resource 3. It can be determined that media resource 2, which ranks first in the ranking result, is the media resource of interest. Media resource 2 can be recommended according to media style 3. That is, the media materials in media resource 2 can be combined with the style file of media style 3 to obtain media resource 2 presented according to media style 3 for recommendation.

[0158] Steps S501-S505 of this embodiment describe the process of determining recommended media resources of interest for a business object, and the recommended media style used to present the media resources of interest, considering object characteristics, material characteristics, and style characteristics of N media styles. Optionally, in the above process, recommendation features related to the business client of the business object can also be considered. The recommendation features are introduced below.

[0159] Recommendation features, also known as contextual features, refer to information used to describe the recommendation characteristics of a business client. Recommendation features can include basic recommendation features and dynamic recommendation features. Basic recommendation features describe the fundamental attributes related to media resource recommendations within the business client of a business object. Basic recommendation features are typically simpler, and feature learning for them is less difficult. Basic recommendation features may include, but are not limited to, at least one of the following: the recommendation location identifier in the business client of the business object, the type of the business client (e.g., the type may include at least one of the business type and operation type mentioned above), and the recommendation location type. The recommendation location refers to the placement of media resources within the business client. For example, if the media resource is an advertisement, the basic recommendation features may include the advertisement placement identifier, the type of the business client, and the advertisement placement type (e.g., splash screen ad or rewarded ad). Dynamic recommendation features refer to information used to describe the dynamic attributes of a business client related to media resources. Dynamic recommendation features are usually quite complex, and feature learning for them is difficult. Dynamic recommendation features may include, but are not limited to, at least one of the following: request information of recommendation requests sent by the business client (for example, request information may include the time when the recommendation request was sent, the location where the recommendation request was sent, and the type of traffic operator used to send the recommendation request, etc.), and media resource interaction information under the type to which the business client belongs (for example, the click-through rate and conversion rate of media resources under the game type to which the business client belongs, etc.).

[0160] After considering the recommendation features, the object data of the business object obtained in step S501 may include the object features of the business object, the material features of the media material in the first media resource, the recommendation features, and the style features of the N media styles corresponding to the first media resource. In other words, the business data of the business object is a feature sequence composed of the object features of the business object, the material features of the media material in the first media resource, the recommendation features, and the style features of the N media styles.

[0161] The first correlation analysis in step S502 refers to the first feature interaction learning of object features, recommendation features and material features. The first feature interaction learning refers to learning the interrelationship and mutual influence between object features, recommendation features and material features, so as to predict the first interaction degree between the business object and the first media resource. In other words, the first correlation analysis can be performed on object features, material features and recommendation features to predict the first interaction degree between the business object and the first media resource.

[0162] In detail, the difference between the first relevance analysis process before and after considering recommendation features lies in the following: Before considering recommendation features, object features and material features can be segmented to obtain first features and second features. For example, sub-features in the first feature can include object attribute features and object device features from the object features, as well as all material features; sub-features in the second feature can include all object features and all material features. After considering recommendation features, object features, material features, and recommendation features can be segmented to obtain first features and second features; for example, sub-features in the first feature can include object attribute features and object device features from the object features, basic recommendation features from the recommendation features, as well as all material features; sub-features in the second feature can include all object features, all recommendation features, and all material features. Apart from this difference, the execution process of other steps in the first relevance analysis is the same before and after considering recommendation features.

[0163] After considering the recommendation features, the second correlation analysis in step S503 refers to the second feature interaction learning of the object features, material features, recommendation features and style features of N media styles. The second feature interaction learning refers to learning the interrelationship and mutual influence between the style features of N media styles and the object features, recommendation features and material features, so as to predict the N second interaction degrees between the business object and the first media resource under N media styles.

[0164] In detail, the difference between the second relevance analysis process before and after considering recommendation features lies in the following: Before considering recommendation features, integrated features obtained by integrating object features and material features can be acquired. After considering recommendation features, integrated features obtained by integrating object features, recommendation features, and material features can be acquired; furthermore, feature integration refers to integrating the first and second cross features introduced in step S502 after introducing recommendation features to obtain N integrated features corresponding to N media styles. Apart from this difference, the execution process of other steps in the second relevance analysis process is the same before and after considering recommendation features.

[0165] Except for steps S501-S503, which change before and after the introduction of recommendation features, the other steps (i.e., steps S504 and S505) remain unchanged. Please refer to the descriptions of steps S504 and S505 above for details, which will not be repeated here. It is easy to see that recommendation features also affect the business's interest in media resources. For example, for advertising, different ad placement types affect the business's click interest (e.g., the business's click interest in rewarded ads is higher than that in splash screen ads). Different business client types also affect the business's click interest (e.g., the business's click interest in ads in game clients is lower than that in e-commerce clients). Therefore, considering the impact of recommendation features on interest during the fine-tuning process can improve the accuracy of interest prediction, thereby improving the accuracy of fine-tuning.

[0166] In this embodiment, the interest of business objects in the presentation of media resources is personalized. The template layout and diverse components in the media style will affect the business object's click interest in the media resources, and even the subsequent conversion interest. Therefore, the style form is considered in the fine-tuning process of media resources. The style form can be combined to predict the business object's interest in the media resources more accurately and improve the accuracy of fine-tuning. Furthermore, the embodiments of this application can increase the computational dimension (i.e., increase the style dimension) without increasing additional resources. When a new dimension is added to the prediction granularity of the fine-ranking model, for example, expanding from object granularity * material granularity to object granularity * material granularity * style granularity, the traditional implementation logic (such as the implementation logic corresponding to Formula 2 above) will experience a Cartesian product-like growth. Currently, the number of requests for scene rendering styles is about 15 billion, the average length of the fine-ranking queue is 130, and the average number of candidate styles per creative is 3.7. If the fine-ranking model is requested directly at object granularity * material granularity * style granularity, the number of fine-ranking model calls will increase by about 5 trillion times, resulting in a significant performance drop. However, the embodiments of this application can encapsulate the style features of multiple media styles together with object features and material features in a feature sequence, realizing that a single request to the fine-ranking model allows the fine-ranking model to predict the interest level under N media styles, instead of requesting the fine-ranking model N times. The embodiments of this application can maintain the original number of requests when a new dimension is added to the prediction granularity without increasing additional resources.

[0167] This application provides a data processing method, which includes a detailed process of a first correlation analysis combining the structure of a predictive model (i.e., a fine-ranking model), and a detailed process of a second correlation analysis combining the structure of the predictive model. This data processing method can be executed by a computer device, such as a fine-ranking device deployed with the predictive model (i.e., the fine-ranking model mentioned above). Figure 9As shown, the data processing method may include, but is not limited to, the following steps S901-S905:

[0168] S901, Obtain business data of the business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, the recommendation characteristics related to the business client of the business object, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1.

[0169] In step S901, the business data of the business object is a feature sequence including the object features of the business object, the material features of the media material in the first media resource to be recommended, the recommendation features related to the business client of the business object, and the style features of N media styles corresponding to the first media resource.

[0170] S902, the material prediction module is called to perform a first correlation analysis on the object features, recommendation features and material features to predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object will interact with the first media resource.

[0171] Before introducing step S902, let's first combine... Figure 10 This section introduces the framework of the content prediction module (Main Tower) in the prediction model. The content prediction module may include a first feature cross-network, a second feature cross-network, and a feature enhancement network. Based on this framework, in step S902, the first relevance analysis process may include: performing feature segmentation on object features, recommendation features, and content features to obtain a first feature (wide feature) and a second feature (deep feature); calling the first feature cross-network to perform first feature cross-processing on the first feature to obtain a first cross feature; calling the second feature cross-network to perform second feature cross-processing on the second feature to obtain a second cross feature; calling the feature enhancement network to perform feature enhancement processing on the second cross feature based on the first cross feature to obtain an enhanced feature; performing feature fusion processing on the first cross feature and the enhanced feature to obtain a fused feature; and predicting the first interaction degree between the business object and the first media resource based on the fused feature.

[0172] The following technical points ①-⑤ are involved in the above-mentioned first correlation analysis:

[0173] ①This application's embodiments use an FM (Factorization Machine) network as the first feature crossover network, a DCNV2 (Deformable Convolution Network V2) network as the second feature crossover network, and an LHUC (Learning Hidden Unit Contributions) network as the feature enhancement network for illustration.

[0174] ② After feature segmentation processing, such as Figure 11 As shown, the first feature (wide feature) can include user-based attributes and device information from the object features, basic recommendation features from the recommendation features (e.g., ad placement ID, media industry ID, and placement type), and all ad side features. Combined with... Figure 11 The detailed structure of the prediction model shown is as follows: the first feature can be processed by a dimensionality reduction layer, and then a first feature cross-processing network is called to perform first feature cross-processing on the dimensionality-reduced first feature to obtain the first cross feature. In this embodiment, the dimensionality reduction layer is sparse embedding as an example. The dimensionality reduction layer can be used to embed a high-dimensional sparse vector into a low-dimensional dense vector; that is, the first feature after dimensionality reduction is a low-dimensional dense vector. The advantage of dimensionality reduction is that using a low-dimensional sparse vector to represent the first feature ensures that only a few dimensions of the first feature have non-zero values, which significantly reduces computational complexity, especially when processing large-scale datasets. Furthermore, low-dimensional sparse vectors typically have better generalization ability and stronger interpretability.

[0175] ③ The second feature can include all user-side features, all context-side features, and all ad-side features. Combined with... Figure 11 The detailed structure of the prediction model shown is as follows: the second feature can be processed by a dimensionality reduction layer, and then a connection layer is called to connect the dimensionality-reduced second feature. The second feature cross network is then called to perform second feature cross processing on the connected second feature to obtain the second cross feature. In this embodiment, the connection layer is Concat as an example. The connection layer can be used to increase the number of channels or feature dimensions so as to better capture the relationship between different features in subsequent layers.

[0176] ④ The feature enhancement process may include: performing a first feature transformation on the first cross feature to obtain a first transformed feature; performing a second feature transformation on the second cross feature to obtain a second transformed feature; and performing a feature multiplication operation between the first transformed feature and the second transformed feature to obtain the enhanced feature. The first feature transformation and the second feature transformation differ in that they combine... Figure 11 The detailed structure of the prediction model shown includes the following: First feature transformation processing refers to applying a first feature transformation to the first cross-feature using a structure of Layer Norm + First Activation Function (ReLU) + Fully Connected Layer (FC) and a structure of Layer Norm + First Activation Function (ReLU) + Fully Connected Layer (FC) + Second Activation Function (Sigmoid)*2. Second feature transformation processing refers to applying a second feature transformation to the second cross-feature using a structure of Layer Norm + Fully Connected Layer (FC) or Layer Norm + First Activation Function (ReLU) + Fully Connected Layer (FC). Feature multiplication refers to element-wise product. Furthermore, the embodiments of this application do not limit the number of feature enhancement processes. Figure 11 The following example illustrates the process of performing three feature enhancement steps. Feature enhancement highlights important features in the second cross-feature analysis, thereby improving the accuracy of the first correlation analysis.

[0177] ⑤ The feature fusion process may include: the result of feature transformation processing of the first cross feature through the structure of normalization layer + first activation function + fully connected layer, and the result of feature transformation processing of the enhanced feature through the structure of normalization layer + first activation function + fully connected layer, and then feature fusion processing is performed to obtain the fused feature.

[0178] S903, the style prediction module is called to perform a second correlation analysis on the object features, material features, recommendation features and style features of N media styles, and predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree is used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style.

[0179] In step S903, as Figure 10 As shown, the style prediction module (Style Tower) in the prediction model is a DNN (Deep Neural Network). Figure 10 With N=3, meaning the first media resource corresponds to 3 media formats (for example, Figure 10Taking media styles 1-3 as an example. In step S903, the second correlation analysis process may include: obtaining the integrated features obtained by integrating the object features, recommendation features and material features; performing feature cross processing on the style features of N media styles respectively to obtain the cross features of N media styles; performing feature fusion processing on the integrated features and the cross features of N media styles respectively to obtain the fused features corresponding to N media styles; and predicting N second interaction degrees between the business object and the first media resource based on the fused features corresponding to N media styles.

[0180] The following technical points ①-④ are involved in the above second correlation analysis:

[0181] ① Figure 11 With N=3, meaning the first media resource corresponds to 3 media formats (for example, Figure 11 Taking media styles 1-3 as an example, the template features of N media styles can be reduced to N dimensionality-reduced template sparse embeddings after passing through a dimensionality reduction layer. Feature cross-processing can then be performed on these N template sparse embeddings to obtain cross-features for the N media styles. Furthermore, since the media styles remain unchanged after being determined, their template features are fixed. All dimensionality-reduced template features in the media style pool can be stored after the prediction model is trained. The N dimensionality-reduced template features can be directly retrieved from storage, ensuring that the template sparse embeddings returned when querying the storage for the same media style are identical. This guarantees that the template sparse embeddings of multiple media styles are within the same value space, reducing the learning cost of the prediction model.

[0182] ② The feature cross-processing of style features of N media styles is performed in parallel by calling N style processing structures in the style prediction module. This improves the learning efficiency of style features of N media styles, thereby improving the prediction efficiency at the style granularity level. This application's embodiment uses an FM network as an example for illustration.

[0183] In one implementation, the style prediction module may include a style processing structure shared by various media styles in the media style pool. In this implementation, the style processing structure can be copied N-1 times, resulting in N-1 copied style processing structures. These N style processing structures (i.e., the original style processing structure in the style prediction module plus the N-1 copied structures) are then used to perform feature cross-processing on the style features of the N media styles, resulting in the cross-features of the N media styles. It can be seen that the style prediction module contains only one style processing structure, which reduces the complexity and number of parameters in the module. Copying is done as needed, which can better adapt to combinations of different media styles. Furthermore, through copying, the N style processing structures share parameters. By sharing parameters, the same set of parameters is used to perform parallel calculations on the N media styles, ensuring that the parameters used when calculating different media styles are completely consistent. This achieves parallel style-granularity prediction of N media styles without increasing the number of parameters or computational complexity.

[0184] In another implementation, the style prediction module can include style processing structures corresponding to each of the P media styles in the media style pool. That is, the style prediction module includes P style processing structures, each corresponding one-to-one with one of the P media styles. The N media styles are the media styles among the P media styles that match the business object, where P is an integer greater than or equal to N. In this implementation, the N style processing structures corresponding to the N media styles can be called to perform feature cross-processing on the style features of the N media styles, resulting in cross-features for the N media styles. In this implementation, different types of media styles correspond to their respective style processing structures, allowing for targeted feature cross-processing for different types of media styles. This enables better learning of the characteristics of different types of media styles and improves the feature learning effect for different types of media styles.

[0185] ③ Feature integration processing refers to integrating the feature learning results of the material prediction module on object features, recommendation features, and material features. Specifically, it involves passing the first cross-feature output from the first feature cross-network and the second cross-feature output from the second feature cross-network in step S902 through two fully connected layers (FC), and then integrating them using an integration layer to obtain N integrated features corresponding to N media styles. This application embodiment uses the Expand layer as an example for illustration.

[0186] ④ The N integrated features can be fused with the cross features of N media styles through an overlay layer to obtain the fused features corresponding to the N media styles. In this embodiment, the overlay layer is an Add layer, which is an element-wise overlay without increasing the feature dimension or the number of feature channels. The fused features corresponding to the N media styles are processed by a structure of two normalization layers (Layer Norm) + a first activation function (ReLU) + a fully connected layer (FC) to predict the N second interaction degrees between the business object and the first media resource.

[0187] S904, perform interaction degree fusion processing on the first interaction degree and N second interaction degrees to obtain N interest degrees of the business object to the first media resource, and the N interest degrees correspond one-to-one with N media styles.

[0188] In this embodiment of the application, the execution process of step S904 is the same as described above. Figure 5 The execution process of step S504 in the illustrated embodiment is the same, and can be found in the above description. Figure 5 The relevant description of step S504 in the illustrated embodiment will not be repeated here.

[0189] S905, determine the recommended media style for recommending the first media resource from N media styles based on N interest levels.

[0190] In this embodiment of the application, the execution process of step S905 is the same as described above. Figure 5 The execution process of step S505 in the illustrated embodiment is the same; please refer to the above for details. Figure 5 The relevant description of step S504 in the illustrated embodiment will not be repeated here.

[0191] Optionally, the N media styles can be candidate media styles selected from the media style pool based on the matching degree between the object characteristics of the business object and the style characteristics of the media styles in the media style pool. Each media resource in the media resource set corresponds to N media styles. For each media resource in the media resource set, a media style that matches each media resource can be selected from the N media styles. The style characteristics of the media style that matches each media resource are used to predict the business object's interest in each media resource.

[0192] Taking the first media resource as an example, M media styles that match the media materials in the first media resource can be selected from N media styles, where M is a positive integer less than or equal to N. Based on this, the execution process of step S903 is adjusted as follows: the style prediction module is called to perform a second correlation analysis on the object features, material features, recommendation features, and style features of the M media styles to predict M second interaction degrees between the business object and the first media resource, with each of the M second interaction degrees corresponding to one of the M media styles. The execution process of step S904 is adjusted as follows: the first interaction degree and the M second interaction degrees are fused to obtain the M interest degrees of the business object towards the first media resource, with each of the M interest degrees corresponding to one of the M media styles. The execution process of step S905 is adjusted as follows: the recommended media style for recommending the first media resource is determined from the M media styles based on the M interest degrees. The execution process of the adjusted step S903 is similar to that of the original step S903. The execution process of the adjusted step S904 is similar to that of the original step S904. The execution process of the adjusted step S905 is similar to that of the original step S905. For details, please refer to the execution process of each step before the adjustment, which will not be repeated here.

[0193] In this context, matching media styles with media materials in the first media resource refers to matching material types. Specifically, the process of selecting M media styles from N media styles that match the media materials in the first media resource can include: obtaining style files for the N media styles; performing material type identification processing on the style files for each of the N media styles to obtain first material type information corresponding to the N media styles, which can be used to indicate the material types allowed to be presented in the corresponding media style; performing material type identification processing on the media materials of the first media resource to obtain second material type information, which can be used to indicate the material types contained in the media materials of the first media resource; and identifying the media styles from the N media styles whose first material type information matches the second material type information as the M media styles that match the media materials in the first media resource.

[0194] Matching the first and second media type information means that the number of identical media types indicated in the first and second media type information exceeds a preset threshold. For example, in media style 1 of N media styles, the allowed media types are images, videos, titles, and descriptive text. The media materials in the first media resource also include images, videos, titles, and descriptive text. If the number of identical media types (4) exceeds the preset threshold (3), then media style 1 can be considered a media style that matches the media materials in the first media resource. Conversely, in media style 2 of N media styles, the allowed media types are images, videos, titles, and descriptive text. The media materials in the first media resource include titles and descriptive text. If the number of identical media types (2) is less than the preset threshold (3), then media style 2 does not match the media materials in the first media resource.

[0195] It's easy to see that by selecting M media styles from N media styles that match the media material of the first media resource, the material type of the first media resource's media material can be made closer to the material types allowed to be filled in the selected M media styles. This can improve the presentation effect of the first media resource. In other words, for each media resource in the media resource set, media styles that produce a better presentation effect for the corresponding media resource can be selected from N media styles to participate in interest prediction. The presentation effect of media resources will affect the business object's interest in media resources. The business object usually has a lower interest in media resources with poor presentation effects. Filtering out some media styles that affect the presentation effect of media resources from participating in interest prediction can speed up the effective interest prediction efficiency.

[0196] In this embodiment, the prediction model is scalable, maintainable, and compatible with existing frameworks. Specifically, this embodiment can build a highly scalable and maintainable style-granularity prediction framework (i.e., style prediction module) within a fine-ranking model structure that lacks style-granularity prediction. Furthermore, even with expansion of the prediction dimensions (i.e., increasing the style-granularity prediction dimension), there is no additional increase in requests or machine costs for the fine-ranking model. Based on the style prediction module, this embodiment can provide personalized predictions with unique style granularity, further leveraging the prediction potential of the fine-ranking model and optimizing its personalized recommendation performance.

[0197] This application provides a data processing method, which includes the training process of a prediction model (i.e., a fine-ranking model). This data processing method can be executed by a computer device, such as a fine-ranking device deployed with the prediction model (i.e., the fine-ranking model mentioned above). Figure 12As shown, the data processing method may include, but is not limited to, the following steps S1201-S1205:

[0198] S1201, Obtain sample data for training the prediction model. The sample data includes sample object features of sample business objects, sample material features of media materials in sample media resources, and style features of target media styles. Among them, sample media resources are media resources recommended to sample business objects, and target media styles are media styles used to recommend sample media resources.

[0199] In step S1201, the sample data may be a feature sequence including sample object features of the sample business object, sample material features of media materials in the sample media resources, and style features of the target media style. Alternatively, if recommendation features are considered, the sample data may be a feature sequence including sample object features of the sample business object, sample recommendation features related to the business client of the sample business object, sample material features of media materials in the sample media resources, and style features of the target media style.

[0200] S1202, call the material prediction module in the prediction model to perform the first correlation analysis on the sample object features and sample material features, and predict the first sample interaction degree between the sample business object and the sample media resource.

[0201] In step S1202, the first correlation analysis process may include: performing feature segmentation processing on sample object features and sample material features to obtain first sample features and second sample features (or, if recommendation features are considered, feature segmentation processing can be performed on sample object features, sample recommendation features, and sample material features to obtain first sample features and second sample features); calling a first feature cross-processing network to perform first feature cross-processing on the first sample features to obtain first sample cross-features; calling a second feature cross-processing network to perform second feature cross-processing on the second sample features to obtain second sample cross-features; calling a feature enhancement network to perform feature enhancement processing on the second sample cross-features based on the first sample cross-features to obtain sample enhanced features; performing feature fusion processing on the first sample cross-features and sample enhanced features to obtain sample fusion features; and predicting the first sample interaction degree between sample business objects and sample media resources based on the sample fusion features.

[0202] The detailed process of the first correlation analysis during model training is as described above. Figure 9 The detailed process of the first correlation analysis in the model application example shown is similar; please refer to the above for details. Figure 9 The relevant descriptions of the embodiments shown will not be repeated here.

[0203] S1203, call the style prediction module in the prediction model to perform a second correlation analysis on the style features of the sample object features, sample material features and target media style, and predict the second sample interaction degree between the sample business object and the sample media resources under the target media style.

[0204] In step S1203, the second correlation analysis process may include: obtaining sample integrated features obtained by integrating the features of sample object features and sample material features (or, if sample recommendation features are considered, obtaining sample integrated features obtained by integrating the features of sample object features, sample recommendation features, and sample material features); performing feature cross processing on the style features of the target media style to obtain the cross features of the target media style; performing feature fusion processing on the sample integrated features and the cross features of the target media style to obtain the fused features corresponding to the target media style; and predicting the second sample interaction degree between the sample business object and the sample media resources under the target media style based on the fused features corresponding to the target media style.

[0205] The style prediction module can include a style processing structure shared by P media styles in the media style pool. It can call the style processing structure shared by P media styles to perform feature cross processing on the style features of the target media style to obtain the cross features of the target media style. Alternatively, the style prediction module can include the style processing structure corresponding to each of the P media styles in the media style pool. It can call the style processing structure corresponding to the target media style to perform feature cross processing on the style features of the target media style to obtain the cross features of the target media style.

[0206] The detailed process of the second correlation analysis during model training is the same as described above. Figure 9 The detailed process of the second correlation analysis in the model application side embodiment shown is similar; please refer to the above for details. Figure 9 The relevant descriptions of the embodiments shown will not be repeated here.

[0207] S1204, perform interaction degree fusion processing on the first sample interaction degree and the second sample interaction degree to obtain the sample interest degree of the sample business object on the sample media resource under the target media style.

[0208] In step S1204, the interaction degree fusion processing refers to the interaction degree addition processing or interaction degree multiplication processing. That is, the interaction degree of the first sample and the interaction degree of the second sample can be added or multiplied to obtain the sample interest degree of the sample business object on the sample media resource under the target media style.

[0209] S1205, train the prediction model based on the sample interest.

[0210] In step S1205, as Figure 13 As shown, training the prediction model based on sample interest can include training the material prediction module and the style prediction module based on sample interest. Training refers to inputting the sample interest into the loss function to calculate the loss value, and adjusting the model parameters of the prediction model in the direction of reducing the loss value, that is, adjusting the parameters of the material prediction module and the parameters of the style prediction module.

[0211] For the case where the style prediction module includes a style processing structure shared by P media styles in the media style pool, training the style prediction module can include: training the style processing structure based on sample interest, in which case the parameters of the style processing structure are shared across the P media styles. Alternatively, for the case where the style prediction module includes style processing structures corresponding to each of the P media styles in the media style pool, the style processing structure corresponding to the target media style can be trained based on sample interest. In this case, the sample interest for each media style is used to train the style processing structure corresponding to each media style, thus enabling targeted training of the feature learning ability of the style processing structure corresponding to each media style for the style features of each media style. Here, the target media style is the media style among the P media styles used to recommend sample media resources.

[0212] In addition, such as Figure 13 As shown, after updating the parameters of the style prediction module, the sparse embedding of the target media style after dimensionality reduction is also updated synchronously. After the model training is completed, the sparse embedding of the P media styles can be stored, which is beneficial for direct reading and use from storage when applying the model.

[0213] The steps S1201-S1205 above describe the training process of the prediction model. The trained prediction model can be used to perform the above operations. Figure 5 and Figure 9The model application side process is shown. By comparing the model application side process and the model training side process, it is not difficult to find that the computational logic of the model training side (i.e. offline) and the model application side (i.e. online) is asymmetrical. Specifically: (1) During offline training, the style granularity prediction process handles the media style in the exposure space. Therefore, there is only one media style that will be exposed in the training samples (e.g., the target media style of the model training side). There is only one style processing structure in the style prediction module (e.g., the style processing structure shared by P media styles or the style processing structure corresponding to the target media style) for forward calculation (i.e., predicting the interaction degree of the second sample) and backward calculation (i.e. adjusting the parameters). (2) During online inference, the media styles in the style granularity prediction processing request space are multiple candidate media styles. When a request is made, a single media resource will correspond to N candidate media styles. Therefore, N style processing structures will be called (for example, the N style processing structures are the style processing structure itself shared by P media styles + N-1 copied style processing structures, or the N style processing structures are the style processing structures corresponding to each of the N media styles). Each style processing structure will use the style features parsed in order as its own feature input to perform forward calculation and obtain N second interaction degrees. In addition, during online inference, multiple second interaction degrees will be calculated in one forward calculation. It is necessary to ensure that the result calculated by each style processing structure is ordered and independent of the index of the style processing structure.

[0214] In this embodiment, a prediction model that considers style granularity in predicting the interest of business objects in media resources can be trained, thereby improving the prediction accuracy of interest. Regarding the model parameters, this embodiment innovatively extends the model structure in a three-dimensional way, expanding the traditional 2D prediction model to a 3D prediction model ([batch_size, embedding_size] → [batch_size, template_num, embedding_size]), thus adding style granularity prediction.

[0215] The methods of the embodiments of this application have been described in detail above. In order to facilitate better implementation of the above solutions of the embodiments of this application, the apparatus of the embodiments of this application is provided below.

[0216] Please see Figure 14 , Figure 14 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. The data processing device can be installed in the computer equipment provided in the embodiment of this application. The computer equipment can be a fine ranking device in a recommendation system. The data processing device is equipped with a prediction model (i.e., a fine ranking model). Figure 14The data processing device shown can be a computer program running on a computer device, which can be used to execute... Figure 5 、and 9 or Figure 12 Some or all of the steps in the method embodiments shown. Please refer to [link / reference]. Figure 14 The data processing apparatus may include the following units:

[0217] The acquisition unit 1401 is used to acquire business data of the business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1.

[0218] The processing unit 1402 is used to perform a first correlation analysis on the object features and material features, and predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object will interact with the first media resource.

[0219] The processing unit 1402 is also used to perform a second correlation analysis on the object features, material features and style features of N media styles, and predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree is used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style.

[0220] The processing unit 1402 is also used to perform interaction degree fusion processing on the first interaction degree and N second interaction degrees to obtain N interest degrees of the business object to the first media resource, and to determine the recommended media style for recommending the first media resource from N media styles based on the N interest degrees, with the N interest degrees corresponding one-to-one with the N media styles.

[0221] In one implementation, the processing unit 1402, when performing a first correlation analysis on object features and material features to predict the first degree of interaction between the business object and the first media resource, specifically executes the following steps:

[0222] The object features and material features are divided into first features and second features. The complexity of the second feature is higher than that of the first feature.

[0223] Perform a first feature crossover process on the first feature to obtain the first crossover feature;

[0224] The second feature is subjected to a second feature crossover process to obtain a second crossover feature; wherein the first feature crossover process is different from the second feature crossover process.

[0225] The first and second cross features are fused to obtain the fused features.

[0226] Based on the integration characteristics, predict the first degree of interaction between the business object and the first media resource.

[0227] In one implementation, the processing unit 1402 is further configured to perform the following steps:

[0228] Based on the first cross feature, the second cross feature is subjected to feature enhancement processing to obtain the enhanced feature;

[0229] Processing unit 1402 is used to perform feature fusion processing on the first cross feature and the second cross feature to obtain the fused feature, specifically by executing the following steps:

[0230] The first cross feature and the enhanced feature are fused together to obtain the fused feature.

[0231] In one implementation, the processing unit 1402 is used to perform feature enhancement processing on the second cross feature based on the first cross feature, and when obtaining the enhanced feature, it specifically performs the following steps:

[0232] The first cross feature is subjected to a first feature transformation process to obtain the first transformed feature;

[0233] The second cross feature is subjected to a second feature transformation process to obtain the second transformed feature; wherein the first feature transformation process is different from the second feature transformation process.

[0234] The enhanced features are obtained by performing feature multiplication between the first transformed features and the second transformed features.

[0235] In one implementation, the processing unit 1402 is used to perform a second correlation analysis on the object features, material features, and style features of N media styles. When predicting the N second interaction degrees between the business object and the first media resource, it is specifically used to execute the following steps:

[0236] Obtain the integrated features obtained by integrating object features and material features;

[0237] The style features of N media styles are subjected to feature cross processing to obtain the cross features of N media styles;

[0238] The integrated features are then fused with the cross features of the N media styles to obtain the fused features corresponding to the N media styles.

[0239] Based on the fusion characteristics corresponding to N media styles, predict N second interaction degrees between the business object and the first media resource.

[0240] In one implementation, the N second interaction degrees are predicted by the style prediction module in the prediction model, which includes a style processing structure shared by various media styles.

[0241] Processing unit 1402 is used to perform feature cross-processing on the style features of N media styles respectively, and to obtain the cross features of N media styles, specifically to perform the following steps:

[0242] The style processing structure is copied N-1 times to obtain N-1 copied style processing structures;

[0243] The N style processing structures are called to perform feature cross-processing on the style features of the N media styles, resulting in the cross-features of the N media styles.

[0244] In one implementation, N second interaction degrees are predicted by the style prediction module in the prediction model. The style prediction module includes P style processing structures, each of which corresponds one-to-one with a P media style. The N media styles are the media styles among the P media styles that match the business object, where P is an integer greater than or equal to N.

[0245] Processing unit 1402 is used to perform feature cross-processing on the style features of N media styles respectively, and to obtain the cross features of N media styles, specifically to perform the following steps:

[0246] Call the N style processing structures corresponding to the N media styles from the P style processing structures, and perform feature cross processing on the style features of the N media styles respectively to obtain the cross features of the N media styles.

[0247] In one implementation, the acquisition unit 1401 is also used to perform the following steps:

[0248] Obtain recommendation features related to the business clients of the business object;

[0249] The processing unit 1402 is used to perform the following steps when performing a first correlation analysis on object features and material features to predict the first interaction degree between the business object and the first media resource: performing a first correlation analysis on object features, material features and recommendation features to predict the first interaction degree between the business object and the first media resource;

[0250] The processing unit 1402 is used to perform a second correlation analysis on object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. Specifically, it is used to perform the following steps: perform a second correlation analysis on object features, material features, recommendation features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource.

[0251] In one implementation, the first media resource is any one of the media resources in the set of media resources to be recommended; the processing unit 1402 is further configured to perform the following steps:

[0252] Based on the interest level of the business users towards each media resource in the media resource set under the corresponding recommended media style, the media resources are sorted.

[0253] Determine the media resources of interest to recommend to business users from the ranking results;

[0254] Recommend media resources of interest to business users based on the recommended media styles corresponding to the media resources of interest.

[0255] In one implementation, the processing unit 1402 is further configured to perform the following steps:

[0256] Select M media styles from N media styles that match the media material in the first media resource, where M is a positive integer less than or equal to N;

[0257] Processing unit 1402 is used to perform a second correlation analysis on object features, material features, and style features of N media styles. When predicting N second interaction degrees between the business object and the first media resource, it specifically performs the following steps:

[0258] A second correlation analysis is performed on the object characteristics, material characteristics, and style characteristics of M media styles to predict N second interaction degrees between the business object and the first media resource.

[0259] In one implementation, when processing unit 1402 filters M media styles from N media styles that match the media material in the first media resource, it specifically performs the following steps:

[0260] Obtain style files for N media styles;

[0261] Perform material type identification processing on the style files of N media styles respectively to obtain the first material type information corresponding to the N media styles. The first material type information is used to indicate the material types that are allowed to be presented in the corresponding media style.

[0262] The media materials in the first media resource are processed to identify the material type, and the second material type information is used to indicate the material type contained in the media materials in the first media resource.

[0263] Among N media styles, the media style that matches the first material type information with the second material type information is determined as the M media styles that match the media material of the first media resource.

[0264] In one implementation, the prediction model includes a material prediction module and a style prediction module; the acquisition unit 1401 is also used to perform the following steps:

[0265] Obtain sample data for training the prediction model. The sample data includes sample object features of sample business objects, sample material features of media materials in sample media resources, and style features of target media styles. Among them, sample media resources are media resources recommended to sample business objects, and target media styles are media styles used to recommend sample media resources.

[0266] Processing unit 1402 is also configured to perform the following steps:

[0267] The material prediction module is invoked to perform a first correlation analysis on the characteristics of the sample object and the characteristics of the sample material, and to predict the first sample interaction degree between the sample business object and the sample media resource.

[0268] The style prediction module is invoked to perform a second correlation analysis on the style features of the sample object, sample material, and target media style, and to predict the second sample interaction degree between the sample business object and the sample media resource under the target media style.

[0269] The interaction degree of the first sample and the interaction degree of the second sample are fused to obtain the sample interest degree of the sample business object on the sample media resource under the target media style, and the prediction model is trained based on the sample interest degree.

[0270] In one implementation, the processing unit 1402, when training the prediction model based on sample interest, specifically performs the following steps:

[0271] The style prediction module includes a style processing structure shared by P types of media styles, and the style processing structure is trained based on the sample interest.

[0272] Alternatively, the style prediction module includes style processing structures corresponding to each of the P media styles, and the style processing structures corresponding to the target media style are trained based on the sample interest.

[0273] Among them, the target media style is the media style used to recommend sample media resources from among the P types of media styles.

[0274] According to one embodiment of this application, Figure 14The data processing apparatus shown can be constructed by combining each unit individually or entirely into one or more other units, or one or more of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above-mentioned units are based on logical function division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the data processing apparatus may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0275] According to another embodiment of this application, the following can be achieved by running on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), a device capable of performing operations such as... Figure 5 , Figure 9 or Figure 12 Computer programs for the steps involved in some or all of the methods shown, to construct, for example... Figure 14 The data processing apparatus shown herein, and the data processing method for implementing the embodiments of this application, are described. A computer program may be recorded on, for example, a computer-readable storage medium, loaded onto the aforementioned computing device via the computer-readable storage medium, and executed therein.

[0276] In this embodiment, business data of the business object can be obtained. This business data may include the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1. Therefore, this embodiment, when predicting the business object's interest in the media resource, considers not only the media material but also the media style, taking into account multiple factors that may affect the interest level. A first correlation analysis can be performed on the object characteristics and material characteristics to predict the first interaction degree between the business object and the first media resource. The first interaction degree can be represented by the probability of the business object interacting with the first media resource. Thus, by analyzing the correlation between the object characteristics and material characteristics, the probability of the business object interacting with the first media resource can be predicted under the influencing factor of media material. A second correlation analysis can be performed on object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles, and any one of the second interaction degrees can be used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style. It can be seen that by analyzing the correlation between object features, material features, and style features, the probability of the business object interacting with the first media resource can be predicted under the two influencing factors of media material and media style. The first interaction degree and N second interaction degrees can be fused to obtain N interest degrees of the business object for the first media resource. Based on the N interest degrees, the recommended media style for recommending the first media resource is determined from N media styles. The N interest degrees correspond one-to-one with the N media styles. It can be seen that the embodiments of this application not only consider the influence of media materials on interest degrees, but also the influence of media materials combined with media styles on interest degrees. This can more accurately predict the business object's interest degree for media resources, which is conducive to determining the media style for recommending media resources based on interest degrees from multiple media styles. This can improve the recommendation effect of media resources.

[0277] Based on the above methods and apparatus embodiments, this application provides a computer device. Please refer to... Figure 15 , Figure 15 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 15 The computer device shown includes at least a processor 1501, an input interface 1502, an output interface 1503, and a computer-readable storage medium 1504. The processor 1501, input interface 1502, output interface 1503, and computer-readable storage medium 1504 can be connected via a bus or other means.

[0278] The computer-readable storage medium 1504 can be stored in the memory of a computer device. The computer-readable storage medium 1504 is used to store computer programs, including computer instructions. The processor 1501 is used to execute the computer program stored in the computer-readable storage medium 1504. The processor 1501 (or CPU (Central Processing Unit)) is the computing and control core of the computer device, suitable for implementing computer programs, specifically for loading and executing computer programs to achieve corresponding method flows or functions.

[0279] This application also provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space for storing the operating system of the computer device. Furthermore, the storage space also stores computer programs suitable for loading and execution by a processor. It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0280] The computer device can be a ranking device in a recommendation system, which deploys a prediction model (i.e., a ranking model). Specifically, the processor 1501 can load and execute the computer program stored in the computer-readable storage medium 1504 to achieve the aforementioned... Figure 5 , Figure 9 or Figure 12 The corresponding steps in the method shown. In a specific implementation, the computer program in the computer-readable storage medium 1504 is loaded by the processor 1501 and executed as follows:

[0281] Obtain business data of the business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1.

[0282] Perform a first correlation analysis on object features and material features to predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object will interact with the first media resource.

[0283] A second correlation analysis is performed on the object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree is used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style.

[0284] The first interaction degree and N second interaction degrees are fused to obtain N interest degrees of the business object for the first media resource. Based on the N interest degrees, the recommended media style for recommending the first media resource is determined from N media styles. The N interest degrees correspond one-to-one with the N media styles.

[0285] In one implementation, when a computer program in computer-readable storage medium 1504 is loaded and executed by processor 1501 to perform a first correlation analysis of object features and material features, and to predict the first degree of interaction between the business object and the first media resource, it is specifically used to perform the following steps:

[0286] The object features and material features are divided into first features and second features. The complexity of the second feature is higher than that of the first feature.

[0287] Perform a first feature crossover process on the first feature to obtain the first crossover feature;

[0288] The second feature is subjected to a second feature crossover process to obtain a second crossover feature; wherein the first feature crossover process is different from the second feature crossover process.

[0289] The first and second cross features are fused to obtain the fused features.

[0290] Based on the integration characteristics, predict the first degree of interaction between the business object and the first media resource.

[0291] In one implementation, the computer program in the computer-readable storage medium 1504 is loaded by the processor 1501 and is also used to perform the following steps:

[0292] Based on the first cross feature, the second cross feature is subjected to feature enhancement processing to obtain the enhanced feature;

[0293] The computer program in the computer-readable storage medium 1504, when loaded and executed by the processor 1501, performs feature fusion processing on the first cross feature and the second cross feature to obtain the fused feature, specifically performs the following steps:

[0294] The first cross feature and the enhanced feature are fused together to obtain the fused feature.

[0295] In one implementation, when the computer program in the computer-readable storage medium 1504 is loaded and executed by the processor 1501 to perform feature enhancement processing on the second cross feature based on the first cross feature to obtain the enhanced feature, it specifically performs the following steps:

[0296] The first cross feature is subjected to a first feature transformation process to obtain the first transformed feature;

[0297] The second cross feature is subjected to a second feature transformation process to obtain the second transformed feature; wherein the first feature transformation process is different from the second feature transformation process.

[0298] The enhanced features are obtained by performing feature multiplication between the first transformed features and the second transformed features.

[0299] In one implementation, when a computer program in computer-readable storage medium 1504 is loaded and executed by processor 1501 to perform a second correlation analysis on object features, material features, and style features of N media styles, and to predict N second interaction degrees between the business object and the first media resource, it is specifically used to perform the following steps:

[0300] Obtain the integrated features obtained by integrating object features and material features;

[0301] The style features of N media styles are subjected to feature cross processing to obtain the cross features of N media styles;

[0302] The integrated features are then fused with the cross features of the N media styles to obtain the fused features corresponding to the N media styles.

[0303] Based on the fusion characteristics corresponding to N media styles, predict N second interaction degrees between the business object and the first media resource.

[0304] In one implementation, the N second interaction degrees are predicted by the style prediction module in the prediction model, which includes a style processing structure shared by various media styles.

[0305] When the computer program in the computer-readable storage medium 1504 is loaded and executed by the processor 1501 to perform feature cross-processing on the style features of N media styles to obtain the cross-features of N media styles, it is specifically used to perform the following steps:

[0306] The style processing structure is copied N-1 times to obtain N-1 copied style processing structures;

[0307] The N style processing structures are called to perform feature cross-processing on the style features of the N media styles, resulting in the cross-features of the N media styles.

[0308] In one implementation, N second interaction degrees are predicted by the style prediction module in the prediction model. The style prediction module includes P style processing structures, each of which corresponds one-to-one with a P media style. The N media styles are the media styles among the P media styles that match the business object, where P is an integer greater than or equal to N.

[0309] When the computer program in the computer-readable storage medium 1504 is loaded and executed by the processor 1501 to perform feature cross-processing on the style features of N media styles to obtain the cross-features of N media styles, it is specifically used to perform the following steps:

[0310] Call the N style processing structures corresponding to the N media styles from the P style processing structures, and perform feature cross processing on the style features of the N media styles respectively to obtain the cross features of the N media styles.

[0311] In one implementation, the computer program in the computer-readable storage medium 1504 is loaded by the processor 1501 and is also used to perform the following steps:

[0312] Obtain recommendation features related to the business clients of the business object;

[0313] When the computer program in the computer-readable storage medium 1504 is loaded and executed by the processor 1501 to perform a first correlation analysis on object features and material features and predict the first interaction degree between the business object and the first media resource, it is specifically used to perform the following steps: perform a first correlation analysis on object features, material features and recommendation features, and predict the first interaction degree between the business object and the first media resource;

[0314] The computer program in the computer-readable storage medium 1504 is loaded and executed by the processor 1501 to perform a second correlation analysis on object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. Specifically, it performs the following steps: performing a second correlation analysis on object features, material features, recommendation features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource.

[0315] In one implementation, the first media resource is any one of the media resources in the set to be recommended; the computer program in the computer-readable storage medium 1504 is loaded by the processor 1501 and is also used to perform the following steps:

[0316] Based on the interest level of the business users towards each media resource in the media resource set under the corresponding recommended media style, the media resources are sorted.

[0317] Determine the media resources of interest to recommend to business users from the ranking results;

[0318] Recommend media resources of interest to business users based on the recommended media styles corresponding to the media resources of interest.

[0319] In one implementation, the computer program in the computer-readable storage medium 1504 is loaded by the processor 1501 and is also used to perform the following steps:

[0320] Select M media styles from N media styles that match the media material in the first media resource, where M is a positive integer less than or equal to N;

[0321] The computer program in the computer-readable storage medium 1504, loaded and executed by the processor 1501, performs the following steps when predicting N second interaction degrees between the business object and the first media resource by performing a second correlation analysis on object features, material features, and style features of N media styles:

[0322] A second correlation analysis is performed on the object characteristics, material characteristics, and style characteristics of M media styles to predict N second interaction degrees between the business object and the first media resource.

[0323] In one implementation, when a computer program in computer-readable storage medium 1504 is loaded and executed by processor 1501 to filter M media styles from N media styles that match the media material in the first media resource, it specifically performs the following steps:

[0324] Obtain style files for N media styles;

[0325] Perform material type identification processing on the style files of N media styles respectively to obtain the first material type information corresponding to the N media styles. The first material type information is used to indicate the material types that are allowed to be presented in the corresponding media style.

[0326] The media materials in the first media resource are processed to identify the material type, and the second material type information is used to indicate the material type contained in the media materials in the first media resource.

[0327] Among N media styles, the media style that matches the first material type information with the second material type information is determined as the M media styles that match the media material of the first media resource.

[0328] In one implementation, the prediction model includes a material prediction module and a style prediction module; the computer program in the computer-readable storage medium 1504 is loaded by the processor 1501 and is also used to perform the following steps:

[0329] Obtain sample data for training the prediction model. The sample data includes sample object features of sample business objects, sample material features of media materials in sample media resources, and style features of target media styles. Among them, sample media resources are media resources recommended to sample business objects, and target media styles are media styles used to recommend sample media resources.

[0330] The material prediction module is invoked to perform a first correlation analysis on the characteristics of the sample object and the characteristics of the sample material, and to predict the first sample interaction degree between the sample business object and the sample media resource.

[0331] The style prediction module is invoked to perform a second correlation analysis on the style features of the sample object, sample material, and target media style, and to predict the second sample interaction degree between the sample business object and the sample media resource under the target media style.

[0332] The interaction degree of the first sample and the interaction degree of the second sample are fused to obtain the sample interest degree of the sample business object on the sample media resource under the target media style, and the prediction model is trained based on the sample interest degree.

[0333] In one implementation, when the computer program in the computer-readable storage medium 1504 is loaded and executed by the processor 1501 to train the prediction model based on sample interest, it specifically performs the following steps:

[0334] The style prediction module includes a style processing structure shared by P types of media styles, and the style processing structure is trained based on the sample interest.

[0335] Alternatively, the style prediction module includes style processing structures corresponding to each of the P media styles, and the style processing structures corresponding to the target media style are trained based on the sample interest.

[0336] Among them, the target media style is the media style used to recommend sample media resources from among the P types of media styles.

[0337] In this embodiment, business data of the business object can be obtained. This business data may include the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1. Therefore, this embodiment, when predicting the business object's interest in the media resource, considers not only the media material but also the media style, taking into account multiple factors that may affect the interest level. A first correlation analysis can be performed on the object characteristics and material characteristics to predict the first interaction degree between the business object and the first media resource. The first interaction degree can be represented by the probability of the business object interacting with the first media resource. Thus, by analyzing the correlation between the object characteristics and material characteristics, the probability of the business object interacting with the first media resource can be predicted under the influencing factor of media material. A second correlation analysis can be performed on object features, material features, and style features of N media styles to predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles, and any one of the second interaction degrees can be used to indicate the probability that the business object will interact with the first media resource presented in the corresponding media style. It can be seen that by analyzing the correlation between object features, material features, and style features, the probability of the business object interacting with the first media resource can be predicted under the two influencing factors of media material and media style. The first interaction degree and N second interaction degrees can be fused to obtain N interest degrees of the business object for the first media resource. Based on the N interest degrees, the recommended media style for recommending the first media resource is determined from N media styles. The N interest degrees correspond one-to-one with the N media styles. It can be seen that the embodiments of this application not only consider the influence of media materials on interest degrees, but also the influence of media materials combined with media styles on interest degrees. This can more accurately predict the business object's interest degree for media resources, which is conducive to determining the media style for recommending media resources based on interest degrees from multiple media styles. This can improve the recommendation effect of media resources.

[0338] This application also provides a computer program product, which includes 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 data processing method described above.

[0339] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0340] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0341] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0342] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data processing method, characterized in that, include: Obtain business data of the business object, the business data including the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1; A first correlation analysis is performed on the object features and the material features to predict the first interaction degree between the business object and the first media resource. The first interaction degree is used to indicate the probability that the business object interacts with the first media resource. A second correlation analysis is performed on the object features, the material features, and the style features of the N media styles to predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles. Any second interaction degree is used to indicate the probability that the business object interacts with the first media resource presented using the corresponding media style. The first interaction degree and the N second interaction degrees are fused to obtain N interest degrees of the business object to the first media resource. Based on the N interest degrees, a recommended media style for recommending the first media resource is determined from the N media styles. The N interest degrees correspond one-to-one with the N media styles.

2. The method as described in claim 1, characterized in that, The step of performing a first correlation analysis on the object features and the material features to predict the first degree of interaction between the business object and the first media resource includes: The object features and the material features are subjected to feature segmentation processing to obtain a first feature and a second feature, wherein the complexity of the second feature is higher than that of the first feature; The first feature is subjected to a first feature crossover process to obtain a first crossover feature; The second feature is subjected to a second feature cross-processing to obtain a second cross feature; wherein the first feature cross-processing is different from the second feature cross-processing. The first cross feature and the second cross feature are subjected to feature fusion processing to obtain the fused feature; Based on the fusion characteristics, predict the first degree of interaction between the business object and the first media resource.

3. The method as described in claim 2, characterized in that, The method further includes: Based on the first cross feature, the second cross feature is subjected to feature enhancement processing to obtain the enhanced feature; The feature fusion process of the first cross feature and the second cross feature to obtain the fused feature includes: The first cross feature and the enhanced feature are subjected to feature fusion processing to obtain the fused feature.

4. The method as described in claim 3, characterized in that, The step of performing feature enhancement processing on the second cross feature based on the first cross feature to obtain enhanced features includes: The first cross feature is subjected to a first feature transformation process to obtain a first transformed feature; The second cross feature is subjected to a second feature transformation process to obtain a second transformed feature; wherein the first feature transformation process is different from the second feature transformation process. The enhanced feature is obtained by performing a feature multiplication operation between the first transformed feature and the second transformed feature.

5. The method as described in claim 1, characterized in that, The second correlation analysis performed on the object features, the material features, and the style features of the N media styles to predict the N second interaction degrees between the business object and the first media resource includes: Obtain the integrated features obtained by performing feature integration processing on the object features and the material features; The style features of the N media styles are subjected to feature cross-processing to obtain the cross features of the N media styles; The integrated features are respectively fused with the cross features of the N media styles to obtain the fused features corresponding to the N media styles; Based on the fusion features corresponding to the N media styles, predict N second interaction degrees between the business object and the first media resource.

6. The method as described in claim 5, characterized in that, The method is executed by a prediction model, and the N second interaction degrees are predicted by the style prediction module in the prediction model. The style prediction module includes a style processing structure shared by various media styles. The step of performing feature cross-processing on the style features of the N media styles to obtain the cross features of the N media styles includes: The style processing structure is copied N-1 times to obtain N-1 copied style processing structures; The N style processing structures are invoked to perform feature cross-processing on the style features of the N media styles respectively, so as to obtain the cross features of the N media styles.

7. The method as described in claim 5, characterized in that, The method is executed by a prediction model. The N second interaction degrees are predicted by the style prediction module in the prediction model. The style prediction module includes P style processing structures, each of which corresponds to one of P media styles. The N media styles are the media styles among the P media styles that match the business object, where P is an integer greater than or equal to N. The step of performing feature cross-processing on the style features of the N media styles to obtain the cross features of the N media styles includes: The N style processing structures corresponding to the N media styles are called from the P style processing structures, and feature cross-processing is performed on the style features of the N media styles respectively to obtain the cross features of the N media styles.

8. The method as described in claim 1, characterized in that, The method further includes: Obtain recommendation features related to the business clients of the business object; The step of performing a first correlation analysis on the object features and the material features to predict the first interaction degree between the business object and the first media resource includes: performing a first correlation analysis on the object features, the material features and the recommendation features to predict the first interaction degree between the business object and the first media resource; The step of performing a second correlation analysis on the object features, the material features, and the style features of the N media styles to predict N second interaction degrees between the business object and the first media resource includes: performing a second correlation analysis on the object features, the material features, the recommendation features, and the style features of the N media styles to predict N second interaction degrees between the business object and the first media resource.

9. The method as described in claim 1, characterized in that, The first media resource is any one of the media resources in the set to be recommended; the method further includes: Based on the interest level of the business object towards each media resource in the media resource set under the corresponding recommended media style, the media resources are sorted. The ranking results are used to determine the media resources of interest to be recommended to the business target. Based on the recommended media style corresponding to the interest-based media resources, the interest-based media resources are recommended to the business objects.

10. The method as described in claim 1, characterized in that, The method further includes: Select M media styles from the N media styles that match the media materials in the first media resource, where M is a positive integer less than or equal to N; The second correlation analysis performed on the object features, the material features, and the style features of the N media styles to predict the N second interaction degrees between the business object and the first media resource includes: A second correlation analysis is performed on the object features, the material features, and the style features of the M media styles to predict N second interaction degrees between the business object and the first media resource.

11. The method as described in claim 10, characterized in that, The step of selecting M media styles from the N media styles that match the media materials in the first media resource includes: Obtain the style files for the N media styles; The style files of the N media styles are processed for material type identification to obtain the first material type information corresponding to the N media styles. The first material type information is used to indicate the material types that are allowed to be presented in the corresponding media style. The media materials in the first media resource are processed to identify the material type, and the second material type information is used to indicate the material type contained in the media materials in the first media resource. Among the N media styles, the media style that matches the first material type information with the second material type information is determined as the M media styles that match the media material of the first media resource.

12. The method according to any one of claims 1-11, characterized in that, The method is executed by a prediction model, which includes a material prediction module and a style prediction module; the method also includes: Obtain sample data for training the prediction model. The sample data includes sample object features of sample business objects, sample material features of media materials in sample media resources, and style features of target media styles. The sample media resources are media resources recommended to the sample business objects, and the target media style is a media style used to recommend the sample media resources. The material prediction module is invoked to perform a first correlation analysis on the features of the sample object and the features of the sample material to predict the first sample interaction degree between the sample business object and the sample media resource. The style prediction module is invoked to perform a second correlation analysis on the features of the sample object, the features of the sample material, and the style features of the target media style, and to predict the second sample interaction degree between the sample business object and the sample media resource under the target media style; The interaction degree of the first sample and the interaction degree of the second sample are fused to obtain the sample interest degree of the sample business object to the sample media resource under the target media style, and the prediction model is trained based on the sample interest degree.

13. The method as described in claim 12, characterized in that, The step of training the prediction model based on the sample interest includes: The style prediction module includes a style processing structure shared by P media styles, and the style processing structure is trained based on the sample interest level; or... The style prediction module includes style processing structures corresponding to each of the P media styles, and trains the style processing structure corresponding to the target media style based on the sample interest degree. The target media style is the media style among the P media styles used to recommend the sample media resources.

14. A data processing apparatus, characterized in that, include: The acquisition unit is used to acquire business data of a business object. The business data includes the object characteristics of the business object, the material characteristics of the media material in the first media resource to be recommended, and the style characteristics of N media styles corresponding to the first media resource, where N is an integer greater than 1. The processing unit is configured to perform a first correlation analysis on the object features and the material features, and predict a first interaction degree between the business object and the first media resource, wherein the first interaction degree is used to indicate the probability that the business object interacts with the first media resource; The processing unit is further configured to perform a second correlation analysis on the object features, the material features, and the style features of the N media styles, and predict N second interaction degrees between the business object and the first media resource. The N second interaction degrees correspond one-to-one with the N media styles, and any one of the second interaction degrees is used to indicate the probability that the business object interacts with the first media resource presented using the corresponding media style. The processing unit is further configured to perform interaction degree fusion processing on the first interaction degree and the N second interaction degrees to obtain N interest degrees of the business object to the first media resource, and determine the recommended media style for recommending the first media resource from the N media styles according to the N interest degrees, wherein the N interest degrees correspond one-to-one with the N media styles.

15. A computer device, characterized in that, The computer device includes: A processor is a tool for implementing computer programs. A computer-readable storage medium storing a computer program adapted to be loaded by the processor and executed as described in any one of claims 1-13.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-13.

17. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the data processing method as described in any one of claims 1-13.

Images