Media resource recommendation method, device, equipment and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for recommending media resources, belonging to the field of artificial intelligence technology. By inputting historical behavior data of a target object and resource data of candidate media resources into a media resource recommendation model, this application takes into account the impact of the object's behavior on clicks and conversions of media resources, thereby increasing the accuracy of at least one of the acquired click-through rate and conversion rate, as well as revenue reference information. Furthermore, based on the acquired data, the application determines a media resource recommended for the target object from the candidate media resources, while simultaneously taking into account the click-through rate, conversion rate, and revenue reference information, thereby effectively improving the accuracy of media resource recommendations.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for recommending media resources. Background Art

[0002] With the advancement of computer technology, online media resources can carry more information and achieve more functions. For example, media resources can be used to promote products. When promoting products to a specific user, algorithms can be used to select and recommend media resources of interest from a candidate set of media resources to achieve the purpose of promoting the product. Related technologies often use click-through rate (CTR) prediction models to predict the click-through rate of each media resource in the candidate set, and recommend media resources with high click-through rates to users.

[0003] In the above technologies, the CTR prediction model only focuses on the click-through rate of media resources. However, even if users click on media resources, they may not necessarily consume the products, so the recommendation accuracy is low. Summary of the Invention

[0004] The present invention provides a method, apparatus, device, and storage medium for recommending media resources, which can improve the accuracy of recommending media resources. The technical solution is as follows:

[0005] In one aspect, a method for recommending media resources is provided, the method comprising:

[0006] Obtaining historical behavior data of the target object, where the historical behavior data represents the target object's browsing behavior and resource transfer behavior based on the media resource;

[0007] Inputting the target object's historical behavior data and the first resource data of the first candidate media resource in the first candidate resource set into the media resource recommendation model to obtain at least one of a first click-through rate and a first conversion rate of each first candidate media resource and revenue reference information, the revenue reference information indicating revenue generated after the target object performs a conversion behavior on the candidate media resource;

[0008] Determining a media resource to be recommended based on at least one of the first click rate and the first conversion rate of each first candidate media resource and the revenue reference information;

[0009] Recommending resources to the target object based on the media resources to be recommended;

[0010] The media resource recommendation model is trained based on historical behavior data of multiple sample objects.

[0011] In one aspect, a method for recommending media resources is provided, the method comprising:

[0012] Displaying a media resource recommendation page, displaying an intelligent recommendation option on the media resource recommendation page, and the intelligent recommendation option is used to select whether to use the media resource recommendation model to recommend media resources;

[0013] When the smart recommendation option is set to use the media resource recommendation model, in response to an upload operation on the media resource on the media resource recommendation page, sending the media resource and a recommendation instruction to the server, the recommendation instruction being used to instruct to recommend the media resource based on the media resource recommendation model;

[0014] Among them, the media resource recommendation model is used to obtain at least one of the first click-through rate and the first conversion rate of the media resource and the revenue reference information based on the historical behavior data of the target object and the first resource data of the media resource. The historical behavior data represents the browsing behavior and resource transfer behavior of the target object based on the media resource, and the revenue reference information represents the revenue generated after the target object performs the conversion behavior on the media resource. The media resource recommendation model is trained based on the historical behavior data of multiple sample objects.

[0015] In one aspect, a device for recommending media resources is provided, the device comprising:

[0016] An acquisition module is used to acquire historical behavior data of a target object, where the historical behavior data represents browsing behavior and resource transfer behavior of the target object based on media resources;

[0017] An input module is configured to input the historical behavior data of the target object and the first resource data of the first candidate media resource in the first candidate resource set into the media resource recommendation model to obtain at least one of a first click-through rate and a first conversion rate of each first candidate media resource and revenue reference information, wherein the revenue reference information represents the revenue generated after the target object performs a conversion behavior on the candidate media resource;

[0018] a determination module, configured to determine a media resource to be recommended based on at least one of a first click-through rate and a first conversion rate of each first candidate media resource and revenue reference information;

[0019] A recommendation module, configured to recommend resources to the target object based on the media resources to be recommended;

[0020] The media resource recommendation model is trained based on historical behavior data of multiple sample objects.

[0021] In some embodiments, the input module includes:

[0022] An input submodule, configured to input the historical behavior data and the first resource data of any first candidate media resource into the media resource recommendation model;

[0023] a processing submodule, configured to process the historical behavior data and the first resource data using the media resource recommendation model and based on the revenue reference space to obtain revenue reference information of the first candidate media resource;

[0024] The processing submodule is used to process the historical behavior data and the first resource data based on at least one of the click rate space and the conversion rate space through the media resource recommendation model to obtain at least one of the first click rate and the first conversion rate of the first candidate media resource.

[0025] In some embodiments, the acquisition module is further used to acquire portrait data and label data of the target object, where the label data is used to represent the preference of the target object;

[0026] Accordingly, the processing submodule includes:

[0027] a first acquisition unit, configured to acquire, based on the media resource recommendation model, an initial object vector of first object data and an initial resource vector of the first resource data, the first object data including historical behavior data, portrait data, and label data of the target object, the initial object vector including subvectors corresponding to multiple categories of data in the first object data, and the initial resource vector including subvectors corresponding to multiple categories of data in the first resource data;

[0028] a fusion unit configured to fuse multiple subvectors of the initial object vector multiple times based on multiple fusion methods to obtain multiple object fusion vectors of the object data, and to fuse multiple subvectors of the initial resource vector multiple times to obtain multiple resource fusion vectors of the resource data;

[0029] a weighting unit, configured to perform weighted summation on the multiple object fusion vectors and the multiple resource fusion vectors based on the revenue reference space to obtain an object vector and a resource vector, wherein the object vector represents an impact of the first object data on the revenue reference information of the first candidate media resource, and the resource vector represents an impact of the first resource data on the revenue reference information of the first candidate media resource;

[0030] The first acquiring unit is configured to acquire revenue reference information of the first candidate media resource based on the object vector and the resource vector.

[0031] In some embodiments, the historical behavior data includes behavior data for multiple historical time periods;

[0032] The first acquisition unit is configured to map the first object data and the first resource data based on the media resource recommendation model to obtain an initial subvector of the historical behavior data, a subvector of the portrait data, a subvector of the label data, and multiple subvectors of the first resource data;

[0033] Extracting relevant information between the behavior data of the multiple historical time periods based on the self-attention mechanism and the initial sub-vector of the historical behavior data to obtain a sub-vector of the historical behavior data;

[0034] Concatenate the sub-vector of the historical behavior data, the sub-vector of the portrait data, and the sub-vector of the label data to obtain the initial object vector;

[0035] The multiple sub-vectors of the first resource data are concatenated to obtain the initial resource vector.

[0036] In some embodiments, the apparatus further comprises:

[0037] a training module for training the media resource recommendation model based on at least one of a click tag and a conversion tag, sample data, and sample revenue reference information, wherein the sample data includes sample historical behavior data of sample objects and sample resource data of sample media resources;

[0038] The training module is used to obtain at least one of the predicted click-through rate and the predicted conversion rate of the sample data and the predicted revenue reference information based on the media resource recommendation model; and train the media resource recommendation model based on at least one of the predicted click-through rate and the predicted conversion rate, at least one of the click label and the conversion label, the predicted revenue reference information and the sample revenue reference information.

[0039] In some embodiments, the sample income reference information is obtained by standardizing the sample income data.

[0040] In some embodiments, the determining module includes:

[0041] A first determining submodule, configured to determine a plurality of first media resources from the first candidate resource set based on a first conversion rate of each of the first candidate media resources;

[0042] A second determining submodule, configured to determine a plurality of second media resources from the first candidate resource set based on the revenue reference information of each of the first candidate media resources;

[0043] The third determining submodule is configured to determine the media resources to be recommended from a second candidate resource set, where the second candidate resource set includes the plurality of first media resources and the plurality of second media resources.

[0044] In some embodiments, the third determining submodule includes:

[0045] a second acquiring unit, acquiring a second click rate and a second conversion rate of each second candidate media resource based on the second object data of the target object and the second resource data of each second candidate media resource in the second candidate resource set;

[0046] a determining unit, configured to determine a media resource to be recommended from the second candidate resource set based on the second click-through rate and the second conversion rate of each of the second candidate media resources;

[0047] The number of data categories in the second object data is greater than the number of data categories in the first object data, and the number of data categories in the second resource data is greater than the number of data categories in the first resource data.

[0048] In some embodiments, the determination unit is used to obtain a correction parameter for each of the second candidate media resources based on the conversion number of each of the second candidate media resources, where the conversion number is the number of conversions of the second candidate media resource within the target historical time period, and the correction parameter is used to correct the second conversion rate of the second candidate media resource; based on the second click-through rate, second conversion rate and correction parameter of each of the second candidate media resources, the media resources to be recommended are determined from the set of second candidate resources.

[0049] In one aspect, a device for recommending media resources is provided, the device comprising:

[0050] A display module is used to display a media resource recommendation page, and display an intelligent recommendation option on the media resource recommendation page, wherein the intelligent recommendation option is used to select whether to use the media resource recommendation model to recommend media resources;

[0051] a sending module configured to, when the intelligent recommendation option is set to use the media resource recommendation model, send the media resource and a recommendation instruction to the server in response to an upload operation of the media resource on the media resource recommendation page, the recommendation instruction being used to instruct to recommend the media resource based on the media resource recommendation model;

[0052] Among them, the resource recommendation model is used to obtain at least one of the first click-through rate and the first conversion rate of the media resource and the revenue reference information based on the historical behavior data of the target object and the first resource data of the media resource. The historical behavior data represents the browsing behavior and resource transfer behavior of the target object based on the media resource, and the revenue reference information represents the revenue generated after the target object performs the conversion behavior on the media resource. The resource recommendation model is trained based on the historical behavior data of multiple sample objects.

[0053] On the one hand, a computer device is provided, which includes one or more processors and one or more memories, wherein the one or more memories store at least one computer program, and the at least one computer program is loaded and executed by the one or more processors to implement the operations performed by the media resource recommendation method.

[0054] In one aspect, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium. The at least one computer program is loaded and executed by a processor to implement the operations performed by the media resource recommendation method.

[0055] In one aspect, a computer program product is provided, comprising at least one computer program stored in a computer-readable storage medium. A processor of a computer device reads the at least one computer program from the computer-readable storage medium and executes the at least one computer program, causing the computer device to implement the operations of the method for recommending media resources.

[0056] The technical solution provided by the present application takes into account the impact of the target object's historical behavior data and the resource data of the candidate media resources into the media resource recommendation model, thereby making at least one of the obtained click-through rate and conversion rate and the revenue reference information more accurate. Further, based on the obtained data, the media resources recommended for the target object are determined from the candidate media resources, while taking into account the click-through rate, conversion rate and revenue reference information, which can effectively improve the accuracy of media resource recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0058] Figure 1 This is a schematic diagram of an implementation environment of a media resource recommendation method provided in an embodiment of the present application;

[0059] Figure 2 This is a flowchart of a method for recommending media resources provided in an embodiment of the present application;

[0060] Figure 3 This is a flowchart of a method for recommending media resources provided in an embodiment of the present application;

[0061] Figure 4This is a flowchart of a method for recommending media resources provided in an embodiment of the present application;

[0062] Figure 5 This is a schematic diagram of a media resource recommendation page provided in an embodiment of the present application;

[0063] Figure 6 This is a schematic diagram of a directional setting page provided in an embodiment of the present application;

[0064] Figure 7 is a schematic diagram of a system optimization method provided in an embodiment of the present application;

[0065] Figure 8 Schematic diagram of an automatic expansion method provided in an embodiment of the present application;

[0066] Figure 9 This is a schematic diagram of a layout setting page provided in an embodiment of the present application;

[0067] Figure 10 This is a schematic diagram of a scheduling setting page provided in an embodiment of the present application;

[0068] Figure 11 is a schematic diagram of a feedback page provided in an embodiment of the present application;

[0069] Figure 12 This is a schematic diagram of a process framework for media resource recommendation provided by an embodiment of the present application;

[0070] Figure 13 This is a flowchart of a method for recommending media resources provided in an embodiment of the present application;

[0071] Figure 14 This is a schematic diagram of a method for recommending media resources provided in an embodiment of the present application;

[0072] Figure 15 This is a schematic diagram of a media resource recommendation model provided in an embodiment of the present application;

[0073] Figure 16 is a schematic diagram of a self-attention sub-model provided in an embodiment of the present application;

[0074] Figure 17 Schematic diagram of an MMoE model provided in an embodiment of the present application;

[0075] Figure 18 is a schematic diagram of a DCN model provided in an embodiment of the present application;

[0076] Figure 19 This is a schematic diagram of an income sub-model provided in an embodiment of the present application;

[0077] Figure 20 is a schematic diagram of a click-through rate sub-model provided in an embodiment of the present application;

[0078] Figure 21 is a schematic diagram of determining a second candidate resource set provided by an embodiment of the present application;

[0079] Figure 22 This is a schematic diagram of obtaining sample data provided by an embodiment of the present application;

[0080] Figure 23 This is a distribution diagram of sample income data provided in an embodiment of the present application;

[0081] Figure 24 This is a schematic diagram of the structure of a media resource recommendation device provided in an embodiment of the present application;

[0082] Figure 25 This is a schematic diagram of the structure of a media resource recommendation device provided in an embodiment of the present application;

[0083] Figure 26 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present application;

[0084] Figure 27 This is a structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0085] To make the objectives, technical solutions, and advantages of this application more clear, the following will further describe the embodiments of this application in detail with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0086] In this application, the terms "first", "second", etc. are used to distinguish identical or similar items with substantially the same effects and functions. It should be understood that there is no logical or temporal dependency between "first", "second", and "nth", nor is there any limitation on the quantity and execution order.

[0087] To facilitate understanding of the technical process of the embodiments of the present application, some terms involved in the embodiments of the present application are explained below:

[0088] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0089] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0090] The technical solutions provided in the embodiments of this application can also be combined with cloud technology. For example, the trained media resource recommendation model can be deployed on a cloud server. Cloud technology refers to a hosting technology that unifies hardware, software, network, and other resources within a wide area network or local area network to achieve data computing, storage, processing, and sharing.

[0091] Blockchain is a new application model for computer technologies, including distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a series of data blocks linked using cryptographic methods. Each block contains information about a batch of online transactions, used to verify the validity of this information (to prevent counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, the platform product and service layer, and the application service layer.

[0092] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0093] Rectified Linear Unit (ReLU), also known as the rectified linear unit, is an activation function commonly used in artificial neural networks. It usually refers to nonlinear functions represented by ramp functions and their variants.

[0094] Approximate Nearest Neighbors (ANN): An algorithm used to quickly find the nearest neighbor vector of a target vector in a set of vectors, which can effectively improve retrieval efficiency.

[0095] The following is an introduction to the implementation environment of the media resource recommendation method provided by this application.

[0096] Figure 1 This is a schematic diagram of an implementation environment of a media resource recommendation method provided in an embodiment of the present application, see Figure 1 The implementation environment includes: a first terminal 101, a second terminal 102, and a media resource recommendation system 103. The first terminal 101 and the second terminal 102 are directly or indirectly connected to the media resource recommendation system 103 via a wired or wireless manner, which is not limited in this embodiment of the application.

[0097] The first terminal 101 and the second terminal 102 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but are not limited thereto.

[0098] The first terminal 101 runs an application that supports media resource uploading. The first terminal 101 is the terminal used by the target. For example, if the media resource is an advertisement, the first terminal 101 is the terminal used by the advertiser. The second terminal 102 runs an application that supports media resource recommendation. For example, the application is an e-commerce application, a social networking application, or a music application. The second terminal 102 is the terminal used by the target.

[0099] It should be understood that the number of first terminals and second terminals may be greater, and the embodiment of the present application only uses the first terminal 101 and the second terminal 102 as examples.

[0100] The media resource recommendation system 103 is used to receive and store media resources uploaded by the first terminal 101, and provide background services for media resource recommendation for applications running in the second terminal 102. In some embodiments, the media resource recommendation system 103 includes a server and a database, the server is used to provide background services related to media resource recommendation, and the database is used to store multiple media resources and related information of media resources. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and basic cloud computing services such as big data and artificial intelligence platforms. Optionally, the number of servers can be more or less, and the embodiments of the present application are not limited to this.

[0101] In some embodiments, the first terminal 101, the second terminal 102 and the server can serve as nodes in a blockchain system.

[0102] based on Figure 1 The implementation environment shown below is combined with Figure 2 , introduces the recommendation method of the media resources provided by this application. Figure 2 Flowchart of the media resource recommendation method provided in the embodiment of the present application, which is executed by the server of the media resource recommendation system 103. Figure 2 As shown, this embodiment includes the following steps 201 to 204.

[0103] 201. The server obtains historical behavior data of a target object, where the historical behavior data indicates browsing behavior and resource transfer behavior of the target object based on media resources.

[0104] In an embodiment of the present application, the target object is an object logged in using a second terminal, and the second application running on the second terminal provides the function of recommending and displaying media resources. During the process of running the second application on the second terminal, the server can obtain the historical behavior data of the target object and the first resource data of the first candidate media resource in the first candidate resource set to recommend media resources to the target object.

[0105] Among them, the historical behavior data includes the media resources that the target object has browsed, clicked, and converted in the historical time period, as well as the applications downloaded by the target user, etc. Among them, the first candidate media resource in the first candidate resource set is used to promote the target object, and the target object includes a virtual target object and a physical target object. For example, the first candidate media resource is an advertisement for a certain game. The first candidate media resource can be a picture, video, or audio, etc. The embodiment of the present application does not limit the type of the first candidate media resource. The first resource data includes the identifier of the first candidate media resource, the identifier of the target object, and content keywords, etc. Optionally, the first resource data also includes the creative identifier and creative keywords, etc. of the first candidate media resource.

[0106] 202. The server inputs the historical behavior data of the target object and the first resource data of the first candidate media resource in the first candidate resource set into the media resource recommendation model to obtain at least one of the first click rate and the first conversion rate of each first candidate media resource and revenue reference information, wherein the revenue reference information indicates the revenue generated after the target object performs a conversion behavior on the candidate media resource.

[0107] The media resource recommendation model is trained based on the historical behavior data of multiple sample objects. In the embodiment of the present application, the media resource recommendation model is also called a full-link object behavior value self-attention multi-task model.

[0108] In this embodiment of the present application, the media resource recommendation model is obtained through multi-task training, where the multi-task refers to at least one of the first click-through rate task and the second click-through rate task, as well as the reference revenue information task. Multi-task training allows the model training process to consider the connections between the various tasks, thereby increasing the accuracy of the trained media resource recommendation model.

[0109] The revenue reference information refers to the amount of resources transferred from the resource transfer behavior of the first candidate media resource. For example, if the first candidate media resource is an advertisement, the revenue reference information would be the order amount of the virtual props in the game; if the first candidate media resource is an e-commerce product advertisement, the revenue reference information would be the order amount of the product.

[0110] 203. The server determines a media resource to be recommended based on at least one of the first click rate and the first conversion rate of each first candidate media resource and the revenue reference information.

[0111] Optionally, the server first determines multiple candidate media resources from multiple media resources stored in the database based on at least one of the first click-through rate and the first conversion rate and the revenue reference information, and then determines the media resources to be recommended from the multiple candidate media resources; or, the server first determines multiple candidate media resources from multiple media resources stored in the database, and then determines the media resources to be recommended from the multiple candidate media resources based on at least one of the first click-through rate and the first conversion rate and the revenue reference information.

[0112] 204. The server recommends resources to the target object based on the media resources to be recommended.

[0113] In the embodiments of the present application, recommending resources to a target object means that the server sends the media resource to be recommended to the first terminal, and the first terminal displays the received media resource to the target object. For example, if the server determines that the resource to be recommended is a game advertisement, the server sends the game advertisement to the first terminal, the second terminal receives the game advertisement, and displays the game advertisement on the page currently viewed by the target object, thereby completing the resource recommendation to the target object.

[0114] The method provided in the embodiment of the present application takes into account the impact of the object's behavior on media resource clicks and conversions by inputting the historical behavior data of the target object and the resource data of the candidate media resources into the media resource recommendation model, thereby making at least one of the obtained click-through rate and conversion rate and the revenue reference information more accurate. Further, based on the obtained data, the media resources recommended for the target object are determined from the candidate media resources, while taking into account the click-through rate, conversion rate and revenue reference information, which can effectively improve the accuracy of media resource recommendations.

[0115] based on Figure 1 The implementation environment shown below is combined with Figure 3 , introduces the recommendation method of the media resources provided by this application. Figure 3 This is a flow chart of a method for recommending media resources provided by an embodiment of the present application, which is executed by the first terminal 101. Figure 3 As shown, the method includes steps 301 to 302.

[0116] 301. A first terminal displays a media resource recommendation page. A smart recommendation option is displayed on the media resource recommendation page. The smart recommendation option is used to select whether to use a media resource recommendation model to recommend media resources.

[0117] In an embodiment of the present application, when the subject has a need to promote a target object, the subject can start a first application in the first terminal that supports the media resource upload function. In response to the subject's start-up operation of the first application, the first terminal displays the media resource recommendation page.

[0118] Among them, the media resource recommendation page is used to provide the function of uploading media resources and the function of selecting the media resource recommendation method. The media resource recommendation page provides the function of selecting the recommendation method of media resources by displaying the smart recommendation option. If the smart recommendation option is set to use the media resource recommendation model, the server will recommend the uploaded media resources through the media resource recommendation model. If the smart recommendation option is set to not use the media resource recommendation model, the server will recommend the uploaded media resources through the recommendation directional data set by the object. Among them, the recommendation directional data is used to describe the basic information, interests, hobbies, behaviors and other characteristics of the object of media resource recommendation. Based on the recommendation directional data, the server can recommend the media resource to the object when receiving a resource acquisition request sent by an object with corresponding characteristics through the second terminal.

[0119] 302. When the smart recommendation option is set to use the media resource recommendation model, the first terminal sends the media resource and a recommendation instruction to the server in response to an upload operation of the media resource on the media resource recommendation page. The recommendation instruction is used to instruct to recommend the media resource based on the media resource recommendation model.

[0120] The method provided in the embodiment of the present application provides a function of selecting a media resource recommendation method by displaying an intelligent recommendation option on a resource recommendation page. Thus, when the media resource is set to be recommended using a media resource recommendation model, at least one of the click-through rate and conversion rate of the media resource relative to the target object and revenue reference information can be obtained through the media resource model, the historical behavior data of the target object and the resource data of the media resource, so as to realize the recommendation of the media resource and achieve the purpose of improving the accuracy of the media resource recommendation.

[0121] above Figure 2 and Figure 3 The corresponding embodiments are only a brief introduction to the media resource recommendation method provided by this application. In some embodiments, the method includes a process of uploading media resources and a process of recommending media resources. The above processes are described below through the following embodiments.

[0122] The following combination Figure 4 The method is described by taking the process of an object uploading media resources through the first terminal 101 as an example. Figure 4 This is a flowchart of a method for recommending media resources provided by an embodiment of the present application. Figure 4 , this embodiment includes the following steps 401 to 403.

[0123] 401. A first terminal displays a media resource recommendation page. The media resource recommendation page displays an intelligent recommendation option, which is used to select whether to use a media resource recommendation model to recommend media resources.

[0124] In some embodiments, the first terminal starts the first application in response to the object's start-up operation on the first application, displays a media resource recommendation page, and displays a smart recommendation option on the media resource recommendation page.

[0125] For example, Figure 5 As shown, the smart recommendation option includes two options: on and off. If the smart recommendation option is set to on, it means that the smart recommendation option is set to use the media resource recommendation model. If the smart option is set to off, the smart recommendation option is set to indicate that the media resource recommendation model is not used.

[0126] In some embodiments, after the subject sets the smart recommendation option to use the media resource recommendation model, the subject can also set recommendation targeting data, thereby enabling the server to recommend the uploaded media resource based on both the recommendation targeting data and the media resource recommendation model. Accordingly, after displaying the smart recommendation option, the step further includes: the first terminal displaying a targeting setting page for setting the recommendation targeting data.

[0127] For example, the orientation setting page is as follows: Figure 6 As shown, the orientation setting page includes multiple object feature setting options, such as geographic location, age, gender and education level, etc. The object can set the recommended orientation data through the multiple object feature setting options. Optionally, the orientation setting page provides two methods for setting recommended orientation data, such as Figure 6 As shown in 601 and 602, it includes a system optimization method and an automatic expansion method, and these two methods are described below respectively.

[0128] (1) System optimization method. The system optimization method refers to the number of object features in the recommended directional data set by the object is less than the number of original object features. The original object features refer to the object features that are predetermined by the object based on prior knowledge and match the features of the media resource. When the system optimization method is used, the objects recommended by the media resource include objects that meet the recommended directional data and objects determined by the media resource recommendation model. For example, Figure 7 As shown, the recommended directional data includes object features A and B, the objects that meet the recommended directional data are the A+B set, and the objects determined by the media resource recommendation model are the F set. Therefore, the objects recommended by the media resource are the A+B set and the F set.

[0129] (2) Automatic expansion method. The automatic expansion method refers to setting recommended directional data based on the original object characteristics and selecting non-breakthrough orientation from the original object characteristics. Non-breakthrough orientation refers to the characteristics that the objects recommended by the media resource must meet. When the automatic expansion method is used, the objects recommended by the media resource include objects that meet the recommended directional data and objects that meet the non-breakthrough orientation among the objects determined by the media resource recommendation model. For example, Figure 8 As shown, the recommended directional data includes object features A, B, C and D, where A and B are set as non-breakthrough directions. The objects that meet the recommended directional data are the A+B+C+D set. The objects determined by the media resource recommendation model are the E set, where the objects that meet the non-breakthrough direction are the A+B+E set. The objects recommended by the media resource are the A+B+C+D set and the A+B+E set.

[0130] It should be noted that, in the traditional media resource recommendation method, when setting the recommendation directional data, the object needs to set a large number of object features, and the recommendation directional data also needs to be manually adjusted according to the click and conversion of the media resource. However, in the method of recommending through the media resource recommendation model in the embodiment of the present application, when setting the recommendation directional data, the object only needs to set a small number of object features, and there is no need to manually optimize the recommendation directional data. The server can recommend the media resource to the appropriate object through the recommendation directional data and the media resource recommendation model, thereby achieving the purpose of improving the click rate, conversion rate and revenue of the media resource, and improving the efficiency of setting the target. Moreover, the media resource recommendation process often includes multiple steps. After determining the candidate media resources by inferring the directional data, it is necessary to further determine the resources to be recommended from the candidate media resources through other models, such as the CTR model or the conversion rate (CVR) model. In the traditional method, manual adjustment of the directional data will cause the objects recommended by the media resource to mutate, causing the model for determining the media resource to be recommended to have a large error. The method of the present application does not require optimization of the recommendation directional data, thereby avoiding large errors.

[0131] 402. When the smart recommendation option is set to use the media resource recommendation model, the first terminal sends the media resource and a recommendation instruction to the server in response to an upload operation of the media resource on the media resource recommendation page. The recommendation instruction is used to instruct to recommend the media resource based on the media resource recommendation model.

[0132] In some embodiments, the recommendation instruction carries recommendation targeting data to instruct the server to recommend the media resource using the recommendation targeting data and the media resource recommendation model.

[0133] In some embodiments, the object can also set the recommended placement, recommended schedule, and bid for the media resource. The recommended placement refers to the application that recommends the media resource, the recommended schedule refers to the time period for recommending the media resource, and the bid refers to the fee paid by the object for recommending the media resource. Accordingly, before sending the media resource and the recommendation instruction to the server, the step also includes: the first terminal displays the placement setting page, and in response to the object's submission operation on the placement setting page, displays the schedule setting page, and in response to the object's submission operation on the schedule setting page, displays the bid setting page, and in response to the object's submission operation on the bid setting page, sends the media resource and recommendation instruction to the server. Accordingly, the recommendation instruction also carries relevant information about the recommended placement, recommended schedule, and bid for the media resource.

[0134] For example, the layout setting page is described. Figure 9 As shown, the placement settings page includes multiple placement options, each of which includes a selection control. In response to clicking on any placement option, the selection control for that placement option appears selected. Selecting a placement indicates that the server will recommend media resources to the application's target in that placement. An application's target refers to the user who registered the application.

[0135] For example, the schedule setting page is described. Figure 10 As shown, the schedule setting page includes a date setting option 1001 and a time setting option 1002. The date setting option is used to set the recommendation date for the media resource. The date setting option includes two options: "Long-term delivery" and "Specify start date and end date". If you select "Long-term delivery", you only need to set the start date, and the server will recommend the media resource from the set start date. If you select "Specify start date and end date", you need to set the start date and end date, and the server will recommend the media resource in the time period between the set start date and end date. The time setting option is used to set the time period for recommending the media resource every day. The time setting option includes three options: "All day", "Specify start time and end time" and "Specify multiple time periods". The object can be set based on demand. After completing the schedule setting for the media resource, the server will recommend the media resource within the set time period.

[0136] For example, the bid setting page includes a conversion goal control and a bid control. The conversion goal control is used to set a conversion goal for a media resource. The conversion goal represents the desired conversion behavior for the media resource. For example, if the media resource is an advertisement for an app, the conversion goal for the media resource could be downloading the app. The bid control is used to set the amount the target pays for each conversion of the media resource.

[0137] Optionally, the object can set shallow conversion goals and deep conversion goals for media resources through the conversion goal control. Shallow conversion goals refer to conversion behaviors that occur on media resources but do not generate actual revenue, while deep conversion goals refer to conversion behaviors that occur on media resources that can generate actual revenue. For example, if the media resource is an advertisement for a game, the shallow conversion goal is downloading the game, and the deep conversion goal is paying for virtual props in the game.

[0138] Optionally, the object can set the billing method of the resource recommendation system through the bid setting page, which includes billing based on the number of clicks (Optimized Cost per Click, oCPC) and billing based on the number of impressions (Optimized Cost per Mille, oCPM).

[0139] 403. The server receives the media resource and the recommendation instruction, and sends a confirmation message to the first terminal. The confirmation message indicates that the media resource has been set to be recommended based on the media resource recommendation model.

[0140] In some embodiments, the server receives the media resource and the recommendation instruction, sets the recommendation method for the media resource as model recommendation, stores the media resource and the recommendation method in a database, and sends a confirmation message to the first terminal. In some embodiments, if the recommendation instruction also carries the recommendation targeting data, recommended placement, recommended schedule, and bid data for the media resource, the server stores the data carried in the recommendation instruction, the media resource, and the recommendation method in a database.

[0141] In some embodiments, after the subject completes uploading the media resource, the subject can view feedback data for the media resource, which indicates how the media resource has been displayed, clicked, and converted. Accordingly, embodiments of the present application further include: the first terminal displays a feedback page in response to the subject's operation to view the feedback data for the media resource, and displays the feedback data for the media resource on the feedback page.

[0142] For example, the feedback page is described. Figure 11 As shown, the feedback page can display the current number of impressions, clicks, conversions, average conversion cost, total cost, and average click cost of the media resource. The conversion volume is the number of times the media resource has been converted to the conversion target, and the average conversion cost is the cost paid by the target each time the media resource is converted. For example, if the conversion target of the media resource is form appointments, the conversion volume is the number of form appointments, and the average conversion cost is the average cost of form appointments. Based on the average conversion cost, the target can determine whether the current cost exceeds the set bid and decide whether to continue with the resource recommendation.

[0143] The method provided in the embodiment of the present application provides a function of selecting a media resource recommendation method by displaying an intelligent recommendation option on a resource recommendation page. Thus, when the media resource is set to be recommended using a media resource recommendation model, at least one of the click-through rate and conversion rate of the media resource relative to the target object and revenue reference information can be obtained through the media resource model, the historical behavior data of the target object and the resource data of the media resource, so as to realize the recommendation of the media resource and achieve the purpose of improving the accuracy of the media resource recommendation.

[0144] The following combination Figure 12 , briefly introduces the overall process framework of media resource recommendation. Figure 12 As shown, media resource recommendation includes four processes: data access 1201, sample construction 1202, model construction 1203 and resource recommendation 1204. Among them, data access 1201 includes the process of obtaining the initial data of the media resource recommendation system and the sample revenue data sent by the object, sample construction 1202 includes the process of obtaining sample data, and model construction 1203 includes the process of training the media resource recommendation model. For detailed descriptions of the above three processes, please refer to Figure 22 The corresponding embodiments are not described here in detail. Resource recommendation 1204 includes a recall process and a recommendation process, wherein the recall process refers to determining a second candidate resource set from a first candidate resource set based on the trained media resource recommendation model, and the recommendation process refers to determining the media resources to be recommended from the second candidate resource set. For a detailed description of the resource recommendation 1204 process, see Figures 13 to 21 The corresponding embodiments will not be described in detail here.

[0145] The following combination Figure 13 , for the above Figure 12 The process of resource recommendation 1204 is described in detail. Figure 13 This is a flow chart of a method for recommending media resources provided by an embodiment of the present application. Figure 13 As shown, the method includes steps 1301 to 1308.

[0146] 1301. A second terminal sends a resource acquisition request to a server, where the resource acquisition request is used to instruct the server to return a media resource recommended to a target object.

[0147] In some embodiments, the second terminal displays a media resource display page in response to the target object's launch operation on the target application. The target object can browse the media resources by clicking, sliding, and other operations on the media resource display page. The second terminal sends a resource acquisition request to the server in response to the target object's operation of browsing the media resources.

[0148] Exemplarily, the media resource display page is a media resource selection page. The second terminal displays the selection page, and the target user enters keywords for the media resource they want to browse on the selection page, or clicks on the resource type of the media resource they want to browse on the selection page. In response to the target user's input operation or click operation, the terminal sends a corresponding resource acquisition request to the server.

[0149] Exemplarily, the media resource display page is a personalized page that includes display locations for multiple media resources. Optionally, in response to the target subject's display operation on the personalized page, the second terminal sends a resource acquisition request to the server to display the personalized page and display the media resources recommended for the target subject on the personalized page. Optionally, the second terminal displays the personalized page, and the target subject can browse the page content by sliding the personalized page. In response to the target subject sliding to any display location, the second terminal sends a resource acquisition request to the server to display the media resources recommended for the target subject at that display location.

[0150] It should be noted that the above are only a few examples of sending resource acquisition requests to the second terminal. The target object can trigger the second terminal to send a resource acquisition request to the server by any operation in any page, and the embodiments of the present application are not limited to this.

[0151] 1302. The server receives a resource acquisition request and acquires historical behavior data of a target object, where the historical behavior data indicates browsing behavior and resource transfer behavior of the target object based on media resources.

[0152] In some embodiments, the server receives the resource acquisition request and acquires historical behavior data of the target object and first resource data of a first candidate media resource in the first resource candidate set from a database, wherein the first candidate media resource is a media resource recommended by a model as a recommendation method.

[0153] In some embodiments, the server also obtains other data, and step 1302 also includes: the server obtains the portrait data and label data of the target object, and the label data is used to represent the preferences of the target object. Among them, the portrait data includes the age, gender and region of the target object, and the label data includes the long-term category and short-term category of the target object. Optionally, the portrait data also includes the consumption status, marital status and educational background of the target object, and the label data also includes the interests and keywords of the target object. By obtaining data of other categories of target objects, the resource recommendation model can obtain the first click-through rate, first conversion rate and revenue reference information of each first candidate media resource based on richer object data, thereby improving the accuracy of recommended media resources.

[0154] Optionally, the server also obtains the device data of the first terminal used by the target object, the display position data in the resource display page browsed by the target object, and the statistical data of the target object in multiple historical time periods. The device data includes the operating system of the first terminal, the type of network connected to the device, etc. The display position data includes the identification of the display position, the size of the display position, the context information of the display position in the resource display page, etc. The statistical features refer to the number of media resources browsed, the number of media resources clicked, the number of media resources followed, the number of media resources set as not interested, etc. by the target user in multiple historical time periods. Optionally, the multiple historical time periods can be the last month, the last three months, and the last six months.

[0155] In some embodiments, as Figure 14 As shown in 1401 to 1403, the server first determines the first candidate resource set from the full media resources based on the received resource acquisition request through quality information to recommend media resources. The quality information is used to indicate the quality of the media resources. The full media resources refer to all media resources stored in the media resource recommendation system. Accordingly, this step also includes: the server determines the media resources at the front target ratio as the first candidate resource set in descending order of the quality information of each media resource in the full media resources. It should be noted that the process of determining the first candidate resource set from the full media resources can be performed at any time after the server stores the media resources sent by the first terminal, and the server can update the first candidate resource set in real time according to the media resources sent by the first terminal. In an embodiment of the present application, the first candidate resource set is also called online index media resources.

[0156] 1303. The server inputs the historical behavior data of the target user and the first resource data of the first candidate media resource in the first candidate media resource set into the media resource recommendation model.

[0157] In some embodiments, the server inputs the first object data of the target object and the first resource data of each first candidate media resource into the media resource recommendation model, where the first object data includes historical behavior data, portrait data, and tag data of the target object.

[0158] Optionally, the first object data also includes device data, display location data and statistical data of the target object.

[0159] 1304. For the first resource data of any first candidate media resource, the server processes the historical behavior data and the first resource data through the media resource recommendation model based on the revenue reference space to obtain revenue reference information of the first candidate media resource.

[0160] The revenue reference space is obtained by training the media resource recommendation model and is used to measure the revenue reference information of the media resource. The training method of the media resource recommendation model is described in detail later in this embodiment. Figure 22 Corresponding embodiments.

[0161] In some embodiments, the server first obtains an initial object vector for the first object data and an initial resource vector for the first resource data using the media resource recommendation model. Then, based on the revenue reference space, the server processes the initial object vector and initial resource vector to obtain revenue reference information. This process is described below based on steps 1304A and 1304B.

[0162] 1304A: The server obtains an initial object vector and an initial resource vector through the media resource recommendation model.

[0163] In some embodiments, as Figure 15 As shown in 1501 to 1502, the server processes the first object data and the first resource data through the embedding vector sub-model and the self-attention sub-model of the media resource recommendation model to obtain the initial object vector and the initial resource vector, which is explained below through the following three processes.

[0164] Process 1: Data mapping.

[0165] In some embodiments, the server inputs the first object data and the first resource data into an embedded vector sub-model, maps the first object data and the first resource data, and obtains an initial sub-vector of the historical behavior data in the first object data, a sub-vector of other category data in the first object data, and multiple sub-vectors of the first resource data. The historical behavior data includes behavior data for multiple historical time periods, and the initial sub-vectors of the historical behavior data include initial sub-sub-vectors corresponding to the behavior data for the multiple historical time periods.

[0166] Process 2: Self-attention mechanism processing.

[0167] In some embodiments, the server inputs the initial subvector of the historical behavior data into the self-attention submodel of the media resource recommendation model. Based on the self-attention mechanism, the server extracts the correlation information between the behavior data of multiple historical time periods to obtain the subvector of the historical behavior data. The self-attention mechanism obtains the correlation between the behaviors of multiple historical time periods, thereby making the obtained first click-through rate, first conversion rate, and reference revenue information more accurate, thereby improving the accuracy of the recommended media resources.

[0168] For example, the process of extracting relevant information between behavioral data of multiple historical time periods based on the self-attention mechanism is described. The server first obtains the word vector (query), key vector (key) and value vector (value) corresponding to each initial sub-sub-vector based on the self-attention sub-model and the behavioral data of multiple historical time periods. For the initial sub-sub-vector of the behavioral data of any historical time period, the server multiplies the word vector of the initial sub-sub-vector with the key vector of each initial sub-sub-vector, normalizes the multiple multiplication results, multiplies the multiple normalized results with the corresponding value vector, and sums the multiple multiplication results to obtain the sub-sub-vector corresponding to the behavioral data of the historical time period. The sub-sub-vector is used to represent the relevant information between the behavioral data of the historical time period and the behavioral data of each historical time period. The server splices the sub-sub-vectors of the behavioral data of each historical time period to obtain the sub-vector of the historical behavioral data. The method for obtaining the sub-vector of the historical behavioral data is shown in formula (1).

[0169]

[0170] Among them, Attention() represents the sub-vector of historical behavior data, Q, K, and V represent the word vector (query), key vector (key), and value vector (value) of the behavior data in each historical time period, and d k is the normalization parameter.

[0171] In some embodiments, as Figure 16 As shown, the server uses a masked self-attention mechanism to obtain sub-sub-vectors of the behavior data for each historical time period. That is, the sub-sub-vectors of the behavior data for any historical time period are obtained only based on the behavior data before that historical time period, thereby ensuring that only the previous behavior can affect the subsequent behavior. Optionally, the server obtains sub-vectors of the historical behavior data based on a multi-head attention mechanism. That is, for the behavior data of each historical time period, multiple relationship information extractions are performed using multiple word vectors, key vectors, and value vectors, and optionally, relationship information extraction is performed twice.

[0172] Process 3: Sub-vector splicing.

[0173] In some embodiments, the server concatenates the subvector of the historical behavior data with subvectors of other categories of data in the first object data to obtain the initial object vector, and concatenates multiple subvectors of the first resource data to obtain the initial resource vector.

[0174] It should be noted that the above-described step 1304A is described using the example of a server receiving a resource acquisition request and obtaining the initial object vector of the target object and the initial resource vector of each first candidate media resource using the media resource recommendation model. In some embodiments, the server can pre-acquire the initial resource vector of each first candidate media resource based on the media resource recommendation model and store the obtained multiple initial resource vectors in a database. Upon receiving the resource acquisition request, the server obtains the initial object vector of the target object using the media resource recommendation model and obtains the initial resource vector of each first candidate media resource from the database, thereby significantly reducing the server's computing pressure and improving the efficiency of media resource recommendation.

[0175] 1304B. The server processes the initial object vector and the initial resource vector based on the media resource recommendation model and the revenue reference space to obtain revenue reference information of the first candidate media resource.

[0176] In some embodiments, as Figure 15 As shown in 1503, the server first obtains a first object vector and a first resource vector from the revenue reference space based on the first multi-task sub-model and the second multi-task sub-model of the media resource recommendation model. The first object vector represents the impact of the first object data on the revenue reference information of the first candidate media resource, and the first resource vector represents the impact of the first resource data on the revenue reference information of the first candidate media resource. Then, based on the revenue sub-model of the media resource recommendation model, the server processes the first object vector and the first resource vector to obtain revenue reference information. The following two processes are explained.

[0177] Process 1: Obtain a first object vector and a first resource vector.

[0178] In some embodiments, the server inputs the initial object vector into the first multi-task sub-model, processes the initial object vector through the first multi-task sub-model and the benefit reference space to obtain the first object vector, inputs the initial resource vector into the second multi-task sub-model, processes the initial resource vector through the second multi-task sub-model and the benefit reference space to obtain the first resource vector.

[0179] For example, the process of obtaining the first object vector is described by taking the first multi-task sub-model as a multi-gate mixture-of-expert network (MMoE). Figure 17As shown, the MMoE model includes multiple expert networks and multiple gating networks. Each expert network is used to fuse multiple sub-vectors of the initial object based on a fusion method to obtain a corresponding fusion vector. Each gating network corresponds to a task and is used to obtain weights corresponding to the fusion features of multiple expert networks based on the corresponding task, so as to weightedly sum the fusion features of multiple expert networks to obtain a vector for processing the corresponding task. The server inputs the initial object vector into multiple expert networks respectively and, based on the multiple expert networks, performs multiple fusions on the multiple sub-vectors of the initial object vector to obtain multiple object fusion vectors for the object data. The server inputs the initial object vector into the gating network corresponding to the benefit reference information task. Based on the gating network, the server obtains the weights corresponding to the fusion features of each object through the benefit reference space. Based on the weights, the server performs a weighted summation of the multiple object fusion vectors to obtain a first object vector. The first multi-task sub-model includes at least two gating networks. Optionally, the first multi-task sub-model includes three expert networks. Through the first multi-task sub-model, multiple sub-vectors in the initial object vector are fused multiple times to obtain richer object features. Further, based on the multiple fused vectors obtained, the first object vector is obtained through the revenue reference space, so that the first object vector can more accurately represent the impact of the first object data on the revenue reference information, making the obtained revenue reference information more accurate, thereby achieving the purpose of improving the accuracy of recommended media resources.

[0180] Optionally, the expert network is a deep cross network (DCN), and the gating network is a softmax function. The structure of the DCN network is as follows: Figure 18 As shown, the DCN network includes two branches, namely the feature cross sub-network and the depth sub-network. The feature cross sub-network includes multiple network layers. The feature cross sub-network is used to linearly fuse multiple sub-vectors in the initial object vector based on formula (2). The depth sub-network includes multiple network layers. The depth sub-network is used to nonlinearly fuse multiple sub-vectors in the initial object vector based on formula (3). The server splices the vectors obtained by the feature cross sub-network and the depth sub-network, and processes the spliced ​​vectors through the sigmoid function to obtain the corresponding object fusion vector.

[0181]

[0182] Among them, x0 represents the initial object vector of the input, x l Represents the input data of the lth layer of the feature cross sub-network, w l and b l They represent the scale parameter and offset parameter of the feature cross sub-network layer l, respectively, l+1Represents the output data of the lth layer of the feature cross subnetwork.

[0183] h l+1 =ReKLu(w' l h l +b' l ) (3)

[0184] Among them, h l Represents the input data of the deep sub-network layer l, w' l and b' l They represent the scale parameter and offset parameter of the lth layer in the deep sub-network, h l+1 Represents the output data of the lth layer of the deep self-network.

[0185] It should be noted that the process of obtaining the first resource vector is the same as the process of obtaining the first object vector, and will not be repeated here.

[0186] Step 2: Obtain reference income information.

[0187] In some embodiments, the server inputs the first object vector and the first resource vector into a revenue sub-model to obtain revenue reference information of the first candidate media resource.

[0188] For example, the process of obtaining revenue reference information based on the revenue sub-model is described. Figure 19 As shown, the revenue sub-model has a dual-tower structure. The two branches of the revenue sub-model are the first neural network and the second neural network. The first neural network and the second neural network both include multiple fully connected layers. Except for the last fully connected layer, each fully connected layer is connected to a ReLU activation function (Parametric Rectified Linear Unit, PReLU) with parameters. The server inputs the first object vector into the first neural network and the first resource vector into the second neural network, performs multiple dimensionality reduction on the first object vector and the first resource vector respectively, performs dot product on the two vectors obtained by dimensionality reduction, and obtains revenue reference information. Optionally, the first neural network and the second neural network both include 3 fully connected layers.

[0189] 1305. The server processes the historical behavior data and the first resource data through the media resource recommendation model based on at least one of the click rate space and the conversion rate space to obtain at least one of the first click rate and the first conversion rate of the first candidate media resource.

[0190] The click rate space and the conversion rate space are obtained by training a media resource recommendation model, and the click rate space and the conversion rate space are used to measure the click rate and the conversion rate of the first candidate media resource, respectively.

[0191] The following describes the process of the server obtaining the first click rate.

[0192] In some embodiments, the server obtains a second object vector and a second resource vector from the click-through rate space using a method similar to that of step 1304 above. The second object vector represents the impact of the first object data on the first click-through rate of the first candidate media resource, and the second resource vector represents the impact of the first resource data on the first click-through rate of the first candidate media resource. The server inputs the second object vector and the second resource vector into the click-through rate sub-model of the media resource recommendation model to obtain the first click-through rate.

[0193] For example, the process of obtaining the first click rate based on the click rate sub-model is described. Figure 20 As shown, the click-through rate sub-model has a dual-tower structure. The two branches of the click-through rate sub-model are the third neural network and the fourth neural network. The third neural network and the fourth neural network both include multiple fully connected layers. Except for the last fully connected layer, each fully connected layer is connected to a ReLU activation function. The server inputs the second object vector into the third neural network and the second resource vector into the fourth neural network. The second object vector and the second resource vector are subjected to multiple dimensionality reduction, and the two vectors obtained by dimensionality reduction are dot-producted. The result of the dot product is mapped by the sigmoid function to obtain the first click-through rate. Optionally, the third neural network and the fourth neural network both include 3 fully connected layers.

[0194] It should be noted that the process of the server obtaining the first conversion rate is similar to the process of obtaining the first click rate, which will not be repeated here.

[0195] It should be noted that for steps 1304 and 1305 above, the server only needs to obtain the initial object vector and initial resource vector once. After obtaining the initial user vector and initial resource vector, the server can obtain at least one of the first click-through rate and the second click-through rate, as well as the revenue reference information, in a serial or parallel manner, although this embodiment of the present application is not limited thereto.

[0196] 1306. The server determines a second candidate resource set from the first candidate resource set based on at least one of the first click rate and the first conversion rate of each first candidate media resource and the revenue reference information.

[0197] In some embodiments, the first candidate resource set includes multiple third candidate media resources with shallow conversion goals and multiple fourth candidate media resources with deep conversion goals. The server determines multiple first media resources from the multiple third candidate media resources based on the first conversion rate. The server determines multiple second media resources from the multiple fourth candidate media resources based on revenue reference information. The multiple first media resources and the multiple second media resources constitute the second candidate resource set. Optionally, the multiple third candidate media resources are partially identical to the multiple fourth candidate media resources.

[0198] Optionally, the server sorts the multiple third candidate media resources from highest to lowest according to the first conversion rate, and determines the third candidate media resource with the highest target ratio as the first media resource. The server also sorts the multiple fourth candidate media resources from highest to lowest according to the revenue reference information, and determines the fourth candidate media resource with the highest target ratio as the second media resource. For example, if the target ratio is 10%, there are 150 third candidate media resources and 100 fourth candidate media resources. After sorting the multiple third candidate media resources and the multiple fourth candidate media resources, the server determines the top 15 third candidate media resources as the first media resource, and the top 10 fourth candidate media resources as the second media resource.

[0199] Alternatively, as Figure 21 As shown, the server uses the first conversion rate as a first search metric to determine multiple first media resources from multiple third candidate media resources based on an ANN search algorithm, and uses revenue reference information as a second search metric to determine multiple second resources from multiple fourth candidate media resources. Optionally, the ANN search algorithm is a Hierarchical Navigable Small World Graphs (HNSW) algorithm.

[0200] It should be noted that, when the ANN algorithm is used to determine the second set of candidate resources, the server does not need to obtain the first conversion rate and revenue reference information of each first candidate media resource. For example, in the process of determining multiple first resources through the ANN algorithm, when the first conversion rate of a third candidate media resource is needed as the first retrieval indicator, based on the media resource recommendation model, the first conversion rate and revenue reference information of the third candidate media resource are obtained, the first conversion rate is used as the first retrieval indicator of the third candidate media resource, and the revenue reference information is stored in the cache. In the process of determining multiple second media resources through the ANN algorithm, when the revenue reference information of a fourth candidate media resource is needed as the second retrieval indicator, the revenue reference information of the fourth candidate media resource is first searched in the cache. If the revenue reference information is obtained in the cache, the retrieval process is continued. If the revenue reference information is not obtained in the cache, the revenue reference information and the first conversion rate of the fourth candidate media resource are obtained based on the media resource recommendation model.

[0201] By determining the second candidate resource set through the ANN algorithm, the server does not need to obtain the first conversion rate and revenue reference information of all first candidate media resources, which greatly reduces the computing pressure of the server and improves the efficiency of recommending media resources.

[0202] In some embodiments, if the recommended orientation data of each first candidate media resource is stored in the database, the server first compares the recommended orientation data of each first candidate media resource with the first object data of the target object, and determines the first candidate media resource whose recommended orientation data matches the first object data as the third media resource, and then compares the first media resource with an unbreakable orientation and each second media resource with the first object data, and takes multiple third media resources, first media resources and second media resources whose unbreakable orientation matches the first object data, and first media resources and second media resources that do not have an unbreakable orientation as the second candidate resource set.

[0203] It should be noted that steps 1302 to 1306 above are the process of determining the second set of candidate resources based on the media recommendation model. In this embodiment of the present application, this process is also referred to as recalling media resources. Optionally, the server uses a multi-threaded service to recall media resources to reduce the server's computing pressure and improve recall efficiency.

[0204] 1307. The server determines a media resource to be recommended from the second candidate media resource set.

[0205] In some embodiments, the server obtains the second click-through rate and the second conversion rate of each second candidate media resource based on the second object data of the target object and the second resource data of each second candidate media resource in the second media resource set, and determines the media resources to be recommended from the second candidate media resource set based on the second click-through rate and the second conversion rate of each second candidate media resource. The number of data categories in the second object data is greater than the number of data categories in the first object data, and the number of data categories in the second resource data is greater than the number of data categories in the first resource data. By obtaining object data and resource data with more data categories, a second click-through rate and a second conversion rate with higher accuracy are obtained, and further determining the media resources to be recommended based on the second click-through rate and the second conversion rate can effectively improve the accuracy of recommended media resources.

[0206] Exemplarily, the process of determining the media resources to be recommended based on the second click-through rate and the second conversion rate is described. The server obtains recommendation reference information based on the second click-through rate, second conversion rate, and bid of each second candidate media resource, sorts the multiple second candidate media resources in descending order of the recommendation reference information, and determines the second candidate media resources in the top N positions as the media resources to be recommended, where N is an integer greater than 0 and less than the number of second candidate media resources. Optionally, N = 1 or 2. The method for obtaining the recommendation reference information of any second candidate media resource is shown in formula (4).

[0207] eCPM1=bid×pCTR×pCVR (4)

[0208] Among them, eCPM1 represents the recommended reference information, bid represents the bid of the second candidate media resource, pCTR represents the second click-through rate, and pCVR represents the second conversion rate.

[0209] In some embodiments, the server first obtains a correction parameter for each second candidate media resource, which is used to correct the second conversion rate of the second candidate media resource, and then obtains the corrected recommendation reference information based on the correction parameter. Accordingly, step 1307 also includes: the server obtains a correction parameter for each second candidate media resource based on the number of conversions of each second candidate media resource, where the conversion number is the number of conversions of the second candidate media resource within the target historical time period. Based on the second click-through rate, second conversion rate, and correction parameter of each second candidate media resource, the server determines the media resource to be recommended from the set of second candidate media resources.

[0210] By obtaining the correction parameter, the second conversion rate of each second candidate media resource can be corrected according to the actual conversion quantity of the second candidate media resource, thereby obtaining corrected recommendation reference information, thereby achieving the purpose of improving the accuracy of recommending media resources.

[0211] In some embodiments, the server obtains correction parameters based on the number of conversions of each second candidate media resource in multiple dimensions. Exemplarily, the process of obtaining correction parameters for any second candidate media resource is described. The server obtains the number of conversions of the second candidate media resource in multiple dimensions within a target historical time period, and based on the number of conversions in the multiple dimensions, obtains correction factors corresponding to the multiple dimensions, each correction factor representing a deviation in the second conversion rate in the corresponding dimension. The correction factors for the multiple dimensions are weighted and summed to obtain the correction parameters for the second candidate media resource.

[0212] The conversion numbers of the above-mentioned multiple dimensions are explained below. The conversion numbers of the multiple dimensions include the conversion number of the second candidate media resource, the conversion number of the target object, the conversion number of the target object, and the conversion number of the target object type. Among them, the target object refers to the object that uploaded the second candidate media resource, and the conversion number of the target object refers to the number of times all the second candidate media resources uploaded by the target object were converted within the target historical time period. The target object represents the target object promoted by the second candidate media resource, and the conversion number of the target object represents the number of times all the second candidate media resources used to promote the target object were converted within the target historical time period. The target object type is the target object type promoted by the second candidate media resource, and the conversion number of the target object type represents the number of times all the second candidate media resources used to promote the target object type were converted within the target historical time period.

[0213] Exemplarily, the method for the server to obtain the correction factor of any dimension is shown in formula (5), and the method for obtaining the correction parameter of the second candidate media resource is shown in formula (6).

[0214] pCVR_bias_factor=conversion / sum(pCVR j ) (5)

[0215]

[0216] Among them, pCVR_bias_factor represents the correction factor of any dimension, conversion represents the number of conversions under any dimension, and pCVR j Indicates the second conversion rate obtained by the server when the dimension is converted for the jth time, sum() represents the summation operation, pCVR_bias_factor i represents the conversion factor of the i-th dimension, D represents the number of dimensions, W i represents the weight of the i-th dimension, and boost_factor represents the correction parameter.

[0217] By way of example, the process of determining a media resource to be recommended based on the second click-through rate, the second conversion rate, and the correction parameter is described. The server obtains the corrected recommendation reference information based on the first click-through rate, the second conversion rate, and the correction parameter. The server sorts the plurality of second candidate media resources in descending order of the corrected recommendation reference information, and determines the top N second candidate media resources as the media resources to be recommended. The method for obtaining the corrected recommendation reference information for any second candidate media resource is shown in Formula (7).

[0218] eCPM1'=bid×pCTR×pCVR×boost_factor (7)

[0219] Wherein, eCPM1' represents the revised recommended reference information.

[0220] In some embodiments, as Figure 14 As shown in 1404 to 1406, the server first determines the third candidate resource set from the second candidate resource set, and then determines the media resources to be recommended from the third candidate resource set. Accordingly, step 1307 can be replaced by: the server obtains the third click-through rate and the third conversion rate of each second candidate resource based on the third object data of the target object and the third resource data of each second candidate media resource, and determines the third candidate resource set from the second candidate resource set based on the third click-through rate and the third conversion rate of each second candidate media resource. The server obtains the second click-through rate and the second conversion rate of each third candidate media resource in the third candidate resource set, and determines the media resources with recommendations from the third candidate resource set based on the second click-through rate and the second conversion rate.

[0221] The number of data categories in the third object data is smaller than the number of data categories in the second object data, and the number of data categories in the third resource data is smaller than the number of data categories in the second resource data. First, a third candidate resource set is selected from the second candidate resource set using the third object data and third resource data, which have fewer data categories. Then, media resources to be recommended are determined from the third candidate resource set using the second object data and second resource data, which have more data categories. This can further improve the accuracy of media resource recommendations. Furthermore, because the object data and resource data with fewer data categories are used first, server computing pressure can be reduced, improving the efficiency of media resource recommendations.

[0222] Exemplarily, the process of determining the third candidate resource set based on the third click-through rate and the third conversion rate is described. The server obtains candidate reference information based on the third click-through rate, the third conversion rate and the bid of each second candidate media resource, sorts the multiple second candidate media resources in descending order of the candidate reference information, and determines the second candidate media resources in the first M positions as the third candidate media resources, where M is an integer greater than N and less than the number of second candidate media resources. Optionally, the server sorts the multiple second candidate media resources by a multi-way merge sorting method. Optionally, M=100. The method for obtaining candidate reference information of any second candidate media resource is shown in formula (8).

[0223] eCPM2=bid×liteCTR×liteCVR (8)

[0224] Among them, eCPM2 represents candidate reference information, liteCTR represents the third click-through rate, and liteCVR represents the third conversion rate.

[0225] In the embodiment of the present application, the process of determining the third resource set from the second candidate resource set is called coarse sorting, and the process of determining the media resources to be recommended from the third candidate resource set is called fine sorting.

[0226] 1308. The server sends a resource acquisition response to the second terminal, where the resource acquisition response carries the media resource to be recommended.

[0227] In some embodiments, the second terminal displays the received media resources to be recommended to the target object. Accordingly, the embodiment of the present application also includes: the second terminal receives the resource acquisition response, obtains the media resources to be recommended, and displays the media resources to be recommended on the resource display page to complete the media resource recommendation to the target object.

[0228] In some embodiments, while the target object is browsing a media resource display page, the server periodically performs a process of determining media resources to be recommended, and sends the media resources to be recommended to the second terminal. The second terminal periodically updates the media resources displayed in the resource display page based on the received media resources to complete the recommendation of media resources to the target object.

[0229] In some embodiments, the server receives an access request for any page from the second terminal, determines the media resources to be recommended based on the same method as steps 1302 to 1307 above, and sends the page information of the page and the media resources to be recommended to the second terminal. The page information is used to indicate the page. The second terminal receives the page information and the media resources, displays the page based on the page information, and displays the received media resources on the page to complete the recommendation of the media resources to the target object.

[0230] Optionally, the method provided in the embodiment of the present application is based on C++ language or Python language and is implemented through the Linux G++ compilation environment and deep learning compilation platform.

[0231] The method provided in the embodiment of the present application takes into account the impact of the object's behavior on media resource clicks and conversions by inputting the historical behavior data of the target object and the resource data of the candidate media resources into the media resource recommendation model, thereby making at least one of the obtained click-through rate and conversion rate and the revenue reference information more accurate. Further, based on the obtained data, the media resources recommended for the target object are determined from the candidate media resources, while taking into account the click-through rate, conversion rate and revenue reference information, which can effectively improve the accuracy of media resource recommendations.

[0232] The following combination Figure 22 , the training process of the media resource recommendation model in the above embodiment is introduced. The training process corresponds to Figure 12 The three processes are data access 1201, sample construction 1202 and model construction 1203.

[0233] In some embodiments, the media resource recommendation model is trained based on at least one of a click tag and a conversion tag, sample data, and sample revenue reference information. The sample data includes sample historical behavior data of sample subjects and sample resource data of sample media resources. Optionally, the sample data also includes other categories of data of the sample subjects, such as profile data, tag data, device data, etc. It should be noted that the data categories of the sample subjects in the sample data are the same as the data categories of the target subjects used in the above-mentioned media resource recommendation process.

[0234] Model training includes two processes: training data preparation and model training. These two processes are explained below.

[0235] Process 1: Training data preparation.

[0236] In some embodiments, the training data includes a plurality of sample data and corresponding click tags, conversion tags, and sample revenue reference information. The plurality of sample data, corresponding click tags, and conversion tags are obtained based on a media resource recommendation system, and the sample revenue reference information corresponding to the plurality of sample data is obtained based on revenue data sent by the subject.

[0237] The following describes the process of obtaining sample data, corresponding click tags, and conversion tags. Figure 22As shown in 2201 to 2202, the server obtains the system log in the media resource recommendation system, parses the system log, and obtains the initial object data of multiple sample objects, the initial resource data of the corresponding multiple sample media resources, and the click status and conversion status of each sample object for the corresponding sample media resources. Multiple categories of data of the sample object are extracted from each initial object data, and multiple categories of data of the sample media resources are extracted from each corresponding initial resource data to obtain multiple sample data. According to the click status and conversion status of each sample object for the corresponding sample media resource, the click label and conversion label corresponding to each sample data are obtained.

[0238] The following describes the process of obtaining sample income reference information for sample data. Figure 22 As shown in 2203 to 2205, the server obtains the sample revenue data of multiple sample media resources sent by the object, wherein each sample media resource corresponds to multiple sample revenue data, and the multiple sample revenue data of each sample media resource are the revenue generated after multiple sample objects have converted the sample media resource. The server pre-processes the received sample revenue data, and based on the sample object and sample media resource corresponding to each sample data, matches the multiple sample revenue data of each sample media resource after pre-processing with each sample data to obtain the sample revenue data corresponding to each sample data, and multiplies the sample revenue data of each sample data by the corresponding object conversion number to obtain the sample revenue reference information of each sample data. The object conversion number refers to the number of times the sample object converts the sample media resource. In an embodiment of the present application, the revenue reference information is also called real user value (RUV).

[0239] Optionally, the server obtains the sample income data once every hour. Optionally, the pre-processing of the sample income data refers to filtering dirty data from the sample income data, where dirty data refers to sample income data with an incorrect data format.

[0240] It should be noted that the sample revenue data for each sample media resource described above is standardized data. In some embodiments, the media resource recommendation system provides a standardized processing function for sample revenue data. Since this sample revenue data is the core data of the object, before sending the sample revenue data to the server, the object can perform standardized processing on the sample revenue data based on the media resource recommendation system and send the processed data to the server, thereby ensuring the security of the core data. The method for standardizing the sample revenue data is shown in Formula (9).

[0241]

[0242] Among them, original_user_value represents the original unprocessed sample income data, and standard_user_value represents the standardized sample income data. Indicates a round-up operation.

[0243] By standardizing the sample revenue data, not only can the security of the core data of the object be guaranteed, but the distribution of the sample revenue data can also be made more balanced, thereby improving the training effect of the media resource recommendation model. Figure 23 As shown, Figure 23 This is a distribution chart of multiple sample revenue data of a sample media resource that has not been standardized. Figure 23 It can be seen that the data distribution of the sample income data is very balanced, and the number of sample users with different sample income data shows an exponential difference. Therefore, by performing a log() operation on the sample income data, the distribution of the sample income data can be made more balanced.

[0244] Process 2: Model training.

[0245] In some embodiments, the training of the media resource recommendation model is achieved through multiple iterations. During any iteration, the server obtains a batch of sample data and corresponding click labels, conversion labels and sample reference information from multiple sample data, and obtains at least one of the predicted click rate and predicted conversion rate of the sample data and the predicted revenue reference information based on the media resource recommendation model, and trains the media resource recommendation model based on at least one of the predicted click rate and the predicted conversion rate, at least one of the click label and the conversion label, the predicted revenue reference information and the sample revenue reference information. Optionally, during the training process, for the first neural network and the second neural network in the revenue sub-model, except for the last fully connected layer, the server standardizes the input data of each fully connected layer through the batch normalization algorithm (BN), thereby accelerating the convergence of the model.

[0246] Optionally, the number of sample data in each batch is 4096, the learning rate (learning_rate) is set to 0.0001, and the number of times the model is trained using all sample data (epoch) is set to 2.

[0247] Exemplarily, the process of training a media resource recommendation model based on at least one of the predicted click-through rate and the predicted conversion rate and the predicted revenue reference information is described. The server obtains the click-through rate loss of each sample data based on the predicted click-through rate and click label of each sample data, obtains the conversion rate loss of each sample data based on the predicted conversion rate and conversion label of each sample data, obtains the revenue loss of each sample data based on the predicted revenue reference information and the sample revenue reference information of each sample data, sums the click-through rate loss, conversion rate loss and revenue loss of each sample data in a batch of sample data respectively, and obtains the click-through rate loss, conversion rate loss and revenue loss of a batch of sample data, and weighted sums the click-through rate loss, conversion rate loss and revenue loss of the batch of sample data to obtain the target loss, and updates the network parameters of the media resource recommendation model based on the target loss. Optionally, the server updates the network parameters based on the target loss by an adaptive moment estimation algorithm (Adam) and a gradient descent method.

[0248] Exemplarily, the method for the server to obtain the click rate loss and conversion rate loss of a batch of sample data is shown in formula (10).

[0249] loss x =-∑[y i ×logP i +(1-y i )×log(1-P i )] (10)

[0250] Among them, loss x represents click rate loss or conversion rate loss, Σ[] represents the summation operation, P i Indicates the predicted click rate or predicted conversion rate of the i-th sample data in a batch of sample data, y i Indicates the click label or conversion label of the i-th sample data, with y i Click the label as an example to illustrate, i =0 or 1, y i = 0 means that the sample media resource of the i-th sample data has not been clicked by the sample object, y i =1 indicates that the sample media resource of the i-th sample data is clicked by the sample object.

[0251] For example, as shown in formula (11), the server obtains the profit loss of any sample data through the Huber loss function. Through the Huber loss function, the profit loss value of each sample data will not be too high or too low, thereby improving the model training effect.

[0252] loss aux=HuberLoss(predict,RUV) (11)

[0253] Among them, loss aux represents the profit loss of any sample data, predict represents the predicted profit reference information of the sample data, RUV represents the sample profit reference information of the sample data, HuberLoss() represents the Huber loss function, and the Huber loss function is shown in formula (12).

[0254]

[0255] Among them, a=predict-RUV, δ is the preset loss parameter, L δ () represents Huber loss, which is the profit loss of sample data.

[0256] Exemplarily, the method for the server to obtain the target loss is shown in formula (13).

[0257] Loss = αloss ctr +βloss cvr +γLoss aux (13)

[0258] Among them, Loss represents the target loss, loss ctr 、loss cvr and Loss aux They represent the click rate loss, conversion rate loss and revenue loss of a batch of sample data respectively, and α, β and γ represent the weight of click rate loss, conversion rate loss and revenue loss respectively.

[0259] It should be noted that the server acquires new training data every day. The server can train the media resource recommendation model based on the newly acquired training data, thereby being able to update the model in a timely manner and improve the accuracy of the media resource recommendation model.

[0260] Figure 24 This is a schematic diagram of the structure of a media resource recommendation device provided in an embodiment of the present application. Figure 24 The device includes: an acquisition module 2401, an input module 2402, a determination module 2403 and a recommendation module 2404.

[0261] An acquisition module 2401 is configured to acquire historical behavior data of a target object, wherein the historical behavior data indicates browsing behavior and resource transfer behavior of the target object based on media resources.

[0262] Input module 2402 is configured to input the historical behavior data of the target subject and the first resource data of the first candidate media resource in the first candidate resource set into the media resource recommendation model to obtain at least one of the first click-through rate and the first conversion rate of each first candidate media resource and revenue reference information, the revenue reference information indicating the revenue generated after the target subject performs a conversion behavior on the candidate media resource;

[0263] A determination module 2403 is configured to determine a media resource to be recommended based on at least one of the first click rate and the first conversion rate of each first candidate media resource and the revenue reference information;

[0264] The recommendation module 2404 is configured to recommend resources to the target object based on the media resources to be recommended;

[0265] The media resource recommendation model is trained based on historical behavior data of multiple sample objects.

[0266] In some embodiments, the input module 2402 includes:

[0267] An input submodule, configured to input the historical behavior data and the first resource data of any first candidate media resource into the media resource recommendation model;

[0268] a processing submodule, configured to process the historical behavior data and the first resource data using the media resource recommendation model and based on the revenue reference space to obtain revenue reference information of the first candidate media resource;

[0269] The processing submodule is used to process the historical behavior data and the first resource data based on at least one of the click rate space and the conversion rate space through the media resource recommendation model to obtain at least one of the first click rate and the first conversion rate of the first candidate media resource.

[0270] In some embodiments, the acquisition module 2401 is further used to acquire portrait data and tag data of the target object, where the tag data is used to represent the preference of the target object;

[0271] Accordingly, the processing submodule includes:

[0272] a first acquisition unit, configured to acquire, based on the media resource recommendation model, an initial object vector of first object data and an initial resource vector of the first resource data, the first object data including historical behavior data, portrait data, and label data of the target object, the initial object vector including subvectors corresponding to multiple categories of data in the first object data, and the initial resource vector including subvectors corresponding to multiple categories of data in the first resource data;

[0273] a fusion unit configured to fuse multiple subvectors of the initial object vector multiple times based on multiple fusion methods to obtain multiple object fusion vectors of the object data, and to fuse multiple subvectors of the initial resource vector multiple times to obtain multiple resource fusion vectors of the resource data;

[0274] a weighting unit, configured to perform weighted summation on the multiple object fusion vectors and the multiple resource fusion vectors based on the revenue reference space to obtain an object vector and a resource vector, wherein the object vector represents an impact of the first object data on the revenue reference information of the first candidate media resource, and the resource vector represents an impact of the first resource data on the revenue reference information of the first candidate media resource;

[0275] The first acquiring unit is configured to acquire revenue reference information of the first candidate media resource based on the object vector and the resource vector.

[0276] In some embodiments, the historical behavior data includes behavior data for multiple historical time periods;

[0277] The first acquisition unit is used to map the first object data and the first resource data respectively based on the media resource recommendation model to obtain the initial sub-vector of the historical behavior data, the sub-vector of the portrait data, the sub-vector of the label data, and multiple sub-vectors of the first resource data; based on the self-attention mechanism and the initial sub-vector of the historical behavior data, extract the relevant information between the behavior data of the multiple historical time periods to obtain the sub-vector of the historical behavior data; splice the sub-vector of the historical behavior data, the sub-vector of the portrait data, and the sub-vector of the label data to obtain the initial object vector; and splice multiple sub-vectors of the first resource data to obtain the initial resource vector.

[0278] In some embodiments, the apparatus further comprises:

[0279] a training module for training the media resource recommendation model based on at least one of a click tag and a conversion tag, sample data, and sample revenue reference information, wherein the sample data includes sample historical behavior data of sample objects and sample resource data of sample media resources;

[0280] The training module is used to obtain at least one of the predicted click-through rate and the predicted conversion rate of the sample data and the predicted revenue reference information based on the media resource recommendation model; and train the media resource recommendation model based on at least one of the predicted click-through rate and the predicted conversion rate, at least one of the click label and the conversion label, the predicted revenue reference information and the sample revenue reference information.

[0281] In some embodiments, the sample income reference information is obtained by standardizing the sample income data.

[0282] In some embodiments, the determining module 2403 includes:

[0283] A first determining submodule, configured to determine a plurality of first media resources from the first candidate resource set based on a first conversion rate of each of the first candidate media resources;

[0284] A second determining submodule, configured to determine a plurality of second media resources from the first candidate resource set based on the revenue reference information of each of the first candidate media resources;

[0285] The third determining submodule is configured to determine the media resources to be recommended from a second candidate resource set, where the second candidate resource set includes the plurality of first media resources and the plurality of second media resources.

[0286] In some embodiments, the third determining submodule includes:

[0287] a second acquiring unit, acquiring a second click rate and a second conversion rate of each second candidate media resource based on the second object data of the target object and the second resource data of each second candidate media resource in the second candidate resource set;

[0288] a determining unit, configured to determine a media resource to be recommended from the second candidate resource set based on the second click-through rate and the second conversion rate of each of the second candidate media resources;

[0289] The number of data categories in the second object data is greater than the number of data categories in the first object data, and the number of data categories in the second resource data is greater than the number of data categories in the first resource data.

[0290] In some embodiments, the determination unit is used to obtain a correction parameter for each of the second candidate media resources based on the conversion number of each of the second candidate media resources, where the conversion number is the number of conversions of the second candidate media resource within the target historical time period, and the correction parameter is used to correct the second conversion rate of the second candidate media resource; based on the second click-through rate, second conversion rate and correction parameter of each of the second candidate media resources, the media resources to be recommended are determined from the set of second candidate resources.

[0291] It should be noted that the media resource recommendation device provided in the above embodiment only uses the division of the above functional modules as an example to illustrate when recommending media resources. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the media resource recommendation device provided in the above embodiment and the media resource recommendation method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0292] Figure 25 This is a schematic diagram of the structure of a media resource recommendation device provided in an embodiment of the present application. Figure 25 As shown, the device includes: a display module 2501 and a sending module 2502.

[0293] A display module is used to display a media resource recommendation page, and display an intelligent recommendation option on the media resource recommendation page, wherein the intelligent recommendation option is used to select whether to use the media resource recommendation model to recommend media resources;

[0294] a sending module configured to, when the intelligent recommendation option is set to use the media resource recommendation model, send the media resource and a recommendation instruction to the server in response to an upload operation of the media resource on the media resource recommendation page, the recommendation instruction being used to instruct to recommend the media resource based on the media resource recommendation model;

[0295] Among them, the resource recommendation model is used to obtain at least one of the first click-through rate and the first conversion rate of the media resource and the revenue reference information based on the historical behavior data of the target object and the first resource data of the media resource. The historical behavior data represents the browsing behavior and resource transfer behavior of the target object based on the media resource, and the revenue reference information represents the revenue generated after the target object performs the conversion behavior on the media resource. The resource recommendation model is trained based on the historical behavior data of multiple sample objects.

[0296] It should be noted that the media resource recommendation device provided in the above embodiment only uses the division of the above functional modules as an example to illustrate when recommending media resources. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the media resource recommendation device provided in the above embodiment and the media resource recommendation method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0297] The embodiments of the present disclosure provide a computer device for executing the above-mentioned method for recommending media resources. In some embodiments, the computer device is provided as a first terminal or a second terminal. Figure 26 The figure is a schematic diagram of the structure of a terminal provided in an embodiment of the present application. Terminal 2600 may be a smartphone, a tablet computer, a Moving Picture Experts Group Audio Layer III (MP3) player, a Moving Picture Experts Group Audio Layer IV (MP4) player, a laptop computer, or a desktop computer. Terminal 2600 may also be referred to as user equipment, a portable terminal, a laptop terminal, a desktop terminal, or other similar names.

[0298] Typically, the terminal 2600 includes: one or more processors 2601 and one or more memories 2602 .

[0299] The processor 2601 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 2601 may be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 2601 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a central processing unit (CPU); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 2601 may be integrated with a graphics processing unit (GPU), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 2601 may also include an artificial intelligence (AI) processor, which is used to process computing operations related to machine learning.

[0300] Memory 2602 may include one or more computer-readable storage media, which may be non-transitory. Memory 2602 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices and flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in memory 2602 is used to store at least one program code, which is used to be executed by processor 2601 to implement the media resource recommendation method provided in the method embodiment of the present application.

[0301] In some embodiments, terminal 2600 may optionally include a peripheral device interface 2603 and at least one peripheral device. Processor 2601, memory 2602, and peripheral device interface 2603 may be connected via a bus or signal lines. Each peripheral device may be connected to peripheral device interface 2603 via a bus, signal lines, or circuit boards. Specifically, the peripheral device may include at least one of a radio frequency circuit 2604, a display screen 2605, a camera assembly 2606, an audio circuit 2607, and a power supply 2609.

[0302] The peripheral device interface 2603 can be used to connect at least one input / output (I / O)-related peripheral device to the processor 2601 and the memory 2602. In some embodiments, the processor 2601, the memory 2602, and the peripheral device interface 2603 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 2601, the memory 2602, and the peripheral device interface 2603 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0303] The radio frequency circuit 2604 is used to receive and transmit radio frequency (RF) signals, also known as electromagnetic signals. The radio frequency circuit 2604 communicates with communication networks and other communication devices via electromagnetic signals. The radio frequency circuit 2604 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. Optionally, the radio frequency circuit 2604 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The radio frequency circuit 2604 can communicate with other terminals via at least one wireless communication protocol. Such wireless communication protocols include, but are not limited to, metropolitan area networks, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or wireless fidelity (WiFi) networks. In some embodiments, the radio frequency circuit 2604 may also include circuits related to near field communication (NFC), which is not limited in this application.

[0304] Display screen 2605 is used to display a user interface (UI). This UI can include graphics, text, icons, videos, or any combination thereof. When display screen 2605 is a touchscreen display, it can also capture touch signals on or above the surface of display screen 2605. These touch signals can be input as control signals to processor 2601 for processing. Display screen 2605 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there can be a single display screen 2605, located on the front panel of terminal 2600. In other embodiments, there can be at least two display screens 2605, located on different surfaces of terminal 2600 or in a foldable design. In some embodiments, display screen 2605 can be a flexible display, located on a curved or foldable surface of terminal 2600. Display screen 2605 can also be configured as a non-rectangular, irregular shape, also known as a special-shaped screen. Display screen 2605 can be made of materials such as liquid crystal display (LCD) and organic light-emitting diode (OLED).

[0305] The camera assembly 2606 is used to capture images or videos. Optionally, the camera assembly 2606 includes a front camera and a rear camera. Typically, the front camera is arranged on the front panel of the terminal, and the rear camera is arranged on the back of the terminal. In some embodiments, there are at least two rear cameras, which are any one of a main camera, a depth of field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth of field camera to realize the background blur function, the fusion of the main camera and the wide-angle camera to realize panoramic shooting and virtual reality (VR) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 2606 may also include a flash. The flash can be a monochrome temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation at different color temperatures.

[0306] The audio circuit 2607 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals that are input into the processor 2601 for processing, or input into the radio frequency circuit 2604 to achieve voice communication. For the purpose of stereo sound collection or noise reduction, there may be multiple microphones, each located in different parts of the terminal 2600. The microphone may also be an array microphone or an omnidirectional collection microphone. The speaker is used to convert electrical signals from the processor 2601 or the radio frequency circuit 2604 into sound waves. The speaker may be a traditional thin film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert electrical signals into sound waves audible to humans, but also convert electrical signals into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 2607 may also include a headphone jack.

[0307] Power supply 2609 is used to power various components in terminal 2600. Power supply 2609 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 2609 includes a rechargeable battery, the rechargeable battery can support wired charging or wireless charging. The rechargeable battery can also be used to support fast charging technology.

[0308] In some embodiments, the terminal 2600 further includes one or more sensors 2610 , including but not limited to: an acceleration sensor 2611 , a gyroscope sensor 2612 , a pressure sensor 2613 , an optical sensor 2615 , and a proximity sensor 2616 .

[0309] The accelerometer 2611 can detect the magnitude of acceleration along the three coordinate axes of the coordinate system established by the terminal 2600. For example, the accelerometer 2611 can be used to detect the components of gravity acceleration along the three coordinate axes. The processor 2601 can control the display screen 2605 to display the user interface in a landscape or portrait view based on the gravity acceleration signal collected by the accelerometer 2611. The accelerometer 2611 can also be used to collect game or user motion data.

[0310] The gyroscope sensor 2612 can detect the orientation and rotation angle of the terminal 2600. It can work in conjunction with the accelerometer 2611 to collect the user's 3D movements on the terminal 2600. Based on the data collected by the gyroscope sensor 2612, the processor 2601 can implement the following functions: motion sensing (for example, changing the UI based on the user's tilt operation), image stabilization during shooting, game control, and inertial navigation.

[0311] The pressure sensor 2613 can be provided on the side frame of the terminal 2600 and / or below the display screen 2605. When the pressure sensor 2613 is provided on the side frame of the terminal 2600, it can detect the user's gripping signal of the terminal 2600. The processor 2601 performs left and right hand recognition or shortcut operations based on the gripping signal collected by the pressure sensor 2613. When the pressure sensor 2613 is provided below the display screen 2605, the processor 2601 controls the operable controls on the UI interface based on the user's pressure operation on the display screen 2605. The operable controls include at least one of a button control, a scroll bar control, an icon control, and a menu control.

[0312] Optical sensor 2615 is used to detect ambient light intensity. In one embodiment, processor 2601 can control the display brightness of display screen 2605 based on the ambient light intensity detected by optical sensor 2615. Specifically, when the ambient light intensity is high, the display brightness of display screen 2605 is increased; when the ambient light intensity is low, the display brightness of display screen 2605 is decreased. In another embodiment, processor 2601 can also dynamically adjust the shooting parameters of camera assembly 2606 based on the ambient light intensity detected by optical sensor 2615.

[0313] Proximity sensor 2616, also known as a distance sensor, is typically located on the front panel of terminal 2600. Proximity sensor 2616 is used to detect the distance between the user and the front of terminal 2600. In one embodiment, when proximity sensor 2616 detects that the distance between the user and the front of terminal 2600 is gradually decreasing, processor 2601 controls display screen 2605 to switch from the screen-on state to the screen-off state. When proximity sensor 2616 detects that the distance between the user and the front of terminal 2600 is gradually increasing, processor 2601 controls display screen 2605 to switch from the screen-off state to the screen-on state.

[0314] Those skilled in the art will understand that Figure 26 The structure shown in the figure does not constitute a limitation on the terminal 2600, and the terminal 2600 may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.

[0315] The embodiments of the present disclosure provide a computer device for executing the above-mentioned method for recommending media resources. In some embodiments, the computer device is provided as a server. Figure 27 This is a schematic diagram of the structure of a server provided in an embodiment of the present application. The server 2700 may have relatively large differences due to different configurations or performances, and may include one or more processors (Central Processing Units, CPU) 2701 and one or more memories 2702, wherein the one or more memories 2702 store at least one program code, and the at least one program code is loaded and executed by the one or more processors 2701 to implement the methods provided in the above-mentioned various method embodiments. Of course, the server 2700 may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server 2700 may also include other components for implementing device functions, which will not be described in detail here.

[0316] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including at least one program code. The at least one program code can be executed by a processor to implement the media resource recommendation method described in the above embodiment. For example, the computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, or an optical data storage device.

[0317] In an exemplary embodiment, a computer program product is also provided, comprising at least one computer program stored in a computer-readable storage medium. A processor of a computer device reads the at least one computer program from the computer-readable storage medium and executes the at least one computer program, causing the computer device to perform the operations performed by the above-described method for recommending media resources.

[0318] In some embodiments, the computer program involved in the embodiments of the present application may be deployed and executed on a computer device, or on multiple computer devices located at one location, or on multiple computer devices distributed at multiple locations and interconnected through a communication network. Multiple computer devices distributed at multiple locations and interconnected through a communication network may constitute a blockchain system.

[0319] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or may be accomplished by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0320] The above are only optional embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for recommending media resources, characterized in that: The method comprises: Inputting first object data of a target object and first resource data of a first candidate media resource in a first candidate resource set into an embedding vector sub-model in a media resource recommendation model, obtaining an initial sub-vector of historical behavior data in the first object data, sub-vectors of other category data in the first object data, and multiple sub-vectors of the first resource data, wherein the historical behavior data represents browsing behavior and resource transfer behavior of the target object based on media resources, and the other category data includes portrait data and label data, wherein the label data represents the preferences of the target object; Inputting the initial sub-vector of the historical behavior data into the self-attention sub-model in the media resource recommendation model to obtain the sub-vector of the historical behavior data; splicing the subvector of the historical behavior data with the subvector of the other category data to obtain an initial object vector; splicing multiple subvectors of the first resource data to obtain an initial resource vector; Inputting the initial object vector and the initial resource vector into a first multi-task sub-model and a second multi-task sub-model in the media resource recommendation model, respectively, and obtaining a first object vector and a first resource vector, a second object vector and a second resource vector, and a third object vector and a third resource vector based on a revenue reference space, a click-through rate space, and a conversion rate space, respectively; Inputting the first object vector and the first resource vector into a revenue sub-model in the media resource recommendation model to obtain revenue reference information of the first candidate media resource, wherein the revenue reference information represents the revenue generated after the target object performs a conversion behavior on the candidate media resource; Inputting the second object vector and the second resource vector into a click-through rate sub-model in the media resource recommendation model to obtain a first click-through rate of the first candidate media resource; Inputting the third object vector and the third resource vector into a conversion rate sub-model in the media resource recommendation model to obtain a first conversion rate of the first candidate media resource; Determining a media resource to be recommended based on at least one of a first click rate and a first conversion rate of each of the first candidate media resources and revenue reference information; Recommending resources to the target object based on the media resources to be recommended; The media resource recommendation model is trained based on historical behavior data of multiple sample objects.

2. The method according to claim 1, characterized in that The process of obtaining the first object vector and the first resource vector based on the benefit reference space includes: Based on multiple fusion methods, multiple subvectors of the initial object vector are fused multiple times to obtain multiple object fusion vectors of the first object data, and multiple subvectors of the initial resource vector are fused multiple times to obtain multiple resource fusion vectors of the first resource data; Based on the revenue reference space, the multiple object fusion vectors and the multiple resource fusion vectors are weightedly summed up respectively to obtain the first object vector and the first resource vector, the first object vector represents the impact of the first object data on the revenue reference information of the first candidate media resource, and the first resource vector represents the impact of the first resource data on the revenue reference information of the first candidate media resource.

3. The method according to claim 1, characterized in that The method further comprises: Training the media resource recommendation model based on at least one of a click tag and a conversion tag, sample data, and sample revenue reference information, wherein the sample data includes sample historical behavior data of sample objects and sample resource data of sample media resources; The training process of the media resource recommendation model includes: Based on the media resource recommendation model, obtaining at least one of a predicted click-through rate and a predicted conversion rate of the sample data and predicted revenue reference information; The media resource recommendation model is trained based on at least one of the predicted click rate and the predicted conversion rate, at least one of the click label and the conversion label, the predicted revenue reference information and the sample revenue reference information.

4. The method according to claim 3, characterized in that The sample income reference information is obtained by standardizing the sample income data.

5. The method according to claim 1, characterized in that The determining of the media resource to be recommended based on at least one of the first click rate and the first conversion rate of each of the first candidate media resources and the revenue reference information includes: determining a plurality of first media resources from the first candidate resource set based on the first conversion rate of each of the first candidate media resources; determining a plurality of second media resources from the first candidate resource set based on the revenue reference information of each of the first candidate media resources; The media resources to be recommended are determined from a second candidate resource set, where the second candidate resource set includes the plurality of first media resources and the plurality of second media resources.

6. The method according to claim 5, characterized in that Determining the media resource to be recommended from the second candidate resource set includes: Based on the second object data of the target object and the second resource data of each second candidate media resource in the second candidate resource set, obtaining a second click rate and a second conversion rate of each second candidate media resource; determining a media resource to be recommended from the second candidate resource set based on the second click rate and the second conversion rate of each second candidate media resource; The number of data categories in the second object data is greater than the number of data categories in the first object data, and the number of data categories in the second resource data is greater than the number of data categories in the first resource data.

7. The method according to claim 6, characterized in that The determining of the media resource to be recommended from the second candidate resource set based on the second click rate and the second conversion rate of each second candidate media resource includes: Obtaining a correction parameter for each second candidate media resource based on the conversion number of each second candidate media resource, where the conversion number is the number of conversions of the second candidate media resource within a target historical time period, and the correction parameter is used to correct the second conversion rate of the second candidate media resource; Based on the second click rate, the second conversion rate, and the correction parameter of each of the second candidate media resources, a media resource to be recommended is determined from the second candidate resource set.

8. A method for recommending media resources, characterized in that: The method comprises: Displaying a media resource recommendation page, displaying an intelligent recommendation option on the media resource recommendation page, the intelligent recommendation option being used to select whether to use the media resource recommendation model to recommend media resources; When the smart recommendation option is set to use the media resource recommendation model, in response to an upload operation of a media resource on the media resource recommendation page, sending the media resource and a recommendation instruction to a server, wherein the recommendation instruction is used to instruct to recommend the media resource based on the media resource recommendation model; The embedding sub-model in the media resource recommendation model is used to obtain, based on first object data of a target object and first resource data of a first candidate media resource in a first candidate resource set, an initial sub-vector of historical behavior data in the first object data, a sub-vector of other category data in the first object data, and multiple sub-vectors of the first resource data, wherein the historical behavior data represents browsing behavior and resource transfer behavior of the target object based on media resources, the other category data includes portrait data and label data, and the label data represents the preferences of the target object; The self-attention sub-model in the media resource recommendation model is used to obtain a sub-vector of the historical behavior data based on the initial sub-vector of the historical behavior data; The first multi-task sub-model and the second multi-task sub-model in the media resource recommendation model are used to obtain a first object vector and a first resource vector, a second object vector and a second resource vector, and a third object vector and a third resource vector based on the initial object vector and the initial resource vector, the revenue reference space, the click-through rate space, and the conversion rate space; the initial object vector is obtained by concatenating a subvector of the historical behavior data and a subvector of the other category data, and the initial resource vector is obtained by concatenating multiple subvectors of the first resource data; The revenue sub-model in the media resource recommendation model is used to obtain revenue reference information of the first candidate media resource based on the first object vector and the first resource vector, wherein the revenue reference information represents the revenue generated after the target object performs a conversion behavior on the candidate media resource; The click-through rate sub-model in the media resource recommendation model is used to obtain a first click-through rate of the first candidate media resource based on the second object vector and the second resource vector; The conversion rate sub-model in the media resource recommendation model is used to obtain a first conversion rate of the first candidate media resource based on the third object vector and the third resource vector; The media resource recommendation model is trained based on historical behavior data of multiple sample objects.

9. A device for recommending media resources, characterized in that: The device comprises: a processing submodule, configured to input first object data of a target object and first resource data of a first candidate media resource in a first candidate resource set into an embedding vector submodel in a media resource recommendation model, to obtain an initial subvector of historical behavior data in the first object data, subvectors of other category data in the first object data, and multiple subvectors of the first resource data, wherein the historical behavior data represents browsing behavior and resource transfer behavior of the target object based on media resources, the other category data includes portrait data and label data, and the label data represents preferences of the target object; The processing submodule is further configured to input the initial subvector of the historical behavior data into the self-attention submodel in the media resource recommendation model to obtain the subvector of the historical behavior data; The processing submodule is further configured to concatenate the subvector of the historical behavior data with the subvector of the other category data to obtain an initial object vector; and concatenate multiple subvectors of the first resource data to obtain an initial resource vector; The processing submodule is further configured to input the initial object vector and the initial resource vector into a first multi-task sub-model and a second multi-task sub-model in the media resource recommendation model, respectively, and obtain a first object vector and a first resource vector, a second object vector and a second resource vector, and a third object vector and a third resource vector based on a revenue reference space, a click-through rate space, and a conversion rate space, respectively; The processing submodule is further configured to input the first object vector and the first resource vector into a revenue submodel in the media resource recommendation model to obtain revenue reference information of the first candidate media resource, wherein the revenue reference information represents the revenue generated after the target object performs a conversion behavior on the candidate media resource; The processing submodule is further configured to input the second object vector and the second resource vector into a click-through rate submodel in the media resource recommendation model to obtain a first click-through rate of the first candidate media resource; The processing submodule is further configured to input the third object vector and the third resource vector into a conversion rate submodel in the media resource recommendation model to obtain a first conversion rate of the first candidate media resource; a determination module, configured to determine a media resource to be recommended based on at least one of a first click rate and a first conversion rate of each of the first candidate media resources and revenue reference information; A recommendation module, configured to recommend resources to the target object based on the media resources to be recommended; The media resource recommendation model is trained based on historical behavior data of multiple sample objects.

10. The device according to claim 9, characterized in that The processing submodule includes: a fusion unit, configured to fuse multiple subvectors of the initial object vector multiple times based on multiple fusion methods to obtain multiple object fusion vectors of the first object data, and to fuse multiple subvectors of the initial resource vector multiple times to obtain multiple resource fusion vectors of the first resource data; A weighting unit is used to perform weighted summation on the multiple object fusion vectors and the multiple resource fusion vectors based on the revenue reference space to obtain the first object vector and the first resource vector, wherein the first object vector represents the influence of the first object data on the revenue reference information of the first candidate media resource, and the first resource vector represents the influence of the first resource data on the revenue reference information of the first candidate media resource.

11. The device according to claim 9, characterized in that The device further comprises: a training module, configured to train the media resource recommendation model based on at least one of a click tag and a conversion tag, sample data, and sample revenue reference information, wherein the sample data includes sample historical behavior data of sample objects and sample resource data of sample media resources; The training module is used to obtain at least one of the predicted click-through rate and the predicted conversion rate of the sample data and the predicted revenue reference information based on the media resource recommendation model; and train the media resource recommendation model based on at least one of the predicted click-through rate and the predicted conversion rate, at least one of the click label and the conversion label, the predicted revenue reference information and the sample revenue reference information.

12. The device according to claim 11, characterized in that The sample income reference information is obtained by standardizing the sample income data.

13. The device according to claim 9, characterized in that The determination module includes: A first determining submodule, configured to determine a plurality of first media resources from the first candidate resource set based on a first conversion rate of each of the first candidate media resources; a second determining submodule, configured to determine a plurality of second media resources from the first candidate resource set based on the revenue reference information of each of the first candidate media resources; The third determining submodule is configured to determine the media resources to be recommended from a second candidate resource set, where the second candidate resource set includes the plurality of first media resources and the plurality of second media resources.

14. The device according to claim 13, characterized in that The third determining submodule includes: a second acquiring unit, acquiring a second click rate and a second conversion rate of each second candidate media resource based on the second object data of the target object and the second resource data of each second candidate media resource in the second candidate resource set; a determining unit, configured to determine a media resource to be recommended from the second candidate resource set based on the second click-through rate and the second conversion rate of each second candidate media resource; The number of data categories in the second object data is greater than the number of data categories in the first object data, and the number of data categories in the second resource data is greater than the number of data categories in the first resource data.

15. The device according to claim 14, characterized in that The determination unit is used to obtain a correction parameter for each second candidate media resource based on the conversion number of each second candidate media resource, where the conversion number is the number of conversions of the second candidate media resource within a target historical time period, and the correction parameter is used to correct the second conversion rate of the second candidate media resource; based on the second click-through rate, second conversion rate and correction parameter of each second candidate media resource, determine the media resource to be recommended from the second candidate resource set.

16. A device for recommending media resources, characterized in that: The device comprises: A display module is used to display a media resource recommendation page, and display an intelligent recommendation option on the media resource recommendation page, wherein the intelligent recommendation option is used to select whether to use the media resource recommendation model to recommend media resources; a sending module configured to, when the intelligent recommendation option is set to use the media resource recommendation model, send the media resource and a recommendation instruction to a server in response to an upload operation of a media resource on the media resource recommendation page, wherein the recommendation instruction is used to instruct to recommend the media resource based on the media resource recommendation model; The embedding sub-model in the media resource recommendation model is used to obtain, based on first object data of a target object and first resource data of a first candidate media resource in a first candidate resource set, an initial sub-vector of historical behavior data in the first object data, a sub-vector of other category data in the first object data, and multiple sub-vectors of the first resource data, wherein the historical behavior data represents browsing behavior and resource transfer behavior of the target object based on media resources, the other category data includes portrait data and label data, and the label data represents the preferences of the target object; The self-attention sub-model in the media resource recommendation model is used to obtain a sub-vector of the historical behavior data based on the initial sub-vector input of the historical behavior data; The first multi-task sub-model and the second multi-task sub-model in the media resource recommendation model are used to obtain a first object vector and a first resource vector, a second object vector and a second resource vector, and a third object vector and a third resource vector based on the initial object vector and the initial resource vector, the revenue reference space, the click-through rate space, and the conversion rate space; the initial object vector is obtained by concatenating a subvector of the historical behavior data and a subvector of the other category data, and the initial resource vector is obtained by concatenating multiple subvectors of the first resource data; The revenue sub-model in the media resource recommendation model is used to obtain revenue reference information of the first candidate media resource based on the first object vector and the first resource vector, wherein the revenue reference information represents the revenue generated after the target object performs a conversion behavior on the candidate media resource; The click-through rate sub-model in the media resource recommendation model is used to obtain a first click-through rate of the first candidate media resource based on the second object vector and the second resource vector; The conversion rate sub-model in the media resource recommendation model is used to obtain a first conversion rate of the first candidate media resource based on the third object vector and the third resource vector; The resource recommendation model is trained based on historical behavior data of multiple sample objects.

17. A computer device, characterized in that: The computer device includes one or more processors and one or more memories, and at least one computer program is stored in the one or more memories. The at least one computer program is loaded and executed by the one or more processors to implement the operations performed by the media resource recommendation method according to any one of claims 1 to 7, or the operations performed by the media resource recommendation method according to claim 8.

18. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the operations performed by the media resource recommendation method according to any one of claims 1 to 7, or the operations performed by the media resource recommendation method according to claim 8.

19. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the media resource recommendation method according to any one of claims 1 to 7, or the media resource recommendation method according to claim 8.

Images