Data processing method and device, equipment and medium

Abstract

The embodiment of the invention provides a data processing method and device, equipment and a medium, the method can be applied to an activity operation scene, and the method comprises the following steps: obtaining a sample training set and an initial prediction network; the initial prediction model comprises an initial branch network of at least one service; performing feature interaction processing on the first object information and the first service information of the service sample data in the sample training set to obtain a sample information interaction feature; outputting a sample global feature of the sample information interaction feature through an initial shared network, and outputting a sample local feature of the sample information interaction feature through a sample branch network corresponding to a sample service generating service sample data; determining a sample prediction result of the business sample data according to the sample global features and the sample local features; and according to the sample prediction result and the object resource contribution value, training an initial prediction model to obtain a resource prediction model. By implementing the embodiment of the invention, the prediction accuracy of the object resource contribution value can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a data processing method, device, equipment and medium. Background Art

[0002] The object resource contribution value, as an indicator for evaluating the contribution of an object to a certain business, plays a vital role in business delivery scenarios. For example, by analyzing and predicting the object resource contribution value, accurate delivery of various businesses can be achieved.

[0003] At present, machine learning or deep learning methods are usually used to build an estimation model for each business to predict the contribution value of object resources. Since the estimation model of each business is carried out independently, the correlation between objects of different businesses is ignored, resulting in the estimation model's ability to extract and identify object information is very limited, reducing the prediction accuracy of object resource contribution value. Summary of the invention

[0004] The embodiments of the present application provide a data processing method, apparatus, device and medium, which can improve the prediction accuracy of object resource contribution value.

[0005] On the one hand, an embodiment of the present application provides a data processing method, including:

[0006] Acquire a sample training set; the sample training set includes business sample data of at least one business, the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object for the sample business, and the at least one business includes a sample business;

[0007] Acquire an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to at least one service;

[0008] Performing feature interaction processing on the first object information and the first service information to obtain sample information interaction features, and determining a sample branch network corresponding to the sample service in an initial branch network corresponding to at least one service;

[0009] The sample global features of the sample information interaction features are output through the initial shared network, and the sample local features of the sample information interaction features are output through the sample branch network;

[0010] Determine the sample prediction result of the business sample data according to the sample global characteristics and sample local characteristics;

[0011] According to the sample prediction results and the object resource contribution value, the network parameters of the initial prediction model are modified to obtain a resource prediction model; the resource prediction model is used to predict the estimated resource contribution value of the object to be deployed for the business to be deployed.

[0012] An embodiment of the present application provides a data processing device, including:

[0013] A first acquisition module is used to acquire a sample training set; the sample training set includes business sample data of at least one business, the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object for the sample business, and the at least one business includes a sample business;

[0014] A second acquisition module is used to acquire an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to at least one service;

[0015] A first feature interaction module is used to perform feature interaction processing on the first object information and the first service information to obtain a sample information interaction feature, and determine a sample branch network corresponding to the sample service in an initial branch network corresponding to at least one service;

[0016] A first feature extraction module, configured to output sample global features of sample information interaction features through an initial shared network, and output sample local features of sample information interaction features through a sample branch network;

[0017] A first prediction module, used to determine a sample prediction result of the business sample data according to the sample global features and the sample local features;

[0018] The parameter correction module is used to correct the network parameters of the initial prediction model according to the sample prediction results and the object resource contribution value to obtain the resource prediction model; the resource prediction model is used to predict the estimated resource contribution value of the object to be launched for the business to be launched.

[0019] Among them, the initial prediction model also includes an initial representation network;

[0020] The first feature interaction module performs feature interaction processing on the first object information and the first business information to obtain a sample information interaction feature for executing the following steps:

[0021] Converting the first object information into an object embedding vector, converting the first business information into a business embedding vector, and inputting the object embedding vector and the business embedding vector into an initial representation network;

[0022] The object implicit features of the object embedding vector and the business implicit features of the business embedding vector are obtained through the initial representation network, and the object implicit features and the business implicit features are subjected to feature interaction processing to obtain sample information interaction features.

[0023] The initial representation network includes an object representation network and a service representation network, the object representation network includes N chain-connected first hidden layers, and the service representation network includes D chain-connected second hidden layers, where N and D are positive integers;

[0024] The first feature interaction module obtains the object implicit features of the object embedding vector and the business implicit features of the business embedding vector through the initial representation network, performs feature interaction processing on the object implicit features and the business implicit features, and obtains sample information interaction features for executing the following steps:

[0025] Extracting features of the service embedding vector through D second hidden layers in the service characterization network to obtain output data of the D second hidden layers, and determining service hidden features in the output data of the D second hidden layers;

[0026] Determine a candidate network layer associated with the service implicit feature among the N first hidden layers included in the object representation network;

[0027] Extract features from the object embedding vector through the object representation network, obtain output data of the candidate network layer, and determine the output data of the candidate network layer as the object implicit features;

[0028] Performing feature interaction processing on the service implicit features and the object implicit features to obtain implicit interaction features; the implicit interaction features are used as input data of the next first implicit layer of the candidate network layer in the object representation network;

[0029] If the candidate network layer is the Nth first hidden layer in the object representation network, the implicit interaction feature is determined as the sample information interaction feature.

[0030] Among them, the sample branch network includes a branch-specific network;

[0031] The first feature extraction module outputs the sample global features of the sample information interaction features through the initial shared network, and outputs the sample local features of the sample information interaction features through the sample branch network, for performing the following steps:

[0032] Inputting sample information interaction features into the initial shared network and the branch dedicated network;

[0033] Through the initial shared network, feature extraction is performed on the sample information interaction features to obtain the sample global features;

[0034] Through the branch-specific network, the interactive features of sample information are extracted to obtain the local features of the sample.

[0035] Among them, the sample branch network also includes a branch gating network;

[0036] The data device also includes:

[0037] The gating module is used to input the sample information interaction feature into the branch gating network, and analyze the sample information interaction feature through the branch gating network to obtain the sample weight feature; the sample weight feature includes a first weight corresponding to the sample global feature and a second weight corresponding to the layout information feature;

[0038] The first prediction module determines the sample prediction result of the business sample data according to the sample global features and the sample local features, and is used to perform the following steps:

[0039] Obtaining a first weight and a second weight in the sample weight feature, merging the product of the first weight and the sample global feature, and the product of the second weight and the sample local feature, to obtain a sample fusion feature;

[0040] Normalize the sample fusion features to obtain the first normalized features corresponding to the business sample data;

[0041] The first normalized feature is classified and processed to obtain a sample prediction result of the business sample data.

[0042] The first prediction module normalizes the sample fusion features to obtain the first normalized features corresponding to the business sample data, which are used to perform the following steps:

[0043] Determine the sample training subset to which the business sample data belongs, determine the sample training group corresponding to the sample business in the sample training subset, and obtain the sample mean and sample variance corresponding to the sample training group; the sample training subset belongs to the sample training set;

[0044] The sample fusion features are normalized according to the sample mean and sample variance to obtain the first normalized features corresponding to the business sample data.

[0045] The initial estimation model further includes a first task output network and a second task output network, and the sample prediction result includes a first task estimation probability and a second task estimation probability;

[0046] The first prediction module performs classification processing on the first normalized feature to obtain a sample prediction result of the business sample data, which is used to perform the following steps:

[0047] Inputting the first normalized feature into the first task output network and the second task output network;

[0048] Through the first task output network, the first normalized feature is subjected to binary classification processing to obtain the first task estimated probability of the business sample data;

[0049] Through the second task output network, multi-classification processing is performed on the first normalized feature to obtain the second task estimated probability of the business sample data.

[0050] The parameter correction module corrects the network parameters of the initial prediction model according to the sample prediction results and the object resource contribution value to obtain a resource prediction model for performing the following steps:

[0051] Determine the first label information and the second label information of the service sample data according to the object resource contribution value; the first label information is used to indicate whether the sample object has made resource contribution to the sample service, and the second label information is used to indicate the resource contribution scope of the sample object to the sample service;

[0052] Determine the first task classification loss according to the first label information and the first task estimated probability, and determine the second task classification loss according to the second label information and the second task estimated probability;

[0053] Linearly weight the first task classification loss and the second task classification loss to obtain the model loss of the initial prediction model;

[0054] The network parameters of the initial prediction model are modified according to the model loss to obtain the resource prediction model.

[0055] The resource prediction model includes an information representation network, a service sharing network, and a service branch network corresponding to at least one service;

[0056] The device also includes:

[0057] A third acquisition module is used to acquire second object information of the candidate object and acquire second business information of the business to be launched;

[0058] A second feature interaction module is used to perform feature interaction processing on the second object information and the second service information through the information representation network to obtain information interaction features, and determine a candidate branch network corresponding to the service to be placed in the service branch network corresponding to at least one service;

[0059] A second feature extraction module is used to output global information features of information interaction features through the business sharing network, and output local information features of information interaction features through the candidate branch network;

[0060] The second prediction module is used to determine the estimated resource contribution of the candidate object to the business to be launched based on the global information characteristics and the local information characteristics.

[0061] The second prediction module determines the estimated resource contribution of the candidate object to the service to be launched based on the global information features and the local information features, and is used to perform the following steps:

[0062] Obtaining business weight features; the business weight features are output by the business gating network in the candidate branch network;

[0063] According to the service weight characteristics, the global information characteristics and the local information characteristics are weighted to obtain the service fusion characteristics;

[0064] The service fusion feature is normalized to obtain a second normalized feature, and the second normalized feature is subjected to binary classification to obtain a third task estimated probability;

[0065] The second normalized feature is subjected to multi-classification processing to obtain the fourth task estimated probability, and the resource contribution estimated value of the candidate object for the to-be-delivered business is determined according to the third task estimated probability and the fourth task estimated probability.

[0066] The fourth task estimated probability includes probability estimated values ​​corresponding to M resource contribution ranges, where M is an integer greater than 1;

[0067] The second prediction module determines the estimated resource contribution value of the candidate object for the service to be launched according to the estimated probability of the third task and the estimated probability of the fourth task, and is used to perform the following steps:

[0068] The product of the probability estimate corresponding to the k-th resource contribution range and the k-th interval resource contribution value is determined as the k-th candidate resource contribution value, where k is a positive integer less than or equal to M;

[0069] Obtain M candidate resource contribution values, and accumulate the M candidate resource contribution values ​​to obtain a cumulative resource contribution value;

[0070] The product of the accumulated resource contribution value and the estimated probability of the third task is determined as the estimated resource contribution value of the candidate object for the business to be launched;

[0071] Among them, the kth interval resource contribution value refers to the interval resource contribution value between the kth resource contribution range and the k-1th resource contribution range; the kth resource contribution range and the k-1th resource contribution range are two adjacent resource contribution ranges in the M resource contribution ranges.

[0072] The data processing device further comprises:

[0073] A delivery module, used for determining the candidate object as a hit object when the estimated resource contribution value is greater than the resource contribution threshold, generating business display information corresponding to the business to be delivered, and sending the business display information to the hit object;

[0074] A sample update module is used to obtain business interaction data of the hit object for business display information;

[0075] The sample updating module is also used to update the sample training set according to the business interaction data to obtain an updated sample training set; the updated sample training set is used to train the resource prediction model.

[0076] On one hand, an embodiment of the present application provides a computer device, including a memory and a processor, wherein the memory is connected to the processor, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method provided in the above aspect of the embodiment of the present application.

[0077] On one hand, an embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored. The computer program is suitable for being loaded and executed by a processor, so that a computer device with a processor executes the method provided in the above aspect of the embodiment of the present application.

[0078] According to one aspect of the present application, a computer program product is provided, which may include a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in the above aspect.

[0079] In an embodiment of the present application, a sample training set consisting of business sample data of at least one business is obtained, and one business sample data includes the first object information of the sample object, the first business information of the sample business, and the object resource contribution value of the sample object to the sample business, and the sample business is any one of the at least one business. An initial prediction model is obtained, and the initial prediction model may include an initial shared network and an initial branch network corresponding to at least one business, and the input data of the initial prediction model includes the first object information and the first business information. In the initial prediction model, feature interaction processing is performed on the first object information and the first business information to obtain sample information interaction features; through the initial shared network and the initial branch network of at least one business, the information of different businesses and the information of different objects are learned in the sample information interaction features of each business sample data, so as to improve the prediction accuracy of the object resource contribution value. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 paying creative work.

[0081] Figure 1It is a structural diagram of a network architecture provided by an embodiment of the present application;

[0082] Figure 2 It is a logic diagram of a hardware environment for implementing a service delivery provided in an embodiment of the present application;

[0083] Figure 3 This is a schematic diagram of an application scenario of a resource prediction model provided in an embodiment of the present application;

[0084] Figure 4 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 1 ;

[0085] Figure 5 is a schematic diagram of a network structure of an initial prediction model provided in an embodiment of the present application;

[0086] Figure 6 This is a schematic diagram of feature interaction processing between first object information and first service information provided in an embodiment of the present application;

[0087] Figure 7 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 2 ;

[0088] Figure 8 This is a schematic diagram of the overall process of a service delivery provided by an embodiment of the present application;

[0089] Fig. 9 is a structural schematic diagram of a data processing device provided in an embodiment of the present application;

[0090] Fig.10 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0091] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0092] To facilitate understanding, the basic concepts involved in the embodiments of the present application are described below.

[0093] Advertising platform: refers to a one-stop advertising platform that can cover the entire advertising process, including creative production of creative materials, advertising, and performance indicator reports in certain business areas, to meet the business party's demand for accurate advertising. Advertising can be considered as a form or carrier of business promotion information, such as game promotion information, product promotion information, concept promotion information, etc.

[0094] Object resource contribution value: It can be called user value (Lifetime Value, LTV). Object resource contribution value can refer to the total value created by a single object for a business or product within a period (for example, a month). In the advertising scenario, object resource contribution value can refer to the consumption value of users attracted by advertising within a period.

[0095] See also Figure 1 , Figure 1 1 is a schematic diagram of a network architecture provided by an embodiment of the present application; the network architecture may include a server 10d and a terminal cluster, the terminal cluster may include one or more terminal devices, and the number of terminal devices included in the terminal cluster and the number of servers included in the network architecture are not limited. Figure 1 As shown, the terminal cluster may specifically include terminal device 10a, terminal device 10b, and terminal device 10c, etc.; all terminal devices in the terminal cluster (for example, including terminal device 10a, terminal device 10b, and terminal device 10c, etc.) may be connected to the server 10d through a network connection, so that each terminal device may exchange data with the server 10d through the network connection.

[0096] in, Figure 1 The terminal devices in the terminal cluster shown may include but are not limited to: smart phones, tablet computers, laptops, PDAs, desktop computers, wearable devices (such as smart watches, smart bracelets, etc.), smart voice interaction devices, smart home appliances (such as smart TVs, etc.), vehicle-mounted devices, aircraft and other electronic devices. This application does not limit the type of terminal devices.

[0097] Figure 1 The server 10d shown can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms. This application does not limit the type of server.

[0098] In the embodiment of the present application, Figure 1 Each terminal device in the terminal cluster shown can run a business platform or a business client. The number of business platforms or business clients running in each terminal device can be one or more; different terminal devices can run the same business platform or business client, or different business platforms or business clients, and this application does not limit this. Among them, the business platform can refer to a comprehensive basic environment or system, and the business platform can be used to carry and run various applications; for example, advertising platforms, game platforms, shopping platforms, etc. Business clients are usually front-end programs developed for specific applications or services, and usually need to be downloaded, installed and updated to adapt to the needs and changes of specific services; for example, game clients, shopping clients, video clients, information clients, etc. This application does not limit the type of business platform or business client running in the terminal device.

[0099] Figure 1 The server 10d shown can be a background server of a service platform or service client. Different service platforms / service clients can correspond to different background servers, or can correspond to the same background server, which is not limited in this application. The service platform / service client running in each terminal device can exchange data with its corresponding background server through the terminal device. Optionally, data can also be exchanged between different service platforms, between different service clients, and between service platforms and service clients, which is not limited in this application.

[0100] In the present application embodiment, by Figure 1 The data interaction between the business platforms or business clients running in each terminal device shown, as well as the data interaction between the business platform or business client and its corresponding background server (for example, server 10d), can realize the delivery of certain specific businesses; for example, game delivery, service delivery, commodity delivery, work delivery (film and television works, music works, cultural products, etc.), scenic spot delivery, educational course delivery, etc. This application does not limit the type of specific business delivered.

[0101] The following uses the advertising delivery scenario as an example to describe the implementation logic of business delivery. Figure 2 , Figure 2 This is a logic diagram of a hardware environment for implementing a service delivery provided by an embodiment of the present application; Figure 2 As shown, the hardware environment implementation logic of business delivery may involve data interaction between business client A, media background service 20a of business client A, business client B, advertisement delivery platform 20b, business platform 20c and resource prediction model 20d.

[0102] Among them, business client A can refer to any client in the terminal device that is used to display the business promotion information of the business to be launched (for example, business B). The media background service 20a of business client A can refer to a collection of a series of technical architectures, systems and functional components that support the operation and function realization of business client A. Business client B can refer to a business application installed in the terminal device, which can be an interface for users to interact with business B; for example, when business B is a game, business client B can refer to the game client of the game, and business client B is the interface for game players to interact with the game. The advertising delivery platform 20b can be a platform that specializes in providing promotion and publicity services for a specific business. The business platform 20c can be a comprehensive basic environment or system for carrying and running various business applications.

[0103] The resource prediction model 20d can be used to predict the estimated resource contribution of any object in one or more businesses; the estimated resource contribution value can refer to the consumption value that the object may generate for a certain business in a future period (for example, 7 days), such as the object resource contribution value estimated by the resource object model. The resource prediction model 20d can be considered as a trained object resource contribution value estimation model, and the resource training model 20d can be trained using sample data of multiple businesses (for example, object information in multiple businesses and business information of multiple businesses); wherein, the training process of the resource training model 20d will be described in the subsequent content.

[0104] For any business that needs to be promoted (for example, business B), a set of objects that may be interested in business B can be selected, and the resource contribution estimate of each object in the object set can be predicted through the resource prediction model 20d. All objects in the object set are sorted in descending order of the estimated resource contribution, and the business advertisement of business B can be placed on the first Y (Y is a positive integer, which can be flexibly set according to the actual application scenario) objects after sorting, so that the precise placement of business advertisements can be achieved, the business promotion cost can be reduced, and the advertising effect of the business can be improved.

[0105] The service client A can exchange data with the media background service 20a of the service client A; for example, the service client A can send an advertisement display data request to the media background service 20a, and after receiving the advertisement display data request from the service client A, the media background service 20a can return the service advertisement to the service client A. Information such as the exposure and click volume of the service advertisement in the service client A can be reported to the media background service 20a. The exposure volume can refer to the number of times the service advertisement is displayed in the service client A; the click volume can refer to the number of times the service advertisement displayed in the service client A is clicked.

[0106] When the object u1 clicks on the service advertisement displayed in the service client A, it can jump from the service client A to the service client B indicated by the service advertisement; the service client B and the service client A can be different clients installed in the terminal device, and the service client A can be any other client except the service client B. The service client B can interact with the service platform 20c for data. For example, the service client B can send login information (for example, account number, password, etc.) to the service platform 20c, and the service platform 20c can return the login result according to the login information. For example, when the login information is verified to be correct, the returned login result is a successful login, and when the login information is verified to be incorrect, the returned login result is a failed login. After the service client B successfully logs in, it can request the service platform 20c to obtain the basic information of the service B. Such as version, update content, server status, etc. The service client B can also request the service platform 20c to obtain the relevant data of the object u1 in the service B, such as level, experience, etc. The service data (for example, task completion status, etc.) generated by the object u1 in the service B can be reported to the service platform 20c in real time.

[0107] The media backend service 20a and the advertising delivery platform 20b can exchange data. The advertising delivery platform 20b can calculate which business to deliver. In the process of calculating which business to deliver, the advertising delivery platform 20b can request the resource prediction model 20d to calculate the delivery value of each business, or determine whether the object is a high-value object. The advertising delivery platform 20b can return the business active objects and consumption information after the advertisement delivery to the resource prediction model 20d. For example, assuming that the resource prediction model 20d determines that the object u1 is a high-value object, the advertising delivery platform 20b can send the business advertisement of business B to the media backend service 20a, and the media backend service 20a can deliver the business advertisement to the business client A of the object u1 for display.

[0108] Among them, high-value objects can refer to objects whose estimated resource contribution is higher than the pre-set resource contribution threshold, and the set of high-value objects can be called a high-contribution object set; the resource contribution threshold can be flexibly set according to the specific needs of the actual application, and this application does not limit this. Business active objects can refer to objects that have logged into the business client and experienced business in the recent period of time (for example, 7 days, half a month); consumption information can be used to characterize the consumption of an object for a certain business, such as whether resource consumption has occurred, the amount of resources consumed, etc.

[0109] The advertising delivery platform 20b and the business platform 20c can interact with each other. The advertising delivery platform 20b can upload interactive data such as advertisement clicks and downloads of the business client indicated by the advertisement to the business platform 20c. The business platform 20c can attribute advertisement conversions based on the interactive data uploaded by the advertising delivery platform 20b, and transmit the business active objects and consumption information after conversion attribution to the advertising delivery platform 20b. Among them, advertisement conversion attribution can refer to the process of which channels lead to the specific operation or conversion of the object for a certain business (such as clicking on a business advertisement, installing a business client, first registration, resource consumption, etc.). For example, when the object u1 installs the business client B indicated by the business advertisement on the terminal device used by the object u1 by clicking on the business advertisement, the advertisement conversion attribution for the object u1 can refer to: the process of determining the channel through which the object u1 installs the business client B. It can be understood that the advertising delivery platform 20b can deliver the same business advertisement to multiple business clients of object u1. Then all business clients that display the business advertisement (for example, business client A) can be used as channels for installing business client B. When object u1 downloads and installs business client B by clicking on the business advertisement displayed in business client A, then the business client B can be used as the result of advertising conversion attribution.

[0110] It is understandable that Figure 2 The data involved in all data interaction processes in the corresponding embodiments are subject to the permission or consent of relevant institutions or departments, or the users themselves, and are carried out in compliance with the requirements of relevant laws and regulations on Internet user data security.

[0111] See also Figure 3 , Figure 3 2 is a schematic diagram of an application scenario of a resource prediction model provided in an embodiment of the present application. The present embodiment of the present application takes the application of resource prediction model 20d in the advertising delivery scenario of business 1 as an example for description. Figure 3 As shown, for any business that needs to be promoted, such as business 1, a candidate object set 30a corresponding to the business 1 can be determined. The candidate objects in the candidate object set 30a may refer to objects that may be of interest to business 1 and are selected using some specific screening methods; that is, all candidate objects in the candidate object set 30a may be considered as objects that may be of interest to business 1. Through the resource prediction model 20d, the estimated resource contribution of each candidate object in the candidate object set 30a in business 1 can be predicted.

[0112] The following description takes any candidate object (e.g., candidate object u2) in the candidate object set 30a as an example. The object information of the candidate object u2 and the business information of business 1 can be obtained; wherein the object information can be considered as the common information of the candidate object u2 in each business, such as basic information of the object, active days of the object, resource consumption amount, resource consumption days, etc.; the business information can include but is not limited to business id (Identity document, identification number), business type (e.g., product types such as electronic products, clothing, food, etc., game types such as shooting, role-playing, card games, etc.), etc.

[0113] The object information of the candidate object u2 and the business information of business 1 are input into the resource prediction model 20d. Through the resource prediction model 20d, the estimated resource contribution of the candidate object u2 for business 1 can be calculated. When the estimated resource contribution of the candidate object u2 is less than or equal to the resource contribution threshold, the candidate object u2 can be added to the low contribution object set 30b. When the estimated resource contribution of the candidate object u2 is greater than the resource contribution threshold, the candidate object u2 can be added to the high contribution object set 30c. Among them, the low contribution object set 30b can refer to a set composed of low-value objects, and the low-value object can refer to an object whose estimated resource contribution is less than or equal to the resource contribution threshold. The calculation process of the estimated resource contribution will be described in the subsequent content.

[0114] After selecting a high-contribution object set 30c from the candidate object set 30a based on the estimated resource contribution of each candidate object, business 1 can be delivered, such as generating business promotion information 30f for business 1, and delivering the business promotion information 30f to the terminal device 30d used by each object in the high-contribution object set 30c, wherein the business promotion information 30f may include a download control of the client corresponding to business 1 (for example, a "download now" control), and welfare description information of business 1 (for example, what welfare can be collected, the time to collect the welfare, etc.). The information content contained in the business promotion information 30f may be designed by the advertising delivery platform for business 1, and the present application does not limit the information content contained in the business promotion information 30f.

[0115] For example, assuming that the high contribution object set 30c includes object u3, the business promotion information 30f can be delivered to the business client C installed in the terminal device 30d, and the business promotion information 30f can be displayed in the interface 30e of the business client C. Among them, the business promotion information 30f can be directly displayed in the main page of the business client C, or in the label page corresponding to the "welfare" label of the main page (such as interface 30e). This application does not limit the display form and display position of the business promotion information 30f in the business client C. The business client C can be any client other than the client of business 1, and this application does not limit the type of business client C to which the business promotion information 30f is delivered.

[0116] Figure 3 The resource prediction model 20d in the corresponding embodiment is obtained by training with the object information and business information of multiple services, and can be used to predict resource contribution estimates for candidate objects of multiple services, which can solve the problem of large differences in object distribution in multiple services; by learning the object feature distribution of different services, the accuracy of resource contribution estimates can be improved. Based on the resource contribution estimates, the set of delivery objects for each service to be delivered is determined, that is, Figure 3 The high contribution object set 30c shown can improve the accuracy of selecting the set of delivery objects. The objects in the delivery object set are more likely to consume resources in the future, thereby improving the effect of business delivery; for example, the effect of business delivery can be expressed in the form of return on investment (ROI), which refers to the percentage between the income and cost of business delivery, and can be used as an important measure of the effect of business delivery.

[0117] It can be understood that, for ease of understanding, the object resource contribution value estimation model that has completed training can be called a resource prediction model; the object resource contribution value estimation model that has not completed training can be called an initial prediction model, that is, the object resource contribution value estimation model before and during training can be called an initial prediction model.

[0118] See also Figure 4 , Figure 4 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 1 It can be understood that the data processing method can be executed by a computer device, which can be a terminal device, such as Figure 1 Any terminal device in the terminal cluster shown, or a server, such as Figure 1 The server 10d shown in the figure is not limited in this application. Figure 4 As shown, the data processing method may include the following steps S101 to S106:

[0119] Step S101, obtaining a sample training set; the sample training set includes business sample data of at least one business, and the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object to the sample business.

[0120] Among them, the sample training set may refer to a data set used to train the network model, and the sample training set may include business sample data collected after at least one business is actually launched. For each business, each business can collect a batch of business sample data, and each business sample data may include the first object information of a sample object, the first business information of a sample business, and the object resource contribution value generated by the sample object for the sample business. The sample business may refer to any business of at least one business that is actually launched to the sample object; the sample object may refer to an object that actually launched any business of at least one business and consumed resources. The object resource contribution value may refer to the amount of resource consumption generated by the sample object for the sample business after the sample business is actually launched to the sample object.

[0121] At least one business may refer to advertising content that needs to be delivered or has actually been delivered. At least one business may include different types of games, different games of the same type, different types of services, different types of products, different types of works, etc. This application does not limit the number and type of at least one business. Different types of games may include, but are not limited to: role-playing games, adventure games, strategy games, shooting games, simulation business games, racing games, card games, music games, casual games, etc. Different types of products may include, but are not limited to: electronic products, clothing products, building materials products, food, household products, etc. Different types of works may include, but are not limited to: film and television works, music works, calligraphy and painting works, book works, training courses, etc.

[0122] The first object information may include basic object information of the sample object and interaction data between the sample object and each service. The basic object information may include information such as the attributes and life span of the sample object, and the interaction data may include information such as the active duration, resource consumption duration, and resource consumption amount of the sample object for each service. The first service information may include information such as the service ID, service type, and average resource contribution value of the sample service.

[0123] Among them, the business id can refer to the information used to uniquely identify the sample business (for example, a number), and the business type can indicate the type of the sample business, such as game type (such as the above-mentioned shooting games, racing games, card games, etc.), product type (such as the above-mentioned electronic products, clothing products, etc.), work type (such as the above-mentioned film and television works, music works, etc.), service type, etc.

[0124] The average resource contribution value may refer to the average value of all object resource contribution values ​​of the sample business. For example, the sample objects that generate object resource contribution values ​​for the sample business include object 1, object 2, and object 3. The object resource contribution value generated by object 1 for the sample business is a1, the object resource contribution value generated by object 2 for the sample business is a2, and the object resource contribution value generated by object 3 for the sample business is a3. Then, the average resource contribution value of the sample objects can be determined to be: (a1+a2+a3) / 3.

[0125] In the business delivery scenario, the sample training set may include business sample data of at least one business, and the first object information in the business sample data may include object basic information, object business cross data, object business sequence data, etc. Among them, the object business cross data may include but is not limited to information such as the active duration (days) of the sample object in one or more businesses, the resource consumption duration (days), and the resource consumption amount. The object business sequence data may include but is not limited to information such as business active sequence and business consumption sequence; business active sequence data may refer to a sequence in which the activity level of the sample object in participating in one or more businesses is arranged in chronological order within a period of time, which can be used to reflect the activity of the sample object in different time periods, such as the active duration of each day, week, and month; the business consumption sequence may refer to a series of data formed in chronological order when the sample object consumes resources for one or more businesses, which can be used to present the frequency, amount, time interval, and other characteristics of the sample object's resource consumption.

[0126] In one or more embodiments, during the model training phase, if some businesses do not have business sample data recovered from actual launches, then it is necessary to collect business sample data for them. The method of collecting business sample data may include, but is not limited to: a cold start object selection method may be adopted to select a set of trial launch objects for each business, and a business trial launch may be performed on each object in the trial launch object set; after the business trial launch, the object consumption samples may be recovered, and the recovered object consumption samples may be used as business sample data and added to the sample training set. If at least one business is a business that has been actually launched, the recovered historical object consumption samples may be used as business sample data and added to the sample training set.

[0127] Among them, the trial launch of a service may refer to a small-scale test launch before the formal large-scale launch of the service. The object consumption sample may refer to the relevant data of the object that has generated resource consumption for the launched service after the service is launched, such as the object basic information of the object, the resource consumption duration of the object, the resource consumption amount, the active duration, the service active sequence, the service consumption sequence, etc. The historical object consumption sample may refer to the relevant data of the object that has generated resource consumption for the service in the past period of time (for example, the past 7 days, the past month, etc.).

[0128] The methods of selecting objects for cold start may include, but are not limited to: random selection, policy-based RFM (Recency, Frequency and Monetary, which are respectively expressed as the time from the last resource consumption to the present, the number of resource consumptions, and the amount of resource consumption) resource contribution value segmentation, etc. Among them, random selection may refer to randomly selecting some objects from the candidate object set of the business (the set of all objects that may be interested in the business) as a trial launch object set, and the trial launch object set may be considered as the initial launch object set of the business. Policy-based RFM resource contribution value segmentation may refer to estimating the object resource contribution value based on the RFM of each object in the candidate object set, segmenting the object resource contribution value of each object, and using the object corresponding to the segment with the highest object resource contribution value as the trial launch object set.

[0129] In the strategy-based RFM resource contribution value segmentation method, it can be considered that the closer the time of the most recent resource consumption is to the present, the more times the resource has been consumed in the past, and the larger the amount of resource consumption, the greater the object resource contribution value. Therefore, the objects in the candidate object set can be divided according to the RFM of each object in the candidate object set, and objects belonging to the same object resource contribution value segment can be divided into the same class of objects; whichever segment has the highest object resource contribution value, the class of objects corresponding to the segment can be used as the trial launch object set.

[0130] For example, assuming that the candidate object set includes object 1, object 2, object 3, object 4, and object 5, based on the RFM of each object, the object resource contribution values ​​of object 1, object 2, and object 5 are divided into segment d1, and the object resource contribution values ​​of object 3 and object 4 are divided into segment d2. When the object resource contribution value of segment d1 is greater than the object resource contribution value of segment d2, objects 1, 2, and 5 corresponding to segment d1 can be used as the trial launch object set. By using the strategy-based RFM resource contribution value segmentation method in the trial launch stage, the accuracy of the trial launch object set can be improved, and thus the launch effect in the trial launch stage can be improved.

[0131] Step S102, obtaining an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to at least one service.

[0132] Among them, the initial prediction model can refer to the object resource contribution value estimation model to be trained. The training goal of the initial prediction model can be to learn the distribution of resource contribution values ​​of different objects in different businesses, which is used to predict the resource contribution estimation value of objects in different businesses (such as games).

[0133] See also Figure 5 , Figure 5 Schematic diagram of the network structure of an initial prediction model provided in the embodiment of the present application. Figure 5 As shown, the initial prediction model may include but is not limited to embedding layer 1, embedding layer 2, object representation network, service representation network, initial shared network, at least one initial branch network corresponding to a service, service normalization network layer, first task output network, second task output network, etc. Among them, the object representation network and the service representation network may be referred to as the initial representation network in the initial prediction model.

[0134] Among them, the embedding layer 1 can be used to learn the feature representation of the first object information, such as converting the discrete categories or symbolic data in the first object information into a continuous vector representation, and reflecting the semantic similarity or correlation between different elements in the first object information in the vector space. The embedding layer 2 can be used to learn the feature representation of the first business information, such as converting the discrete categories or symbolic data in the first business information into a continuous vector representation, and reflecting the semantic similarity or correlation between different elements in the first business information in the vector space. In the model training stage, the embedding layer 1 and the embedding layer 2 can be initialized with a specific distribution (for example, uniform distribution, normal distribution, etc.), and use it as a learnable parameter, which can be learned during the model training process. For example, the first object information can be initialized by uniform distribution, normal distribution, etc. to obtain an object initialization vector; similarly, the first business information can be initialized by uniform distribution, normal distribution, etc. to obtain a business initialization vector.

[0135] Embedding layer 1 and embedding layer 2 can be a layer of network structure in the initial prediction model, such as a fully connected layer, a convolutional layer, etc. Optionally, embedding layer 1 and embedding layer 2 can also be a multi-layer network structure in the initial prediction model, such as a convolutional neural network, a deep learning model, transformers (a deep learning model using an attention mechanism), etc. Embedding layer 1 and embedding layer 2 can have the same network structure or different network structures. This application does not limit the specific structure of embedding layer 1 and embedding layer 2.

[0136] In one or more embodiments, embedding layer 1 and embedding layer 2 can be initialized using structures such as Word2Vec (a word embedding model), BERT (Bidirectional Encoder Representation from Transformers, a pre-trained model for bidirectional encoding), and GloVe (global word frequency statistics). Structures such as Word2Vec, BERT, and GloVe usually require separate model training. After the training is completed, they can be used as the initialization of embedding layer 1 and embedding layer 2. At this time, the initialization parameters of embedding layer 1 and embedding layer 2 can be set to non-learnable parameters.

[0137] In the model training stage, whether embedding layer 1 and embedding layer 2 in the initial prediction network are set as learnable parameters depends on the initialization method of embedding layer 1 and embedding layer 2. For ease of understanding, the embodiment of the present application is described by taking embedding layer 1 and embedding layer 2 as learnable parameters as an example.

[0138] The object representation network can be used to extract and identify features of the first object information; the business representation network can be used to extract and identify features of the first business information. Both the object representation network and the business representation network can be represented by a deep learning neural network (DNN), such as the object representation network and the business representation network can be a multi-layer perceptron (MLP), or can be other deep learning model structures. This application does not limit the structure of the object representation network and the business representation network.

[0139] The initial shared network can also be called a shared expert network. The initial shared network can be used to extract global information from the business sample data of all businesses. Figure 5 As shown, the initial branch network corresponding to at least one service may include an initial branch network 40a corresponding to service 1, an initial branch network 40b corresponding to service 2, and an initial branch network 40c corresponding to service 3, etc. One service corresponds to an initial branch network, for example, one game corresponds to an initial branch network, and the initial branch network of each service may include a branch-specific network and a branch gated network (Gate Control network). For example, the initial branch network 40a may include a branch-specific network 1 and a branch gated network 1, the initial branch network 40b may include a branch-specific network 2 and a branch gated network 2, and the initial branch network 40c may include a branch-specific network 3 and a branch gated network 3, etc.

[0140] The branch-specific network of each service can be used to extract local information from the service sample data of the service; that is, the branch-specific network corresponding to each service is only used to process the service sample data of the service, and is a network structure independently constructed for a service. The output results of the initial shared network and the output results of the branch-specific network of each service can be weighted by the branch gating network of the service to calculate the fusion vector output of each service. The input data of the initial shared network, the branch-specific network corresponding to each service, and the branch gating network can all be the output results of the information representation network.

[0141] Among them, the initial shared network, the branch-specific network of each business, and the branch gating network can all adopt a multi-layer perceptron (MLP) structure, or other deep learning network models can be adopted, which is not limited in this application. It should be noted that compared with the initial shared network, the branch-specific network of each business can adopt a Bool Mask structure to perform forward calculations on each business sample data. For example, if the currently calculated business sample data data1 is the sample data of business 1, then the Bool mask structure in the branch-specific network 1 of business 1 can output 1, and the Bool mask structure in the remaining branch-specific networks can output 0; that is, in the initial shared network and the initial branch networks of all businesses contained in the initial prediction model, in essence only the initial shared network, and the branch-specific network 1 and the branch gating network 1 of business 1 perform forward calculations on the business sample data data1.

[0142] The business normalization network layer can be used to normalize the business sample data of each business separately, capture the differences in object distribution of different businesses, and thus improve the prediction accuracy of object resource contribution value. For example, in the game delivery scenario, the business normalization network layer can be called the game partition normalization layer (Game Partitioned Normalization Layer, GPN), through which the differences in object distribution of different games can be captured.

[0143] The initial prediction model may include two task output networks, respectively referred to as the first task output network and the second task output network. In an embodiment of the present application, the estimation of the object resource contribution value may be converted into two tasks: the first task may be a binary classification task of whether resource consumption has occurred; the second task may be a multi-classification task of the amount of resource consumption. The first task output network may be used to output the estimated probability of the first task (which may be referred to as the first task estimated probability), and the second task output network may be used to output the estimated probability of the second task (which may be referred to as the second task estimated probability).

[0144] The first task output network can adopt an MLP network structure carrying an activation function, or can adopt other network structures, which is not limited in this application. Among them, the MLP network structure may include an input layer, a hidden layer, and an output layer; the input layer can be used to receive the output data of the business normalization layer, the output layer can output the corresponding result according to the requirements of the first task, the hidden layer can be used to process the data transmitted by the input layer, and each hidden layer can carry an activation function. For example, when the first task output network contains only one hidden layer, the activation function carried by the hidden layer can be a sigmoid function. When the first task output network contains multiple hidden layers, the activation function carried by the last hidden layer can be a sigmoid function, and the activation functions carried by other hidden layers except the last hidden layer can be any one of sigmoid, tanh, relu and other functions.

[0145] The second task output network can be converted into multiple binary classification tasks, and an ordered classification network structure can be used to output the results of multiple binary classification tasks; the second task output network can also use an MLP network structure carrying an activation function (for example, a sigmoid function), or can use other network structures, which are not limited in this application. The ordered classification network structure helps the initial prediction model learn the orderliness of the resource consumption amount and the cumulativeness of the resource consumption amount. This application does not limit the form of the ordered classification network structure.

[0146] The following takes a business sample data in the sample training set as an example to describe the forward calculation process of the business sample data in the initial prediction model.

[0147] Step S103, performing feature interaction processing on the first object information and the first service information to obtain sample information interaction features, and determining a sample branch network corresponding to the sample service in an initial branch network corresponding to at least one service.

[0148] The sample information interaction feature may refer to a fusion feature that fuses the first object information and the first service information. The sample branch network may refer to an initial branch network corresponding to the currently calculated service sample data. For example, Figure 5 As shown, when the currently calculated service sample data is service sample data of service 1 (sample service), the sample branch network may be the initial branch network 40a corresponding to service 1.

[0149] Specifically, the first object information can be converted into an object embedding vector according to the embedding layer 1 in the initial prediction model; and the first business information can be converted into a business embedding vector according to the embedding layer 2 in the initial prediction model. The first object information and the first business information can include numerical data (e.g., resource consumption amount, years, average resource contribution value, etc.), discrete data (e.g., object attributes), sequence data (e.g., business activity sequence, business consumption sequence), text data, and other information.

[0150] For the numerical data in the first object information and the first business information, maximum and minimum normalization processing can be performed to map the numerical data to the interval [0,1]. In one or more embodiments, the minimum and maximum values ​​of the same type of numerical data in all business sample data of the sample training set can be obtained; the current numerical data in the first object information is subtracted from the minimum value, and then divided by the difference between the maximum and minimum values ​​to obtain the normalized numerical feature. For example, the minimum value min and the maximum value max of the resource consumption amount in all business sample data can be obtained, and the result of normalization of the resource consumption amount m1 in the first object information can be expressed as: (m1-min) / (max-min).

[0151] For the discrete data in the first object information, one-hot encoding can be used to obtain the one-hot code of the discrete data, and then the one-hot code can be mapped to the initialization vector of the discrete data; the one-hot code can refer to a vector that contains only one value 1 and the rest are 0. For the sequence data in the first business information, after converting it into a vector sequence, the attention mechanism can be used to weight the elements in the vector sequence to obtain an initialization vector sequence. Among them, the attention mechanism here can be a fully connected network layer, or it can be other network structures, and the present application does not limit the structure of the attention mechanism. For the text data in the first object information (or the first business information), the text initialization vector of the text data can be obtained in the same way as the discrete data.

[0152] After obtaining the initialization vectors of the numerical data, discrete data, sequence data, and text data in the first object information, the normalized numerical features, the initialization vectors corresponding to the discrete data, the initialization vector sequences corresponding to the sequence data, and the text initialization vectors corresponding to the text data can be subjected to feature fusion processing (concatenation, addition, etc.) to obtain the object initialization vector corresponding to the first object information, and then the object initialization vector can be processed through the embedding layer 1 to obtain the object embedding vector of the first object information. The same processing method as the first object information can be used to obtain the business embedding vector corresponding to the first business information through the embedding layer 2, which will not be repeated here.

[0153] Optionally, content interest weight features can be added to the object embedding vector, which can include text, short video and long video content interest, such as text and picture content interest features of business advertisements. By adding content interest weight features to the object embedding vector, the feature diversity of the object embedding vector can be improved.

[0154] After obtaining the object embedding vector and the business embedding vector, the object embedding vector and the business embedding vector can be input into the initial representation network in the initial prediction model. Through the initial representation network, the object implicit features of the object embedding vector and the business implicit features of the business embedding vector are obtained, and the object implicit features and the business implicit features are subjected to feature interaction processing to obtain sample information interaction features. Specifically, the object embedding vector can be input into the object representation network contained in the initial representation network, and the business embedding vector can be input into the business representation network contained in the initial representation network. Among them, the object representation network can include N chain-connected first hidden layers (network layers), and the business representation network can include D chain-connected second hidden layers (network layers), D and N are both positive integers, D and N can be the same value or different values, and this application does not limit this. That is to say, both the object representation network and the business representation network can include one or more network layers.

[0155] Through the D second hidden layers in the business characterization network, feature extraction is performed on the business embedding vector to obtain the output data of the D second hidden layers, and the business implicit features are determined in the output data of the D second hidden layers. Among the N first hidden layers contained in the object characterization network, the candidate network layer associated with the business implicit features is determined; feature extraction is performed on the object embedding vector through the object characterization network to obtain the output data of the candidate network layer, and the output data of the candidate network layer is determined as the object implicit feature. The business implicit features and the object implicit features are subjected to feature interaction processing to obtain implicit interaction features; the implicit interaction features are used as input data for the next first hidden layer of the candidate network layer in the object characterization network; if the candidate network layer is the Nth first hidden layer in the object characterization network, the implicit interaction features are determined as sample information interaction features.

[0156] Among them, the object implicit feature can be the final output data of the object representation network, such as the output data of the last first hidden layer of the object representation network; the business implicit feature can be the final output data of the business representation network, such as the output data of the last second hidden layer of the business representation network. Optionally, when the object representation network and the business representation network both include multiple network layers (hidden layers), the object implicit feature can include the output data of some or all network layers (hidden layers) in the object representation network, and the business implicit feature can include the output data of some or all network layers (hidden layers) in the business representation network. Each feature interaction processing is for one business implicit feature and one object implicit feature, that is, the business implicit features and the object implicit features have the same number. The candidate network layer can refer to the first hidden layer in the object representation network whose output data is selected as the first hidden layer for feature cross processing with the business implicit feature.

[0157] For example, when the second output data of three second hidden layers are selected as the service implicit features in the service representation network, the output data of three first hidden layers can be selected as the object implicit features in the object representation network. When the output data of the first first hidden layer, the fifth first hidden layer and the seventh first hidden layer in the object representation network are used as the object implicit features, the first first hidden layer, the fifth first hidden layer and the seventh first hidden layer in the object representation network can be called candidate network layers in the object representation network.

[0158] Feature interaction processing may include but is not limited to: using element product method to perform feature interaction on the above-mentioned object implicit features and business implicit features, using concat method to perform feature interaction on the above-mentioned object implicit features and business implicit features, using matrix product method to perform feature interaction on the above-mentioned object implicit features and business implicit features, etc.

[0159] In one or more embodiments, assuming that both the object representation network and the service representation network include multiple network layers (hidden layers), the object implicit feature is the output data of the last network layer of the object representation network, and the service implicit feature is the output data of the last network layer of the service representation network; then the object implicit feature and the service implicit feature can be subjected to feature interaction processing to obtain sample information interaction features. That is to say, in this case, in the object representation network and the service representation network, feature interaction processing can be performed only once.

[0160] Optionally, assuming that the object implicit features include the output data of each network layer in the object representation network, and the business implicit features include the output data of each network layer in the business representation network; then the result after the feature interaction processing can be input to the next network layer of the object representation network, and the input of the next network layer of the business representation network is the business implicit features output by the previous network layer. That is to say, in this case, multiple feature interaction processes can be performed in the object representation network and the business representation network. The embodiments of the present application are not limited to the number of feature interaction processes in the object representation network and the business representation network.

[0161] For example, suppose that the object representation network includes 10 network layers (hidden layers), the business representation network includes 5 network layers (hidden layers), the object implicit features include the output data of any 5 network layers in the 10 network layers, and the business implicit features include the output data corresponding to the 5 network layers. Then the feature interaction processing in the object representation network and the business representation network may include: ① Performing feature interaction processing on the output data of the 2nd network layer in the object representation network and the output data of the 1st network layer in the business object network to obtain the first feature interaction result, which can be input to the 3rd network layer of the object representation network, and the input of the next network layer of the business object network is the output data of the previous network layer; ② Performing feature interaction processing on the output data of the 4th network layer in the object representation network and the output data of the 2nd network layer in the business object network to obtain the second feature interaction result, which can be input to the 5th network layer of the object representation network; ③ Performing feature interaction processing on the 6th network layer in the object representation network and the output data of the 2nd network layer in the business object network to obtain the second feature interaction result, which can be input to the 5th network layer of the object representation network; ④ Perform feature interaction processing on the output data of the 8th network layer in the object representation network and the output data of the 4th network layer in the business object network to obtain the fourth feature interaction result, which can be input into the 9th network layer of the object representation network; ⑤ Perform feature interaction processing on the output data of the 10th network layer in the object representation network and the output data of the 5th network layer in the business object network to obtain the fifth feature interaction result, which can be used as the sample information interaction feature.

[0162] Optionally, the object representation network and the business representation network may also include the same number of network layers. The object implicit features and business implicit features for feature interaction processing may be output data of any number and any position of network layers in the object representation network, as well as output data of any number and any position of network layers in the business representation network. This application does not limit this.

[0163] It should be noted that if the element-wise multiplication method is used for feature interaction processing in the object representation network and the business representation network, the object implicit features and business implicit features for feature interaction processing need to have the same dimension. The process of feature interaction processing by element-wise multiplication can be found in Figure 6 , Figure 6 1 is a schematic diagram of a feature interaction process between first object information and first service information provided by an embodiment of the present application. Figure 6 As shown, assuming that both the object representation network 50a and the service representation network 50b adopt a multi-layer perceptron structure, the object representation network 50a may include N chain-connected first hidden layers, and the service representation network 50b may include N chain-connected second hidden layers, where N is a positive integer. The N first hidden layers included in the object representation network 50a and the N second hidden layers included in the service representation network 50b may both be fully connected layers.

[0164] After obtaining the object embedding vector of the first object information and the business embedding vector of the first business information, the object embedding vector can be used as input data of the object representation network 50a, and the business embedding vector can be used as input data of the business representation network 50b. The input data of the i-th first hidden layer in the object representation network 50a can be obtained, and the input data of the i-th second hidden layer in the business representation network 50b can be obtained. When i is 1, the input data of the i-th first hidden layer (i.e., the 1st first hidden layer) is the object embedding vector, and the input data of the i-th second hidden layer (i.e., the 1st second hidden layer) is the business embedding vector, and i is a positive integer less than or equal to N.

[0165] When i is not 1, feature extraction can be performed on the input data of the i-th first hidden layer according to the i-th first hidden layer to obtain the output data of the i-th first hidden layer; feature extraction can be performed on the input data of the i-th second hidden layer according to the i-th second hidden layer to obtain the output data of the i-th second hidden layer. The output data of the i-th second hidden layer is used as the input data of the i+1-th second hidden layer in the service characterization network. The output data of the i-th first hidden layer is multiplied (element-wise multiplication) with the output data of the i-th second hidden layer to obtain the i-th implicit interaction feature (the i-th feature interaction result); the i-th implicit interaction feature is used as the input data of the i+1-th first hidden layer in the object characterization network.

[0166] By analogy, feature interaction processing can be continued in the object representation network 50a and the business representation network 50b until the Nth implicit interaction feature is obtained, and the Nth implicit interaction feature can be determined as the sample information interaction feature; the Nth implicit interaction feature is obtained by performing a dot product operation on the output data of the Nth first hidden layer and the output data of the Nth second hidden layer.

[0167] like Figure 6 As shown, the object embedding vector can be input into the first first hidden layer in the object representation network 50a, and the business embedding vector can be input into the first second hidden layer in the business representation network 50b; the object embedding vector can be point-multiplied with the output data of the first second hidden layer after passing through the output data of the first first hidden layer to obtain the first implicit interaction feature. The first implicit interaction feature can be input into the second first hidden layer, and the output data of the first second hidden layer can be input into the second second hidden layer. Similarly, the output data of the second first hidden layer can be point-multiplied with the output data of the second second hidden layer to obtain the second implicit interaction feature. The second implicit interaction feature can be input into the third first hidden layer, and the output data of the second second hidden layer can be input into the third second hidden layer. This is done by analogy until the output data of the Nth first hidden layer is multiplied by the output data of the Nth second hidden layer to obtain the Nth implicit interaction feature, and the Nth implicit interaction feature is determined as the sample information interaction feature. By performing feature interaction processing on the first object information and the first business information, the extraction and recognition of business information can be enhanced, the feature interaction between the object information and the business information can be realized, and the representation capability of the sample information interaction feature can be improved.

[0168] After obtaining the sample information interaction features of the business sample data, the sample information interaction features can be input into the initial shared network in the initial prediction model, as well as the branch-specific network and branch gating network corresponding to each business. According to the relevant features of the business information carried by the sample information interaction features (for example, business ID), the sample branch business that needs to be forward-calculated for the sample information interaction features is determined from all the initial branch networks included in the initial prediction model; that is, the branch-specific network and branch gating network corresponding to the sample object are determined.

[0169] Step S104: outputting sample global features of sample information interaction features through the initial shared network, and outputting sample local features of sample information interaction features through the sample branch network.

[0170] Specifically, after the sample information interaction feature is input into the initial shared network, and the branch dedicated network and branch gating network corresponding to the sample business, the sample information interaction feature can be extracted (forward calculation) through the initial shared network to obtain the sample global feature. The sample global feature can be used to represent the global information of the current business sample data, such as learning the object distribution difference between the current business sample data and the business sample data of other businesses. The sample information interaction feature can be extracted through the branch dedicated network to obtain the sample local feature, which can be used to represent the local information of the current business sample data, such as learning the object distribution in the business sample data of the sample business. The sample information interaction feature is input into the branch gating network, and the sample information interaction feature is analyzed through the branch gating network to obtain the sample weight feature; the sample weight feature can include a first weight corresponding to the sample global feature and a second weight corresponding to the layout information feature. The larger the weight, the more important the corresponding feature; if the first weight is greater than the second weight, it means that the sample global feature is more important than the global information feature; if the first weight is less than the second weight, it means that the sample local feature is more important than the global information feature.

[0171] Step S105, determining the sample prediction result of the business sample data according to the sample global features and the sample local features.

[0172] Specifically, the first weight and the second weight in the sample weight feature output by the branch gating network can be obtained, and the product between the first weight and the sample global feature and the product between the second weight and the sample local feature can be merged to obtain the sample fusion feature; the sample fusion feature can be considered as the gated merging result of the sample business. The sample fusion feature is normalized through the business normalization network layer in the initial prediction model to obtain the first normalized feature corresponding to the business sample data; the first normalized feature is classified to obtain the sample prediction result of the business sample data. Among them, the sample prediction result can be used to represent the classification result of the business sample data, and can be used to calculate the object resource contribution value estimated for the sample object.

[0173] In one or more embodiments, the normalization processing of the sample fusion feature may include: determining the sample training subset to which the business sample data belongs, determining the sample training group corresponding to the sample business in the sample training subset, and obtaining the sample mean and sample variance corresponding to the sample training group; the sample training subset belongs to the sample training set; according to the sample mean and sample variance, the sample fusion feature is normalized to obtain the first normalized feature corresponding to the business sample data.

[0174] Among them, the sample training subset may refer to the set of business sample data in the batch to which the current business sample data belongs. In the model training stage, the sample training set may be trained in batches, and the sample training set may be divided into multiple batches, each of which may be called a sample training subset. Each batch may include business sample data corresponding to one or more businesses, and the sample training subset may be divided into multiple sample training groups according to different businesses; the sample training group corresponding to the business sample may be determined from the multiple sample training groups contained in the sample training subset. The mean and variance may be calculated separately for each business in each batch, such as calculating the sample mean and sample variance corresponding to the sample business in the sample training subset to which the business sample data belongs, subtracting the sample mean from the sample fusion feature, and dividing it by the sample variance to obtain the first normalized feature.

[0175] In one or more embodiments, the initial estimation model may include a first task output network and a second task output network, and the sample prediction result includes a first task estimated probability and a second task estimated probability, wherein the first task estimated probability may be a classification result of the first task output by the first task output network, and the second task estimated probability may be a classification result of the second task output by the second task output network. The first normalized feature may be input to the first task output network and the second task output network; the first normalized feature may be subjected to binary classification processing through the first task output network to obtain the first task estimated probability of the business sample data; the first normalized feature may be subjected to multi-classification processing through the second task output network to obtain the second task estimated probability of the business sample data.

[0176] Among them, the first task estimated probability can be expressed as a two-dimensional vector, one element in the two-dimensional vector can represent the probability value that the sample object consumes resources for the sample business, and another element in the two-dimensional vector can represent the probability value that the sample object does not consume resources for the sample business.

[0177] The second task may be to convert the object resource contribution value estimation into multiple binary classification tasks, such as setting M preset values ​​for the continuous object resource contribution value, such as [0, LTV 1 , LTV 2 , …, LTV K , …, LTV M ], LTV k It represents the kth preset value arranged in order from small to large, M is a positive integer, and k is a positive integer less than or equal to M. Multiple binary classification tasks can be represented as whether the resource consumption amount exceeds LTV KThe second task estimated probability output by the second task output network may include probability estimated values ​​corresponding to M resource contribution ranges.

[0178] The classification result of each binary classification task in the second task can be understood as: the object resource contribution value predicted based on the business sample data, and the probability estimated value of belonging to the M resource contribution ranges. The above M resource contribution ranges can be expressed as [0, LTV 1 ), [LTV 1 , LTV 2 ), [LTV 2 , LTV 3 ), …, [LTV M-1 , LTV M ]. The multiple binary classification tasks in the second task can adopt a multi-layer perceptron network structure carrying an activation function (for example, sigmoid), and the present application does not limit the structure of the second task output network.

[0179] The second task estimated probability can be expressed as an M-dimensional vector. The first element of the M-dimensional vector can be expressed as the resource consumption amount generated by the sample object in the sample business belongs to the resource contribution range [0, LTV 1 ), that is, whether it exceeds LTV 1 The probability estimate of the probability of exceeding LTV is greater. 1 The second element in the M-dimensional vector can be expressed as the amount of resource consumption generated by the sample object in the sample business belongs to the resource contribution range [LTV 1 , LTV 2 ), that is, whether it exceeds LTV 2 The probability estimate of the probability of exceeding LTV is greater. 2 By analogy, the Mth element in the M-dimensional vector can be expressed as the amount of resource consumption generated by the sample object in the sample business that falls within the resource contribution range [LTV M-1 , LTV M ], that is, whether it exceeds LTV M The probability estimate of the probability of exceeding LTV is greater. M The less likely it is.

[0180] Step S106, based on the sample prediction results and the object resource contribution value, the network parameters of the initial prediction model are modified to obtain a resource prediction model.

[0181] Specifically, the first label information and the second label information of the business sample data are determined according to the object resource contribution value; the first label information is used to characterize whether the sample object has made resource contributions to the sample business, and the second label information is used to characterize the resource contribution range of the sample object to the sample business. For example, if the object resource contribution value carried by the business sample data is 1000, then the first label information can be represented as a vector [0, 1], indicating that the sample object has generated resource consumption in the sample business. When 1000 is less than LTV 2 and greater than LTV 1 If the value of , the second label information can be represented as a vector [1, 1, 0, 0, ..., 0], and the vector corresponding to the second label information is an M-dimensional vector, except that the first and second elements are 1, the remaining elements are 0. If the object resource contribution value carried by the service sample data is 0, the first label information can be represented as a vector [1, 0], indicating that the sample object does not consume resources in the sample service, and the second label information can be represented as a vector [1, 0, 0, 0 ..., 0], and the vector corresponding to the second label information is an M-dimensional vector, except that the first element is 1, the remaining elements are 0.

[0182] The first task classification loss can be determined based on the first label information and the first task estimated probability; the first task classification loss can be a cross entropy loss, or it can be other types of losses, which is not limited in this application. The second task classification loss is determined based on the second label information and the second task estimated probability; the second task classification loss can be a binary cross entropy loss, or it can be other types of losses, which is not limited in this application. It can be understood that the sample training set may include business sample data that has generated resource consumption, and may also include business sample data that has not generated resource consumption (for example, the object resource contribution value carried by such business sample data is 0). In the process of calculating the first task classification loss and the second task classification loss, the first task loss is calculated for all business sample data in the sample training set, and the second task loss is only calculated for business sample data with an object resource contribution value greater than 0.

[0183] The first task classification loss and the second task classification loss are linearly weighted to obtain the model loss of the initial prediction model; for example, assuming that the first task loss is recorded as L1 and the second task loss is recorded as L2, then the model loss can be expressed as L = L1 + α × L2, where α is a hyperparameter that can be flexibly set according to the actual application to control the importance of the second task loss L2 in the model loss. According to the model loss, the network parameters of the initial prediction model are modified to obtain the resource prediction model; that is, by minimizing the model loss and performing reverse gradient backpropagation based on the model loss, the network parameters of the initial prediction model can be continuously iteratively updated; when the initial prediction model reaches the training stop condition, the training is stopped and the model parameters at the time of stopping the training are saved. At this time, the initial prediction model can be used as the resource prediction model after training.

[0184] Among them, the training stop condition can be that the model loss reaches convergence, such as when the error between the model loss of the previous training and the model loss of the next training is less than or equal to a preset error threshold (which can be set according to the actual application), it indicates that the model loss reaches convergence, and the initial prediction model at this time meets the training stop condition. Optionally, the training stop condition can also be that the number of training times of the initial prediction model reaches a preset maximum number of training times; this application does not limit the setting of the training stop condition.

[0185] In one or more embodiments, the model training offline evaluation index of the initial prediction model can adopt two indicators: auc (Area Under the Curve) of whether resource consumption (first task) is performed and the gini coefficient of the amount of resource consumption; among which the gini coefficient of the amount of resource consumption is the main offline evaluation index. auc can be used to measure the performance of the first task output network. The larger the auc value, the better the performance of the first task output network. The gini coefficient refers to the ratio of the area enclosed by the absolute fairness line (line of equality) and the Lorenz curve (Lorenz Curve) to the area below the absolute fairness line; the larger the gini coefficient, the stronger the distinguishing ability of the second task output network; when the ROC curve (Receiver Operating Characteristic Curve, a curve with false positive rate as the horizontal axis and true positive rate as the vertical axis) and the gini curve overlap, the gini coefficient is twice the auc minus 1. Experiments in a certain delivery scenario found that by using the resource prediction model to predict the estimated resource contribution of each object, the AUC indicator of whether to consume resources (the first task) increased by 1%, and the Gini coefficient indicator of the resource consumption amount increased by 2%, which improved the prediction accuracy of the object's resource contribution value and improved the business delivery effect.

[0186] In an embodiment of the present application, a sample training set consisting of business sample data of at least one business is obtained, and one business sample data includes the first object information of the sample object, the first business information of the sample business, and the object resource contribution value of the sample object to the sample business, and the sample business is any one of the at least one business. An initial prediction model is obtained, and the initial prediction model may include an initial shared network and an initial branch network corresponding to at least one business, and the input data of the initial prediction model includes the first object information and the first business information. In the initial prediction model, feature interaction processing is performed on the first object information and the first business information to obtain sample information interaction features; through the initial shared network and the initial branch network of at least one business, the information of different businesses and the information of different objects are learned in the sample information interaction features of each business sample data, so as to improve the prediction accuracy of the object resource contribution value.

[0187] pass Figures 4 to 6 In the implementation process of the corresponding embodiment, after the resource prediction model is trained, the resource prediction model can be applied in the advertising delivery platform. The resource prediction model can be used to predict the object resource contribution value of each object of multiple services. Figure 7 , taking a business to be launched and an object as an example, the application process of the resource prediction model is described in detail. The resource prediction model refers to the trained initial prediction model. The resource prediction model and the initial prediction model have the same network structure, but the model parameters they carry are different. In the initial prediction model, the trained initial representation model can be called an information representation model, and the information representation network can be used to extract and identify relevant information of object information and business information; the trained initial shared network can be called a business sharing network, and the business sharing network can be used to extract the object distribution differences between different businesses; the trained initial branch network can be called a business branch network, and each business branch network can be used to extract the object distribution information of its corresponding business; the network structure of the resource prediction model can be as follows Figure 5 As shown, no further description is given here.

[0188] Figure 7 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 2 It can be understood that the data processing method can be executed by a computer device, which can be a terminal device, such as Figure 1 Any terminal device in the terminal cluster shown, or a server, such as Figure 1 The server 10d shown in the figure is not limited in this application. Figure 7 As shown, the method may include the following steps S201 to S209:

[0189] Step S201, obtaining second object information of a candidate object, and obtaining second business information of a business to be launched.

[0190] The candidate object may be any business that may be interested in the business to be launched, such as Figure 3 The candidate objects in the candidate object set 30a in the corresponding embodiment. The business to be launched may refer to any business that needs to be promoted and advertised, such as any game, activity, product, etc. that needs to be promoted and advertised.

[0191] The second object information can be used to represent relevant information of the candidate object, such as the basic information of the candidate object, the active duration of the candidate object for the above-mentioned service to be launched or other services, the resource consumption duration, the resource consumption amount, the service active sequence, the service consumption sequence, etc. The information content contained in the second object information can be found in Figure 4 The description of the first object information in the corresponding embodiment will not be repeated here.

[0192] The second service information may include relevant information of the service to be launched, such as the service ID, service type, average resource contribution value, etc. of the service to be launched. The information content included in the second service information can be found in Figure 4 The description of the first service information in the corresponding embodiment will not be repeated here.

[0193] Step S202: Perform feature interaction processing on the second object information and the second service information through the information representation network to obtain information interaction features, and determine a candidate branch network corresponding to the service to be launched in the service branch network corresponding to at least one service.

[0194] The feature interaction processing process of the second object information and the second business information in the information representation network of the resource prediction model can be seen in Figure 4 In the corresponding embodiment, the feature interaction processing process of the first object information and the second object information in the initial prediction model is not repeated here. The candidate branch network may refer to a business branch network corresponding to the business to be launched, and the business branch network may include a business-specific network and a business gating network for processing the business to be launched.

[0195] Step S203: outputting global information features of information interaction features through the service sharing network, and outputting local information features of information interaction features through the candidate branch network.

[0196] Step S204, obtaining a service weight feature; the service weight feature is output by the service gating network in the candidate branch network.

[0197] The calculation process of information interaction features in the business sharing network and the candidate branch network can be found in Figure 4 In the corresponding embodiment, the calculation process of the sample information interaction feature in the initial shared network and the sample branch network. In other words, the method of obtaining the global information feature, the local information feature and the service weight feature can be found in Figure 4 The method for obtaining the sample global features, sample local features and sample weight features in the corresponding embodiment will not be described in detail here.

[0198] Step S205 , weighting the global information feature and the local information feature according to the service weight feature to obtain the service fusion feature.

[0199] Specifically, if in the service weight feature, the weight for the global information feature is expressed as a1, the weight for the local information feature is expressed as a2, and the global information feature is expressed as x1, and the local information feature is expressed as x2, then the service fusion feature can be expressed as: a1×x1+a2×x2.

[0200] Step S206, normalizing the service fusion feature to obtain a second normalized feature, and performing binary classification on the second normalized feature to obtain a third task estimated probability.

[0201] Specifically, the business mean and business variance of the business to be launched can be obtained, the sample mean is subtracted from the business fusion feature, and then divided by the business variance, and the second normalized feature can be calculated. Among them, the business mean may refer to the average value of the sample mean of the business to be launched in all batches during the last training in the model training stage; the business variance may refer to the average value of the sample variance of the business to be launched in all batches during the last training in the model training stage; or it may be other values ​​set according to the training results of the model training stage, which is not limited in this application. In the application stage of the resource prediction model, the business mean and business variance of each business to be launched can be fixed values.

[0202] The third task estimated probability may be the classification result of the first task output by the first task output network in the resource prediction model; the third task estimated probability may be represented as a two-dimensional vector. The meaning of the two-dimensional vector may be referred to in Figure 4 In the corresponding embodiment, the relevant description of the estimated probability of the first task will not be repeated here.

[0203] Step S207, performing multi-classification processing on the second normalized feature to obtain the fourth task estimated probability, and determining the resource contribution estimated value of the candidate object for the to-be-delivered business according to the third task estimated probability and the fourth task estimated probability.

[0204] The fourth task estimated probability can be the classification result of the second task output by the second task output network in the resource prediction model; the third task estimated probability can be expressed as an M-dimensional vector. The meaning of the M-dimensional vector can be seen in Figure 4 In the corresponding embodiment, the relevant description of the estimated probability of the second task will not be repeated here.

[0205] Specifically, the estimated probability of the fourth task may include probability estimated values ​​corresponding to M resource contribution ranges, and the product of the probability estimated value corresponding to the kth resource contribution range and the kth interval resource contribution value is determined as the kth candidate resource contribution value, where k is a positive integer less than or equal to M; M candidate resource contribution values ​​are obtained, and the M candidate resource contribution values ​​are accumulated to obtain a cumulative resource contribution value; the product of the cumulative resource contribution value and the estimated probability of the third task is determined as the resource contribution estimated value of the candidate object for the business to be launched.

[0206] The kth interval resource contribution value refers to the interval resource contribution value between the kth resource contribution range and the k-1th resource contribution range; the kth resource contribution range and the k-1th resource contribution range are two adjacent resource contribution ranges in the M resource contribution ranges. For example, when k is 1, the first interval resource contribution value can be LTV 1 -0; when k is 2, the second interval resource contribution value can be LTV 2 -LTV 1 wait.

[0207] Assuming the estimated probability of the third task is expressed as p, whether the resource consumption amount exceeds the LTV k The probability estimate (which can be understood as the contribution range of the kth resource) is expressed as p k , then the estimated resource contribution of the candidate object to the business to be launched can be expressed as the following formula (1):

[0208]

[0209] Wherein, pLTV in formula (1) represents the estimated resource contribution value calculated for the candidate object by the resource prediction model. For all candidate objects of each service to be launched, the estimated resource contribution value of each candidate object can be calculated in the above manner, which will not be described one by one here.

[0210] Step S208: when the estimated resource contribution value is greater than the resource contribution threshold, the candidate object is determined as a hit object, business display information corresponding to the to-be-delivered business is generated, and the business display information is sent to the hit object.

[0211] In one or more embodiments, the hit object may be considered to be an object for which the resource contribution estimate is calculated by the resource prediction model and is determined to be a target for which the above-mentioned service to be delivered can be delivered. The hit object may also be called a high-value object. The service display information may refer to the service promotion information delivered to the hit object, or may be called a service advertisement, such as Figure 3 The business promotion information 30f in the corresponding embodiment.

[0212] Specifically, when the resource contribution estimate is greater than the resource contribution threshold, the candidate object can be added to a high contribution object set, all of which are hit objects for the services to be launched, and the service display information can be sent to all objects in the high contribution object set.

[0213] When the estimated resource contribution value is less than or equal to the resource contribution threshold, the candidate object is determined as a missed object. A missed object may refer to an object whose estimated resource contribution value is calculated by a resource prediction model and is determined to be unnecessary for the above-mentioned business to be placed. A missed object may also be called a low-value object.

[0214] For example, when the service to be launched is a game, the service display information can be game promotion information (game trailer, game animation, game poster, etc.), and the game promotion information is sent to each hit object in the high contribution object set, such as being displayed in one or more service clients used by the hit object. When the service to be launched is a product, the service display information can be a product advertisement, and the product advertisement is sent to each hit object in the high contribution object set, such as being displayed in a shopping client, video client, etc. used by the hit object.

[0215] Step S209: obtaining the business interaction data of the hit object with respect to the business display information, and updating the sample training set according to the business interaction data to obtain an updated sample training set.

[0216] Specifically, after delivering the business display information to all objects (hit objects) in the high contribution object set, the hit objects can click on the business display information, download and install the client corresponding to the above-mentioned business to be delivered, and perform registration, consumption, trial play and other operations in the client corresponding to the business to be delivered. The above data can be called the business interaction data of the hit object for the business display information. In other words, the business interaction data can refer to the interaction data generated by the hit object for the business to be delivered, such as the active duration (days) of the hit object for the business to be delivered, the amount of resource consumption, the duration (days) of resource consumption, the business activity sequence, the business consumption sequence and other data. This application does not limit the information content contained in the interaction data.

[0217] The business interaction data of the hit object in the business to be launched is obtained, and new business sample data is obtained based on the business interaction data; the sample training set is updated based on the new business sample data to obtain an updated sample training set, and the updated sample training set is used to update the resource prediction model.

[0218] See also Figure 8 , Figure 8 This is a schematic diagram of the overall process of a service delivery provided by an embodiment of the present application. Figure 8 As shown, the complete process of the business delivery scenario involved in the embodiment of the present application may include: determining a set of candidate objects for multiple businesses, adopting a cold start object selection method, and determining a batch of trial delivery objects for each business (the trial delivery object set can be considered as the initial delivery object set selected for the business). Then, the business delivery can be carried out to the trial delivery objects, and the object consumption samples are recovered based on the business interaction data of the trial delivery objects in the actual delivered business. These recovered object consumption samples can be used as business sample data for training the initial prediction model. For example, these recovered object consumption samples can be added as business sample data to a sample training set, and the initial prediction model is trained through the business sample data in the sample training set. The initial prediction model after the training is completed can be called a resource prediction model.

[0219] Through the trained resource prediction model, the resource contribution estimates of other objects in the candidate object set of multiple businesses except the trial launch objects are calculated, and then the hit objects can be selected from the candidate object set of multiple businesses in order of the resource contribution estimates from large to small. These selected hit objects can constitute a high-contribution object set.

[0220] The high contribution object set can be considered as the updated set of delivery objects, that is, the trial delivery object set is replaced with the high contribution object set, and business delivery is performed on the hit objects in the high contribution object set. Based on the business interaction data of the hit objects in the actual delivered business, delivery feedback samples are recovered; these recovered delivery feedback samples can be used to update the sample training set, and the updated sample training set can be used to update the training resource prediction model. Delivery feedback samples can refer to the relevant data of objects that interact with the delivered business after the business is delivered, such as the object basic information of the object, the resource consumption time of the object, the resource consumption amount, the active time, the business active sequence, the business consumption sequence, etc. Delivery feedback samples can include object consumption samples, that is, delivery feedback samples can include relevant data of objects that have generated resource consumption in the delivered business, and can also include relevant data of objects that have not generated resource consumption in the delivered business.

[0221] It is understandable that the resource prediction model can be regularly trained and updated in combination with the collected delivery feedback samples (for example, with daily granularity, weekly granularity, etc.). After the resource prediction model is updated, the delivery object set can be updated synchronously. The object selection ratio of the delivery object set can be adjusted in combination with ROI. When the ROI exceeds expectations, the object selection ratio of the delivery object set can be increased; when the ROI is lower than expected, the object selection ratio of the delivery object set can be reduced.

[0222] In the embodiment of the present application, the object information and business information of multiple services can be used to train and predict resource contribution estimates for candidate objects of multiple services, which can solve the problem of large differences in object distribution in multiple services; by learning the object feature distribution of different services, the accuracy of resource contribution estimates can be improved. Based on the resource contribution estimate, the set of delivery objects for each service to be delivered is determined, such as Figure 3 The high contribution object set 30c shown can improve the accuracy of selecting the delivery object set. The objects in the delivery object set are more likely to consume services in the future, thereby improving the service delivery effect.

[0223] It is understandable that in the specific implementation of the present application, the user's personal basic information and the user's interactive data in the business client (for example, game client, shopping client, etc.) may be involved, such as the user's active time (days) in the game client, the amount of resource consumption, the resource consumption time (days), resource consumption records, game task records, etc. When the above embodiments of the present application are applied to specific products or technologies, it is necessary to obtain the permission or consent of the relevant institutions or departments, or the users themselves, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant regions. For example, an authorization prompt interface for user-related data can be displayed in the interface of the business client. When the user confirms the authorization, the relevant data of the user can be collected, used and processed; when the user refuses to authorize, the above embodiments of the present application cannot collect, use and process the relevant data of the user.

[0224] See also Fig. 9 , Fig. 9 Schematic diagram of a data processing device provided in an embodiment of the present application. Fig. 9 As shown, the data processing device 1 may include: a first acquisition module 101, a second acquisition module 102, a first feature interaction module 103, a first feature extraction module 104, a first prediction module 105, and a parameter correction module 106;

[0225] The first acquisition module 101 is used to acquire a sample training set; the sample training set includes business sample data of at least one business, the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object for the sample business, and the at least one business includes a sample business;

[0226] The second acquisition module 102 is used to acquire an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to at least one service;

[0227] A first feature interaction module 103 is used to perform feature interaction processing on the first object information and the first service information to obtain a sample information interaction feature, and determine a sample branch network corresponding to the sample service in an initial branch network corresponding to at least one service;

[0228] A first feature extraction module 104, configured to output sample global features of sample information interaction features through an initial shared network, and output sample local features of sample information interaction features through a sample branch network;

[0229] The first prediction module 105 is used to determine the sample prediction result of the business sample data according to the sample global characteristics and the sample local characteristics;

[0230] The parameter correction module 106 is used to correct the network parameters of the initial prediction model according to the sample prediction results and the object resource contribution value to obtain a resource prediction model; the resource prediction model is used to predict the resource contribution estimate of the object to be deployed for the business to be deployed.

[0231] In one or more embodiments, the initial prediction model further includes an initial representation network;

[0232] The first feature interaction module 103 performs feature interaction processing on the first object information and the first service information to obtain sample information interaction features for executing the following steps:

[0233] Converting the first object information into an object embedding vector, converting the first business information into a business embedding vector, and inputting the object embedding vector and the business embedding vector into an initial representation network;

[0234] The object implicit features of the object embedding vector and the business implicit features of the business embedding vector are obtained through the initial representation network, and the object implicit features and the business implicit features are subjected to feature interaction processing to obtain sample information interaction features.

[0235] In one or more embodiments, the initial representation network includes an object representation network and a service representation network, the object representation network includes N chain-connected first hidden layers, the service representation network includes D chain-connected second hidden layers, and N and D are positive integers;

[0236] The first feature interaction module 103 obtains the object implicit features of the object embedding vector and the business implicit features of the business embedding vector through the initial representation network, performs feature interaction processing on the object implicit features and the business implicit features to obtain sample information interaction features for executing the following steps:

[0237] Extracting features of the service embedding vector through D second hidden layers in the service characterization network to obtain output data of the D second hidden layers, and determining service hidden features in the output data of the D second hidden layers;

[0238] Determine a candidate network layer associated with the service implicit feature among the N first hidden layers included in the object representation network;

[0239] Extract features from the object embedding vector through the object representation network, obtain output data of the candidate network layer, and determine the output data of the candidate network layer as the object implicit features;

[0240] Performing feature interaction processing on the service implicit features and the object implicit features to obtain implicit interaction features; the implicit interaction features are used as input data of the next first implicit layer of the candidate network layer in the object representation network;

[0241] If the candidate network layer is the Nth first hidden layer in the object representation network, the implicit interaction feature is determined as the sample information interaction feature.

[0242] In one or more embodiments, the sample branch network includes a branch-specific network;

[0243] The first feature extraction module 104 outputs sample global features of sample information interaction features through the initial shared network, and outputs sample local features of sample information interaction features through the sample branch network, for performing the following steps:

[0244] Inputting sample information interaction features into the initial shared network and the branch dedicated network;

[0245] Through the initial shared network, feature extraction is performed on the sample information interaction features to obtain the sample global features;

[0246] Through the branch-specific network, the interactive features of sample information are extracted to obtain the local features of the sample.

[0247] In one or more embodiments, the sample branch network further includes a branch gating network;

[0248] The data device 1 further includes: a gate control module 107;

[0249] The gating module 107 is used to input the sample information interaction feature into the branch gating network, and analyze the sample information interaction feature through the branch gating network to obtain the sample weight feature; the sample weight feature includes a first weight corresponding to the sample global feature and a second weight corresponding to the layout information feature;

[0250] In one or more embodiments, the first prediction module 105 determines the sample prediction result of the business sample data according to the sample global feature and the sample local feature, and is used to perform the following steps:

[0251] Obtaining a first weight and a second weight in the sample weight feature, merging the product of the first weight and the sample global feature, and the product of the second weight and the sample local feature, to obtain a sample fusion feature;

[0252] Normalize the sample fusion features to obtain the first normalized features corresponding to the business sample data;

[0253] The first normalized feature is classified and processed to obtain a sample prediction result of the business sample data.

[0254] In one or more embodiments, the first prediction module 105 performs normalization processing on the sample fusion feature to obtain a first normalized feature corresponding to the business sample data, which is used to perform the following steps:

[0255] Determine the sample training subset to which the business sample data belongs, determine the sample training group corresponding to the sample business in the sample training subset, and obtain the sample mean and sample variance corresponding to the sample training group; the sample training subset belongs to the sample training set;

[0256] The sample fusion features are normalized according to the sample mean and sample variance to obtain the first normalized features corresponding to the business sample data.

[0257] In one or more embodiments, the initial estimation model further includes a first task output network and a second task output network, and the sample prediction result includes a first task estimation probability and a second task estimation probability;

[0258] The first prediction module 105 performs classification processing on the first normalized feature to obtain a sample prediction result of the business sample data, which is used to perform the following steps:

[0259] Inputting the first normalized feature into the first task output network and the second task output network;

[0260] Through the first task output network, the first normalized feature is subjected to binary classification processing to obtain the first task estimated probability of the business sample data;

[0261] Through the second task output network, multi-classification processing is performed on the first normalized feature to obtain the second task estimated probability of the business sample data.

[0262] In one or more embodiments, the parameter modification module 106 modifies the network parameters of the initial prediction model according to the sample prediction results and the object resource contribution value to obtain a resource prediction model for performing the following steps:

[0263] Determine the first label information and the second label information of the service sample data according to the object resource contribution value; the first label information is used to indicate whether the sample object has made resource contribution to the sample service, and the second label information is used to indicate the resource contribution scope of the sample object to the sample service;

[0264] Determine the first task classification loss according to the first label information and the first task estimated probability, and determine the second task classification loss according to the second label information and the second task estimated probability;

[0265] Linearly weight the first task classification loss and the second task classification loss to obtain the model loss of the initial prediction model;

[0266] The network parameters of the initial prediction model are modified according to the model loss to obtain the resource prediction model.

[0267] In one or more embodiments, the resource prediction model includes an information representation network, a service sharing network, and a service branch network corresponding to at least one service;

[0268] The data processing device 1 further includes: a third acquisition module 108, a second feature interaction module 109, a second feature extraction module 110, and a second prediction module 111;

[0269] The third acquisition module 108 is used to acquire the second object information of the candidate object and acquire the second business information of the business to be launched;

[0270] The second feature interaction module 109 is used to perform feature interaction processing on the second object information and the second service information through the information representation network to obtain information interaction features, and determine a candidate branch network corresponding to the service to be placed in the service branch network corresponding to at least one service;

[0271] A second feature extraction module 110, configured to output global information features of information interaction features through a service sharing network, and output local information features of information interaction features through a candidate branch network;

[0272] The second prediction module 111 is used to determine the estimated resource contribution of the candidate object to the service to be launched according to the global information features and the local information features.

[0273] In one or more embodiments, the second prediction module 111 determines the estimated resource contribution of the candidate object to the service to be placed according to the global information features and the local information features, and is used to perform the following steps:

[0274] Obtaining business weight features; the business weight features are output by the business gating network in the candidate branch network;

[0275] According to the service weight characteristics, the global information characteristics and the local information characteristics are weighted to obtain the service fusion characteristics;

[0276] The service fusion feature is normalized to obtain a second normalized feature, and the second normalized feature is subjected to binary classification to obtain a third task estimated probability;

[0277] The second normalized feature is subjected to multi-classification processing to obtain the fourth task estimated probability, and the resource contribution estimated value of the candidate object for the to-be-delivered business is determined according to the third task estimated probability and the fourth task estimated probability.

[0278] In one or more embodiments, the fourth task estimated probability includes probability estimated values ​​corresponding to M resource contribution ranges, where M is an integer greater than 1;

[0279] The second prediction module 111 determines the estimated resource contribution value of the candidate object to the service to be placed according to the estimated probability of the third task and the estimated probability of the fourth task, and is used to perform the following steps:

[0280] The product of the probability estimate corresponding to the k-th resource contribution range and the k-th interval resource contribution value is determined as the k-th candidate resource contribution value, where k is a positive integer less than or equal to M;

[0281] Obtain M candidate resource contribution values, and accumulate the M candidate resource contribution values ​​to obtain a cumulative resource contribution value;

[0282] The product of the accumulated resource contribution value and the estimated probability of the third task is determined as the estimated resource contribution value of the candidate object for the business to be launched;

[0283] Among them, the kth interval resource contribution value refers to the interval resource contribution value between the kth resource contribution range and the k-1th resource contribution range; the kth resource contribution range and the k-1th resource contribution range are two adjacent resource contribution ranges in the M resource contribution ranges.

[0284] In one or more embodiments, the data processing device 1 further includes: a delivery module 112 , a sample update module 113 ;

[0285] The delivery module 112 is used to determine the candidate object as a hit object when the estimated resource contribution value is greater than the resource contribution threshold, generate business display information corresponding to the business to be delivered, and send the business display information to the hit object;

[0286] The sample updating module 113 is used to obtain the business interaction data of the hit object with respect to the business display information;

[0287] The sample updating module 113 is also used to update the sample training set according to the business interaction data to obtain an updated sample training set; the updated sample training set is used to train the resource prediction model.

[0288] According to an embodiment of the present application, the above Figure 4 The relevant steps involved in the data processing method shown can be represented by Fig. 9 The data processing device 1 shown in the figure is executed by each module. For example, Figure 4 Step S101 shown can be performed by Fig. 9 The first acquisition module 101 shown is used to execute, Figure 4 Step S102 shown can be performed by Fig. 9 The second acquisition module 102 shown is used to execute, Figure 4 Step S103 shown can be performed by Fig. 9 The first feature interaction module 103 is shown to execute, Figure 4 Step S104 shown can be performed by Fig. 9 The first feature extraction module 104 shown is used to perform the above.

[0289] According to one embodiment of the present application, Fig. 9 The various modules in the data processing device 1 shown can be respectively or all combined into one or several modules to constitute, or some of the modules can also be split into at least two smaller units in function, and the same operation can be achieved without affecting the realization of the technical effects of the embodiments of the present application. The above modules are divided based on logical functions. In practical applications, the function of one module can also be realized by at least two units, or the function of at least two modules can be realized by one module. In other embodiments of the present application, the data processing device 1 can also include other modules or units. In practical applications, these functions can also be implemented with the assistance of other modules, and can be implemented by the collaboration of at least two modules. In addition, the description of the beneficial effects of adopting the same method will not be repeated.

[0290] See also Fig.10 , Fig.10 Schematic diagram of the structure of a computer device provided in an embodiment of the present application. Fig.10 As shown, the computer device 1000 may be a terminal device, for example, Figure 1The terminal device 10a in the corresponding embodiment may also be a server, for example, Figure 1 The server 10d in the corresponding embodiment will not be limited here. For ease of understanding, this application takes a computer device as an example of a terminal device. The computer device 1000 may include: a processor 1001, a network interface 1004 and a memory 1005. In addition, the computer device 1000 may also include: a user interface 1003, and at least one communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. Among them, the user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or it may be a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 1005 may optionally also be at least one storage device located away from the aforementioned processor 1001. As Fig.10 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application program.

[0291] The network interface 1004 in the computer device 1000 can also provide a network communication function, and the optional user interface 1003 can also include a display screen (Display) and a keyboard (Keyboard). Fig.10 In the computer device 1000 shown, the network interface 1004 can provide a network communication function; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005 to achieve:

[0292] Acquire a sample training set; the sample training set includes business sample data of at least one business, the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object for the sample business, and the at least one business includes a sample business;

[0293] Acquire an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to at least one service;

[0294] Performing feature interaction processing on the first object information and the first service information to obtain sample information interaction features, and determining a sample branch network corresponding to the sample service in an initial branch network corresponding to at least one service;

[0295] Output the sample global features of the sample information interaction features through the initial shared network, and output the sample local features of the sample information interaction features through the sample branch network;

[0296] Determine the sample prediction result of the business sample data according to the sample global characteristics and sample local characteristics;

[0297] According to the sample prediction results and the object resource contribution value, the network parameters of the initial prediction model are modified to obtain a resource prediction model; the resource prediction model is used to predict the estimated resource contribution value of the object to be deployed for the business to be deployed.

[0298] It should be understood that the computer device 1000 described in the embodiment of the present application can execute the above Figure 4 , Figure 7 The description of the data processing method in any of the embodiments may also be performed as described above. Fig. 9 The description of the data processing device 1 in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of the same method will not be repeated here either.

[0299] In addition, it should be pointed out here that: the embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program executed by the data processing device 1 mentioned above, and the computer program includes computer instructions. When the processor executes the computer instructions, it can execute the above-mentioned Figure 4 , Figure 7 The description of the data processing method in any of the embodiments will not be repeated here. In addition, the description of the beneficial effects of the same method will not be repeated. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc. For technical details not disclosed in the computer-readable storage medium embodiment involved in this application, please refer to the description of the method embodiment of this application. As an example, program instructions can be deployed on a computer device for execution, or on multiple computer devices located at one location, or on multiple computer devices distributed at multiple locations and interconnected by a communication network. Multiple computer devices distributed at multiple locations and interconnected by a communication network can constitute a blockchain system.

[0300] In addition, it should be noted that: the embodiment of the present application also provides a computer program product, which may include a computer program, and the computer program may be stored in a computer-readable storage medium. The processor of the computer device reads the computer program from the computer-readable storage medium, and the processor may execute the computer program, so that the computer device performs the above Figure 4 , Figure 7 The description of the data processing method in any of the embodiments will not be repeated here. In addition, the description of the beneficial effects of the same method will not be repeated. For technical details not disclosed in the computer program product or computer program embodiment involved in the present application, please refer to the description of the method embodiment of the present application.

[0301] The terms "first", "second", etc. in the description, claims, and drawings of the embodiments of the present application are used to distinguish different media contents, rather than to describe a specific order. In addition, the term "comprising" and any variation thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product, or equipment comprising a series of steps or units is not limited to the listed steps or modules, but may optionally include steps or modules that are not listed, or may optionally include other step units inherent to these processes, methods, devices, products, or equipment.

[0302] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0303] The method and related apparatus provided by the embodiment of the present application are described with reference to the method flow chart and / or structural diagram provided by the embodiment of the present application. Specifically, each process and / or box in the method flow chart and / or structural diagram, as well as the combination of the processes and / or boxes in the flow chart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the process. Figure 1 A process or multiple processes and / or structures Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or structures Figure 1These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A flow or multiple flows and / or structures illustrate the steps of the functions specified in one block or multiple blocks.

[0304] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0305] The above disclosure is only the preferred embodiment of the present application, which certainly cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A data processing method, characterized in that: include: Acquire a sample training set; the sample training set includes business sample data of at least one business, the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object for the sample business, and the at least one business includes the sample business; Acquire an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to the at least one service; Perform feature interaction processing on the first object information and the first service information to obtain sample information interaction features, and determine a sample branch network corresponding to the sample service in the initial branch network corresponding to the at least one service; Outputting sample global features of the sample information interaction features through the initial shared network, and outputting sample local features of the sample information interaction features through the sample branch network; Determining a sample prediction result of the business sample data according to the sample global feature and the sample local feature; According to the sample prediction results and the object resource contribution value, the network parameters of the initial prediction model are modified to obtain a resource prediction model; the resource prediction model is used to predict the estimated resource contribution value of the object to be launched for the business to be launched.

2. The method according to claim 1, characterized in that The initial prediction model also includes an initial representation network; The performing feature interaction processing on the first object information and the first service information to obtain sample information interaction features includes: Converting the first object information into an object embedding vector, converting the first service information into a service embedding vector, and inputting the object embedding vector and the service embedding vector into the initial representation network; The object implicit features of the object embedding vector and the business implicit features of the business embedding vector are obtained through the initial representation network, and the object implicit features and the business implicit features are subjected to feature interaction processing to obtain sample information interaction features.

3. The method according to claim 2, characterized in that The initial representation network includes an object representation network and a service representation network, the object representation network includes N chain-connected first hidden layers, and the service representation network includes D chain-connected second hidden layers, where N and D are positive integers; The obtaining, through the initial representation network, object implicit features of the object embedding vector and business implicit features of the business embedding vector, performing feature interaction processing on the object implicit features and the business implicit features to obtain sample information interaction features, includes: Performing feature extraction on the service embedding vector through the D second hidden layers in the service characterization network to obtain output data of the D second hidden layers, and determining service hidden features in the output data of the D second hidden layers; Determining a candidate network layer associated with the service implicit feature among the N first hidden layers included in the object representation network; Performing feature extraction on the object embedding vector through the object representation network, obtaining output data of the candidate network layer, and determining the output data of the candidate network layer as an object implicit feature; Performing feature interaction processing on the service implicit feature and the object implicit feature to obtain an implicit interaction feature; the implicit interaction feature is used as input data of the next first implicit layer of the candidate network layer in the object representation network; If the candidate network layer is the Nth first hidden layer in the object representation network, the implicit interaction feature is determined as the sample information interaction feature.

4. The method according to claim 1, characterized in that: The sample branch network includes a branch-specific network; The step of outputting the sample global features of the sample information interaction features through the initial shared network, and outputting the sample local features of the sample information interaction features through the sample branch network, comprises: Inputting the sample information interaction characteristics into the initial shared network and the branch dedicated network; Extracting the sample information interaction features through the initial shared network to obtain sample global features; The sample information interaction features are extracted through the branch-specific network to obtain sample local features.

5. The method according to claim 4, characterized in that The sample branch network also includes a branch gating network; The method further comprises: The sample information interaction feature is input into the branch gating network, and the sample information interaction feature is analyzed through the branch gating network to obtain a sample weight feature; the sample weight feature includes a first weight corresponding to the sample global feature and a second weight corresponding to the layout information feature; Wherein, determining the sample prediction result of the business sample data according to the sample global feature and the sample local feature includes: Obtaining a first weight and a second weight in the sample weight feature, and merging a product of the first weight and the sample global feature and a product of the second weight and the sample local feature to obtain a sample fusion feature; Normalizing the sample fusion feature to obtain a first normalized feature corresponding to the business sample data; The first normalized features are classified and processed to obtain a sample prediction result of the business sample data.

6. The method according to claim 5, characterized in that The normalizing the sample fusion feature to obtain a first normalized feature corresponding to the business sample data includes: Determine a sample training subset to which the service sample data belongs, determine a sample training group corresponding to the sample service in the sample training subset, and obtain a sample mean and a sample variance corresponding to the sample training group; the sample training subset belongs to the sample training set; The sample fusion feature is normalized according to the sample mean and the sample variance to obtain a first normalized feature corresponding to the business sample data.

7. The method according to claim 5, characterized in that The initial estimation model further includes a first task output network and a second task output network, and the sample prediction result includes a first task estimation probability and a second task estimation probability; The classifying process of the first normalized feature to obtain the sample prediction result of the business sample data includes: Inputting the first normalized features into the first task output network and the second task output network; Through the first task output network, the first normalized feature is subjected to binary classification processing to obtain the first task estimated probability of the business sample data; The first normalized features are processed by the second task output network to obtain the second task estimated probability of the business sample data.

8. The method according to claim 7, characterized in that The method of modifying the network parameters of the initial prediction model according to the sample prediction result and the object resource contribution value to obtain a resource prediction model includes: Determine first label information and second label information of the service sample data according to the object resource contribution value; the first label information is used to indicate whether the sample object has made resource contribution to the sample service, and the second label information is used to indicate the resource contribution range of the sample object to the sample service; Determine a first task classification loss according to the first label information and the first task estimated probability, and determine a second task classification loss according to the second label information and the second task estimated probability; Linearly weighting the first task classification loss and the second task classification loss to obtain a model loss of the initial prediction model; The network parameters of the initial prediction model are modified according to the model loss to obtain the resource prediction model.

9. The method according to any one of claims 1 to 8, characterized in that: The resource prediction model includes an information representation network, a service sharing network, and a service branch network corresponding to at least one service; The method further comprises: Acquire the second object information of the candidate object, and acquire the second business information of the business to be launched; Through the information representation network, the second object information and the second service information are subjected to feature interaction processing to obtain information interaction features, and a candidate branch network corresponding to the service to be placed is determined in the service branch network corresponding to the at least one service; Outputting the global information features of the information interaction features through the business sharing network, and outputting the local information features of the information interaction features through the candidate branch network; According to the global information feature and the local information feature, an estimated resource contribution of the candidate object to the service to be launched is determined.

10. The method according to claim 9, characterized in that The determining, according to the global information feature and the local information feature, an estimated resource contribution of the candidate object to the service to be placed, includes: Acquire a service weight feature; the service weight feature is output by a service gating network in the candidate branch network; According to the service weight feature, weighted processing is performed on the global information feature and the local information feature to obtain a service fusion feature; Normalizing the service fusion feature to obtain a second normalized feature, and performing binary classification on the second normalized feature to obtain a third task estimated probability; Multi-classification processing is performed on the second normalized feature to obtain a fourth task estimated probability, and an estimated resource contribution value of the candidate object for the to-be-delivered business is determined according to the third task estimated probability and the fourth task estimated probability.

11. The method according to claim 10, characterized in that The fourth task estimated probability includes probability estimated values ​​corresponding to M resource contribution ranges, where M is an integer greater than 1; The determining, according to the third task estimated probability and the fourth task estimated probability, an estimated resource contribution value of the candidate object for the to-be-delivered service includes: The product of the probability estimate corresponding to the k-th resource contribution range and the k-th interval resource contribution value is determined as the k-th candidate resource contribution value, where k is a positive integer less than or equal to M; Acquire M candidate resource contribution values, and accumulate the M candidate resource contribution values ​​to obtain a cumulative resource contribution value; Determine the product of the accumulated resource contribution value and the estimated probability of the third task as the estimated resource contribution value of the candidate object for the service to be placed; Among them, the kth interval resource contribution value refers to the interval resource contribution value between the kth resource contribution range and the k-1th resource contribution range; the kth resource contribution range and the k-1th resource contribution range are two adjacent resource contribution ranges in the M resource contribution ranges.

12. The method according to claim 9, characterized in that The method further comprises: When the resource contribution estimate is greater than the resource contribution threshold, the candidate object is determined as a hit object, business display information corresponding to the to-be-delivered business is generated, and the business display information is sent to the hit object; Obtaining business interaction data of the hit object with respect to the business display information; The sample training set is updated according to the business interaction data to obtain an updated sample training set; the updated sample training set is used to train the resource prediction model.

13. A data processing device, characterized in that: include: A first acquisition module is used to acquire a sample training set; the sample training set includes business sample data of at least one business, the business sample data includes first object information of a sample object, first business information of a sample business, and an object resource contribution value of the sample object for the sample business, and the at least one business includes the sample business; A second acquisition module is used to acquire an initial prediction model; the initial prediction model includes an initial shared network and an initial branch network corresponding to the at least one service; A first feature interaction module is used to perform feature interaction processing on the first object information and the first service information to obtain a sample information interaction feature, and determine a sample branch network corresponding to the sample service in an initial branch network corresponding to the at least one service; A first feature extraction module, configured to output a sample global feature of the sample information interaction feature through the initial shared network, and output a sample local feature of the sample information interaction feature through the sample branch network; A first prediction module, used to determine a sample prediction result of the business sample data according to the sample global feature and the sample local feature; The parameter correction module is used to correct the network parameters of the initial prediction model according to the sample prediction results and the object resource contribution value to obtain a resource prediction model; the resource prediction model is used to predict the estimated resource contribution value of the object to be launched for the business to be launched.

14. A computer device, characterized in that: including memory and processor; The memory is connected to the processor, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method according to any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded and executed by a processor, so that a computer device having the processor performs the method according to any one of claims 1 to 12.

16. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 12 when being executed by a processor.

Images