Business processing method, apparatus, device, and storage medium

Abstract

Embodiments of the present application disclose a service processing method, device and equipment, and a storage medium. The method can comprise: obtaining M user features of a target user, the M user features being shared by N user interaction data extraction networks; calling the N user interaction data extraction networks to analyze and process the M user features respectively, to obtain user interaction data corresponding to each interaction operation; obtaining N resource interaction data of a target resource, and performing a service related to resource recommendation based on the N user interaction data and the N resource interaction data. Embodiments of the present application can improve the accuracy of resource recommendation and the power consumption of resource recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a business processing method and device, equipment and storage medium. BACKGROUND

[0002] With the continuous development of the Internet, users tend to interact with various resources on the network, such as reading various articles on the network and interacting with the articles by operations such as liking, commenting or forwarding; or buying goods on the network, sharing the goods with friends or forwarding the goods to a content sharing platform.

[0003] Based on this, there are more and more businesses related to resource recommendation, such as recommending matching resources to users according to the attribute information and historical resource interaction of the users; or training a resource recommendation model based on a plurality of training sample pairs (each training sample pair includes a user and a resource), and subsequently recommending matching resources to any user based on the resource recommendation model. Therefore, how to execute the business related to resource recommendation has become a hot issue in current research. SUMMARY

[0004] The embodiments of the present application provide a business processing method, device, equipment and storage medium, which improve the accuracy of the execution of the business related to resource recommendation.

[0005] In one aspect, the embodiments of the present application provide a business processing method, comprising:

[0006] Obtaining M user features of a target user, the M user features being shared by N user interaction data extraction networks corresponding to N interaction operations, N being an integer greater than 1 and M being an integer greater than 0;

[0007] Calling the N user interaction data extraction networks to analyze and process the M user features respectively to obtain user interaction data corresponding to each interaction operation; one user interaction data being used to represent the execution degree of the target user on one interaction operation;

[0008] Obtaining N resource interaction data of a target resource, one resource interaction data being used to represent the execution degree of the target resource on one interaction operation;

[0009] Executing a business related to resource recommendation based on the N user interaction data and the N resource interaction data.

[0010] In one aspect, the embodiments of the present application provide an interaction device, comprising:

[0011] An acquisition unit is configured to acquire M user features of a target user, the M user features being shared by N user interaction data extraction networks corresponding to N interaction operations, N being an integer greater than 1, and M being an integer greater than 0;

[0012] A processing unit is configured to call the N user interaction data extraction networks to respectively analyze and process the M user features, to obtain user interaction data corresponding to each interaction operation, one user interaction data being used to represent an execution degree of the target user on one interaction operation;

[0013] The acquisition unit is further configured to acquire N resource interaction data of a target resource, one resource interaction data being used to represent an execution degree of the target resource on one interaction operation;

[0014] An execution unit is configured to execute a resource recommendation related service based on the N user interaction data and the N resource interaction data.

[0015] In one aspect, an embodiment of the present application provides a terminal, comprising:

[0016] a processor adapted to implement one or more computer programs; and

[0017] a computer storage medium, the computer storage medium storing one or more computer programs, the one or more computer programs being adapted to be loaded and executed by the processor to perform the following steps:

[0018] acquire M user features of a target user, the M user features being shared by N user interaction data extraction networks corresponding to N interaction operations, N being an integer greater than 1, and M being an integer greater than 0;

[0019] call the N user interaction data extraction networks to respectively analyze and process the M user features, to obtain user interaction data corresponding to each interaction operation, one user interaction data being used to represent an execution degree of the target user on one interaction operation;

[0020] acquire N resource interaction data of a target resource, one resource interaction data being used to represent an execution degree of the target resource on one interaction operation; and execute a resource recommendation related service based on the N user interaction data and the N resource interaction data.

[0021] In one aspect, an embodiment of the present application provides a computer storage medium, characterized in that the computer storage medium stores a computer program, and the computer program is executed by a processor to perform the following steps:

[0022] M user features of a target user are acquired, the M user features are shared by N user interaction data extraction networks corresponding to N interaction operations, N is an integer greater than 1, and M is an integer greater than 0;

[0023] The N user interaction data extraction networks are called to analyze and process the M user features respectively, and user interaction data corresponding to each interaction operation is obtained; one user interaction data is used to represent an execution degree of the target user on one interaction operation;

[0024] N resource interaction data of a target resource are acquired, one resource interaction data is used to represent an execution degree of the target resource on one interaction operation; and a service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0025] In one aspect, the embodiment of the present application provides a computer program product or a computer program, the computer program product comprises a computer program, and the computer program is stored in a computer storage medium; a processor of a service processing device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the service processing device executes the following steps:

[0026] M user features of a target user are acquired, the M user features are shared by N user interaction data extraction networks corresponding to N interaction operations, N is an integer greater than 1, and M is an integer greater than 0;

[0027] The N user interaction data extraction networks are called to analyze and process the M user features respectively, and user interaction data corresponding to each interaction operation is obtained; one user interaction data is used to represent an execution degree of the target user on one interaction operation;

[0028] N resource interaction data of a target resource are acquired, one resource interaction data is used to represent an execution degree of the target resource on one interaction operation; and a service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0029] In the embodiment of the present application, M user features of a target user are acquired, and N user interaction data extraction networks are called to analyze and process the M user features, so that N user interaction data of the target user is obtained, one user interaction data extraction network is used to extract one user interaction data; then, N resource interaction data of a target resource are acquired; finally, a service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0030] In the above service processing process, the M user features are shared by the N interaction data extraction networks, and the user features do not need to be repeatedly extracted for each interaction data extraction network, so that the service processing resources can be saved. In addition, one user interaction data can represent the execution degree of a target user on an interaction operation, and one resource interaction data is used to represent the execution degree of a target resource on an interaction operation, that is, in the execution of the service related to the resource recommendation, the execution procedure of the target user on each interaction operation and the execution procedure of the target resource on each interaction operation are simultaneously referred to, so that the accuracy of service execution can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0032] Figure 1a is a flowchart of a service related to resource recommendation provided by an embodiment of the present application;

[0033] Figure 1b is another flowchart of a service related to resource recommendation provided by an embodiment of the present application;

[0034] Figure 2 is a flowchart of a service processing method provided by an embodiment of the present application;

[0035] Figure 3a is a structural diagram of a user information management module provided by an embodiment of the present application;

[0036] Figure 3b is a structural diagram of a resource information management module provided by an embodiment of the present application;

[0037] Figure 4 is a structural diagram of a resource recommendation model provided by an embodiment of the present application;

[0038] Figure 5 is another flowchart of a service processing method provided by an embodiment of the present application;

[0039] Figure 6 is a diagram of training a resource recommendation model provided by an embodiment of the present application;

[0040] Figure 7 is another flowchart of a service processing method provided by an embodiment of the present application;

[0041] Figure 8is a schematic diagram for recommending a matching resource for a user provided by an embodiment of the present application;

[0042] Figure 9 is another schematic diagram for recommending a resource for a user provided by an embodiment of the present application;

[0043] Figure 10 is a structural schematic diagram of a service processing device provided by an embodiment of the present application;

[0044] Figure 11 is a structural schematic diagram of a service processing device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0046] Artificial intelligence (AI) is to use a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making.

[0047] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation, etc.

[0048] An embodiment of the present application proposes a service processing scheme based on artificial intelligence technology. The service processing scheme is executed by a service processing device, which can be a terminal such as a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, a vehicle-mounted terminal, a smart television, etc. The service processing device can also be a server, such as a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0049] In its specific implementation, the above business processing scheme obtains M user features of the target user (which can be any user). These M user features can be shared by N user interaction data extraction networks. Each of these N user interaction data extraction networks analyzes and processes the M user features to obtain N user interaction data points. Each user interaction data extraction network extracts one user interaction data point. Since the N user interaction data extraction networks share the same M user features, there is no need to provide user features separately for each network, thus saving resource consumption in business processing. Furthermore, N resource interaction data points of the target resource are obtained, and then resource recommendation-related business is executed based on these N user interaction data points and N resource interaction data points.

[0050] A user interaction data point represents the degree to which a target user performs an interactive action. This action can include any one or more of actions such as commenting, liking, forwarding, and sharing. The degree of performance reflects the target user's interest in the interactive action; for example, a higher degree of performance indicates a higher level of interest. In other words, a user interaction data point represents the target user's preference for a particular interactive action. A resource interaction data point represents the degree to which an interactive action is performed on a target resource. This degree reflects the user's preference for performing an interactive action on that target resource. A higher degree of performance for a resource interaction data point indicates a greater preference for performing the corresponding interactive action on that resource resource; that is, the more often the user may perform that type of interactive action. Executing resource recommendation-related business based on N user interaction data points and N resource interaction data points can improve the accuracy of business execution.

[0051] Optionally, the aforementioned resource recommendation-related services may include recommending matching resources to a target user. When performing the service of recommending matching resources to a target user, the target resource may be any one of multiple candidate resources recommended to the target user. The N resource interaction data points of the target resource may be pre-acquired and stored in a resource recommendation reference database. This resource recommendation reference database may include N resource interaction data points corresponding to each of the multiple candidate resources, including the target resource. The N resource interaction data points of the target resource may be obtained by pre-calling an N resource interaction data extraction network to analyze and process the resource characteristics of the target resource.

[0052] The general process for recommending matching resources to target users using the above business processing solution can be found in [link to relevant documentation]. Figure 1a As shown, in Figure 1a The text assumes that interactive actions include commenting and sharing. Figure 1aThe illustrated business processing scheme includes a comment corresponding user interaction data extraction network 1 and a forwarding corresponding user interaction data extraction network 2. The user interaction data extraction network 1 analyzes and processes M user features to obtain user interaction data 1, and the user interaction data extraction network 1 analyzes and processes M user features to obtain user interaction data 2.

[0053] After obtaining the user interaction data 1 and the user interaction data 2, the business processing device selects a candidate resource as a target resource from the resource recommendation reference database (the target resource can be randomly selected or selected according to a certain selection rule, which will be described in detail in the embodiments), such as candidate resource 0 as the target resource, and obtains two resource interaction data corresponding to the target resource, i.e., resource interaction data 1 and resource interaction data 2. The resource interaction data 1 is obtained by analyzing and processing L resource features by the comment corresponding resource interaction data extraction network 1, and the resource interaction data 2 is obtained by analyzing and processing L resource features by the forwarding corresponding resource interaction data extraction network 2.

[0054] Then, the N user interaction data and the N resource interaction data of the target user are compared in similarity. If the similarity is greater than or equal to a similarity threshold, it means that the target resource can be recommended to the target user as a resource, and the target resource is output to the target user.

[0055] As an optional implementation, the similarity calculation between the N resource interaction data of the N target resources and the N user interaction data of the target user can include: splicing the N user interaction data to obtain a first splicing result; splicing the N resource interaction data to obtain a second splicing result; and calculating the similarity between the first splicing result and the second splicing result. It should be noted that in the resource recommendation reference database, the N resource interaction data corresponding to each candidate resource can be spliced, and then the splicing result obtained by splicing is stored in the resource recommendation reference database. According to each splicing result, the N resource interaction data corresponding to each candidate resource can be parsed. Therefore, the N resource interaction data of the target resource obtained from the resource recommendation reference database can be directly stored in the resource recommendation reference database, or can be stored in the resource recommendation reference database in the form of a splicing result. It should be understood that according to the execution degree of the target user to various interactive operations and the execution degree of the target resource preferred by the user to each interactive operation, it can be determined whether the target resource is a resource matched with the target user, which can improve the accuracy of resource recommendation.

[0056] Optionally, the resource recommendation-related service can further include training a resource recommendation model, which can include N user interaction data extraction networks and N resource interaction data extraction networks, and the N resource interaction data of the target resource is obtained by analyzing and processing the L resource features of the target resource by the N resource interaction data extraction networks in real time. For example, referring to Figure 1b An execution diagram of another resource recommendation-related service provided by an embodiment of the present application is shown in Figure 1b An embodiment of the present application has the same assumption as Figure 1a An embodiment of the present application has the same assumption as

[0057] Optionally, training the resource recommendation model based on the N user interaction data and the N resource interaction data can include: performing dot product operation on the user interaction data and the resource interaction data corresponding to the same interaction operation to obtain prediction information of the target user performing each interaction operation on the target resource, such as performing dot product operation on the user interaction data 1 corresponding to the comment and the resource interaction data 1 corresponding to the comment to obtain prediction information 1 of the target user performing the comment on the target resource; for another example, performing dot product operation on the user interaction data 2 corresponding to the forwarding and the resource interaction data 2 corresponding to the forwarding to obtain prediction information 2 of the target user performing the forwarding on the target resource; further, obtaining label information of the target user performing each interaction operation on the target resource; determining the value of the loss function corresponding to each interaction operation based on the prediction information and the label information; then adding the values of the loss functions to obtain the loss function value of the resource recommendation model, and then adjusting the parameters of the resource recommendation model in the direction of reducing the loss function value. The parameters herein can include the parameters in the user interaction data extraction network and the parameters in the resource interaction data extraction network.

[0058] In the above resource recommendation model training, multi-task learning is performed based on the same resource features or user features in one resource recommendation model (analyzing the execution degree of the user on the multiple interaction operations, or the execution procedure of the target resource being executed by multiple interaction operations), which realizes feature sharing between multiple tasks and saves the power consumption overhead of separately extracting features for each task.

[0059] Based on the above service processing scheme, an embodiment of the present application provides a service processing method, referring to Figure 2 A flow diagram of a service processing method provided by an embodiment of the present application is shown in Figure 2 The flow processing method shown can be executed by a service processing device, and specifically can be executed by a processor of the service processing device. Figure 2 The service processing method shown can include the following steps:

[0060] Step S201, obtaining M user features of a target user.

[0061] The M user features can be determined based on user attribute information, and the M user features can be understood as being obtained by fusing and crossing the user attribute information. Alternatively, the M user features can be obtained by calling an M user feature processing network to process user attribute information of the target user. In a specific implementation, obtaining the M user features of the target user can include: obtaining user attribute information of the target user, the user attribute information including any one or more of the following: age, gender, and historical preference information for N interactive operations of the target user; performing vector conversion processing on the user attribute information to obtain a user embedding vector; and calling the M user feature processing network to cross and abstract the user embedding vector to obtain the M user features.

[0062] The historical preference information for N interactive operations can refer to resources that the target user likes, or resources that the target user likes to perform any one of the N interactive operations, and the like. The vector conversion processing on the user attribute information actually refers to Embedding of the user attribute information. Embedding refers to converting sparse and high-dimensional original features into low-dimensional and dense vector representations. The number of user attribute information can be one or more. When the number of user attribute information is more than one, the vector conversion processing on the user attribute information can include: performing vector conversion processing on each user attribute information to obtain a user sub-embedding vector corresponding to each user attribute information; and splicing the user sub-embedding vectors to obtain a user embedding vector.

[0063] For example, assuming that the number of user attribute information of the target user is W, the i-th user attribute information can be represented as X i , where i is an integer greater than or equal to 1 and less than or equal to W, and the vector conversion processing on the i-th user attribute information to obtain a user sub-embedding vector corresponding to the i-th user attribute information can be represented as e i = emb(X i ), e i represents the user sub-embedding vector corresponding to the i-th user attribute information, and emb() represents Embedding processing. Splicing all user sub-embedding vectors to obtain user features of the target user can be represented as: represents the user features, and concat() represents splicing processing.

[0064] The M user feature processing networks can have the same network structure or different network structures. In the embodiments of the present application, it is assumed that the M user feature processing networks have the same structure. The structure of each user feature processing network in the M user feature processing networks in the embodiments of the present application is described below by taking any user feature processing network as an example. Any user feature processing network can be denoted as ExpertU. Any user feature processing network can be a multilayer perceptron (MLP), which can also be called an artificial neural network (ANN), and includes an input layer, an output layer and a hidden layer. The layers of the multilayer perceptron are fully connected. The bottom layer is the input layer, the middle layer is the hidden layer, and the last layer is the output layer. Any user feature processing network can include an input layer, three hidden layers and an output layer, and the layers are fully connected. The three hidden layers in the MLP are fully connected, so the user feature processing network can be simply understood as including three fully connected layers (mainly referring to including three fully connected hidden layers). The user feature processing network is called to perform feature cross and feature abstraction processing on the user latent vector to obtain the user feature of the target user, and specifically includes: calling the input layer of the MLP to obtain the user latent vector; the input layer transmits the user latent vector to the hidden layer, each hidden layer performs nonlinear transformation operation on the user latent vector, the calculation result of the previous hidden layer is used as the input of the next hidden layer, and the output result of the last hidden layer is transmitted to the output layer of the MLP, and the output result is transmitted to the network or module connected with the MLP by the output layer.

[0065] For example, any user feature processing network performs feature cross and feature abstraction processing on the user latent vector to obtain a user feature, which can be represented by the following formula (1):

[0066]

[0067] In formula (1), represents a user feature obtained by performing feature cross and feature abstraction processing on the user latent vector by the jth user feature processing network, the jth user feature processing network is any one of the at least one user feature processing network, and MLP() represents nonlinear transformation operation of a hidden layer.

[0068] In step S202, N user interaction data extraction networks are called to analyze and process the M user features respectively to obtain user interaction data corresponding to each interaction operation.

[0069] The M user features obtained by the step S201 can be shared by the N user interaction data extraction networks, so that the M user features of the target user can be calculated for each user interaction data extraction network, and the power consumption of the service processing device is saved.

[0070] In one embodiment, the N user interaction data extraction networks and the M user feature processing networks can be deployed in a user information management module, and the N user interaction data extraction networks are connected in parallel with the user feature processing networks. Referring to Figure 3a A structural diagram of a user information management module is provided for the embodiments of the present application, 301 represents the user information management module, 302 represents the M user feature processing networks, and 303 represents the N user interaction data extraction networks. From Figure 3a It can be seen that each user feature processing network is connected with the N user interaction data extraction networks, that is, the user features obtained by each user feature processing network are output to each user interaction data extraction network.

[0071] For example, M is equal to 2, N is equal to 2, the M user feature processing networks include a user feature processing network 1, which can be represented as Expert1U, and a user feature processing network 2, which can be represented as Expert2U; the N user interaction data extraction networks include a user interaction data extraction network A, which can be represented as Tower AU, and a user interaction data extraction network B, which can be represented as Tower BU. It is assumed that the interaction operation corresponding to the user interaction data extraction network A is a comment, and the interaction operation corresponding to the user interaction data extraction network B is a forward. The user features output by the Expert1U are output to the Tower AU and the Tower BU, and the user features output by the Expert2U are also output to the Tower AU and the Tower BU. The user features are converted into user latent vectors, and the user latent vectors are used as inputs of the Expert1U and the Expert2U.

[0072] In one embodiment, the user information management module further includes N user interaction weight management networks such as Figure 3a For example, the N user interaction weight management networks can be represented as Gate AU and Gate BU. Each user interaction weight management network corresponds to an interaction operation object. The user interaction weight management network corresponding to each interaction operation is used to generate a user interaction weight value for the corresponding interaction operation based on the user latent vector. The user interaction weight value corresponding to each interaction operation acts on the input of the user interaction data extraction network corresponding to the corresponding interaction operation. For example, in Figure 3aIn some embodiments, the user interaction weight management network Gate AU generates a user interaction weight value corresponding to the comment based on the user latent vector, and the user interaction weight value acts on the input of the user interaction data extraction network Tower AU. Figure 3a In some embodiments, the user interaction weight management network Gate BU generates a user interaction weight value corresponding to the forwarding based on the user latent vector, and the user interaction weight value acts on the input of the user interaction data extraction network Tower BU.

[0073] Based on the above, when the service processing device calls N user interaction data extraction networks to analyze and process M user features respectively, to obtain user interaction data corresponding to each interaction operation, the service processing device can specifically include: calling each user interaction weight management network in the N user interaction weight management networks to perform a weight generation operation on the user latent vector, to obtain a user interaction weight value corresponding to each interaction operation; multiplying the user interaction weight value corresponding to the i-th interaction operation with each user feature in the M user features; adding each multiplication result, and inputting the addition result into the i-th user interaction data extraction network corresponding to the i-th interaction operation, i being an integer greater than or equal to 1 and less than or equal to N; calling the i-th user interaction data extraction network to analyze and process the addition result, to obtain user interaction data corresponding to the i-th interaction operation.

[0074] In other words, the input of the i-th user interaction data extraction network is the operation result of adding the M user features based on the user interaction weight value corresponding to the i-th interaction operation. Assuming that the i-th user interaction data extraction network can be referred to as the user interaction data extraction network Tower AU, the input of the i-th user interaction data extraction network can be represented by the following formula (2): Figure 3a

[0075]

[0076] In formula (2), the input of the i-th user interaction data extraction network is represented by , the j-th user feature output by the j-th user feature processing network is represented by , the i-th user feature processing network can be any one of the M user feature processing networks, , and the user interaction weight value corresponding to the i-th interaction operation is represented by

[0077] For example, referring to Figure 3a ​In this context, assuming the user feature output by the user feature processing network Expert1U is x1, and the user feature output by the user feature processing network Expert2U is x2, the user interaction weight management network Gate AU, which generates the user interaction weight value for the comment interaction based on the user's latent vector, is represented as w1, and the user interaction weight authority module Gate BU, which generates the user interaction weight value for the forwarding interaction based on the user's latent vector, is represented as w2, then the input to the user interaction data extraction network Tower AU should be: x1*w1+x2*w1; and the input to the user interaction data extraction network Tower BU should be x1*w2+x2*w2.

[0078] As mentioned above, after obtaining the input to each user interaction data extraction network, each network is called to analyze and process its own input. Specifically, the user interaction data extraction network Tower AU analyzes and processes x1*w1+x2*w1 to obtain the user interaction data corresponding to the comment. Optionally, any user interaction data extraction network can be a multilayer perceptron, and may include a hidden layer. The network analyzes and processes the sum of the weighted multiplications of the aforementioned user features to obtain the user interaction data corresponding to any interaction operation. This can include: calling the hidden layer in any user interaction data extraction network to perform a nonlinear transformation on the summation result to obtain the user interaction data corresponding to any interaction operation.

[0079] For example, the network for extracting user interaction data for the i-th user can be Figure 3a The Tower AU in the formula (2) processes each user feature input to the i-th user interaction data extraction network using formula (2) to obtain the input of the i-th user interaction data extraction network, which is the input in formula (2). As the input of the i-th user interaction data extraction network, the i-th user interaction data extraction network performs nonlinear transformation on the input to obtain the user interaction data corresponding to the i-th interaction operation, which can be specifically expressed as the following formula (3):

[0080]

[0081] In formula (3), This represents the user interaction data corresponding to the i-th type of interactive operation, and MLP() represents nonlinear transformation processing.

[0082] Step S203: Obtain N resource interaction data of the target resource.

[0083] Optionally, the N resource interaction data can be obtained by analyzing and processing the L resource features by calling the N resource interaction data extraction network in real time; or the N resource interaction data can be obtained by analyzing and processing the L resource features by calling the N resource interaction data extraction network in advance, and pre-stored in the resource recommendation reference database. The resource recommendation reference database can include N resource interaction data of a plurality of candidate resources, and the target resource can be any one of the plurality of candidate resources.

[0084] A resource interaction data extraction network is used to analyze the L resource features to obtain resource interaction data corresponding to a resource interaction. That is, a resource interaction data extraction network corresponds to an interaction operation, and a resource interaction data corresponds to an interaction operation.

[0085] The L resource features can be obtained by processing the historical interaction information of the target resource by calling the L resource feature processing network. In a specific implementation, the L resource features can be obtained by processing the historical interaction information by calling the L resource feature processing network, which can include: obtaining the historical interaction information of the target resource, the historical interaction information including information indicating that the target resource is executed N interaction operations in a historical time; performing vector conversion on the historical interaction information to obtain a resource hidden vector; and calling the L resource feature processing network to perform feature intersection and feature abstraction processing on the resource hidden vector to obtain the L resource features. When obtaining the historical interaction information of the target resource, the resource identifier of the target resource can also be obtained.

[0086] The number of historical interaction information is at least one, and the resource hidden vector can be obtained by performing vector conversion on the historical interaction information, which can include: performing Embedding processing on each historical interaction information to obtain a resource sub-vector; and performing splicing processing on a plurality of resource hidden vectors to obtain a resource hidden vector. For example, the number of historical interaction information is H, and any historical interaction information can be represented as y i , i is an integer greater than or equal to 1 and less than or equal to P. Any historical interaction information can be obtained by performing Embedding to obtain a resource sub-vector, which can be represented by the formula: e i = emb(y i ), wherein e i represents any resource sub-vector.

[0087] The H historical interaction information is processed by performing Embedding to obtain H resource sub-vectors, and the H resource sub-vectors are spliced to obtain a resource hidden vector, which can be represented as: Herein represents the resource hidden vector.

[0088] The structure of the N resource interaction data extraction network is the same as the structure of the N user interaction data extraction network in step S202, and can also be an input layer, three (or other number, three is taken as an example in the embodiment of the application) hidden layers, and an output layer. The N resource interaction data extraction network is called to analyze and process the M resource features to obtain the resource interaction data corresponding to each interaction operation, which can include: each hidden layer of each resource interaction data extraction network is called to perform nonlinear transformation operation on the M resource features, the nonlinear transformation processing result of the previous hidden layer is taken as the input of the next hidden layer, and the nonlinear transformation operation of the last hidden layer is taken as a resource interaction data.

[0089] The calculation formula of this part is the same as formula (1) above, and the difference from formula (1) is that the input hidden vector in this formula is a resource hidden vector, not a user hidden vector. The parameters for nonlinear transformation operation of the resource hidden vector and the user hidden vector in the same hidden layer can be different.

[0090] In an embodiment, the L resource feature processing networks and the N resource interaction data extraction networks can be deployed in the resource information management module. Referring to Figure 3b A schematic diagram of a resource information management module provided in an embodiment of the application is shown in Figure 3b In FIG. 31, 31 represents a resource information management module, 32 represents N resource interaction data extraction networks, and 33 represents L resource feature processing networks. Figure 3b It can be seen that each resource feature processing network is connected with the N resource interaction data extraction networks, that is, the resource features obtained by each resource feature processing network are output to each resource interaction data extraction network.

[0091] For example, L is equal to 2, N is equal to 2, the L resource feature processing networks can include a resource feature processing network 1 represented as Expert1I and a resource feature processing network 2 represented as Expert2I, and the N resource interaction data extraction networks include a resource interaction data extraction network A represented as Tower AI and a resource interaction data extraction network B represented as Tower BI. It is assumed that the interaction operation corresponding to the resource interaction data extraction network A is a comment, and the interaction operation corresponding to the resource interaction data extraction network B is a forward. The resource features output by Expert1I are output to Tower AI and Tower BI respectively, and similarly, the resource features output by Expert2I are also output to Tower AI and Tower BI respectively. It should be understood that, here, Tower AI and Tower BI are the same as the above-mentioned Figure 3aTower AU and Tower BU can be networks with the same structure, and Expert1I and Expert2I can also be networks with the same structure as those mentioned above. Figure 3a Expert1U and Expert2U can be networks with the same structure, or they can be different networks.

[0092] In one embodiment, the resource information management module may further include N resource interaction weight management networks, such as... Figure 3b As shown in Figure 34, for example, a network for managing the interaction weights of N resources can be represented as Gate AI and Gate BI. Here, Gate AI and Gate BI are the same as those mentioned above. Figure 3a In this context, Gate AU and Gate BU can be the same network or different networks. Each resource interaction weight management network corresponds to a specific interaction operation; for example, Gate AI corresponds to the comment operation, and Gate BI corresponds to the share operation. Each resource interaction weight management network is used to generate resource interaction weight values ​​for each interaction operation based on the resource latent vector.

[0093] Each resource interaction weight management network generates a resource interaction weight value for each interaction operation, which is then applied to the input of the corresponding resource interaction data extraction network. For example, in... Figure 3b In the first example, the resource interaction weight management network Gate AI generates a resource interaction weight value corresponding to a comment based on the resource latent vector, and this resource interaction weight value is applied to the input of the resource interaction data extraction network 1; similarly, the resource interaction weight management network Gate BI generates a resource interaction weight value corresponding to a repost based on the resource latent vector, and this resource interaction weight value is applied to the input of the resource interaction data extraction network 2.

[0094] Based on this, when the aforementioned business processing equipment calls N resource interaction data extraction networks to analyze and process L resource features respectively, and obtains the resource interaction data corresponding to each interaction operation, it may specifically include:

[0095] Each of the N resource interaction weight management networks is invoked to generate weights for the resource latent vectors, obtaining the resource interaction weight value corresponding to each interaction operation. The resource interaction weight value corresponding to the j-th interaction operation is multiplied by each of the L resource features. The results of each multiplication operation are added together, and the result is input into the j-th resource interaction data extraction network corresponding to the j-th interaction operation, where j is an integer greater than or equal to 1 and less than or equal to N. The j-th resource interaction data extraction network is invoked to analyze and process the addition result to obtain the resource interaction data corresponding to the j-th interaction operation.

[0096] In other words, the input of the jth resource interaction data extraction network is the operation result of adding the L resource features based on the resource interaction weight value corresponding to the jth interaction operation. Assuming that the jth resource interaction data extraction network can be referred to as Tower AI, the input of the jth resource interaction data extraction network can be represented by the following formula (4): Figure 3b In the formula (4),

[0097]

[0098] In the formula (4), represents the input of the jth resource interaction data extraction network, represents the resource feature output by the nth resource feature processing network, n is an integer greater than or equal to 1 and less than or equal to L, represents the resource interaction weight value generated by the resource interaction weight management network Gate AI based on the resource hidden vector.

[0099] For example, referring to Figure 3b In the formula (4), the resource feature output by the resource feature processing network Expert1I is y1, the resource feature output by the resource feature processing network Expert2I is y2, the resource interaction weight value generated by the resource interaction weight management network Gate AI based on the resource hidden vector for the comment interaction operation is represented as w3, and the resource interaction weight value generated by the resource interaction weight management network Gate BI based on the resource hidden vector for the forwarding interaction operation is represented as w4. Then, the input of the resource interaction data extraction network Tower AI should be y1*w3+y2*w3, and the input of the resource interaction data extraction network Tower BI should be y1*w4+y2*w4.

[0100] As described above, after obtaining the input of each resource interaction data extraction network, each resource interaction data extraction network is called to analyze and process the respective input, i.e., the resource interaction data extraction network Tower AI analyzes and processes y1*w3+y2*w3 to obtain the resource interaction data corresponding to the comment, and the resource interaction data extraction network Tower BI analyzes and processes y1*w4+y2*w4 to obtain the resource interaction data corresponding to the forwarding.

[0101] Optionally, any resource interaction data extraction network can be a multi-layer perception machine, any resource interaction data extraction network can include a hidden layer, and any resource interaction data extraction network analyzes and processes the addition operation result of the above-mentioned various resource features after weight multiplication to obtain the resource interaction data corresponding to any interaction operation, which can include: calling the hidden layer in any resource interaction data extraction network to perform nonlinear transformation processing on the addition operation result to obtain the resource interaction data corresponding to any interaction operation.

[0102] For example, for the resource interaction data extraction network Tower AU, the input of each resource feature input to the Tower AU is processed by formula (4) to obtain the input of the resource interaction data extraction network Tower AU, that is, the input of formula (4) is Figure 3b As the input of the Tower AU, the Tower AU performs nonlinear transformation processing on the input to obtain the resource interaction data corresponding to the comment interaction operation, which can be specifically represented by the following formula (5):

[0103]

[0104] In formula (5), represents the resource interaction data corresponding to the comment interaction operation, and MLP() represents nonlinear transformation processing.

[0105] Step S204, performing a service related to resource recommendation based on the N user interaction data and the N resource interaction data.

[0106] Optionally, the service related to resource recommendation can include recommending a matched resource to the user, in which case the N resource interaction data of the target resource obtained in step S203 can be obtained from a resource recommendation reference database. The resource recommendation reference database stores N resource interaction data of a plurality of candidate resources, or the resource recommendation reference database can also store a splicing result of the N resource interaction data corresponding to each candidate resource.

[0107] In an embodiment, if the service related to resource recommendation refers to recommending a matched resource to the user, performing the service related to resource recommendation based on the N user interaction data and the N resource interaction data includes: splicing the N user interaction data to obtain a first splicing result, and splicing the N resource interaction data to obtain a second splicing result; if the similarity between the first splicing result and the second splicing result is greater than or equal to a similarity threshold, the target resource is determined as the resource matched to the target user.

[0108] Optionally, the service related to resource recommendation can refer to training a resource recommendation model, and the user information management module and the resource information management module described above can be deployed in the resource recommendation model, at which time the resource recommendation model constituted by the user information management module and the resource information management module can be a double-tower model. The so-called double-tower model refers to modeling the user information and the resource information respectively, and representing them as low-dimensional vectors to facilitate fast calculation of the similarity between them. See Figure 4 FIG. 1 is a schematic diagram of a resource recommendation model provided by an embodiment of the present application, which can include the user information management module 301 and the resource information management module 31 described above.​

[0109] The output of the N user interaction data extraction network in the user information management module 301 and the output of the N resource interaction data extraction network in the resource information management module 31 are used to predict the prediction information of the target user performing each interaction operation on the target resource, and then based on the prediction information of each interaction operation and the label information of the target user performing each interaction operation on the target resource, the loss function value of the resource recommendation model is obtained, and the resource recommendation model is optimized in the direction of reducing the loss function value. The above process is as shown in 401 of Figure 4 .

[0110] When the business related to resource recommendation refers to training the resource recommendation model, the N resource interaction data can be obtained by real-time calling the N resource interaction data extraction network based on the analysis and processing of the M resource characteristics, at this time, the target user and the target resource form a training sample pair, and the label information of the target user performing each interaction operation on the target resource is given. The label information of each interaction operation is used to indicate the situation of the target user performing the interaction operation on the target resource, such as the target user performing the interaction operation on the target resource, or the target user not performing the interaction operation on the target resource.

[0111] In this case, the business related to resource recommendation is performed based on the N user interaction data and the N resource interaction data, and the approximate process can include: determining the prediction information of the target user performing each interaction operation on the target resource based on the N user interaction data and the N resource interaction data; then determining the value of the loss function corresponding to the corresponding interaction operation based on the difference between the prediction information of the target user performing each interaction operation on the target resource and the label information of the target user performing the corresponding interaction operation on the target resource; adding the value of the loss function corresponding to each interaction operation to obtain the loss function value of the resource recommendation model, and adjusting the parameters of the resource recommendation model in the direction of reducing the loss function value of the resource recommendation model.

[0112] Optionally, when the business related to resource recommendation is to recommend matching resources to the target user, the user information management module and the resource information management module can also be deployed in the same resource recommendation module, that is, the resource recommendation model structure as shown in Figure 4 can be used to perform the business of training the resource recommendation model, and can also perform the business of recommending matching resources to the target user, as shown in 402 of Figure 4 .

[0113] In this embodiment of the invention, user characteristics of the target user are obtained, and N user interaction data extraction networks are invoked to analyze and process these user characteristics to obtain N user interaction data of the target user, with each user interaction data extraction network used to extract one user interaction data. Then, N resource interaction data of the target resource are obtained. Finally, based on the N user interaction data and the N resource interaction data, business related to resource recommendation is executed.

[0114] In the aforementioned business processing, a user feature is shared by N interaction data extraction networks, eliminating the need to extract a corresponding user feature for each network, thus saving business processing resources. Furthermore, a user interaction data point can represent the degree to which a target user performs an interaction, while a resource interaction data point represents the degree to which an interaction is performed on a target resource. When performing resource recommendation-related business based on N user interaction data points and N resource interaction data points, the accuracy of business execution can be improved.

[0115] Based on the above-described business processing method, this embodiment of the invention provides another business processing method, see [link to relevant documentation]. Figure 5 This is a flowchart illustrating another business processing method provided in an embodiment of the present invention. Figure 5 The flowchart of the business processing method shown can be executed by the business processing device, specifically by the processor of the business processing device. Figure 5 The business processing methods shown may include training a resource recommendation model, the structure of which can be as follows: Figure 4 As shown, it includes a user information management module and a resource information management module. The user information management module includes M user feature processing modules and N user interaction data extraction networks, and the resource information management module includes L resource feature processing modules and N resource interaction data extraction networks. Figure 5 The business processing method may include the following steps:

[0116] Step S501: Obtain M user characteristics of the target user.

[0117] Step S502: Call the N user interaction data extraction network to analyze and process the M user features to obtain user interaction data corresponding to each interaction operation. One user interaction data is used to represent the degree of execution of an interaction operation by the target user.

[0118] Step S503: Obtain L resource features of the target resource.

[0119] Step S504, calling the N resource interaction data extraction networks to analyze and process the L resource features, to obtain resource interaction data corresponding to each interaction operation, one resource interaction data being used to represent the executed degree of the target resource being executed by the target user in one interaction operation.

[0120] In one embodiment, some feasible implementation manners included in steps S501-S504 can refer to Figure 2 The description of related steps in the embodiments will not be repeated here.

[0121] Step S505, determining the prediction information of the target user performing each interaction operation on the target resource according to the N user interaction data and the N resource interaction data.

[0122] The prediction information of the target user performing each interaction operation on the target resource is used to represent the probability of the target user possibly performing one interaction operation on the target resource. Next, taking the ith interaction operation as an example, how to determine the prediction information of the target user performing each interaction operation on the target resource is introduced. In a specific implementation, the following can be included: performing dot product operation on the ith user interaction data and the ith resource interaction data; performing nonlinear conversion on the dot product operation result to obtain the prediction information of the target user performing the ith interaction operation on the target resource; wherein i is an integer greater than or equal to 1 and less than or equal to N.

[0123] For example, refer to Figure 6 A schematic diagram of training a resource recommendation model provided by the embodiment of the present application is shown in FIG. 6, 600 represents the resource recommendation model, the resource recommendation model includes a user information management module 601 and a resource information management module 602, the structure of the user information management module 601 can be as shown in the above Figure 3a The structure of the resource information management module 602 can be as shown in the above Figure 3b Suppose that in the Figure 6 user information management module 601, the user interaction data extraction network corresponding to the ith interaction operation is Tower AU, and in the resource information management module 602, the user interaction data extraction network corresponding to the ith interaction operation is Tower AI, the outputs of Tower AU and Tower AI are subjected to dot product operation, which can be represented by the following formula (6):

[0124]

[0125] In formula (6), represents the output of the user interaction data extraction network corresponding to the ith interaction operation being Tower Au, represents the output of the user interaction data extraction network corresponding to the i-th interactive operation Tower AI, and log its(A) represents the dot product operation result.

[0126] Alternatively, after obtaining the dot product operation result, further nonlinear operation can be performed on the dot product operation result to obtain the prediction information of the target user performing the i-th interactive operation on the target resource. In a specific implementation, the nonlinear conversion of the dot product operation result can be implemented by a Sigmoid function, which is a nonlinear function, and its expression can be: sigmoid(x) = 1 / (1+e -x ), where x represents the input of the function. Then, the prediction information of the target user performing the i-th interactive operation on the target resource obtained by performing nonlinear conversion on the dot product operation result based on the Sigmoid function can be represented as formula (7) as follows:

[0127] y(A) = Sigmoid(log its(A)) (7)

[0128] In formula (7), y(A) represents the prediction information of the target user performing the i-th interactive operation on the target resource.

[0129] Step S506, obtaining the label information of the target user performing each interactive operation on the target resource.

[0130] The label information of the target user performing each interactive operation on the target resource is pre-set, and the label information of the target user performing each interactive operation on the target resource is used to represent the condition of the target user performing an interactive operation on the target resource, such as the target user performing a comment interactive operation on the target resource, or the target user not performing a comment interactive operation on the target resource, etc.

[0131] Step S507, obtaining the loss function corresponding to each interactive operation, and determining the value of the loss function corresponding to the corresponding interactive operation between the prediction information of the target user performing each interactive operation on the target resource and the label information of the target user performing the corresponding interactive operation on the target resource.

[0132] Wherein, the loss functions corresponding to various interactive operations can be the same, which can be Cross Entropy, or can be different. In the embodiment of the application, it is assumed that the loss functions corresponding to each interactive operation are the same. Cross Entropy is a loss function used to measure the difference between two probability distributions.

[0133] After obtaining the predicted information of the target user performing the i-th interaction operation on the target resource based on the above formula (7), the label information of the target user performing the i-th interaction operation on the target resource and the predicted information are substituted into the above loss function, and the value of the loss function corresponding to the i-th interaction operation can be obtained, which can be specifically expressed as formula (8):

[0134]

[0135] In formula (8), loss(A) represents the value of the loss function corresponding to the i-th interaction operation, represents the label information of the target user performing the i-th interaction operation on the target resource, and y(A) represents the predicted information of the target user performing the i-th interaction operation on the target resource; it is assumed that y(A) is simply written as y, and is simply written as Then the calculation of can be expressed as the following formula:

[0136]

[0137] Step S508: Adding the values of the loss functions corresponding to each interaction operation to obtain the loss function value of the resource recommendation model, and adjusting the parameters of the resource recommendation model in the direction of reducing the loss function value of the resource recommendation model.

[0138] Based on the above formula (8), the value of the loss function corresponding to each interaction operation can be calculated, and the values of two loss functions are obtained in FIG. 6. Then, the values of the loss functions corresponding to each interaction operation are added to obtain the loss function value of the resource recommendation model.

[0139] Then, the parameters in the resource recommendation model are adjusted in the direction of reducing the loss function value of the resource recommendation model. The parameters in the resource recommendation model can include the parameters of each network in the user information management module 601, such as the parameters of the M user feature processing networks, the parameters of the N user interaction weight management networks, and the parameters of the N user interaction data extraction networks; the parameters in the resource recommendation model can also include the parameters of each network in the resource information management module 602, such as the parameters of the L resource feature processing networks, the parameters of the N resource interaction weight management networks, and the parameters of the N resource interaction data extraction networks, and the like.

[0140] As can be known from the above, when the service related to the resource recommendation is training the resource recommendation model, the target user and the target resource form a training sample pair, and the above only takes this training sample pair as an example to introduce the training process of the resource recommendation model based on this training sample pair. In practice, the service processing device also needs to obtain other training sample pairs to train the resource recommendation model. When the training sample pair is determined, the obtained samples need to be preprocessed, such as filtering out samples with too short exposure time from the resource samples, and the like.

[0141] In the embodiment of the application, the user features of the target user are obtained, and N user interaction data extraction networks are called to analyze and process the M user features to obtain N user interaction data of the target user. One user interaction data extraction network is used to extract one user interaction data. Then, L resource features of the target resource are obtained, and N resource interaction data extraction networks are called to analyze and process the M resource features to obtain resource interaction data corresponding to each interaction operation. Finally, the prediction information of the target user performing each interaction operation on the target resource is determined based on the N user interaction data and the N resource interaction data, and the label information of the target user performing each interaction operation on the target resource is obtained. The value of the loss function corresponding to each interaction operation is determined based on the prediction information and the label information of the target user performing each interaction operation on the target resource. Finally, the values of the loss functions corresponding to each interaction operation are added to obtain the loss function value of the resource recommendation model, and the parameters of the resource recommendation model are adjusted in the direction of reducing the loss function value of the resource recommendation model. In the above service execution process, the M user features are shared by the N interaction data extraction networks, and there is no need to extract user features for each interaction data extraction network. In this way, the service processing resources can be saved.

[0142] Based on the above service processing method, another service processing method is provided in the embodiment of the application, which is described with reference to Figure 7 The flowchart of another service processing method provided in the embodiment of the application is shown in FIG. 8. Figure 7 The service processing method is executed by a service processing device, and can be executed by a processor of the service processing device. Figure 7 In the service processing method shown in FIG. 8, the service related to the resource recommendation can include recommending a matched resource to the target user. Figure 7 The service processing method can include the following steps:

[0143] In step S701, M user features of a target user are obtained, and the M user features are shared by N user interaction data extraction networks. N is an integer greater than 1.

[0144] Step S702, calling the N user interaction data extraction networks to respectively analyze and process the M user features to obtain user interaction data corresponding to each interaction operation. One user interaction data is used to represent the execution degree of the target user on one interaction operation.

[0145] In one embodiment, some feasible implementations included in steps S701 and S702 can refer to the description of related steps in the embodiments, which will not be repeated here. Figure 2 In one embodiment, some feasible implementations included in steps S701 and S702 can refer to the description of related steps in the embodiments, which will not be repeated here.

[0146] Step S703, obtaining a resource recommendation reference database, and obtaining N resource interaction data corresponding to the target resource from the resource recommendation reference database.

[0147] As described above, the resource recommendation reference database can store N resource interaction data corresponding to each candidate resource, or can store splicing results of N resource interaction data of each candidate resource. The N resource interaction data corresponding to each candidate resource can also be obtained through the splicing result of each candidate resource.

[0148] Optionally, obtaining the N resource interaction data corresponding to the target resource from the resource recommendation reference database includes: randomly determining a candidate resource that has not been selected as the target resource from the resource recommendation reference database, and obtaining the N resource interaction data corresponding to the target resource.

[0149] Optionally, the plurality of candidate resources are divided into P resource groups, each resource group includes at least one candidate resource and N resource interaction data corresponding to each candidate resource, each resource group corresponds to an index identifier, and the index identifier corresponding to any resource group is determined based on the N resource interaction data corresponding to each candidate resource in the any resource group; obtaining the N resource interaction data corresponding to the target resource from the resource recommendation reference database includes: splicing the N user interaction data to obtain a first splicing result; selecting a target resource group with an index identifier matching the first splicing result from the P resource groups; selecting any candidate resource that has not been selected as the target resource from the target resource group, and obtaining the N resource interaction data corresponding to the target resource.

[0150] In one embodiment, the index identifier of any resource group is determined based on the N resource interaction data corresponding to each candidate resource in the any resource group, including: splicing the N resource interaction data corresponding to each candidate resource in the any resource group to obtain a plurality of splicing results; and then performing an average operation on the plurality of splicing results to obtain the index identifier corresponding to the any resource group.

[0151] It should be understood that, first, based on the first splicing result, one resource group is determined from the plurality of resource groups, and then a target sub-resource is acquired from the resource group, so as to judge whether the target resource matches the target user, which can reduce the search range of the matching resource, improve the search efficiency, and save the power consumption of the search resource.

[0152] In step S704, the N user interaction data are spliced to obtain a first splicing result, and the N resource interaction data are spliced to obtain a second splicing result.

[0153] The splicing of the N user interaction data to obtain the first splicing result can include: acquiring a splicing weight value corresponding to each interaction operation; multiplying each user interaction data and the splicing weight value of the interaction operation corresponding to the corresponding user interaction data; and splicing the results of the multiplication operations to obtain the first splicing result.

[0154] It should be understood that, according to the business requirements, different splicing weight values can be set for different interaction operations, such as hoping to recommend more resources that can trigger user sharing to the user, in which case, the splicing weight value corresponding to the sharing interaction operation needs to be increased. Therefore, in general, the corresponding splicing weight values can be set for different interaction operations according to the business requirements, and each user interaction data is spliced according to the splicing weight value corresponding to each interaction operation, so as to recommend resources that are more in line with the business requirements to the user.

[0155] As described above, when the resource recommendation business refers to recommending matching resources to the target user, the user information management module and the resource information management module can also be deployed in the same resource recommendation model, and the specific structure can be as shown in the above Figure 6 As described above, when the resource recommendation business refers to recommending matching resources to the target user, the user information management module and the resource information management module can also be deployed in the same resource recommendation model, and the specific structure can be as shown in the above

[0156] To be consistent with the prediction target of the fine ranking model, when recommending the target resource for the target user, a recall comprehensive score can be calculated according to the N user interaction data of the target user and the N resource interaction data of the target resource, and the recall score can be used as a reference for subsequent coarse ranking and fine ranking.

[0157] Specifically, the user interaction data and the resource interaction data under the same interaction operation can be subjected to dot product operation, and then a recommendation weight value is determined for the dot product operation of one interaction operation, and the recall comprehensive score is obtained by adding the dot product operation results of each interaction operation based on the recommendation weight value of each interaction operation. Here, the recommendation weight value corresponding to each interaction operation is determined based on the prediction target of the fine ranking model. For example, if the prediction target of the fine ranking model is to recommend more resources that can trigger user sharing for the user, then the recommendation weight value corresponding to the sharing interaction operation is greater than the recommendation weight values corresponding to other interaction operations.

[0158] For example, in the resource recommendation model shown in the above Figure 6 The user interaction data corresponding to the comment interaction operation output by the user interaction data extraction network Tower AU included in the user information management module is represented as The resource interaction data corresponding to the comment interaction operation output by the resource interaction data extraction network Tower AI included in the resource information management module is represented as The user interaction data corresponding to the forwarding interaction operation output by the user interaction data extraction network Tower BU is represented as The resource interaction data corresponding to the comment interaction operation output by the resource interaction data extraction network Tower BI is represented as

[0159] Suppose the recommendation weight value set for the comment interaction operation is w a The recommendation weight value set for the forwarding interaction operation is w b Then the recall comprehensive score can be calculated by the following formula (9):

[0160]

[0161] Step S705, if the similarity between the first splicing result and the second splicing result is greater than or equal to the similarity threshold, the target resource is determined as the resource matched with the target user.

[0162] In an embodiment, the business processing device can calculate the similarity between the first spliced result and the second spliced result based on a cosine similarity calculation method. The similarity threshold value can be preset by the business processing device, and in the cosine similarity calculation method, the greater the distance between the first spliced result and the second spliced result, the more similar they are. For example, the first spliced result indicates that the target user is a rational person and will not easily comment on or forward some resources and other interactive operations, and the second spliced result indicates that the target resource has been interacted by many users. At this time, the similarity between the first spliced result and the second spliced result is large, and the target resource can be recommended to the target user.

[0163] As described above, when the business related to resource recommendation is to recommend matching resources to users, the user information management module and the resource information management module are also deployed in a resource recommendation model. Based on the description of steps S701-S705, and assuming that the user information management module and the resource information management module are deployed in the same resource recommendation model, an embodiment of performing resource recommendation for a target user is introduced. Referring to Figure 8 A schematic diagram of a resource recommendation model provided by an embodiment of the present application for performing resource recommendation business is shown in FIG. 8. 800 represents a resource recommendation model. 801 represents a user information management module, and 802 represents a resource information management module. The N resource interaction data of the target resource are obtained by the resource information management module 802 in the resource recommendation model based on the historical interaction information of the target resource in advance and stored in a resource recommendation reference database. Specifically, the resource interaction data under each interactive operation of the target resource can be spliced to obtain a second spliced result. The resource recommendation reference database can also include spliced results corresponding to multiple other candidate resources. The business processing device can divide the multiple candidate resources according to the similarity between the spliced results to obtain multiple resource groups, and then obtain an index identifier corresponding to each resource group based on the multiple candidate spliced results included in each resource group. The index identifiers corresponding to the multiple resource groups can constitute an index library. When it is necessary to recommend matching resources to the target user, the user information management module 801 in the resource recommendation model is called to process the user attribute information of the target user to obtain N user interaction data, splice the N user interaction data to obtain a first spliced result, and then select a second spliced result of the target resource from the resource group corresponding to the matching index identifier according to the matching relationship between the first spliced result and the index identifier in the index library. Whether the target resource is a matching resource of the target user is determined based on the similarity between the first spliced result and the second spliced result.

[0164] In short, a schematic diagram of recommending resources to users provided by an embodiment of the present application can be as shown in FIG. 8. Figure 9As shown, first, a plurality of preprocessed training samples are used to train a resource recommendation model. After the resource recommendation model is trained, the historical interaction information and resource identifiers of a plurality of candidate resources are input into the resource recommendation model, the resource information management module in the resource recommendation model is called to generate N resource interaction data for each candidate resource based on the historical interaction information of each candidate resource, then splicing is performed, resource groups are divided according to the splicing result, and then an index library is constructed; when a target user requests a resource, the user attribute information of the target user is input into the resource recommendation model, the user information management module in the resource recommendation model is called to obtain N user interaction data of the target user based on the user attribute information, then the N user interaction data are spliced, matching retrieval is performed in the index library according to the splicing result, and finally the result is output.

[0165] In the embodiment of the application, M user features of a target user are obtained, and N user interaction data extraction networks are called to analyze and process the M user features to obtain N user interaction data of the target user, one user interaction data extraction network being used to extract one user interaction data; then, N resource interaction data of a target resource are obtained from a resource recommendation reference database; then, the N user interaction data are spliced to obtain a first splicing result, and the N resource interaction data are spliced to obtain a second splicing result, and the similarity between the first splicing result and the second splicing result is compared, and if the similarity is higher than a similarity threshold, it is determined that the target resource is a resource matched with the target user. One user interaction data can represent the execution degree of the target user on one interactive operation, and one resource interaction data is used to represent the execution degree of the target resource on one interactive operation, and when a resource recommendation-related service is executed based on the N user interaction data and the N resource interaction data, the accuracy of service execution can be improved.

[0166] Based on the above-mentioned embodiment of the service processing method, the embodiment of the application further provides a service processing device. Referring to Figure 10 A structural schematic diagram of a service processing device provided by the embodiment of the application is shown in FIG. 10. The service processing device shown in FIG. 10 can run the following units:

[0167] The obtaining unit 1001 is configured to obtain M user features of a target user, the M user features being shared by N user interaction data extraction networks corresponding to N interactive operations, N being an integer greater than 1, and M being an integer greater than 0;

[0168] The processing unit 1002 is configured to call the N user interaction data extraction networks to analyze and process the M user features respectively to obtain user interaction data corresponding to each interactive operation; one user interaction data is used to represent the execution degree of the target user on one interactive operation.

[0169] The acquisition unit 1001 is further configured to acquire N resource interaction data of a target resource, one resource interaction data being used to represent an executed degree of performing one kind of interaction operation on the target resource.

[0170] The execution unit 1003 is configured to perform a service related to resource recommendation based on the N user interaction data and the N resource interaction data.

[0171] In one embodiment, the N user interaction data extraction networks are deployed in a user information management module, the user information management module further comprises M user feature processing networks, the M user feature processing networks are respectively connected with the N user interaction data extraction networks, and the acquisition unit 1001 performs the following steps when acquiring M user features of a target user:

[0172] acquiring user attribute information of the target user, the user attribute information comprising any one or more of the following: age, gender and historical preference information of the target user on N kinds of interaction operations;

[0173] performing vector conversion processing on the user attribute information to obtain a user latent vector;

[0174] calling the M user feature processing networks to respectively perform feature cross and feature abstraction processing on the user latent vector to obtain the M user features of the target user.

[0175] In one embodiment, the user information management module further comprises N user interaction weight management networks, and the processing unit 1002 performs the following steps when calling the N user interaction data extraction networks to respectively analyze and process the M user features to obtain user interaction data corresponding to each kind of interaction operation:

[0176] calling each user interaction weight management network in the N user interaction weight management networks to perform weight generation operation on the user latent vector to obtain a user interaction weight value corresponding to each kind of interaction operation;

[0177] multiplying the user interaction weight value corresponding to the i-th kind of interaction operation with each user feature in the M user features;

[0178] adding each multiplication result, and calling an i-th user interaction data extraction network corresponding to the i-th kind of interaction operation to analyze and process a result of the addition operation to obtain user interaction data corresponding to the i-th kind of interaction operation; wherein i is an integer greater than or equal to 1 and less than or equal to N.

[0179] In an embodiment, the service related to the resource recommendation comprises training a resource recommendation model, the resource recommendation model comprises N resource interaction data extraction networks, and the acquisition unit 1001 performs the following steps when acquiring N resource interaction data of a target resource:

[0180] acquiring L resource features of the target resource, L being an integer greater than or equal to 1; calling the N resource interaction data extraction networks to respectively analyze and process the L resource features, to obtain resource interaction data corresponding to each interaction operation.

[0181] In an embodiment, the N resource interaction data extraction networks are included in a resource information management module, the resource information management module is deployed in the resource recommendation model, and the resource information management module further comprises L resource feature processing networks, the L resource feature processing networks are respectively connected with the N resource interaction data extraction networks, and the acquisition unit 1001 performs the following steps when acquiring L resource features of a target user:

[0182] acquiring historical interaction information of the target resource, the historical interaction information comprising information indicating that the target resource has been executed N interaction operations in a historical time; performing vector conversion on the historical interaction information to obtain a resource latent vector; calling the L resource feature processing networks to perform feature cross and feature abstraction processing on the resource latent vector, to obtain the L resource features of the target resource.

[0183] In an embodiment, the resource information management module further comprises N resource interaction weight management networks, and the acquisition unit 1001 performs the following steps when calling the N resource interaction data extraction networks to respectively analyze and process the L resource features to obtain resource interaction data corresponding to each interaction operation:

[0184] calling each resource interaction weight management network in the N resource interaction weight management networks to perform weight generation operation on the resource latent vector, to obtain resource interaction weight values corresponding to each interaction operation; multiplying the resource interaction weight values corresponding to the jth interaction operation with each resource feature in the M resource features;

[0185] adding each multiplication result, and calling the jth resource interaction data extraction network corresponding to the jth interaction operation to analyze and process the result of the addition, to obtain resource interaction data corresponding to the jth interaction operation, j being an integer greater than or equal to 1 and less than or equal to N.

[0186] In an embodiment, the execution unit 1003 performs the following steps when performing a service related to resource recommendation based on the N user interaction data and the N resource interaction data:

[0187] According to the N user interaction data and the N resource interaction data, determine the prediction information of the target user performing each of the N interaction operations on the target resource; obtain the label information of the target user performing each of the N interaction operations on the target resource;

[0188] Obtain the loss function corresponding to each interaction operation; determine the value of the loss function corresponding to the corresponding interaction operation based on the difference between the prediction information of the target user performing each interaction operation on the target resource and the label information of the target user performing the corresponding interaction operation on the target resource;

[0189] Add the values of the loss functions corresponding to each interaction operation to obtain the loss function value of the resource recommendation model, and adjust the parameters of the resource recommendation model in the direction of reducing the loss function value of the resource recommendation model.

[0190] In an embodiment, when determining the prediction information of the target user performing each of the N interaction operations on the target resource according to the N user interaction data and the N resource interaction data, the execution unit 1003 performs the following steps:

[0191] Dot product operation is performed on the i-th user interaction data and the i-th resource interaction data; nonlinear conversion is performed on the dot product operation result to obtain the prediction information of the target user performing the i-th interaction operation on the target resource; wherein i is an integer greater than or equal to 1 and less than or equal to N.

[0192] In an embodiment, the service related to resource recommendation includes recommending matching resources to the target user, and the obtaining unit 1001 performs the following steps when obtaining the N resource interaction data corresponding to the target resource:

[0193] Obtain a resource recommendation reference database, the resource recommendation reference database including N resource interaction data corresponding to each candidate resource in a plurality of candidate resources, the plurality of candidate resources including the target resource; obtain the N resource interaction data corresponding to the target resource from the resource recommendation reference database.

[0194] In one embodiment, the plurality of candidate resources are divided into P resource groups, each resource group including at least one candidate resource and N resource interaction data corresponding to each candidate resource, each resource group corresponding to an index identifier, and the index identifier corresponding to any resource group is determined based on the N resource interaction data corresponding to each candidate resource in the resource group. When the acquisition unit 1001 acquires the N resource interaction data of the target resource from the resource recommendation reference database, it performs the following steps:

[0195] The N user interaction data are concatenated to obtain a first concatenation result; the target resource group whose index identifier matches the first concatenation result is searched from the P resource groups; any unselected candidate resource is selected from the target resource group as the target resource, and the N resource interaction data corresponding to the target resource are obtained.

[0196] In one embodiment, when execution unit 1003 performs resource recommendation-related business based on the N user interaction data and the N resource interaction data, it performs the following steps:

[0197] The N resource interaction data are concatenated to obtain a second concatenation result; if the similarity between the first concatenation result and the second concatenation result is greater than or equal to the similarity threshold, then the target resource is determined to be a resource that matches the target user.

[0198] In one embodiment, when the acquisition unit 1001 splices the N user interaction data to obtain a first splicing result, it performs the following steps: acquiring the splicing weight value corresponding to each interaction operation.

[0199] Multiply each user interaction data with the corresponding interaction operation weight value; then concatenate the results of each multiplication to obtain the first concatenation result.

[0200] According to one embodiment of the present invention, Figure 2 , Figure 5 as well as Figure 7 The steps involved in the business processing method shown can be derived from... Figure 10 The tasks are performed by the various units in the business processing apparatus shown. For example, steps S201 and S203 shown in Figure 2 can both be performed by... Figure 10 The business processing device shown is executed by the acquisition unit 1001. Step S202 can be performed by... Figure 10 The processing unit 1002 in the business processing device shown is responsible for executing step S204; step S204 can be performed by... Figure 10 The execution unit 1003 in the illustrated business processing device performs the execution; for example, Figure 5Steps S501, S503, S506, and S507 shown can all be... Figure 10 The acquisition unit 1001 in the business processing device shown executes the steps S502, S504, and S505, which can all be performed by... Figure 10 The processing unit 1002 in the aforementioned business processing device executes step S508, which can be performed by... Figure 10 The execution unit 1003 in the aforementioned business processing device performs the execution; for example, Figure 7 In the business processing method shown, steps S701 and S703 can be derived from... Figure 10 The acquisition unit 1001 in the aforementioned business processing device executes the steps S702-S705, which can all be performed by... Figure 10 The processing unit 1002 in the aforementioned business processing device performs the execution.

[0201] According to another embodiment of the present invention, Figure 10 The various units in the illustrated business processing apparatus can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The above-mentioned units are divided based on logical functions. In practical applications, the function of one unit can also be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, the business processing apparatus may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0202] According to another embodiment of the present invention, the following can be performed by running on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). Figure 2 , Figure 5 as well as Figure 7 The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 10 The present invention describes a business processing apparatus and a business processing method for implementing embodiments thereof. The computer program may be recorded on, for example, a computer-readable storage medium, loaded onto the aforementioned computing device via the computer-readable storage medium, and run therein.

[0203] In the embodiment of the present application, M user features of a target user are acquired, and N user interaction data extraction networks are called to analyze and process the M user features to obtain N user interaction data of the target user, one user interaction data extraction network being used to extract one user interaction data; then, N resource interaction data of a target resource are acquired; finally, a service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0204] In the service processing process, the M user features are shared by the N interaction data extraction networks, and it is not necessary to repeatedly extract user features for each interaction data extraction network, so that service processing resources can be saved. In addition, one user interaction data can represent the execution degree of the target user on one interactive operation, and one resource interaction data is used to represent the execution degree of the target resource on one interactive operation, that is, when the service related to resource recommendation is executed, the execution procedure of the target user on each interactive operation and the execution procedure of the target resource on each interactive operation are simultaneously referred to, so that the accuracy of service execution can be improved.

[0205] Based on the above-mentioned service processing method embodiment and service processing device embodiment, the embodiment of the present application provides a service processing device. Referring to Figure 11 A structural schematic diagram of a service processing device provided by the embodiment of the present application is shown in FIG. 1. Figure 11 The service processing device shown in the figure at least includes a processor 1101, an input interface 1102, an output interface 1103 and a computer storage medium 1104. The processor 1101, the input interface 1102, the output interface 1103 and the computer storage medium 1104 can be connected by bus or other means.

[0206] The computer storage medium 1104 can be stored in the memory of the service processing device, and the computer storage medium 1104 is used to store a computer program, and the processor 1101 is used to execute the computer program stored in the computer storage medium 1104. The processor 1101 (or CPU (Central Processing Unit, Central Processor)) is the computing core and control core of the service processing device, which is suitable for implementing one or more computer programs, and is particularly suitable for loading and executing one or more computer programs to implement corresponding method processes or corresponding functions.

[0207] In one embodiment, the processor 1101 in the embodiment of the present application can be used to perform: obtaining M user features of a target user, the M user features being shared by N user interaction data extraction networks corresponding to N interaction operations, N being an integer greater than 1, and M being an integer greater than 0; calling the N user interaction data extraction networks to analyze and process the M user features respectively to obtain user interaction data corresponding to each interaction operation; one user interaction data being used to represent an execution degree of the target user on one interaction operation; obtaining N resource interaction data of a target resource, one resource interaction data being used to represent an execution degree of the target resource on one interaction operation; and performing a resource recommendation related service based on the N user interaction data and the N resource interaction data.

[0208] In the embodiment of the present application, the M user features of the target user are obtained, and the N user interaction data extraction networks are called to analyze and process the M user features to obtain N user interaction data of the target user, one user interaction data extraction network being used to extract one user interaction data; then, the N resource interaction data of the target resource are obtained; and finally, a resource recommendation related service is performed based on the N user interaction data and the N resource interaction data.

[0209] In the above service processing process, the M user features are shared by the N interaction data extraction networks, and it is not necessary to repeatedly extract user features for each interaction data extraction network, so that service processing resources can be saved. In addition, one user interaction data can represent an execution degree of the target user on one interaction operation, and one resource interaction data is used to represent an execution degree of the target resource on one interaction operation, that is, when performing the resource recommendation related service, the execution procedure of the target user on each interaction operation and the execution procedure of the target resource on each interaction operation are simultaneously referred to, so that the accuracy of service execution can be improved.

[0210] The embodiment of the present application further provides a computer storage medium (Memory), which is a memory device in a service processing device and is used to store programs and data. It can be understood that the computer storage medium herein can include a built-in storage medium of the service processing device, and of course can also include an extended storage medium supported by the service processing device. The computer storage medium provides a storage space, and the storage space stores an operating system of the service processing device. In addition, one or more computer programs suitable for being loaded and executed by the processor 1001 are also stored in the storage space; it should be noted that the computer storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory; and optionally can also be at least one computer storage medium located away from the aforementioned processor.

[0211] In an embodiment, one or more computer programs in the computer storage medium can be loaded and executed by the processor 1101 to implement the corresponding steps in the business processing method embodiments described above. Figure 2 , Figure 4 and Figure 7 In a specific implementation, one or more computer programs in the computer storage medium are loaded and executed by the processor 1101 to implement the following steps:

[0212] Obtain M user features of a target user, the M user features are shared by N user interaction data extraction networks corresponding to N interaction operations, N is an integer greater than 1, and M is an integer greater than 0; call the N user interaction data extraction networks to analyze and process the M user features respectively, to obtain user interaction data corresponding to each interaction operation; one user interaction data is used to represent the execution degree of the target user on one interaction operation; obtain N resource interaction data of a target resource, one resource interaction data is used to represent the execution degree of the target resource on one interaction operation; and perform a resource recommendation related business based on the N user interaction data and the N resource interaction data.

[0213] In an embodiment, the N user interaction data extraction networks are deployed in a user information management module, the user information management module further includes M user feature processing networks, the M user feature processing networks are connected with the N user interaction data extraction networks respectively, and the processor 1101 performs the following steps when obtaining the M user features of the target user:

[0214] Obtain user attribute information of the target user, the user attribute information includes any one or more of the following: age, gender, and historical preference information of the target user on N interaction operations;

[0215] Perform vector conversion processing on the user attribute information to obtain a user latent vector;

[0216] Call the M user feature processing networks to perform feature cross and feature abstraction processing on the user latent vector respectively to obtain the M user features of the target user.

[0217] In an embodiment, the user information management module further includes N user interaction weight management networks, and the processor 1101 performs the following steps when calling the N user interaction data extraction networks to analyze and process the M user features respectively to obtain user interaction data corresponding to each interaction operation:

[0218] The N user interaction weight management networks are called to perform weight generation operations on the user hidden vector to obtain user interaction weight values corresponding to each interaction operation.

[0219] The user interaction weight value corresponding to the i-th interaction operation is multiplied with each of the M user features.

[0220] The multiplication results are added, and the i-th user interaction data extraction network corresponding to the i-th interaction operation is called to analyze and process the addition result to obtain user interaction data corresponding to the i-th interaction operation; and 1 < < i < < N.

[0221] In an embodiment, the resource recommendation-related service includes training a resource recommendation model, the resource recommendation model includes N resource interaction data extraction networks, and the processor 1101 performs the following steps when obtaining N resource interaction data of a target resource:

[0222] L resource features of the target resource are obtained, L being an integer greater than or equal to 1.

[0223] The N resource interaction data extraction networks are called to analyze and process the L resource features respectively to obtain resource interaction data corresponding to each interaction operation.

[0224] In an embodiment, the N resource interaction data extraction networks are included in a resource information management module, the resource information management module is deployed in the resource recommendation model, the resource information management module further includes L resource feature processing networks, the L resource feature processing networks are connected with the N resource interaction data extraction networks respectively, and the processor 1101 performs the following steps when obtaining L resource features of a target resource:

[0225] Historical interaction information of the target resource is obtained, the historical interaction information includes information indicating that the target resource is executed with N interaction operations in a historical time; the historical interaction information is converted into a vector to obtain a resource hidden vector; and the L resource feature processing networks are called to perform feature cross and feature abstraction processing on the resource hidden vector to obtain the L resource features of the target resource.

[0226] In an embodiment, the resource information management module further includes N resource interaction weight management networks, and the processor 1101 performs the following steps when calling the N resource interaction data extraction networks to analyze and process the L resource features respectively to obtain resource interaction data corresponding to each interaction operation:

[0227] The N resource interaction weight management networks are called to perform weight generation operations on the resource hidden vector, and resource interaction weight values corresponding to each interaction operation are obtained.

[0228] The resource interaction weight values corresponding to the jth interaction operation are multiplied with each of the M resource features.

[0229] The multiplication results are added, and the jth resource interaction data extraction network corresponding to the jth interaction operation is called to analyze and process the addition result, to obtain resource interaction data corresponding to the jth interaction operation, j being an integer greater than or equal to 1 and less than or equal to N.

[0230] In one embodiment, the processor 1101 performs the following steps when performing a service related to resource recommendation based on the N user interaction data and the N resource interaction data:

[0231] According to the N user interaction data and the N resource interaction data, the prediction information of the target user performing each interaction operation on the target resource is determined.

[0232] The label information of the target user performing each interaction operation on the target resource is obtained.

[0233] The loss function corresponding to each interaction operation is obtained.

[0234] The value of the loss function corresponding to each interaction operation is determined based on the difference between the prediction information of the target user performing each interaction operation on the target resource and the label information of the target user performing the corresponding interaction operation on the target resource.

[0235] The values of the loss functions corresponding to each interaction operation are added to obtain the loss function value of the resource recommendation model, and the parameters of the resource recommendation model are adjusted in the direction of reducing the loss function value of the resource recommendation model.

[0236] In one embodiment, the processor 1101 performs the following steps when determining the prediction information of the target user performing each interaction operation on the target resource based on the N user interaction data and the N resource interaction data:

[0237] The ith user interaction data and the ith resource interaction data are dot multiplied, and the dot multiplication result is nonlinearly converted to obtain the prediction information of the target user performing the ith interaction operation on the target resource; wherein i is an integer greater than or equal to 1 and less than or equal to N.

[0238] In an embodiment, the service related to resource recommendation comprises recommending a matching resource to the target user, and the processor 1101 performs the following steps when obtaining the N resource interaction data corresponding to the target resource:

[0239] obtaining a resource recommendation reference database, wherein the resource recommendation reference database comprises N resource interaction data corresponding to each candidate resource in a plurality of candidate resources, and the plurality of candidate resources comprises the target resource;

[0240] obtaining the N resource interaction data corresponding to the target resource from the resource recommendation reference database.

[0241] In an embodiment, the plurality of candidate resources are divided into P resource groups, each resource group comprising at least one candidate resource and N resource interaction data corresponding to each candidate resource, each resource group corresponding to an index identifier, and the index identifier corresponding to any resource group being determined based on the N resource interaction data corresponding to each candidate resource in the any resource group;

[0242] The processor 1101 performs the following steps when obtaining the N resource interaction data corresponding to the target resource from the resource recommendation reference database:

[0243] concatenating the N user interaction data to obtain a first concatenation result;

[0244] searching the P resource groups for a target resource group whose index identifier matches the first concatenation result;

[0245] selecting any unselected candidate resource in the target resource group as a target resource and obtaining N resource interaction data corresponding to the target resource.

[0246] In an embodiment, the processor 1101 performs the following steps when performing a service related to resource recommendation based on the N user interaction data and the N resource interaction data:

[0247] concatenating the N resource interaction data to obtain a second concatenation result;

[0248] If the similarity between the first concatenation result and the second concatenation result is greater than or equal to a similarity threshold, the target resource is determined as a resource matching the target user.

[0249] In an embodiment, the processor 1101 performs the following steps when concatenating the N user interaction data to obtain a first concatenation result:

[0250] obtaining a concatenation weight value corresponding to each interaction operation;

[0251] each user interaction data is multiplied by the splicing weight value of the corresponding interaction operation of the user interaction data;

[0252] the results of each multiplication operation are spliced to obtain a first splicing result.

[0253] In the embodiment of the application, M user features of a target user are obtained, and N user interaction data extraction networks are called to analyze and process the M user features to obtain N user interaction data of the target user, one user interaction data extraction network being used to extract one user interaction data. Then, N resource interaction data of a target resource are obtained. Finally, a service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0254] In the above service processing process, the M user features are shared by the N user interaction data extraction networks, and it is not necessary to repeatedly extract user features for each user interaction data extraction network, so that service processing resources can be saved. In addition, one user interaction data can represent the execution degree of the target user on one kind of interaction operation, and one resource interaction data is used to represent the execution degree of the target resource on one kind of interaction operation, that is, when the service related to resource recommendation is executed, the execution procedure of the target user on each kind of interaction operation and the execution procedure of the target resource on each kind of interaction operation are simultaneously referred to, so that the accuracy of service execution can be improved.

[0255] According to one application of the application, the embodiment of the application further provides a computer product or a computer program, the computer product comprising a computer program, and the computer program is stored in a computer readable storage medium. The processor 1101 reads the computer program from the computer readable storage medium, so that the service processing device loads and executes:

[0256] M user features of a target user are obtained, the M user features being shared by N user interaction data extraction networks corresponding to N kinds of interaction operations, N being an integer greater than 1 and M being an integer greater than 0;

[0257] The N user interaction data extraction networks are called to analyze and process the M user features respectively to obtain user interaction data corresponding to each kind of interaction operation, one user interaction data being used to represent the execution degree of the target user on one kind of interaction operation;

[0258] N resource interaction data of a target resource are obtained, one resource interaction data being used to represent the execution degree of the target resource on one kind of interaction operation;

[0259] A service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0260] In the embodiments of the present application, M user features of a target user are acquired, and N user interaction data extraction networks are called to analyze and process the M user features to obtain N user interaction data of the target user, one user interaction data extraction network being used to extract one user interaction data; then, N resource interaction data of a target resource are acquired; and finally, a service related to resource recommendation is executed based on the N user interaction data and the N resource interaction data.

[0261] In the service processing process, the M user features are shared by the N interaction data extraction networks, and it is not necessary to repeatedly extract user features for each interaction data extraction network, so that the service processing resources can be saved. In addition, one user interaction data can represent the execution degree of the target user on one kind of interaction operation, and one resource interaction data is used to represent the execution degree of the target resource on one kind of interaction operation, that is, when the service related to resource recommendation is executed, the execution procedure of the target user on each kind of interaction operation and the execution procedure of the target resource on each kind of interaction operation are simultaneously referred to, so that the accuracy of service execution can be improved.

Claims

1. A service processing method characterized by, The method comprises the following steps: obtaining M user features of a target user, the M user features being shared by N user interaction data extraction networks corresponding to N interaction operations, N being an integer greater than 1, and M being an integer greater than 0; calling the N user interaction data extraction networks to analyze and process the M user features respectively to obtain user interaction data corresponding to each interaction operation; one user interaction data is used to represent the execution degree of the target user on one interaction operation, and the execution degree reflects the interest degree of the target user on one interaction operation; obtaining N resource interaction data of a target resource, one resource interaction data being used to represent the execution degree of the target resource on one interaction operation, and the execution degree reflecting the degree of user preference for executing one interaction operation on the target resource; performing a resource recommendation-related service based on the N user interaction data and the N resource interaction data; the resource recommendation-related service comprises recommending a matched resource to the target user and training a resource recommendation model; if the resource recommendation-related service comprises training a resource recommendation model, then performing the resource recommendation-related service based on the N user interaction data and the N resource interaction data comprises: determining prediction information of the target user on each interaction operation of the target resource based on the N user interaction data and the N resource interaction data; obtaining label information of the target user on each interaction operation of the target resource; obtaining a loss function corresponding to each interaction operation; determining the value of the loss function corresponding to each interaction operation based on the difference between the prediction information of the target user on each interaction operation of the target resource and the label information of the target user on the corresponding interaction operation; adding the values of the loss functions corresponding to each interaction operation to obtain the loss function value of the resource recommendation model, and adjusting the parameters of the resource recommendation model in the direction of reducing the loss function value of the resource recommendation model.

2. The method of claim 1, wherein, The N user interaction data extraction networks are deployed in a user information management module, the user information management module further comprises M user feature processing networks, the M user feature processing networks are connected with the N user interaction data extraction networks respectively, and the M user features of the target user are obtained by: obtaining user attribute information of the target user, the user attribute information comprising any one or more of the following: the age, gender and historical preference information of the target user on N interaction operations; performing vector conversion processing on the user attribute information to obtain a user latent vector; calling the M user feature processing networks to perform feature cross and feature abstraction processing on the user latent vector respectively to obtain the M user features of the target user.

3. The method of claim 2, wherein, The user information management module further includes N user interaction weight management networks, the N user interaction data extraction networks are respectively called to analyze and process the M user features, and user interaction data corresponding to each interaction operation is obtained, including: Each user interaction weight management network in the N user interaction weight management networks is called to perform a weight generation operation on the user hidden vector, and a user interaction weight value corresponding to each interaction operation is obtained; The user interaction weight value corresponding to the i-th interaction operation is multiplied by each user feature in the M user features; The results of the respective multiplication operations are added, and the i-th user interaction data extraction network corresponding to the i-th interaction operation is called to analyze and process the result of the addition operation, and user interaction data corresponding to the i-th interaction operation is obtained; wherein i is an integer greater than or equal to 1 and less than or equal to N.

4. The method of claim 1, wherein, The resource recommendation-related business includes training a resource recommendation model, the resource recommendation model includes N resource interaction data extraction networks, and the N resource interaction data of the target resource are obtained, including: L resource features of the target resource are obtained, L is an integer greater than or equal to 1; The N resource interaction data extraction networks are respectively called to analyze and process the L resource features, and resource interaction data corresponding to each interaction operation is obtained.

5. The method of claim 4, wherein, The N resource interaction data extraction networks are included in a resource information management module, the resource information management module is deployed in the resource recommendation model, and the resource information management module further includes L resource feature processing networks, the L resource feature processing networks are respectively connected with the N resource interaction data extraction networks, and the L resource features of the target resource are obtained, including: Historical interaction information of the target resource is obtained, and the historical interaction information includes information indicating that the target resource is executed N interaction operations in a historical time; The historical interaction information is vector converted to obtain a resource hidden vector; The L resource feature processing networks are called to perform feature cross and feature abstraction processing on the resource hidden vector to obtain the L resource features of the target resource.

6. The method of claim 5, wherein, The resource information management module further includes N resource interaction weight management networks, the N resource interaction data extraction networks are respectively called to analyze and process the L resource features, and resource interaction data corresponding to each interaction operation is obtained, including: Each resource interaction weight management network in the N resource interaction weight management networks is called to perform a weight generation operation on the resource hidden vector, and a resource interaction weight value corresponding to each interaction operation is obtained; The resource interaction weight value corresponding to the j-th interaction operation is multiplied by each resource feature in the L resource features; The results of the respective multiplication operations are added, and the j-th resource interaction data extraction network corresponding to the j-th interaction operation is called to analyze and process the result of the addition operation, and resource interaction data corresponding to the j-th interaction operation is obtained, j is an integer greater than or equal to 1 and less than or equal to N.

7. The method of claim 1, wherein, The determining, according to the N user interaction data and the N resource interaction data, of prediction information of the target user performing each of the N interaction operations on the target resource comprises: performing dot product operation on the ith user interaction data and the ith resource interaction data; performing nonlinear conversion on the result of the dot product operation to obtain the prediction information of the target user performing the ith interaction operation on the target resource; wherein i is an integer greater than or equal to 1 and less than or equal to N.

8. The method of claim 1, wherein, The resource recommendation-related business comprises recommending a matching resource to the target user, and the obtaining of the N resource interaction data corresponding to the target resource comprises: obtaining a resource recommendation reference database, wherein the resource recommendation reference database comprises N resource interaction data corresponding to each candidate resource in a plurality of candidate resources, and the plurality of candidate resources comprises the target resource; obtaining the N resource interaction data corresponding to the target resource from the resource recommendation reference database.

9. The method of claim 8, wherein, The plurality of candidate resources are divided into P resource groups, each resource group comprising at least one candidate resource and N resource interaction data corresponding to each candidate resource, and each resource group corresponding to an index identifier, wherein the index identifier corresponding to any resource group is determined based on the N resource interaction data corresponding to each candidate resource in the any resource group. The obtaining of the N resource interaction data corresponding to the target resource from the resource recommendation reference database comprises: performing splicing processing on the N user interaction data to obtain a first splicing result; finding a target resource group with an index identifier matching the first splicing result from the P resource groups; selecting any one unselected candidate resource in the target resource group as a target resource and obtaining the N resource interaction data corresponding to the target resource.

10. The method of claim 9, wherein, The performing of the resource recommendation-related business based on the N user interaction data and the N resource interaction data comprises: performing splicing processing on the N resource interaction data to obtain a second splicing result; if the similarity between the first splicing result and the second splicing result is greater than or equal to a similarity threshold, determining the target resource as a resource matching the target user.

11. The method of claim 9, wherein, The performing of the splicing processing on the N user interaction data to obtain the first splicing result comprises: obtaining a splicing weight value corresponding to each interaction operation; performing multiplication operation on each user interaction data and the splicing weight value corresponding to the interaction operation of the corresponding user interaction data; performing splicing processing on the results of the multiplication operations to obtain the first splicing result.

12. A service processing apparatus characterized by comprising: The method comprises: obtaining an M user feature of a target user, wherein the M user feature is shared by N user interaction data corresponding to N interaction operations, N is an integer greater than 1, and M is an integer greater than 0; performing analysis and processing on the M user features by calling the N user interaction data extraction network respectively to obtain user interaction data corresponding to each interaction operation; and One user interaction data is used to represent an execution degree of the target user on an interaction operation, and the execution degree reflects an interest degree of the target user on an interaction operation; The acquisition unit is further configured to acquire N resource interaction data of the target resource, and one resource interaction data is used to represent an execution degree of the target resource on an interaction operation, and the execution degree reflects a degree of user preference on the execution of the target resource on an interaction operation; The execution unit is configured to execute a resource recommendation-related service based on the N user interaction data and the N resource interaction data; the resource recommendation-related service includes recommending a matched resource to the target user, and also includes training a resource recommendation model; If the resource recommendation-related service includes training a resource recommendation model, the execution unit executes the following steps when executing the resource recommendation-related service based on the N user interaction data and the N resource interaction data: determining predicted information of the target user on the execution of each interaction operation in the N interaction operations on the target resource according to the N user interaction data and the N resource interaction data; acquiring label information of the target user on the execution of each interaction operation in the N interaction operations on the target resource; acquiring a loss function corresponding to each interaction operation; determining a value of the loss function corresponding to each interaction operation based on a difference between the predicted information of the target user on the execution of each interaction operation on the target resource and the label information of the target user on the execution of the corresponding interaction operation on the target resource; adding the values of the loss functions corresponding to each interaction operation to obtain a loss function value of the resource recommendation model, and adjusting parameters of the resource recommendation model in a direction of reducing the loss function value of the resource recommendation model.

13. A service processing device, characterized by comprise: a processor adapted to implement one or more computer programs; and a computer storage medium storing one or more computer programs adapted to be loaded and executed by the processor to perform the method of any one of claims 1-11.

14. A computer storage medium, characterized in that, The computer storage medium stores a computer program, and the computer program is executed by the processor to perform the method of any one of claims 1-11. The computer storage medium stores a computer program, and the computer program is executed by the processor to perform the method of any one of claims 1-11.

Images