Method for determining delivery information and related device

Abstract

This application relates to the field of artificial intelligence technology and discloses a method and related equipment for determining delivery information. The method includes: outputting a first user feature of the target user relative to candidate merchants by a first graph attention network; outputting a first merchant feature of candidate merchants relative to the target user by a second graph attention network; determining a second user feature of the target user and a second merchant feature of candidate merchants by a dual attention network based on an interaction relationship graph; outputting an interaction probability of the target user relative to candidate merchants by an interaction prediction network based on the global user feature of the target user and the global merchant feature of candidate merchants; and filtering among multiple candidate merchants based on the interaction probability of the target user relative to candidate merchants to determine the target merchant, thereby identifying the delivery information of the target merchant as the delivery information to be delivered to the target user. This solution can improve the efficiency of users obtaining delivery information.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method for determining delivery information and related equipment. Background Technology

[0002] With the development of internet technology, the internet has gradually become an important channel for information transmission. However, due to the vast variety and sheer volume of information on the internet, users may need to interact with the server multiple times and filter information repeatedly to find the information they need or are interested in from this massive amount of data. Because users need to interact with the server multiple times to obtain the information they need or are interested in, related technologies suffer from low efficiency in information retrieval. Summary of the Invention

[0003] In view of the above problems, this application proposes a method for determining delivery information and related equipment to improve the above problems.

[0004] According to one aspect of the embodiments of this application, a method for determining delivery information is provided, comprising: outputting a first user feature of the target user relative to the candidate merchants by a first graph attention network based on a friend relationship graph of the target user and merchant embedding features of candidate merchants; outputting a first merchant feature of the candidate merchants relative to the target user by a second graph attention network based on a neighbor merchant relationship graph of the candidate merchants, the distance between each neighbor merchant and the candidate merchant, and the user embedding features of the target user; and determining a second user feature of the target user and a second merchant feature of the candidate merchants by a dual attention network based on an interaction relationship graph, wherein the interaction relationship graph is based on user interactions with merchants. The interaction behavior is constructed by fusing the target user's user embedding features, first user features, and second user features to obtain the target user's global user features; fusing the candidate merchant's merchant embedding features, first merchant features, and second merchant features to obtain the candidate merchant's global merchant features; the interaction prediction network outputs the interaction probability of the target user relative to the candidate merchant based on the target user's global user features and the candidate merchant's global merchant features; and the target merchant is determined by filtering among multiple candidate merchants based on the target user's interaction probability relative to the candidate merchant, so that the target merchant's delivery information is determined as the delivery information to be delivered to the target user.

[0005] According to one aspect of the embodiments of this application, an apparatus for determining delivery information is provided, comprising: a first user feature determination module, configured to output a first user feature of the target user relative to the candidate merchants by a first graph attention network based on a friend relationship graph of the target user and merchant embedding features of candidate merchants; a first merchant feature determination module, configured to output a first merchant feature of the candidate merchants relative to the target user by a second graph attention network based on a neighbor merchant relationship graph of the candidate merchants, the distance between each neighbor merchant and the candidate merchants, and the user embedding features of the target user; and a second feature determination module, configured to determine a second user feature of the target user and a second merchant feature of the candidate merchants by a dual attention network based on an interaction relationship graph, wherein the interaction relationship graph is constructed based on user-triggered interaction behaviors with merchants. The system includes: a global user feature determination module, used to fuse the user embedding features, first user features, and second user features of the target user to obtain the global user features of the target user; a global merchant feature determination module, used to fuse the merchant embedding features, first merchant features, and second merchant features of the candidate merchants to obtain the global merchant features of the candidate merchants; an interaction probability determination module, used by the interaction prediction network to output the interaction probability of the target user relative to the candidate merchants based on the global user features of the target user and the global merchant features of the candidate merchants; and a target merchant determination module, used to filter among multiple candidate merchants based on the interaction probability of the target user relative to the candidate merchants to determine the target merchant, so as to determine the delivery information of the target merchant as the delivery information to be delivered to the target user.

[0006] In some embodiments, the first graph attention network includes a first encoding layer, a first attention layer, and a first linear layer; the first user feature determination module includes: a friend encoding feature determination unit, configured to encode the friends in the friend relationship graph of the target user by the first encoding layer to obtain the friend encoding features of each friend; a friend influence weight determination unit, configured to perform attention calculation by the first attention layer based on the friend encoding features of each friend, the user embedding features of the target user, and the merchant embedding features of the candidate merchant to obtain the friend influence weight of each friend for the target user to trigger interactive behavior with the candidate merchant; and a first linear weighting unit, configured to perform linear weighting by the first linear layer based on the friend influence weight corresponding to each friend and the friend encoding features of each friend to obtain the first user feature of the target user relative to the candidate merchant.

[0007] In some embodiments, the first attention layer includes a cascaded first splicing layer, a first multilayer perceptron, and a normalization layer; the friend influence weight determination unit includes: a first splicing unit, configured to splice the user embedding feature of the target user and the merchant embedding feature of the candidate merchant with the friend encoding feature of each friend respectively by the first splicing layer to obtain the splicing feature corresponding to each friend; an initial friend influence weight determination unit, configured to process the splicing feature corresponding to each friend by the first multilayer perceptron to obtain the initial friend influence weight of each friend; and a first normalization processing unit, configured to normalize the initial friend influence weight of each friend by the normalization layer to obtain the friend influence weight of each friend for triggering interactive behavior of the candidate merchant for the target user.

[0008] In some embodiments, the second graph attention network includes a second encoding layer, a second attention layer, and a second linear layer; the first merchant feature determination module includes: a neighbor encoding feature determination unit, configured to encode each neighbor merchant in the neighbor merchant relationship graph of the candidate merchant by the second encoding layer to obtain the neighbor encoding features of each neighbor merchant; a neighbor influence weight determination unit, configured to perform attention calculation by the second attention layer based on the neighbor encoding features of each neighbor merchant, the distance between each neighbor merchant and the candidate merchant, the merchant embedding features of the candidate merchant, and the user embedding features of the target user to obtain the neighbor influence weight of each neighbor merchant for the target user to trigger interactive behavior on the candidate merchant; and a second linear weighting unit, configured to perform linear weighting by the second linear layer based on the neighbor influence weights corresponding to each neighbor merchant and the neighbor encoding features of each neighbor merchant to obtain the first merchant feature of the candidate merchant relative to the target user.

[0009] In some embodiments, the dual attention network includes a third encoding layer, a third attention network, and a fourth attention network; the second feature determination module includes: an encoding unit, configured to encode the users and merchants in the interaction relationship graph by the third encoding layer respectively, to obtain merchant encoding features of each merchant and user encoding features of each user in the interaction relationship graph; a second merchant feature determination unit, configured to process the third attention network according to the merchant encoding features of each merchant in the interaction relationship graph, the user encoding features of each user, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain second merchant features of the candidate merchants; and a second user feature determination unit, configured to process the fourth attention network according to the user encoding features of each user in the interaction relationship graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain second user features of the target user.

[0010] In some embodiments, the third attention network includes a pooling layer, a third attention layer, and a third linear layer; the second merchant feature determination unit includes: a pooling processing unit, used by the pooling layer to perform pooling processing based on the user coding features of all users in the interaction relationship graph to obtain intermediate user features; a merchant weight determination unit, used by the third attention layer to perform attention calculation based on the intermediate user features, the merchant coding features of each merchant in the interaction relationship graph, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the merchant weight of each merchant in the interaction relationship graph, wherein the merchant weight refers to the weight of the corresponding merchant in triggering an interaction behavior with the candidate merchant for the target user; and a third linear weighting unit, used by the third linear layer to perform linear weighting based on the merchant weights and merchant coding features of each merchant in the interaction relationship graph to obtain the second merchant features of the candidate merchants.

[0011] In some embodiments, the fourth attention network includes a fourth attention layer and a fourth linear layer; the second user feature determination unit includes: a user weight determination unit, configured to perform attention calculation by the fourth attention layer based on the user coding features of each user in the interaction relationship graph, the second merchant features of the candidate merchant, the user embedding features of the target user, and the merchant embedding features of the candidate merchant, to obtain the user weight of each user in the interaction relationship graph; the user weight refers to the weight of the corresponding user in triggering the interaction behavior of the candidate merchant for the target user; and a fourth linear weighting unit, configured to perform linear weighting by the fourth linear layer based on the user weight of each user in the interaction relationship graph and the user coding features of each user, to obtain the second user feature of the target user.

[0012] In some embodiments, the interaction prediction network includes a second stitching layer, a second multilayer perceptron, and a classification layer; the interaction probability determination module includes: a second stitching unit, configured to stitch together the global user features of the target user and the global merchant features of the candidate merchants by the second stitching layer to obtain target stitched features; an interaction score determination unit, configured to transform the target stitched features by the multilayer perceptron to obtain the interaction score of the target user with the candidate merchants; and an interaction probability determination unit, configured to normalize the interaction score by the classification layer to obtain the interaction probability of the target user with the candidate merchants.

[0013] In some embodiments, the device for determining the delivery information further includes: a training data acquisition module, configured to acquire training data, the training data including multiple training samples, the training samples indicating sample users, sample merchants, and sample tags, the sample tags being used to indicate whether the sample user triggers an interactive behavior with the sample merchant; a first output module, configured to output a first user feature of the sample user relative to the sample merchant by the first graph attention network based on the friend relationship graph of the sample user and the merchant embedding features of the sample merchant; a second output module, configured to output a first merchant feature of the sample merchant relative to the sample user by the second graph attention network based on the neighbor merchant relationship graph of the sample merchant, the distance between each neighbor merchant and the sample merchant, and the user embedding features of the sample user; and a feature determination module, configured to determine a second user feature of the sample user and a second user feature of the sample user by the dual attention network based on the sample interaction relationship graph. The sample merchant's second merchant feature; a first fusion module, used to fuse the sample user's user embedding feature, first user feature, and second user feature to obtain the sample user's global user feature; a second fusion module, used to fuse the sample merchant's merchant embedding feature, first merchant feature, and second merchant feature to obtain the sample merchant's global merchant feature; a third output module, used by the interaction prediction network to output the interaction probability of the sample user with the sample merchant based on the sample user's global user feature and the sample merchant's global merchant feature; a loss value calculation module, used to calculate the loss value of the loss function based on the sample user's interaction probability with the sample merchant and the sample label; and a back-adjustment module, used to back-adjust the parameters of the first graph attention network, the second graph attention network, the dual attention network, and the interaction prediction network based on the loss value until the training termination condition is met.

[0014] In some embodiments, the device for determining delivery information further includes: a distance calculation module for calculating the distance between the candidate merchant and other merchants; a filtering module for filtering out the first number of merchants closest to the candidate merchant based on the distance between the candidate merchant and other merchants, as the neighbor merchants of the candidate merchant; and a neighbor merchant relationship graph construction module for constructing a neighbor merchant relationship graph of the candidate merchant based on the neighbor merchants of the candidate merchant.

[0015] In some embodiments, the device for determining the delivery information further includes: a first interaction record acquisition module, used to acquire interaction records of each merchant; and an interaction relationship graph construction module, used to construct the interaction relationship graph based on the merchants and users with interaction relationships in the interaction records.

[0016] In some embodiments, the device for determining the delivery information further includes: a second interaction record acquisition module, used to acquire interaction records of a specified merchant; a reference merchant determination module, used to select a second number of neighboring merchants of the specified merchant as reference merchants if the number of interactions with the specified merchant is less than a set threshold based on the interaction records of the specified merchant; a user score determination module, used to calculate the user score corresponding to each reference user based on the interaction probability of each reference user in the reference user set for each reference merchant; the reference user refers to a user whose interaction probability with the reference merchant exceeds a probability threshold; a target reference user determination module, used to select a third number of reference users with the highest user scores from the reference user set as target reference users based on the user scores corresponding to each reference user; and a sending module, used to return the delivery information of the specified merchant to the target reference user if an information request initiated by the target reference user is received.

[0017] According to one aspect of the embodiments of this application, an electronic device is provided, including: a processor; a memory, the memory storing computer-readable instructions, which, when executed by the processor, implement the method for determining the delivery of information as described above.

[0018] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, on which computer-readable instructions are stored, which, when executed by a processor, implement the method for determining the delivery of information as described above.

[0019] According to one aspect of the embodiments of this application, a computer program product is provided, including computer instructions that, when executed by a processor, implement the method for determining the delivery of information as described above.

[0020] In this application, a first graph attention network determines the target user's first user characteristics under the influence of friends based on the target user's friend relationship graph. A second graph attention network determines the candidate merchant's first merchant characteristics under the influence of neighboring merchants based on the candidate merchant's neighbor merchant relationship graph. A dual attention network determines the target user's second user characteristics and the candidate merchant's second merchant characteristics based on the interaction relationship graph. The global user characteristics of the target user, obtained by combining the target user's first user characteristics, second user characteristics, and user embedding characteristics, can reflect the target user's user characteristics in multiple dimensions, and the expressed user characteristics are more accurate and comprehensive. Similarly, the global merchant characteristics of the candidate merchants, obtained by combining the candidate merchants' first merchant characteristics, second merchant characteristics, and merchant embedding characteristics, reflect the candidate merchants' merchant characteristics in multiple dimensions, and the expressed merchant characteristics are more accurate and comprehensive. Based on this, the prediction of the target user's interaction probability with the candidate merchants based on the target user's global user characteristics and the candidate merchants' global merchant characteristics can ensure the accuracy of the predicted interaction probability.

[0021] Since interaction probability reflects the probability that a target user is interested in a candidate merchant, screening candidate merchants and determining target merchants based on interaction probability can ensure the matching between the determined target merchants and the target users. This, in turn, ensures the matching between the target merchant's advertising information and the target users. Therefore, using the target merchant's advertising information as the information to be advertised to the target users means that the target merchant's advertising information has a high probability of being the information the target users need. As a result, users do not need to repeatedly request information or repeatedly modify their search terms to obtain the information they need, which can effectively improve the efficiency of users obtaining information. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0023] Figure 1 A is a schematic diagram of a friend relationship graph according to one embodiment.

[0024] Figure 1 B is a schematic diagram of a neighboring merchant relationship diagram according to one embodiment.

[0025] Figure 1 C is a schematic diagram of an interaction relationship diagram according to one embodiment.

[0026] Figure 2This is a schematic diagram illustrating an application scenario of a method for determining delivery information according to an embodiment of this application.

[0027] Figure 3 This is a flowchart illustrating a method for determining delivery information according to an embodiment of this application.

[0028] Figure 4 This is a flowchart illustrating the prediction of interaction probability according to an embodiment of this application.

[0029] Figure 5 This is a flowchart illustrating step 310 according to an embodiment of this application.

[0030] Figure 6 This is a flowchart illustrating step 320 according to an embodiment of this application.

[0031] Figure 7 This is a flowchart illustrating step 330 according to an embodiment of this application.

[0032] Figure 8 This is a schematic diagram illustrating a dual attention network used in an embodiment of this application to determine the second merchant features of a candidate merchant and the second user features of a target user.

[0033] Figure 9 This is a flowchart illustrating a method for determining delivery information according to another embodiment of this application.

[0034] Figure 10 This is a flowchart illustrating a method for determining delivery information according to another embodiment of this application.

[0035] Figure 11 This is a flowchart illustrating a method for determining delivery information according to a specific embodiment of this application.

[0036] Figure 12 This is a block diagram of a device for determining delivery information according to an embodiment of this application.

[0037] Figure 13 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0038] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0039] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0040] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0041] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0042] It should be noted that "multiple" in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0043] It is understood that in the specific implementation of this application, data related to users' friend relationships and user access records of merchants are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0044] Before providing a detailed description of the solution in this application, the terms used are explained as follows:

[0045] Graph Attention Networks (GAT) are neural networks that combine attention mechanisms with Graph Neural Networks (GNNs). GATs are primarily used to process graph-structured data, which consists of nodes and edges connecting them. Graph-structured data is used to depict the relationships between different nodes. Examples of graph-structured data include social networks, knowledge graphs, complex file systems, friend graphs, neighbor-business graphs, and interaction graphs (described below).

[0046] When aggregating node information in graph structure data, the graph attention network uses an attention mechanism to assign importance weights to neighboring nodes for the current target node. Then, based on the importance weights of each neighboring node for the target node, it aggregates the feature representations of the current target node and the feature representations of the neighboring nodes to obtain the final feature representation of the target node.

[0047] Dual Attention Networks (DANs) are networks that collaboratively learn the influence of dual contextual information and assign attention weights based on each layer of contextual information, thereby mining the features of nodes under the mutual influence of these two layers of contextual information. For example, the interaction graph in this application includes user nodes representing users and merchant nodes representing merchants. The feature representation of any node in the interaction graph is influenced not only by its neighboring user nodes but also by its neighboring merchant nodes. The influence of neighboring user nodes can be considered as one layer of contextual information, while the influence of neighboring merchant nodes can be considered as another layer. In this scheme, a dual attention network is used to extract the features of nodes under the mutual influence of dual contextual information from merchant nodes and user nodes in the interaction graph. In a specific embodiment, the dual attention network may include two graph attention networks, whereby one graph attention network focuses on features under the influence of one layer of contextual information.

[0048] Friend relationship graph: This is a relationship graph constructed based on the friend relationships between users. A node in the friend relationship graph represents a user P. i The edge connecting two adjacent nodes (P) i P j ) represents user P i and user P jThey are friends with each other. In this context, two users who are friends with each other in the friend relationship graph can be users who are friends within the same application, or users who are friends in different applications. Figure 1 A is a schematic diagram of a friend relationship graph according to one embodiment, such as... Figure 1 As shown in Figure A, users P1 and P2, P1 and P3, P1 and P4, and P2 and P3 are friends.

[0049] Neighbor Merchant Relationship Graph: This refers to a network of relationships built based on the distance between merchants. A node in the neighbor merchant relationship graph represents a merchant v. a The edge connecting two adjacent nodes (S) a S b ) indicates merchant S b It is the distance from merchant S a One of the k most recent merchants, where k is a positive integer. k can be set according to actual needs, for example, k can be set to 30, etc. Figure 1 Example B illustrates a schematic diagram of the neighboring merchant relationship graph.

[0050] Interaction Relationship Graph: This is a graph depicting the relationships between users and merchants based on user interactions. These interactions can include visiting a merchant's page, buying products from the merchant, leaving a review, liking or favorited the merchant's page, or entering the merchant's live stream, among other things. No specific limitations are specified here. Nodes in the interaction relationship graph include user nodes representing users and merchant nodes representing merchants. Edges connect user nodes and merchant nodes. An edge (P...) represents a user node and a merchant node. i S j ) represents user P i For merchant S j Interactive behavior was triggered. Figure 1 C exemplarily illustrates a schematic diagram of an interaction relationship graph.

[0051] Audience targeting: Before placing an ad, a merchant (also known as a Point of Interest, POI) needs to define the target audience as a targeting condition on the ad placement platform. This process of defining the target audience is called audience targeting.

[0052] In related technologies, with the development of internet technology, the internet has gradually become an important channel for information transmission. However, due to the vast variety and large volume of information on the internet, users may need to interact with the server multiple times and filter information repeatedly to find the information they need or are interested in from the massive amount of data. Because users need to interact with the server multiple times to obtain the information they need or are interested in, related technologies suffer from low efficiency in information retrieval. This application addresses this problem by incorporating deep learning technology from artificial intelligence.

[0053] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0054] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0055] Figure 2 This is a schematic diagram illustrating an application scenario of a method for determining delivery information according to an embodiment of this application, such as... Figure 2 As shown, this application scenario includes terminal 210 and information delivery platform 220. Terminal 210 can establish a communication connection with information delivery platform 220 via wired or wireless network.

[0056] Terminal 210 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, in-vehicle terminal, smart TV, or other interactive electronic device, without specific limitations.

[0057] The information delivery platform 220 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0058] The information delivery platform 220 can, according to the method of this application, construct a friend relationship graph for each user, a neighbor merchant relationship graph for merchants, and an interaction relationship graph based on the interaction records of merchants. It then deploys a first attention network, a second attention network, and a dual attention network. The first attention network determines the first user characteristics of the target user based on the friend relationship graph, the second attention network determines the first merchant characteristics of candidate merchants based on the neighbor merchant relationship graph, and the dual attention network determines the second user characteristics of the target user and the second merchant characteristics of candidate merchants based on the interaction relationship graph. Based on the user embedding characteristics, first user characteristics, and second user characteristics of the target user, as well as the merchant embedding characteristics, first merchant characteristics, and second merchant characteristics of the candidate merchants, the interaction probability is predicted to obtain the interaction probability of the target user triggering an interaction behavior with the candidate merchants. Finally, the target merchants are selected and determined for the user based on the interaction probability of the target user with the candidate merchants.

[0059] The information delivery platform 220 can associate and store the user identifier of the target user with the merchant identifier of the target merchant determined for the target user. Thus, after receiving an information request initiated by the terminal 210, the information delivery platform 220 obtains the user identifier carried in the information request, determines the merchant identifier associated with the user identifier, and then sends the delivery information of the merchant indicated by the determined merchant identifier to the terminal 210.

[0060] The merchant's delivery information can be introductory information (such as introductory text, introductory video, etc.), the merchant's page, the merchant's advertisement, etc., without specific limitations. After detecting the triggered information acquisition operation, terminal 210 sends an information request to information delivery platform 220. Information acquisition operations include page refresh operations, opening a specified page operations, video playback operations, etc., without specific limitations.

[0061] The information delivery platform 220 can also communicate and connect with merchant platforms, thereby obtaining delivery information from various merchants. Merchant platforms can be online food delivery platforms, online ride-hailing platforms, online shopping platforms, online ticketing service platforms, or comprehensive lifestyle service platforms. Comprehensive lifestyle service platforms are merchant platforms that provide various types of services such as food, movies, beauty, nail salons, and sports.

[0062] In some embodiments, the merchant platform can also be used as an information delivery platform 220, thereby enabling the merchant platform to perform the methods of this application.

[0063] In some embodiments, the information delivery platform 220 can be an advertising delivery platform, and correspondingly, the merchant's delivery information can be the merchant's advertisement.

[0064] It is worth mentioning that the method of this application is not limited to being executed by the information delivery platform 220, but can also be executed by the terminal 210 with sufficient computing power, or can be executed jointly by the information delivery platform 220 and the terminal 210.

[0065] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0066] Figure 3 This is a flowchart illustrating a method for determining delivery information according to an embodiment of this application. This method can be executed by a computer device with processing capabilities, such as a server, and is not specifically limited thereto. (Refer to...) Figure 3 As shown, the method includes at least steps 310 to 370, which are described in detail below:

[0067] Step 310: The first graph attention network outputs the first user feature of the target user relative to the candidate merchants based on the target user's friend relationship graph and the merchant embedding features of the candidate merchants.

[0068] The target user refers to the user whose information is to be targeted, and the target user's friend relationship graph indicates other users who have a friend relationship with the target user.

[0069] In the scheme of this application, the information to be delivered can be a merchant's merchant page (such as a merchant's product page or merchant's introduction page), a merchant's push article, a merchant's advertisement, etc. The merchants associated with the delivered information can be regarded as candidate merchants in this application.

[0070] In some embodiments, to ensure the quality of the information delivered, merchants can be pre-screened, and the screened merchants can be used as candidate merchants in this application. In practice, the number of times a merchant is interacted with can reflect the overall quality of the merchant to a certain extent (such as the quality of the merchant's service, the quality of the merchant's products, etc.). Therefore, merchants can be screened based on the number of times they are interacted with, for example, merchants whose number of interactions exceeds the threshold for the first interaction can be used as candidate merchants in this application.

[0071] The number of times a merchant is interacted with can be statistically calculated based on the merchant's interaction records. These records refer to the data generated after a user triggers an interaction with the merchant's advertising information. Such user interactions can include clicking on the merchant's advertising information (e.g., clicking on the merchant's advertisement or clicking on the merchant's page), adding the merchant to favorites, placing an order with the merchant, or liking the merchant's comments.

[0072] The first graph attention network refers to a graph attention network used to mine user features of target users from the friend relationship graph. Correspondingly, the user features identified by the first graph attention network for the target user are called the target user's first user features. The attention mechanism set in the first graph attention network can be a single-head attention mechanism or a multi-head attention mechanism, without specific limitations here.

[0073] In some embodiments of this application, before step 310, the method further includes: obtaining the target user's friend information; and constructing a friend relationship graph of the target user based on the target user's friend information. Specifically, after obtaining authorization from the target user, the method may request the target user's friend address book to obtain the target user's friend information.

[0074] As described above, in processing graph-structured data, the graph attention network first assigns weights to the neighboring nodes of the node to be processed (referred to as the target node for ease of description). Then, it aggregates the features of the target node's neighboring nodes, their corresponding weights, and the target node's features to obtain the target node's features. Specifically, in this application, the first graph attention network also first assigns weights to the target user's friends in the target user's friend relationship graph. Then, it combines the friends' weights and the friends' feature representations in the friend graph to perform feature aggregation to determine the target user's first user features.

[0075] In this application's scheme, the weights assigned to the target user's friends by the first graph attention network (i.e., the friend influence weights hereinafter) are used to indicate the influence weights of friends on the target user's interaction behavior triggered by the candidate merchant. It is understood that the influence weights of the target user's friends on the target user's interaction behavior triggered by the candidate merchant will differ depending on the candidate merchant. Therefore, in this application's scheme, the merchant embedding features of the candidate merchants are also used as one of the inputs to the first graph attention network, so that the first graph attention network combines the target user's friends and the merchant embedding features of the candidate merchants to comprehensively determine the target user's feature representation in the friend relationship graph, i.e., the target user's first user feature.

[0076] The merchant embedding features of a candidate merchant refer to the merchant characteristics determined based on the candidate merchant's merchant information. Merchant information includes, for example, the merchant's name, the goods or services offered, the industry to which the goods or services belong, the merchant's address, and the merchant's location, etc., without specific limitations. It is understood that, in this application, since the merchant embedding features of a candidate merchant are determined based on the candidate merchant's merchant information, these features are independent of the candidate merchant's contextual information, such as the candidate merchant's neighboring merchants in the neighbor merchant relationship graph and / or the users or merchants adjacent to the candidate merchant in the interaction relationship graph.

[0077] In some embodiments, the merchant embedding features of the candidate merchants can be obtained by initializing the basic information of the candidate merchants; in other embodiments, the merchant embedding features can be obtained by encoding the merchant information, for example, by performing one-hot encoding on the merchant information; in other embodiments, other encoder networks can be used to encode the merchant information, which is not specifically limited here.

[0078] Step 320: The second graph attention network outputs the first merchant feature of the candidate merchant relative to the target user based on the neighbor merchant relationship graph of the candidate merchant, the distance between each neighbor merchant and the candidate merchant, and the user embedding features of the target user.

[0079] The second graph attention network refers to a graph attention network used to determine merchant features based on the neighbor merchant relationship graph of candidate merchants. Correspondingly, the merchant features determined by the second graph attention network for candidate merchants are called the first merchant features of the candidate merchants. Similar to the first graph attention network, the attention mechanism in the second graph attention network can be a single-head attention layer or multiple attention layers.

[0080] In determining the second merchant feature of a candidate merchant, the second graph attention network first assigns weights (i.e., neighbor influence weights below) to each neighbor merchant in the neighbor merchant relationship graph of the candidate merchant, and then aggregates the weights and feature representations of the neighbor merchants to determine the second merchant feature of the candidate merchant.

[0081] The weights assigned to each neighboring merchant in the second graph attention network are used to indicate the influence weight of the neighboring merchants on the interactive behavior triggered by the target user. Since the needs or preferences of different target users are different, the influence weights of each neighboring merchant on the interactive behavior triggered by the target user are different. Therefore, in this application, the user embedding features of the target user are also used as one of the inputs of the second graph attention network.

[0082] The user embedding features of a target user refer to user characteristics determined based on the target user's basic user information, such as the user's place of residence, age, tag information, nickname, and location, etc., without specific limitations. It is understood that, in this application, since the user embedding features of a target user are determined based on the target user's user information, these features are independent of the target user's contextual information, such as the target user's friends in the friend relationship graph and / or adjacent users or merchants in the interaction relationship graph.

[0083] In some embodiments, the user embedding features of the target user can be obtained by encoding the user's basic information, for example, by performing one-hot encoding on the user's basic information. In other embodiments, the user's basic information can also be encoded using other encoder networks, which are not specifically limited here.

[0084] In practice, the distance between neighboring merchants and candidate merchants also affects the influence weight of neighboring merchants on the target user's interaction with the candidate merchant. Theoretically, all other things being equal, the farther a neighboring merchant is from the candidate merchant, the smaller its influence weight on the target user's interaction with the candidate merchant. Based on this, in the scheme of this application, the distance between the candidate merchant and each neighboring merchant is also used as one of the inputs of the second graph attention network. Thus, the weights assigned to neighboring merchants by the second graph attention network (i.e., the neighbor influence weights below) simultaneously take into account the characteristics of the target user and the distance between neighboring merchants and candidate merchants.

[0085] In some embodiments of this application, before step 330, the method further includes: calculating the distance between the candidate merchant and other merchants; selecting the first number of merchants closest to the candidate merchant based on the distance between the candidate merchant and other merchants, as the candidate merchant's neighbor merchants; and constructing a neighbor merchant relationship graph of the candidate merchant based on the candidate merchant's neighbor merchants.

[0086] The initial number can be set according to actual needs, such as 30. By selecting the first number of merchants closest to the candidate merchants as the candidate merchants' neighbors, the amount of data processing can be reduced while fully considering the influence of neighboring merchants on the candidate merchants.

[0087] In other embodiments, a distance threshold can be preset to designate merchants whose distance from the candidate merchant is less than the distance threshold as the candidate merchant's neighbor merchants, thereby constructing a neighbor merchant relationship graph for the candidate merchant.

[0088] Step 330: The dual attention network determines the second user characteristics of the target user and the second merchant characteristics of the candidate merchants based on the interaction relationship graph, which is constructed based on the user's interactive behavior triggered by the merchant.

[0089] As described above, the nodes in the interaction graph include user nodes representing users and merchant nodes representing merchants. An edge connecting a user node and a merchant node indicates that the user indicated by the user node triggered an interaction with the merchant represented by the merchant node. The characteristics of a node in the interaction graph are influenced not only by the characteristics of the user nodes in the graph but also by the characteristics of the merchant nodes. Therefore, in this application, a dual attention network is used to determine the user characteristics of the target user (i.e., the second user characteristics of the target user) and the merchant characteristics of the candidate merchants (i.e., the merchant characteristics of the candidate merchants) under the dual influence of the user nodes and merchant nodes in the interaction graph.

[0090] In this application, the interaction relationship graph is also a type of graph structure data; therefore, the dual attention network can also be a dual graph attention network. In some embodiments, the dual attention network may include two graph attention networks, one of which is used to determine the second user characteristics of the target user under the dual influence of user nodes and merchant nodes in the interaction relationship graph, and the other graph attention network is used to determine the second merchant characteristics of the candidate merchant under the dual influence of user nodes and merchant nodes in the interaction relationship graph.

[0091] In some embodiments of this application, before step 330, the method further includes: obtaining the interaction records of each merchant; and constructing an interaction relationship graph based on the merchants and users with interactive relationships in the interaction records.

[0092] The interaction record of a merchant indicates the user identifier that triggered the interaction with that merchant. Based on this interaction record, the merchant node representing that merchant and the user node representing the user indicated by the user identifier are determined in the interaction relationship graph, and the determined merchant node and user node are connected. This process is repeated for each interaction record of each merchant to construct the corresponding interaction relationship graph.

[0093] Step 340: The user embedding features, first user features, and second user features of the target user are fused to obtain the global user features of the target user.

[0094] In some embodiments, the user embedding features, first user features, and second user features of the target user can be concatenated, and the concatenated user features obtained can be used as the global user features of the target user.

[0095] In some embodiments, the user embedding features, first user features, and second user features of the target user can be superimposed, and the resulting first superposition result can be used as the global user features of the target user. For example, if the target user P i User embedding features are u i Target user P i The first user characteristic is Target user P i The second user characteristic is Then the global user characteristics of the target user uvec i for:

[0096]

[0097] In other embodiments, weighting coefficients can be set for the user embedding feature, the first user feature, and the second user feature respectively, so that during the superposition process, the user embedding feature, the first user feature, and the second user feature are linearly weighted according to the set weighting coefficients to obtain the first superposition result.

[0098] In some embodiments, the spliced ​​user features (or the first overlay result) can be further fused through a first fusion network to extract deep features, and the extracted deep features can be used as the global user features of the target user. The first fusion network can be constructed through a fully connected network or the like, and is not specifically limited here.

[0099] Step 350: The merchant embedding features, first merchant features, and second merchant features of the candidate merchants are fused to obtain the global merchant features of the candidate merchants.

[0100] In some embodiments, the merchant embedding features, first merchant features, and second merchant features of the candidate merchant can be concatenated, and the concatenated merchant features obtained can be used as the global merchant features of the candidate merchant.

[0101] In some embodiments, the merchant embedding features, first merchant features, and second merchant features of candidate merchants can be superimposed, and the resulting second superposition result can be used as the global user features of the target user. For example, if candidate merchant S j The merchant embedding feature is v j Candidate Merchant S j First Merchant Characteristics Candidate Merchant S j The second merchant characteristic is Candidate Merchant S j Global Merchant Features vvec j for:

[0102]

[0103] In other embodiments, weighting coefficients can be set for the merchant embedding feature, the first merchant feature, and the second merchant feature respectively. In this way, during the superposition process, the merchant embedding feature, the first merchant feature, and the second merchant feature are linearly weighted according to the set weighting coefficients to obtain the corresponding second superposition result.

[0104] In some embodiments, the spliced ​​merchant features (or the second overlay result) can be further fused through a second fusion network to extract deep features, and the extracted deep features can be used as the global merchant features of the candidate merchants. The second fusion network can be constructed through a fully connected network or the like, and is not specifically limited here.

[0105] Step 360: The interaction prediction network outputs the interaction probability of the target user relative to the candidate merchants based on the global user characteristics of the target user and the global merchant characteristics of the candidate merchants.

[0106] The probability of interaction between a target user and a candidate merchant refers to the probability that a target user will trigger an interactive behavior with a candidate merchant.

[0107] An interaction prediction network is a neural network model used to predict the probability of interaction. In some embodiments, the interaction prediction network may include one or more fully connected networks. These networks perform nonlinear transformations based on the global user characteristics of the target user and the global merchant characteristics of the candidate merchants. Then, a classification layer classifies and outputs the interaction probability of the target user relative to the candidate merchants.

[0108] In some embodiments, the interaction prediction network includes a second stitching layer, a second multilayer perceptron, and a classification layer; in this embodiment, step 360 includes: stitching together the global user features of the target user and the global merchant features of the candidate merchants by the second stitching layer to obtain target stitched features; transforming the target stitched features by the multilayer perceptron to obtain the interaction score of the target user with the candidate merchants; and normalizing the interaction score by the classification layer to obtain the interaction probability of the target user with the candidate merchants.

[0109] A multi-layered perceptron (MLP) is a fully connected feedforward neural network that includes an input layer, an output layer, and one or more hidden layers between the input and output layers. These hidden layers can be viewed as stacked layers of nonlinear transformations, progressively learning the nonlinear features of the data. In this application, for ease of distinction, the multi-layered perceptron in the interactive prediction network is referred to as a second multi-layered perceptron. The number of hidden layers in the second multi-layered perceptron can be set according to actual needs.

[0110] After the target stitched features are input into the input layer of the second multilayer perceptron, the hidden layers in the second multilayer perceptron perform nonlinear transformations layer by layer based on these features, and the output layer of the second multilayer perceptron outputs the nonlinear transformation result obtained after linear transformation. Then, the classification layer outputs the interaction probability of the target user with the candidate merchant based on the nonlinear transformation result. In some embodiments, the classification layer can use the classification function of Formula 3 below to perform classification, obtaining an interaction probability of the target user with the candidate merchant within the range of 0 to 1.

[0111] σ(x)=1 / (1+e -x ); (Formula 3)

[0112] Where x is the nonlinear transformation result output by the second multilayer perceptron, and σ(x) is the interaction probability.

[0113] Step 370: Based on the interaction probability of the target user relative to the candidate merchants, filter among multiple candidate merchants to determine the target merchant, so as to determine the target merchant's advertising information as the advertising information to be advertised to the target user.

[0114] Understandably, when there are multiple candidate merchants, repeating steps 310-360 above for each candidate merchant can yield the probability of interaction between the target user and each candidate merchant.

[0115] In some embodiments, for the same target user, after obtaining the interaction frequency of the target user with each of the multiple candidate merchants through steps 310-360 above, candidate merchants can be filtered according to a set frequency threshold. Specifically, candidate merchants with interaction frequencies higher than the frequency threshold can be identified as target merchants. The number of identified target merchants is not limited; it can be one or more.

[0116] In other embodiments, candidate merchants can be filtered according to a set number of merchants, such as N, where N is a positive integer. Specifically, the top N candidate merchants with the highest interaction frequency can be determined as target merchants.

[0117] The information that can be used to target merchants can be their merchant links, introductory information (such as introductory text, introductory videos, etc.), or advertisements, without any specific limitations.

[0118] Similarly, when there are multiple target users, for each target user, follow steps 310-360 above to obtain the interaction probability of each target user with the candidate merchant. In this case, the target users for the candidate merchant can also be filtered based on the interaction frequency of multiple target users with the same candidate merchant. The filtered target users are then used as the target users for the candidate merchant's advertising. The specific process for filtering target users is similar to that for filtering candidate merchants, and will not be repeated here. After determining one or more target users for a candidate merchant, the advertising information for the candidate merchant can be delivered to the target users.

[0119] In some embodiments, after determining the target merchant for the target user (and the target user for the target merchant), the user identifier of the target user can be associated with and stored with the merchant identifier of the target merchant (or the merchant identifier of the target merchant can be associated with and stored with the user identifier of the target user). Thus, after receiving an information request initiated by the terminal, the associated merchant identifier can be determined based on the user identifier carried in the information request, and the delivery information corresponding to the determined merchant identifier can be returned to the initiator of the information request.

[0120] Figure 4 This is a flowchart illustrating the prediction of interaction probabilities according to an embodiment of this application, such as... Figure 4 As shown, the first graph attention network outputs the first user feature of the target user based on the friend relationship graph, the dual attention network outputs the second user feature of the target user and the second merchant feature of the candidate merchant based on the interaction relationship graph, and the second graph attention network outputs the first merchant feature of the candidate merchant based on the neighbor merchant relationship graph. Then, based on the first user feature, user embedding feature, and second user feature of the target user, the global user feature of the target user is determined; and based on the first merchant feature, merchant embedding feature, and second merchant feature of the candidate merchant, the global merchant feature of the candidate merchant is determined. Finally, the global user feature of the target user and the global merchant feature of the candidate merchant are input into the second multilayer perceptron for nonlinear transformation, and then the classification layer outputs the interaction probability of the target user triggering an interaction behavior with the candidate merchant. Figure 4 In This represents the predicted probability of interaction between sample users and sample merchants.

[0121] In this application, a first graph attention network determines the target user's first user characteristics under the influence of friends based on the target user's friend relationship graph. A second graph attention network determines the candidate merchant's first merchant characteristics under the influence of neighboring merchants based on the candidate merchant's neighbor merchant relationship graph. A dual attention network determines the target user's second user characteristics and the candidate merchant's second merchant characteristics based on the interaction relationship graph. The global user characteristics of the target user, obtained by combining the target user's first user characteristics, second user characteristics, and user embedding characteristics, can reflect the target user's user characteristics in multiple dimensions, and the expressed user characteristics are more accurate and comprehensive. Similarly, the global merchant characteristics of the candidate merchants, obtained by combining the candidate merchants' first merchant characteristics, second merchant characteristics, and merchant embedding characteristics, reflect the candidate merchants' merchant characteristics in multiple dimensions, and the expressed merchant characteristics are more accurate and comprehensive. Based on this, the prediction of the target user's interaction probability with the candidate merchants based on the target user's global user characteristics and the candidate merchants' global merchant characteristics can ensure the accuracy of the predicted interaction probability.

[0122] Since interaction probability reflects the probability that a target user is interested in a candidate merchant, screening candidate merchants and determining target merchants based on interaction probability can ensure the matching between the determined target merchants and target users, and thus ensure the matching between the target merchant's advertising information and the target user. Therefore, using the target merchant's advertising information as the information to be advertised to the target user means that the target merchant's advertising information has a high probability of being the information the target user needs. As a result, users do not need to make multiple repeated information requests or screenings to obtain the information they need, which can effectively improve the efficiency of users obtaining information.

[0123] In some embodiments of this application, the first graph attention network includes a first encoding layer, a first attention layer, and a first linear layer; such as Figure 5 As shown, step 310 includes:

[0124] Step 510: The first encoding layer encodes the friends in the target user's friend relationship graph to obtain the friend encoding features of each friend.

[0125] Friend encoding features refer to the feature representations of friends obtained by encoding each friend through the first encoding layer. Based on the target friend's friend relationship graph, the first encoding layer extracts nodes representing the target user's friends, and then encodes these extracted friend-representing nodes to obtain the friend encoding features of that node.

[0126] Step 520: The first attention layer performs attention calculation based on the friend encoding features of each friend, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain the friend influence weight of each friend on the interactive behavior triggered by the target user to the candidate merchants.

[0127] In this application, the first attention layer combines the user embedding features of the target user, the merchant embedding features of the candidate merchants, and the friend encoding features of each friend to calculate the friend influence weight corresponding to each friend. The first attention layer can be a single-head attention layer or a multi-head attention layer; no specific limitation is made here.

[0128] For users in social networks, the preferences or social activities of their friends can influence them to some extent. Therefore, in this application, the second user characteristics of the target user under the influence of friends are determined based on the target user's friend relationship graph.

[0129] In some embodiments of this application, the first attention layer includes a cascaded first splicing layer, a first multilayer perceptron, and a first normalization layer; step 520 includes: the first splicing layer splicing the user embedding features of the target user and the merchant embedding features of the candidate merchants with the friend encoding features of each friend to obtain splicing features corresponding to each friend; the first multilayer perceptron processing the splicing features corresponding to each friend to obtain the initial friend influence weight of each friend; and the first normalization layer normalizing the initial friend influence weight of each friend to obtain the friend influence weight of each friend for triggering interactive behavior of the candidate merchant for the target user.

[0130] For example, if the target user P i User embedding features are u i Candidate Merchant S j The merchant embedding feature is v j If the target user P i The friend relationship graph includes the target user P i |F i | friends, vector set Each element in the table represents the target user P. i The friend coding characteristics of a friend, i.e., the target user P i The kth friend P i,k The friend coding feature is f i,k Candidate Merchant S j The merchant embedding feature is v j Then the first splicing layer can obtain the target user P. i The kth friend P i,k Corresponding splicing feature A k for:

[0131] Ak =[f i,k ;u i ;v j ]; (Formula 4)

[0132] The target user P obtained by the first multilayer perceptron i The kth friend P i,k The initial friend influence weight a(i,k) is:

[0133]

[0134] Where φ(x) = max(0,x), W f Let b be the weight matrix of the first multilayer perceptron. f Let be the bias matrix of the second multilayer perceptron. For W f The reverse, W f and b f It can be determined through model training.

[0135] The first normalization layer can be normalized according to the following formula to obtain the target user P. i The kth friend P i,k The influence weight of friends α(i,k):

[0136]

[0137] Step 530: The first linear layer performs linear weighting based on the friend influence weights of each friend and the friend coding features of each friend to obtain the first user features of the target user relative to the candidate merchants.

[0138] Specifically, the friend influence weight of each friend is used as a weighting coefficient to linearly weight the friend coding features of all the target user's friends, thus obtaining the target user's first user feature relative to the candidate merchants.

[0139] Continuing with the example above, the target user P can be calculated using the following formula. i Compared to candidate merchant S j First user characteristics

[0140]

[0141] Through steps 510-530 above, the first user characteristic of the target user is determined by combining the influence weight of each friend of the target user on the interactive behavior triggered by the target user to the candidate merchant and the friend coding characteristics of each friend.

[0142] In some embodiments of this application, the second graph attention network includes a second encoding layer, a second attention layer, and a second linear layer; such as Figure 6 As shown, step 320 includes:

[0143] Step 610: The second coding layer encodes each neighboring merchant in the neighboring merchant relationship graph of the candidate merchant to obtain the neighbor coding features of each neighboring merchant.

[0144] Neighbor coding features refer to the feature representations of neighboring merchants obtained by encoding each neighboring merchant of a candidate merchant by the second coding layer. The second coding layer extracts nodes representing the neighboring merchants of a candidate merchant from the neighboring merchant relationship graph, and then encodes the extracted nodes accordingly to obtain the neighbor coding features of that node.

[0145] Step 620: The second attention layer performs attention calculation based on the neighbor encoding features of each neighbor merchant, the distance between each neighbor merchant and the candidate merchant, the merchant embedding features of the candidate merchant, and the user embedding features of the target user, to obtain the neighbor influence weight of each neighbor merchant for the target user's interaction behavior with the candidate merchant.

[0146] Similar to the first attention layer, the second attention layer can be either a single-head attention layer or a multi-head attention layer; no specific limitations are made here.

[0147] For users, the neighboring merchants of a merchant have a certain influence on whether users trigger interactive behavior with that merchant. In this solution, the neighboring merchant relationship graph of the candidate merchants is combined to calculate the neighboring influence weight of the neighboring merchants on the user's interactive behavior with the candidate merchant, and then determine the first merchant feature of the candidate merchant under the influence of the neighboring merchants.

[0148] In some embodiments, the second attention layer includes a second splicing layer, a third multilayer perceptron, and a second normalization layer. In this embodiment, step 620 further includes: the second splicing layer splicing the merchant embedding features of the candidate merchant and the user embedding features of the target user with the neighbor encoding features of each neighbor merchant and the distance features corresponding to each neighbor merchant to obtain the neighbor merchant splicing features corresponding to each neighbor merchant; wherein the distance features corresponding to the neighbor merchants are determined based on the distance between the neighbor merchants and the candidate merchants; the second multilayer perceptron processes the neighbor merchant splicing features corresponding to each neighbor merchant to obtain the initial neighbor influence weights corresponding to each neighbor merchant; and the second normalization layer normalizes the initial merchant influence weights corresponding to each neighbor merchant to obtain the neighbor influence weights of each neighbor merchant for triggering interactive behavior of the target user on the candidate merchant.

[0149] Step 630: The second linear layer performs linear weighting based on the neighbor influence weights of each neighbor merchant and the neighbor coding features of each neighbor merchant to obtain the first merchant features of the candidate merchant relative to the target user.

[0150] Continuing with the example above, if candidate merchant S j The neighbor merchant relationship diagram includes candidate merchant S j |L j |Nearby merchants, candidate merchant S j The kth neighboring merchant S j,k The neighbor encoding feature is l j,k Using vector sets Indicates candidate merchant S j |L j A set of neighbor coding features for each neighboring merchant.

[0151] First, calculate the candidate merchant S according to the following formula 8. j The kth neighboring merchant S j,k The corresponding distance feature g(j,k):

[0152] g(j,k)=1 / (0.5+d(S j S j,k ); (Formula 8)

[0153] Wherein d(S j S j,k ) indicates candidate merchant S j The kth neighboring merchant S j,k With candidate merchant S j The distance between them.

[0154] The second splicing layer obtains the candidate merchant S according to the following formula 9. j The kth neighboring merchant S j,k Corresponding splicing feature B j,k :

[0155] B j,k =[l j,k ;v j ;u i ;g(j,k)]; (Formula 9)

[0156] The third multilayer perceptron outputs candidate merchant S according to the following formula 10. j The kth neighboring merchant S j,k The corresponding initial neighbor influence weight b(j,k) is:

[0157]

[0158] Among them, W lLet b be the weight matrix of the third multilayer perceptron. l This is the bias matrix of the third multilayer perceptron. For W l The reverse, W l and b l It can be determined through training.

[0159] The second normalization layer outputs candidate merchant S according to the following formula 11. j The kth neighboring merchant S j,k The corresponding neighbor influence weight β(j,k):

[0160]

[0161] Then, the second linear layer outputs candidate merchant S according to the following formula 12. j Relative to target user P i First Merchant Characteristics

[0162]

[0163] Through steps 610-630 above, the first merchant feature of the candidate merchant under the influence of neighboring merchants is determined by combining the neighbor influence weight of each neighboring merchant for the target user's interactive behavior towards the candidate merchant and the neighbor coding features of each neighboring merchant.

[0164] In this scheme, the first graph attention network calculates the influence weight of friends and the second graph attention network calculates the influence weight of neighbors. It uses not only the context-independent features of the target user (i.e., user embedding features) and the context-independent features of the candidate merchants (i.e., merchant embedding features), but also the context-independent features of the candidate merchants (i.e., merchant embedding features). Therefore, this weight calculation method can be called the calculation method of context-aware attention network.

[0165] In some embodiments of this application, the dual attention network includes a third coding layer, a third attention network, and a fourth attention network; such as Figure 7 As shown, step 330 includes:

[0166] Step 710: The third coding layer encodes the users and merchants in the interaction relationship graph respectively, to obtain the merchant coding features of each merchant and the user coding features of each user in the interaction relationship graph.

[0167] Step 720: The third attention network processes the merchant coding features of each merchant, the user coding features of each user, the user embedding features of the target user, and the merchant embedding features of the candidate merchants in the interaction relationship graph to obtain the second merchant features of the candidate merchants.

[0168] In some embodiments of this application, the third attention network includes a pooling layer, a third attention layer, and a third linear layer; step 720 includes: the pooling layer performing pooling processing based on the user coding features of all users in the interaction relationship graph to obtain intermediate user features; the third attention layer performing attention calculation based on the intermediate user features, the merchant coding features of each merchant in the interaction relationship graph, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the merchant weight of each merchant in the interaction relationship graph, where the merchant weight refers to the weight of the corresponding merchant in triggering the interaction behavior of the candidate merchant for the target user; and the third linear layer performing linear weighting based on the merchant weight of each merchant in the interaction relationship graph and the merchant coding features of each merchant to obtain the second merchant features of the candidate merchants.

[0169] In some embodiments, the pooling layer can perform average pooling based on the user coding features of all users in the interaction graph to obtain intermediate user features. (User-POI summary vector):

[0170]

[0171] The network structure of the third attention layer is similar to that of the first attention layer mentioned above. The third attention layer includes a third splicing layer, a fourth multilayer perceptron, and a third normalization layer. The merchant coding features of each merchant in the third splicing layer are spliced ​​with the intermediate user features, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the second merchant splicing features corresponding to each merchant.

[0172] If using vector sets This represents the user coding characteristics of each user in the interaction graph, where p i,k This represents the user-coded features (User-POI embeddings) of the k-th user in the interaction graph; using a vector set. Let q represent the merchant encoding features (POI-User embeddings) of the k-th merchant in the interaction graph, where q j,k This represents the merchant coding feature of the k-th merchant in the interaction graph.

[0173] The third splicing layer outputs the splicing feature C of the second merchant corresponding to the k-th merchant in the interaction relationship graph. k for:

[0174]

[0175] The fourth multilayer perceptron splices the feature C of the second merchant corresponding to the k-th merchant. k The initial merchant weight c(j, k) corresponding to the kth merchant is:

[0176]

[0177] Among them, W q Let b be the weight matrix of the fourth multilayer perceptron. q This is the bias matrix of the fourth multilayer perceptron. For W q The reverse, W q and b q It can be determined through training.

[0178] The third normalization layer normalizes the initial merchant weights to obtain the merchant weight γ(j,k) corresponding to the k-th merchant:

[0179]

[0180] The third linear layer linearly weights the merchant weight γ(j,k) and the merchant coding features corresponding to each merchant to obtain the second merchant features of the candidate merchants. (POI-User vector):

[0181]

[0182] Step 730: The fourth attention network processes the user coding features of each user in the interaction relationship graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the second user features of the target user.

[0183] In some embodiments of this application, the fourth attention network includes a fourth attention layer and a fourth linear layer; step 730 includes: the fourth attention layer performing attention calculation based on the user coding features of each user in the interaction relationship graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the user weight of each user in the interaction relationship graph; the user weight refers to the weight of the corresponding user in triggering the interaction behavior of the candidate merchant for the target user; the fourth linear layer performing linear weighting based on the user weight of each user in the interaction relationship graph and the user coding features of each user to obtain the second user features of the target user.

[0184] Similar to the first attention layer, the fourth attention layer includes a fourth splicing layer, a fifth multilayer perceptron, and a fourth normalization layer. The fourth splicing layer outputs the splicing feature D of the second user corresponding to the k-th user in the interaction graph. k for:

[0185]

[0186] The fifth multilayer perceptron splices the features D of the second user corresponding to the k-th user. k The initial user weight d(j, k) corresponding to the k-th user is:

[0187]

[0188] The fourth normalization layer normalizes the initial user weights corresponding to the k-th user, resulting in the user weight ε(i,k) for the k-th user:

[0189]

[0190] The fourth linear layer performs a linear weighting based on the user weights and user coding features corresponding to each user, resulting in the second user feature (User-POI vector) for the target user. for:

[0191]

[0192] Through the above process, the second user characteristics of the target user and the second user characteristics of the candidate merchants are determined based on the interaction relationship graph.

[0193] In some embodiments of this application, the first, third, fourth, and fifth perceptrons in the above embodiments can be replaced with fully connected networks. The number of layers in the fully connected network and the number of neurons in each layer can be set according to actual needs.

[0194] Figure 8 This is a schematic diagram illustrating a dual attention network used in an embodiment of this application to determine the second merchant characteristics of candidate merchants and the second user characteristics of target users. The dual attention network includes a third graph attention network and a fourth graph attention network. Figure 8 As shown, the user coding features of all users in the interaction graph are pooled to obtain intermediate user features. Then, the intermediate user features and the merchant coding features of merchants in the interaction graph are input into the third graph attention network, which outputs the second merchant features of candidate merchants. After that, the user coding features of users in the interaction graph and the second merchant features of candidate merchants are input into the fourth graph attention network, which outputs the second user features of the target user.

[0195] In other embodiments of this application, the merchant coding features of all merchants in the interaction relationship graph can be pooled first to obtain intermediate merchant features; then the intermediate merchant features and the user coding features of users in the interaction relationship graph can be input into the fourth graph attention network, and the second user features of the target user can be output from the second merchant features of the candidate merchants; then the second user features of the target user and the merchant coding features of the merchants in the interaction relationship graph can be input into the third graph attention network, and the second merchant features of the candidate merchants can be output from the third graph attention network.

[0196] In some embodiments of this application, since the process of determining the first user characteristics of the target user requires the use of the friend encoding characteristics of each friend of the target user, and the process of determining the first merchant characteristics of the candidate merchant requires the use of the neighbor encoding characteristics of each neighbor merchant of the candidate merchant, and since this application may need to perform interaction probability calculations for multiple candidate merchants, in order to reduce the processing load of the first graph attention network and the second graph attention network, each node in the friend relationship graph and each node in the neighbor merchant relationship graph can be pre-encoded to obtain the encoding characteristics of each node in the friend relationship graph and the encoding characteristics of each node in the neighbor merchant relationship graph.

[0197] Therefore, in subsequent calculations, the encoded features of the target user's friends can be used as the friend encoded features in the above embodiment, and the encoded features of the candidate merchant's neighboring merchants can be used as the neighbor encoded features in the above embodiment. This eliminates the need to deploy corresponding encoding layers in the first and second graph attention networks. Consequently, the encoded features of each node in the friend relationship graph and the neighboring merchant relationship graph can be shared, i.e., used to determine the second user features of other target users or the second merchant features of other candidate merchants. Furthermore, this reduces the complexity of the first and second graph attention networks and improves training speed.

[0198] In some embodiments of this application, an encoder network can be used to encode nodes in a friend relationship graph and a neighbor merchant relationship graph using the Line algorithm, thereby obtaining the encoding features of each node. The Line algorithm describes the similarity between two nodes through their second-order similarity, where the second-order similarity is represented by the first-order similarity of each node in the two nodes. Specifically, if there is a direct edge between two nodes x and y, then the edge weight w x,y Let O be the first-order similarity between the two nodes; if there is no direct edge between the two nodes, the first-order similarity is zero; if O is not the first-order similarity, then the first-order similarity is zero. x =(w x,1 ,w x,2 ,...,w x,|X| If y represents the first-order similarity between vertex x and all other vertices, then the second-order similarity between nodes x and y is O(n). xand O y The similarity is represented by the fact that if there are no identical neighboring nodes, the second-order similarity is 0.

[0199] In some embodiments of this application, such as Figure 9 As shown, the method also includes:

[0200] Step 910: Obtain training data. The training data includes multiple training samples. The training samples indicate the sample user, sample merchant, and sample label. The sample label is used to indicate whether the sample user triggers an interactive behavior with the sample merchant.

[0201] The training samples include positive samples and negative samples. Each negative sample can be represented as (sample user ID, sample merchant ID, 0), where 0 is the sample label indicating that the sample user did not trigger an interaction with the sample merchant. Each positive sample can be represented as (sample user ID, sample merchant ID, 1), where 1 is the sample label indicating that the sample user triggered an interaction with the sample merchant.

[0202] In a specific embodiment, the interaction records of each merchant can be obtained from the merchant platform, and the feedback records of users regarding the advertising information of the merchants can be obtained from the user terminal. For example, if a user clicks on the advertising information of a merchant, a positive sample can be constructed based on the feedback record of clicking the advertising information; if the feedback record indicates that the user did not click on the advertising information of a merchant, a negative sample can be constructed based on the feedback record of not clicking the advertising information.

[0203] In some embodiments, to ensure training effectiveness and avoid excessive negative samples affecting training results, negative samples can be downsampled to reduce their number. In a specific embodiment, the ratio of positive to negative samples can be preset, for example, a ratio of 1:5, and the number of positive and negative samples can be determined based on this ratio.

[0204] Step 920: The first graph attention network outputs the first user feature of the sample user relative to the sample merchant based on the friend relationship graph of the sample user and the merchant embedding features of the sample merchant.

[0205] Step 930: The second graph attention network outputs the first merchant feature of the sample merchant relative to the sample user based on the neighbor merchant relationship graph of the sample merchant, the distance between each neighbor merchant and the sample merchant, and the user embedding features of the sample user.

[0206] Step 940: The dual attention network determines the second user characteristics of the sample users and the second merchant characteristics of the sample merchants based on the sample interaction relationship graph.

[0207] Step 950: The user embedding features, first user features, and second user features of the sample users are fused to obtain the global user features of the sample users.

[0208] Step 960: The merchant embedding features, first merchant features, and second merchant features of the sample merchants are fused to obtain the global merchant features of the sample merchants.

[0209] Step 970: The interaction prediction network outputs the interaction probability of the sample user to the sample merchant based on the global user characteristics of the sample user and the global merchant characteristics of the sample merchant.

[0210] Step 980: Calculate the loss value of the loss function based on the interaction probability of sample users with sample merchants and sample tags.

[0211] In some embodiments, the loss function can be the cross-entropy loss function:

[0212]

[0213] in, Represents the set of positive and negative samples. y represents the predicted probability of interaction between sample users and sample merchants. ij This indicates the actual interaction probability indicated by the sample label.

[0214] In other embodiments, the loss function may also be other absolute value loss functions, squared loss functions, etc., and no specific limitation is made here.

[0215] Step 990: Adjust the parameters of the first image attention network, the second image attention network, the dual attention network, and the interaction prediction network in reverse according to the loss value until the training termination condition is met.

[0216] The training termination condition can be either the convergence of the loss function or the reaching of a set iteration threshold. After adjusting the parameters of the first graph attention network, the second graph attention network, the dual attention network, and the interaction prediction network, the adjusted first graph attention network, the second graph attention network, the dual attention network, and the interaction prediction network re-predict the interaction probability. This process is repeated, enabling the first graph attention network, the second graph attention network, the dual attention network, and the interaction prediction network to learn the ability to accurately predict the interaction probability.

[0217] In some embodiments of this application, such as Figure 10 As shown, the method further includes:

[0218] Step 1010: Obtain the interaction records of the specified merchant.

[0219] Step 1020: If the number of interactions with a specified merchant is less than a set threshold based on the interaction records of the specified merchant, select the second number of neighboring merchants of the specified merchant as reference merchants.

[0220] In some embodiments, if the number of interactions with a specified merchant is not less than a set threshold, the specified merchant can be directly used as a candidate merchant, and the probability of the target user interacting with the specified merchant can be predicted according to the method described in the above embodiments. It is understood that when the number of interactions with a specified merchant is not less than the set threshold, it indicates that the quality (service quality, product quality, etc.) of the specified merchant is recognized by the user. Therefore, the specified merchant is used as a candidate merchant, and the advertising information of the specified merchant is used as candidate advertising information. This ensures the quality of the advertising information determined for the target user according to the method of this application.

[0221] It is understandable that if the number of interactions with a specified merchant is less than the set threshold, it may be because the specified merchant is a new merchant, and therefore the current number of interactions with the specified merchant is less than the set threshold. In this case, the method of this embodiment can be used to determine candidate users for the specified merchant, thereby increasing the probability that the specified merchant, which is a new merchant, will be discovered.

[0222] In some embodiments, in step 1020, filtering conditions can be set to select one or more neighboring merchants of a specified merchant as reference merchants. The filtering conditions may include: the distance from the specified merchant is less than a set distance threshold (or the third nearest neighboring merchant to the specified merchant); and / or the number of interactions with the neighboring merchants is not less than a second threshold. For example, the five nearest neighboring merchants to the specified merchant can be selected as reference merchants, or neighboring merchants of the specified merchant that have been accessed more than 1000 times can be selected as reference merchants.

[0223] Step 1030: Calculate the user score for each reference user based on the interaction probability of each reference user with each reference merchant in the reference user set; a reference user is a user whose interaction probability with a reference merchant exceeds a probability threshold.

[0224] Step 1040: Based on the user scores of each reference user, select the third number of reference users with the highest user scores from the reference user set as target reference users.

[0225] Step 1050: If an information request is received from the target reference user, return the advertising information of the specified merchant to the target reference user.

[0226] In this embodiment, user scores are calculated based on the user's interaction probability with reference merchants. Target reference users are then identified for designated merchants based on these user scores, and the advertising information for these designated merchants is sent to the target reference users. This provides advertising opportunities to designated merchants with fewer interactions, and correspondingly increases the diversity of advertising information delivered to target users. This enriches the advertising information delivered while ensuring the matching degree between the advertising information delivered to target users and the target users.

[0227] Figure 11 This is a flowchart illustrating a method for determining delivery information according to a specific embodiment. For example... Figure 11 As shown, including:

[0228] Step 1110: Obtain the merchant's access records.

[0229] Step 1120: Construct a friend relationship graph, a neighbor / merchant relationship graph, and an interaction relationship graph.

[0230] Step 1130: Construct an interaction probability prediction model. The interaction probability prediction model may include... Figure 4 The first graph shows the attention network, the second graph shows the attention network, the dual attention network, and the interaction prediction network.

[0231] Step 1140: Train the interaction probability prediction model. For details on the training process, please refer to [link / reference needed]. Figure 9 The process in.

[0232] Step 1150: Determine audience targeting based on the predicted interaction probability.

[0233] In this embodiment, users whose interaction probability with a merchant is higher than a second probability threshold can be identified as targeted users of that merchant. Thus, for users identified as targeted users, the merchant's delivery information is the delivery information to be delivered to those users.

[0234] The solution in this embodiment can be applied to advertising placement scenarios, where the merchant's placement information can be an advertisement. In advertising placement scenarios, audience targeting is generally required first, i.e., pre-determining the target audience for each merchant. In related technologies, staff typically analyze historical advertising data to assess advertising feedback, and then use this feedback to determine audience targeting.

[0235] The method described in this application can be used to automatically target audiences for merchants' advertisements. Specifically, for a merchant (e.g., merchant A), based on the predicted probability of user interaction with the merchant, users whose interaction probability with merchant A is higher than a second probability threshold can be identified as the targeted audience for merchant A. It can be seen that using the method described in this application eliminates the need for manual audience targeting by staff analyzing advertising feedback, achieving automated audience targeting while ensuring accuracy.

[0236] Actual testing revealed that 105 merchants using this solution to generate audience targeting saw an average 30% increase in user click-through rate and visit rate (number of visitors / number of impressions) compared to audience targeting methods using related technologies. This demonstrates that, compared to existing audience targeting solutions, this solution can improve the advertising performance of merchants with existing campaign records.

[0237] When the method of this application is applied to advertising, for merchants with high access records and user visit records, such as merchants with more than 1,000 clicks or visits in the past 3 days, the interaction probability prediction model provided in this application can calculate the interaction probability of the merchant with the advertising plan among all users. Then, for each merchant, a specified number of users with an interaction probability higher than the second probability threshold are selected as the target users of the merchant, and the user IDs of the determined target users are output to a distributed storage path for storage, so as to deliver the merchant's advertisement to the target users.

[0238] For merchants with fewer interaction records (such as visit records or user visit records), such as those with fewer than 1,000 visits or clicks in the past 3 days, if the merchant (e.g., Merchant B) has a campaign plan for the day, then they can follow the guidelines. Figure 10 In the corresponding embodiment, reference merchants are first determined for merchant B. For example, the five most recently visited or clicked merchants closest to merchant B are determined as reference merchants. Then, the interaction probability prediction model is used to predict the interaction probability of each user with each reference merchant. Then, the user score of each user is calculated by combining the interaction probability of each user with all reference merchants. Finally, a specified number of users with the highest scores are selected as the targeted users of merchant B.

[0239] As can be seen, the method described in this application can automatically target audiences not only to merchants with a large number of historical interaction records, but also to merchants with fewer historical interaction records or new merchants. Furthermore, this solution provides a way to target audiences by combining user relationship information and merchant location information.

[0240] In some embodiments, to support the online advertising system in recalling merchant advertisements, the user IDs of users targeted by the merchant can be further written into the online data engine. When a targeted user initiates an information request, such as opening a public account, an interactive page, or a video playback page, the data engine can retrieve the merchant corresponding to the user's user ID, target that user, and send the corresponding merchant's advertisement to that user.

[0241] In the context of ad targeting, the solution proposed in this application can be applied to a Data Management Platform (DMP). The DMP first constructs a friend relationship graph, a neighbor merchant relationship graph, and an interaction relationship graph. It then integrates the feature representations of each node in these graphs to determine the global user characteristics of the target user and the global merchant characteristics of the candidate merchants. Finally, the global user characteristics of the target user and the global merchant characteristics of the candidate merchants are combined with the information of all user and merchant nodes in the three graphs. Based on the global user characteristics of the target user and the global merchant characteristics of the candidate merchants, the interaction probability is predicted and the target audience for each candidate merchant is determined.

[0242] The following describes an apparatus embodiment of this application, which can be used to perform the methods described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments described in the above embodiments of this application.

[0243] Figure 12 This is a block diagram of a device for determining delivery information according to one embodiment, such as... Figure 12As shown, the device for determining the delivery information includes: a first user feature determination module 1210, used by a first graph attention network to output a first user feature of the target user relative to the candidate merchants based on the target user's friend relationship graph and the merchant embedding features of the candidate merchants; a first merchant feature determination module 1220, used by a second graph attention network to output a first merchant feature of the candidate merchants relative to the target user based on the candidate merchants' neighbor merchant relationship graph, the distance between each neighbor merchant and the candidate merchant, and the target user's user embedding features; a second feature determination module 1230, used by a dual attention network to determine the second user feature of the target user and the second merchant feature of the candidate merchants based on an interaction relationship graph, wherein the interaction relationship graph is constructed based on the user's interactive behavior triggered by the merchant; and a global user feature determination module 1230. The module 1240 is used to fuse the user embedding features, first user features, and second user features of the target user to obtain the global user features of the target user; the global merchant feature determination module 1250 is used to fuse the merchant embedding features, first merchant features, and second merchant features of the candidate merchants to obtain the global merchant features of the candidate merchants; the interaction probability determination module 1260 is used by the interaction prediction network to output the interaction probability of the target user relative to the candidate merchants based on the global user features of the target user and the global merchant features of the candidate merchants; the target merchant determination module 1270 is used to filter among multiple candidate merchants based on the interaction probability of the target user relative to the candidate merchants to determine the target merchant, so as to determine the target merchant's delivery information as the delivery information to be delivered to the target user.

[0244] In some embodiments, the first graph attention network includes a first encoding layer, a first attention layer, and a first linear layer; the first user feature determination module 1210 includes: a friend encoding feature determination unit, used to encode the friends in the friend relationship graph of the target user by the first encoding layer to obtain the friend encoding features of each friend; a friend influence weight determination unit, used to perform attention calculation by the first attention layer based on the friend encoding features of each friend, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the friend influence weight of each friend for the target user to trigger interactive behavior with the candidate merchants; and a first linear weighting unit, used to perform linear weighting by the first linear layer based on the friend influence weight of each friend and the friend encoding features of each friend to obtain the first user feature of the target user relative to the candidate merchants.

[0245] In some embodiments, the first attention layer includes a cascaded first stitching layer, a first multilayer perceptron, and a normalization layer; the friend influence weight determination unit includes: a first stitching unit, configured to stitch together the user embedding features of the target user and the merchant embedding features of the candidate merchant with the friend encoding features of each friend, respectively, by the first stitching layer to obtain the stitched features corresponding to each friend; an initial friend influence weight determination unit, configured to process the stitched features corresponding to each friend by the first multilayer perceptron to obtain the initial friend influence weight of each friend; and a first normalization processing unit, configured to normalize the initial friend influence weight of each friend by the normalization layer to obtain the friend influence weight of each friend for triggering interactive behavior of the candidate merchant for the target user.

[0246] In some embodiments, the second graph attention network includes a second encoding layer, a second attention layer, and a second linear layer; the first merchant feature determination module 1220 includes: a neighbor encoding feature determination unit, used to encode each neighbor merchant in the neighbor merchant relationship graph of the candidate merchant by the second encoding layer to obtain the neighbor encoding features of each neighbor merchant; a neighbor influence weight determination unit, used to perform attention calculation by the second attention layer based on the neighbor encoding features of each neighbor merchant, the distance between each neighbor merchant and the candidate merchant, the merchant embedding features of the candidate merchant, and the user embedding features of the target user to obtain the neighbor influence weight of each neighbor merchant for triggering interactive behavior of the candidate merchant by the target user; and a second linear weighting unit, used to perform linear weighting by the second linear layer based on the neighbor influence weights corresponding to each neighbor merchant and the neighbor encoding features of each neighbor merchant to obtain the first merchant feature of the candidate merchant relative to the target user.

[0247] In some embodiments, the dual attention network includes a third encoding layer, a third attention network, and a fourth attention network; the second feature determination module 1230 includes: an encoding unit, configured to encode users and merchants in the interaction relationship graph by the third encoding layer respectively, to obtain merchant encoding features of each merchant and user encoding features of each user in the interaction relationship graph; a second merchant feature determination unit, configured to process the third attention network according to the merchant encoding features of each merchant, the user encoding features of each user, the user embedding features of the target user, and the merchant embedding features of the candidate merchants in the interaction relationship graph, to obtain second merchant features of the candidate merchants; and a second user feature determination unit, configured to process the fourth attention network according to the user encoding features of each user, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants in the interaction relationship graph, to obtain second user features of the target user.

[0248] In some embodiments, the third attention network includes a pooling layer, a third attention layer, and a third linear layer; the second merchant feature determination unit includes: a pooling processing unit, used by the pooling layer to perform pooling processing based on the user coding features of all users in the interaction relationship graph to obtain intermediate user features; a merchant weight determination unit, used by the third attention layer to perform attention calculation based on the intermediate user features, the merchant coding features of each merchant in the interaction relationship graph, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the merchant weight of each merchant in the interaction relationship graph, where the merchant weight refers to the weight of the corresponding merchant in triggering the interaction behavior of the candidate merchant for the target user; and a third linear weighting unit, used by the third linear layer to perform linear weighting based on the merchant weight of each merchant in the interaction relationship graph and the merchant coding features of each merchant to obtain the second merchant features of the candidate merchants.

[0249] In some embodiments, the fourth attention network includes a fourth attention layer and a fourth linear layer; the second user feature determination unit includes: a user weight determination unit, used by the fourth attention layer to perform attention calculation based on the user coding features of each user in the interaction relationship graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain the user weight of each user in the interaction relationship graph; the user weight refers to the weight of the corresponding user in triggering the interaction behavior of the candidate merchant for the target user; and a fourth linear weighting unit, used by the fourth linear layer to perform linear weighting based on the user weight of each user in the interaction relationship graph and the user coding features of each user, to obtain the second user features of the target user.

[0250] In some embodiments, the interaction prediction network includes a second stitching layer, a second multilayer perceptron, and a classification layer; the interaction probability determination module 1260 includes: a second stitching unit, used to stitch together the global user features of the target user and the global merchant features of the candidate merchants by the second stitching layer to obtain target stitched features; an interaction score determination unit, used to transform the target stitched features by the multilayer perceptron to obtain the interaction score of the target user with the candidate merchants; and an interaction probability determination unit, used to normalize the interaction score by the classification layer to obtain the interaction probability of the target user with the candidate merchants.

[0251] In some embodiments, the device for determining the delivery information further includes: a training data acquisition module, configured to acquire training data, the training data including multiple training samples, the training samples indicating sample users, sample merchants, and sample tags, the sample tags being used to indicate whether a sample user triggers an interactive behavior with a sample merchant; a first output module, configured to output a first user feature of a sample user relative to a sample merchant by a first graph attention network based on the sample user's friend relationship graph and the merchant embedding features of the sample merchant; a second output module, configured to output a first merchant feature of a sample merchant relative to a sample user by a second graph attention network based on the sample merchant's neighbor merchant relationship graph, the distance between each neighbor merchant and the sample merchant, and the sample user's user embedding features; and a feature determination module, configured to determine a second user feature of the sample user and a second user feature of the sample user based on a sample interaction relationship graph by a dual attention network. The system comprises the following modules: a second merchant feature for sample merchants; a first fusion module, which fuses the user embedding feature, the first user feature, and the second user feature of sample users to obtain the global user feature of sample users; a second fusion module, which fuses the merchant embedding feature, the first merchant feature, and the second merchant feature of sample merchants to obtain the global merchant feature of sample merchants; a third output module, which outputs the interaction probability of sample users to sample merchants based on the global user feature and the global merchant feature of sample merchants; a loss value calculation module, which calculates the loss value of the loss function based on the interaction probability of sample users to sample merchants and the sample label; and a back-adjustment module, which back-adjusts the parameters of the first graph attention network, the second graph attention network, the dual attention network, and the interaction prediction network based on the loss value until the training termination condition is met.

[0252] In some embodiments, the device for determining delivery information further includes: a distance calculation module for calculating the distance between candidate merchants and other merchants; a filtering module for filtering out the first number of merchants closest to the candidate merchant based on the distance between the candidate merchant and other merchants, as the candidate merchant's neighbor merchants; and a neighbor merchant relationship graph construction module for constructing a neighbor merchant relationship graph of the candidate merchant based on the candidate merchant's neighbor merchants.

[0253] In some embodiments, the device for determining the delivery information further includes: a first interaction record acquisition module, used to acquire the interaction records of each merchant; and an interaction relationship graph construction module, used to construct an interaction relationship graph based on the merchants and users with interaction relationships in the interaction records.

[0254] In some embodiments, the device for determining the delivery information further includes: a second interaction record acquisition module, used to acquire the interaction records of a specified merchant; a reference merchant determination module, used to select a second number of neighboring merchants of the specified merchant as reference merchants if the number of interactions with the specified merchant is less than a set threshold based on the interaction records of the specified merchant; a user score determination module, used to calculate the user score corresponding to each reference user based on the interaction probability of each reference user to each reference merchant in the reference user set; a reference user refers to a user whose interaction probability with a reference merchant exceeds a probability threshold; a target reference user determination module, used to select a third number of reference users with the highest user scores from the reference user set as target reference users based on the user scores corresponding to each reference user; and a sending module, used to return the delivery information of the specified merchant to the target reference user if an information request initiated by the target reference user is received.

[0255] Figure 13 The following is a schematic diagram showing the structure of a computer system suitable for implementing an electronic device according to an embodiment of the present application. Figure 13 The computer system 1300 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0256] like Figure 13 As shown, the computer system 1300 includes a processor, which may be a Central Processing Unit (CPU) 1301. The CPU can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on a program stored in Read-Only Memory (ROM) 1302 or a program loaded from storage portion 1308 into Random Access Memory (RAM) 1303. The RAM 1303 also stores various programs and data required for system operation. The CPU 1301, ROM 1302, and RAM 1303 are interconnected via a bus 1304. An input / output (I / O) interface 1305 is also connected to the bus 1304.

[0257] The following components are connected to I / O interface 1305: an input section 1306 including a keyboard, mouse, etc.; an output section 1307 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1308 including a hard disk, etc.; and a communication section 1309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1309 performs communication processing via a network such as the Internet. A drive 1310 is also connected to I / O interface 1305 as needed. Removable media 1311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1310 as needed so that computer programs read from them can be installed into storage section 1308 as needed.

[0258] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1309, and / or installed from removable medium 1311. When the computer program is executed by central processing unit (CPU) 1301, it performs various functions defined in the system of this application.

[0259] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0260] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0261] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0262] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries computer-readable instructions that, when executed by a processor, implement the methods in any of the above embodiments.

[0263] According to one aspect of this application, an electronic device is also provided, comprising: a processor; and a memory storing computer-readable instructions that, when executed by the processor, implement the methods of any of the above embodiments.

[0264] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the methods of any of the above embodiments.

[0265] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0266] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of this application.

[0267] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0268] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for determining the delivery of information, characterized in that, include: The first graph attention network outputs the first user feature of the target user relative to the candidate merchants based on the target user's friend relationship graph and the merchant embedding features of the candidate merchants; The second graph attention network outputs the first merchant feature of the candidate merchant relative to the target user based on the neighbor merchant relationship graph of the candidate merchant, the distance between each neighbor merchant and the candidate merchant, and the user embedding features of the target user; The dual attention network determines the second user characteristics of the target user and the second merchant characteristics of the candidate merchants based on the interaction graph, wherein the interaction graph is constructed based on the user's interactive behavior triggered by the merchant; The target user's embedded features, first user features, and second user features are fused to obtain the target user's global user features; The merchant embedding features, first merchant features, and second merchant features of the candidate merchants are fused together to obtain the global merchant features of the candidate merchants. The interaction prediction network outputs the interaction probability of the target user relative to the candidate merchants based on the global user characteristics of the target user and the global merchant characteristics of the candidate merchants. Based on the interaction probability of the target user relative to the candidate merchants, multiple candidate merchants are filtered to determine the target merchant, so that the delivery information of the target merchant is determined as the delivery information to be delivered to the target user.

2. The method according to claim 1, characterized in that, The first graph attention network includes a first encoding layer, a first attention layer, and a first linear layer; The first graph attention network outputs a first user feature of the target user relative to the candidate merchants based on the target user's friend relationship graph and the merchant embedding features of the candidate merchants, including: The first encoding layer encodes the friends in the target user's friend relationship graph to obtain the friend encoding features of each friend; The first attention layer performs attention calculations based on the friend encoding features of each friend, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain the friend influence weight of each friend for the target user's interaction behavior with the candidate merchants. The first linear layer performs linear weighting based on the friend influence weights of each friend and the friend coding features of each friend to obtain the first user feature of the target user relative to the candidate merchant.

3. The method according to claim 2, characterized in that, The first attention layer includes a cascaded first splicing layer, a first multilayer perceptron, and a first normalization layer; The first attention layer performs attention calculations based on the friend encoding features of each friend, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the friend influence weight of each friend for the interactive behavior triggered by the target user on the candidate merchant, including: The first splicing layer splices the user embedding feature of the target user and the merchant embedding feature of the candidate merchant with the friend encoding feature of each friend to obtain the spliced ​​feature corresponding to each friend. The first multilayer perceptron processes the spliced ​​features corresponding to each friend to obtain the initial friend influence weight of each friend. The initial friend influence weights of each friend are normalized by the first normalization layer to obtain the friend influence weights of each friend for the target user to trigger interactive behavior with the candidate merchant.

4. The method according to claim 1, characterized in that, The second graph attention network includes a second encoding layer, a second attention layer, and a second linear layer; The second graph attention network outputs a first merchant feature of the candidate merchant relative to the target user based on the neighbor merchant relationship graph of the candidate merchant, the distance between each neighbor merchant and the candidate merchant, and the user embedding features of the target user. This includes: The second encoding layer encodes each neighboring merchant in the neighboring merchant relationship graph of the candidate merchant to obtain the neighbor encoding features of each neighboring merchant; The second attention layer performs attention calculations based on the neighbor coding features of each neighbor merchant, the distance between each neighbor merchant and the candidate merchant, the merchant embedding features of the candidate merchant, and the user embedding features of the target user, to obtain the neighbor influence weight of each neighbor merchant for the target user's interaction behavior with the candidate merchant. The second linear layer performs linear weighting based on the neighbor influence weights and neighbor coding features of each neighbor merchant to obtain the first merchant feature of the candidate merchant relative to the target user.

5. The method according to claim 1, characterized in that, The dual attention network includes a third encoding layer, a third attention network, and a fourth attention network; The step of determining the second user characteristics of the target user and the second merchant characteristics of the candidate merchants by the dual attention network based on the interaction relationship graph includes: The third encoding layer encodes the users and merchants in the interaction relationship graph respectively, to obtain the merchant encoding features of each merchant and the user encoding features of each user in the interaction relationship graph; The third attention network processes the merchant coding features of each merchant in the interaction relationship graph, the user coding features of each user, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the second merchant features of the candidate merchants. The fourth attention network processes the user coding features of each user in the interaction relationship graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the second user features of the target user.

6. The method according to claim 5, characterized in that, The third attention network includes a pooling layer, a third attention layer, and a third linear layer; The third attention network processes the merchant coding features of each merchant in the interaction graph, the user coding features of each user, the user embedding features of the target user, and the merchant embedding features of the candidate merchants to obtain the second merchant features of the candidate merchants, including: The pooling layer performs pooling processing based on the user coding features of all users in the interaction relationship graph to obtain the intermediate user features; The third attention layer performs attention calculations based on the intermediate user features, the merchant coding features of each merchant in the interaction graph, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain the merchant weight of each merchant in the interaction graph. The merchant weight refers to the weight of the corresponding merchant in triggering the interaction behavior of the target user on the candidate merchant. The third linear layer performs linear weighting based on the merchant weights and merchant coding features of each merchant in the interaction relationship graph to obtain the second merchant features of the candidate merchants.

7. The method according to claim 5 or 6, characterized in that, The fourth attention network includes a fourth attention layer and a fourth linear layer; The process by the fourth attention network, based on the user coding features of each user in the interaction graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain the second user features of the target user includes: The fourth attention layer performs attention calculations based on the user coding features of each user in the interaction relationship graph, the second merchant features of the candidate merchants, the user embedding features of the target user, and the merchant embedding features of the candidate merchants, to obtain the user weight of each user in the interaction relationship graph; the user weight refers to the weight of the corresponding user in triggering the interaction behavior of the candidate merchant for the target user; The fourth linear layer performs linear weighting based on the user weights and user coding features of each user in the interaction graph to obtain the second user features of the target user.

8. The method according to claim 1, characterized in that, The interactive prediction network includes a second splicing layer, a second multilayer perceptron, and a classification layer; The interaction prediction network outputs the interaction probability of the target user relative to the candidate merchants based on the global user characteristics of the target user and the global merchant characteristics of the candidate merchants, including: The second splicing layer splices the global user features of the target user and the global merchant features of the candidate merchants to obtain the target spliced ​​features; The multilayer perceptron transforms the target splicing features to obtain the interaction score of the target user for the candidate merchant; The interaction score is normalized by the classification layer to obtain the probability of the target user interacting with the candidate merchant.

9. The method according to claim 1, characterized in that, The method further includes: Acquire training data, which includes multiple training samples. The training samples indicate sample users, sample merchants, and sample tags. The sample tags are used to indicate whether the sample user triggers an interactive behavior with the sample merchant. The first graph attention network outputs the first user feature of the sample user relative to the sample merchant based on the friend relationship graph of the sample user and the merchant embedding feature of the sample merchant; The second graph attention network outputs a first merchant feature of the sample merchant relative to the sample user based on the neighbor merchant relationship graph of the sample merchant, the distance between each neighbor merchant and the sample merchant, and the user embedding features of the sample user. The dual attention network determines the second user characteristics of the sample user and the second merchant characteristics of the sample merchant based on the sample interaction relationship graph; The user embedding features, first user features, and second user features of the sample user are fused to obtain the global user features of the sample user. The merchant embedding features, the first merchant features, and the second merchant features of the sample merchants are fused together to obtain the global merchant features of the sample merchants. The interaction prediction network outputs the interaction probability of the sample user with the sample merchant based on the global user characteristics of the sample user and the global merchant characteristics of the sample merchant; The loss value of the loss function is calculated based on the interaction probability of the sample users with the sample merchants and the sample tags; The parameters of the first graph attention network, the second graph attention network, the dual attention network, and the interaction prediction network are adjusted in reverse based on the loss value until the training termination condition is met.

10. The method according to claim 1, characterized in that, Before the second graph attention network outputs the first merchant feature of the candidate merchant relative to the target user based on the neighbor merchant relationship graph of the candidate merchant, the distance between each neighbor merchant and the candidate merchant, and the user embedding features of the target user, the method further includes: Calculate the distance between the candidate merchant and other merchants; Based on the distance between the candidate merchants and other merchants, the first number of merchants closest to the candidate merchants are selected as the neighbor merchants of the candidate merchants; Construct a neighbor merchant relationship graph for the candidate merchant based on its neighbor merchants.

11. The method according to claim 1, characterized in that, Before the dual attention network determines the second user features of the target user and the second merchant features of the candidate merchants based on the interaction graph, the method further includes: Obtain the interaction records of each merchant; The interaction relationship graph is constructed based on the merchants and users with interactive relationships in the interaction records.

12. The method according to claim 1, characterized in that, The method further includes: Get the interaction records of a specified merchant; If, based on the interaction records of the specified merchant, it is determined that the number of interactions with the specified merchant is less than a set threshold, a second number of neighboring merchants of the specified merchant are selected as reference merchants. Calculate the user score for each reference user based on the interaction probability of each reference user with each of the reference merchants in the reference user set; The reference user refers to a user whose probability of interacting with the reference merchant exceeds a probability threshold. Based on the user scores corresponding to each reference user, select the third number of reference users with the highest user scores from the set of reference users as target reference users; If an information request is received from the target reference user, the advertising information of the designated merchant is returned to the target reference user.

13. A device for determining the delivery of information, characterized in that, include: The first user feature determination module is used to output the first user feature of the target user relative to the candidate merchants by the first graph attention network based on the friend relationship graph of the target user and the merchant embedding features of the candidate merchants; The first merchant feature determination module is used by the second graph attention network to output the first merchant feature of the candidate merchant relative to the target user based on the neighbor merchant relationship graph of the candidate merchant, the distance between each neighbor merchant and the candidate merchant, and the user embedding feature of the target user. The second feature determination module is used to determine the second user features of the target user and the second merchant features of the candidate merchants by the dual attention network based on the interaction relationship graph, wherein the interaction relationship graph is constructed based on the user's interactive behavior triggered by the merchant; The global user feature determination module is used to fuse the user embedding features, the first user feature, and the second user feature of the target user to obtain the global user feature of the target user. The global merchant feature determination module is used to fuse the merchant embedding features, the first merchant feature, and the second merchant feature of the candidate merchant to obtain the global merchant feature of the candidate merchant. The interaction probability determination module is used by the interaction prediction network to output the interaction probability of the target user relative to the candidate merchant based on the global user characteristics of the target user and the global merchant characteristics of the candidate merchant. The target merchant determination module is used to filter among multiple candidate merchants based on the interaction probability of the target user relative to the candidate merchants, and determine the target merchant so as to determine the delivery information of the target merchant as the delivery information to be delivered to the target user.

14. An electronic device, characterized in that, include: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1-12.

15. A computer-readable storage medium storing computer-readable instructions thereon, characterized in that, When the computer-readable instructions are executed by a processor, the method as described in any one of claims 1-12 is implemented.

16. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the method of any one of claims 1-12.

Images