Model training method, resource pushing method, device, medium, and computer program product

Abstract

A model training method, executed by an electronic device, and comprising: acquiring scene interaction graphs respectively corresponding to a plurality of information flow scenes, wherein the scene interaction graphs are used for indicating historical interaction data between a plurality of object nodes and a plurality of resource nodes in the information flow scenes, the object nodes represent objects, and the resource nodes represent interactive resources provided to the objects in the information flow scenes (101); on the basis of the object nodes and the resource nodes, connecting the scene interaction graphs to obtain a global scene interaction graph (102); constructing a global object feature of each object node and a global resource feature of each resource node in the global scene interaction graph (103); and on the basis of the global object features and the global resource features, training initial scene graph networks corresponding to the scene interaction graphs to obtain a scene graph network, the scene graph network being used for determining a push resource for a target object, so as to push the push resource to the target object (104).

Description

Model training methods, resource delivery methods, equipment, media, and computer program products

[0001] Related applications

[0002] This application claims priority to Chinese patent application filed on November 29, 2024, with application number 202411757211.X, entitled "Model Training Method, Resource Recommendation Method and Related Device", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of data processing technology, specifically to a model training method, a resource push method, a device, a medium, and a computer program product. Background Technology

[0004] Personalized push systems leverage users' historical interaction data, employing specific rules or machine learning algorithms to construct user interest preferences and achieve efficient distribution of items. This effectively filters information, improves user experience, and has seen significant development in areas such as information feeds. However, in the information feed field, there are often multiple scenarios requiring efficient information distribution. For example, for some online interactive platforms, their corresponding information feed push scenarios may include multiple scenarios such as subscribing to articles, short videos, live streams, and mini-programs.

[0005] Traditional push algorithms mainly focus on single-scenario modeling, using historical user interaction data within each scenario to build push algorithms for item distribution. However, the development of single-scenario push algorithms has the following problems: (1) Affected by information patterns (such as text / images / videos / live broadcasts, etc.), historical user interaction data in a single scenario has pattern bias / data gaps, making it impossible to re-mind the user's comprehensive information flow preferences, resulting in poor accuracy of the personalized push system; (2) The scarcity of user interaction data in a single scenario creates data bias, making it a huge challenge to build an accurate personalized push system, which seriously affects the user's product experience; (3) Single-scenario modeling requires developers to maintain multiple push systems simultaneously, greatly increasing operation and maintenance costs and R&D costs. Summary of the Invention

[0006] This application provides a model training method, a resource push method, a device, a medium, and a computer program product.

[0007] In a first aspect, embodiments of this application provide a model training method, executed by an electronic device, comprising:

[0008] Obtain scene interaction diagrams corresponding to multiple information flow scenarios. The scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. The object nodes represent objects, and the resource nodes represent interactive resources provided to the objects in the information flow scenario.

[0009] Based on the object nodes and the resource nodes, the scene interaction graph is connected to obtain a global scene interaction graph;

[0010] In the global scene interaction graph, global object features of each object node and global resource features of each resource node are constructed; and

[0011] Based on the global object features and the global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain the scene graph network. The scene graph network is used to determine the push resources for the push object, so as to push the push resources to the push object.

[0012] Accordingly, embodiments of this application provide a model training apparatus, including:

[0013] The acquisition unit is used to acquire scene interaction diagrams corresponding to multiple information flow scenarios. The scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. The object nodes represent objects, and the resource nodes represent interactive resources provided to the objects in the information flow scenario.

[0014] A connection unit is used to connect the scene interaction graph based on the object node and the resource node to obtain a global scene interaction graph;

[0015] The construction unit is configured to construct the global object features of each object node and the global resource features of each resource node in the global scene interaction graph; and

[0016] The training unit is used to train the initial scene graph network corresponding to the scene interaction graph based on the global object features and the global resource features to obtain the scene graph network. The scene graph network is used to determine the push resources for the push object so as to push the push resources to the push object.

[0017] Secondly, embodiments of this application also provide a resource push method, executed by an electronic device, including:

[0018] When a resource push request is received for a target object, the target information stream scenario in which the target object is currently located is detected.

[0019] The target scene graph network corresponding to the target information flow scene is selected from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes.

[0020] Based on the target scene graph network, at least one push resource associated with the push recipient is determined, and the push resource is pushed to the push recipient.

[0021] Accordingly, embodiments of this application provide a resource push device, including:

[0022] The receiving unit is used to detect the target information stream scenario in which the push object is currently located when a resource push request for the push object is received.

[0023] A filtering unit is used to filter out the target scene graph network corresponding to the target information flow scene from the scene graph network. The scene graph network includes global object features and global resource features constructed based on a global scene interaction graph, and a graph network trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes.

[0024] The push unit is used to determine at least one push resource associated with the push object based on the target scene graph network, and push the push resource to the push object.

[0025] Furthermore, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to execute the model training method and resource push method provided in embodiments of this application.

[0026] Furthermore, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in the model training method and resource push method provided in embodiments of this application.

[0027] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in the model training method and resource push method provided in embodiments of this application.

[0028] Details of one or more embodiments of this application are set forth in the following drawings and description. Other features, objects, and advantages of this application will become apparent from the specification, drawings, and claims. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the published drawings without creative effort.

[0030] Figure 1 is a schematic diagram of graph structure modeling based on graph neural networks provided in an embodiment of this application;

[0031] Figure 2 is a schematic diagram of the application scenario of the model training method provided in the embodiments of this application;

[0032] Figure 3 is a flowchart illustrating the model training method provided in an embodiment of this application;

[0033] Figure 4A is a schematic diagram of an information feed scenario with a subscription article type;

[0034] Figure 4B is a schematic diagram of a short video-type information flow scenario;

[0035] Figure 4C is a schematic diagram of a live streaming information flow scenario;

[0036] Figure 4D is a schematic diagram of an information flow scenario for mini-programs;

[0037] Figure 5 is a scene interaction diagram corresponding to an information flow scenario provided in an embodiment of this application;

[0038] Figure 6 is a schematic diagram of the processing method involved in the model training method provided in the embodiments of this application;

[0039] Figure 7 is a schematic diagram of a global scene interaction diagram provided in an embodiment of this application;

[0040] Figure 8 is a schematic diagram of constructing virtual object connection nodes and virtual resource connection nodes in any scene interaction diagram provided by an embodiment of this application;

[0041] Figure 9 is a schematic diagram of the training process starting from the global scene interaction graph provided in an embodiment of this application;

[0042] Figure 10 is a schematic diagram of the generation of the enhanced interaction diagram provided in an embodiment of this application;

[0043] Figure 11 is a flowchart illustrating the resource push method provided in an embodiment of this application;

[0044] Figure 12A is a schematic diagram of point-based modeling involved in traditional mainstream push algorithms;

[0045] Figure 12B is a schematic diagram of sequence modeling involved in traditional mainstream push algorithms;

[0046] Figure 13A is a schematic diagram of the auxiliary paradigms involved in the joint construction of existing multi-scenario push systems;

[0047] Figure 13B is a schematic diagram of the multi-tower paradigm involved in the existing multi-scenario joint construction of push systems;

[0048] Figure 13C is a schematic diagram of the expert paradigms involved in the joint construction of existing multi-scenario push systems;

[0049] Figure 14 is another flowchart illustrating the resource push method provided in an embodiment of this application;

[0050] Figure 15 is a schematic diagram of the model training device provided in an embodiment of this application;

[0051] Figure 16 is a schematic diagram of the resource push device provided in an embodiment of this application;

[0052] Figure 17 is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0054] Push systems are one of the core businesses of modern Internet companies. They push personalized content or products to users by analyzing information such as user behavior and content characteristics. Push systems mainly include the following basic components: (1) Users: users in the system, which can be individual users or enterprise users, etc.; (2) Items: goods, services, content, etc. in the system; (3) User behavior: various operations of users in the system, such as browsing, purchasing, liking, etc.; (4) Content characteristics: various attributes of items, such as the price, brand, category, etc. of goods.

[0055] Graph Neural Networks (GNNs) are an emerging artificial intelligence technology that can learn structural information on graphs in an unsupervised manner and perform tasks such as node classification and graph classification in a supervised manner. In push notification systems, GNN-based graph structure modeling (as shown in Figure 1) can be used to learn graph structures such as object behavior and item features, and can fully explore users' higher-order interests (e.g., user-user, user-item, user-item-user, etc.) to capture more comprehensive user preferences, thereby improving push notification quality. GNN-based graph structure modeling aggregates neighbor information of nodes through a message passing mechanism and updates the node's feature identifiers to learn graph structures such as object behavior and item features. Specifically, for nodes corresponding to object behavior and item features, in each iteration, the node receives information from its neighboring nodes, aggregates this information, and updates its own feature representation. After multiple iterations, each node's feature identifier will contain information from its neighbors and more distant nodes, thus uncovering the user's higher-order interests. GNNs are a class of deep learning models used to process graph structure data. Entities in graph structure data are represented as nodes, and relationships between entities are represented as edges. The main components of a GNN include: (1) a graph: representing relationships, connections, and other information within the system; (2) nodes: representing entities on the graph, such as users and items; (3) edges: representing relationships, connections, and other information between nodes; (4) message passing: representing the process of information passing between nodes on the graph; and (5) aggregation: representing the process of information aggregation on nodes. The goal of a GNN is to learn useful labels from graph-structured data and use these labels for various tasks, such as node classification, graph classification, and prediction. The core idea of ​​a GNN is to aggregate the neighbor information of a node and update the node's feature labels through a message-passing mechanism. This process usually involves multiple iterations to capture information from further distances in the graph. Ultimately, the feature label of each node will contain information about its neighbors and more distant nodes.

[0056] In push notification systems, information such as object behavior and content characteristics can be represented as nodes and edges of a graph. Therefore, graph neural networks can be used to learn this information and generate personalized push notifications. Specifically, graph neural networks can be used to learn information such as the similarity between users, the similarity between items, and the relevance between users and items.

[0057] As mentioned earlier, traditional push algorithms mainly focus on single-scenario modeling, but single-scenario push algorithms have many problems (such as data bias, pattern bias, etc.), making it difficult to accurately mine users' information flow preferences. To address this, this application provides a model training method, a resource push method, a device, a medium, and a computer program product. More specifically, the device, medium, and computer program product are electronic devices, computer-readable storage media, and computer program products. The model training device can be integrated into an electronic device, which can be a server or a terminal, etc.

[0058] The server can be a standalone 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, content delivery network (CDN) acceleration services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these.

[0059] Figure 2 illustrates an application scenario of the model training method provided in this application embodiment. As shown in Figure 2, taking the model training device integrated into an electronic device, with the electronic device serving as a server, as an example, the user terminal (i.e., the client) where the user (including the pushed object) is located can establish a communication connection with the server. Based on this, the server can obtain scene interaction graphs corresponding to multiple information flow scenarios. The scene interaction graphs are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario. Then, based on the object nodes and resource nodes, the scene interaction graphs are connected to obtain a global scene interaction graph. After that, global object features of each object node and global resource features of each resource node are constructed in the global scene interaction graph. Then, based on the global object features and global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain a scene graph network. The scene graph network is used to determine the push resources for the pushed object so as to push the push resources to the pushed object.

[0060] It should be noted that there can be multiple clients in Figure 2, and multiple clients can send data access requests to the server at the same time or at different times.

[0061] It is understood that, in the specific implementation of this application, related data such as (object description information, resource description information, historical interaction data between object nodes and resource nodes) are involved. When the following embodiments of this application are applied to specific products or technologies, 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.

[0062] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0063] This embodiment will be described from the perspective of a model training device, which can be integrated into an electronic device, such as a server.

[0064] A model training method includes: acquiring scene interaction graphs corresponding to multiple information flow scenarios, wherein the scene interaction graphs are used to indicate historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario; object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario; connecting the scene interaction graphs based on object nodes and resource nodes to obtain a global scene interaction graph; constructing global object features for each object node and global resource features for each resource node in the global scene interaction graph; and training an initial scene graph network corresponding to the scene interaction graph based on the global object features and global resource features to obtain a scene graph network, wherein the scene graph network is used to determine push resources for the push recipient, so as to push the push resources to the push recipient.

[0065] Figure 3 shows a flowchart of the model training method provided in this embodiment. As shown in Figure 3, the specific process of the model training method is as follows:

[0066] 101. Obtain the scene interaction diagrams corresponding to each of the multiple information flow scenarios. The scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario.

[0067] In the context of information flow, an information flow scenario is a specific business context defined within the information flow domain based on business functions and resource push objectives. Within this context, the system pushes suitable interactive resources to specific types of objects. Different scenarios generate independent interactive data due to differences in resource types, object characteristics, and interaction rules. An information flow scenario can be understood as any scenario within the information flow domain involving push tasks. For example, for a web interaction platform, the information flow scenarios it encompasses may include subscribing to articles (as shown in Figure 4A), short videos (as shown in Figure 4B), live streaming (as shown in Figure 4C), and mini-programs (as shown in Figure 4D). In the subscribing to articles scenario, users can be pushed articles they are interested in; in the short video scenario, they can be pushed short videos they are interested in; in the live streaming scenario, they can be pushed products they are interested in; in the mini-program scenario, they can be pushed mini-programs they are interested in, and so on. The scenario interaction diagram corresponding to any information flow scenario can include two different types of nodes: object nodes and resource nodes. Object nodes can be understood as users entering the information flow scenario through application software. Resource nodes can be understood as interactive resources provided to users in the information flow scenario. For example, in an article subscription scenario, resource nodes can be various articles available for users to browse; in a short video scenario, resource nodes can be video resources available for users to watch; in a live streaming scenario, resource nodes can be live video resources available for users to watch; and in a mini-program scenario, resource nodes can be various types of mini-programs available for users to use.

[0068] The scene interaction graph is a bipartite graph formed by modeling the interaction relationships between objects and resources in an information flow scenario using a graph neural network (GNN). In the graph, object nodes represent objects participating in the interaction, resource nodes represent resources available for interaction, and edges between nodes represent historical interaction data between objects and resources, reflecting the object-resource interaction structure in a specific scenario. As shown in Figure 1, the scene interaction graph can be understood as a graph model obtained by modeling users and items based on a graph structure using GNN in a push system. Since the functions of different information flow scenarios differ, the corresponding in-scene interaction information can be independent of each other. Taking the article subscription scenario and the short video scenario as examples, the article subscription scenario mainly pushes article resources, while the short video scenario mainly pushes video resources; the types of resources pushed are different. Furthermore, the users (objects) involved in the article subscription scenario may be different from those involved in the short video scenario; that is, users who use the short video function may not also use the article subscription function, and vice versa. Moreover, due to the different types of resources pushed and the potential differences in users, the interaction information generated in different information flow scenarios also differs. Therefore, different information flow scenarios can generate different interaction logs, resulting in different scenario interaction diagrams.

[0069] In any scene interaction diagram, object nodes representing users and resource nodes representing items can be distinguished by different shapes, and connecting lines can be used to represent historical interaction data between object nodes and resource nodes. Historical interaction data can be understood as the actual interaction relationships that have occurred between object nodes and resource nodes in the information flow scene. For example, in Figure 1, "circles" are used to represent object nodes corresponding to users, and "squares" are used to represent resource nodes corresponding to items. The connecting lines between "circles" and "squares" represent the interaction relationships between object nodes and resource nodes.

[0070] The process of obtaining scene interaction graphs corresponding to multiple information flow scenarios may include: obtaining object nodes, resource nodes, and historical interaction data between object nodes and resource nodes corresponding to multiple information flow scenarios; and constructing a bipartite graph based on object nodes, resource nodes, and historical interaction data to obtain scene interaction graphs.

[0071] In this embodiment, user information, resource information, and historical interaction data between users and resources corresponding to each information flow scenario can be obtained first; user information can be represented as object nodes and resource information can be represented as resource nodes; then, the connection relationship between object nodes and resource nodes can be established based on the historical interaction data, so that the resulting bipartite graph can be used as a scenario interaction graph.

[0072] A bipartite graph, also known as a bipartite graph, is a special model in graph theory. It is defined as a graph where vertices can be divided into two mutually exclusive sets, such that every edge connects vertices in both sets, without connecting vertices within the same set. For example, Figure 5 shows a scene interaction graph corresponding to an information flow scenario provided in an embodiment of this application. As shown in Figure 5, u represents an object node in the information flow scenario, and v represents a resource node. To easily distinguish multiple object nodes in the information flow scenario, Arabic numerals are added after the object node u for differentiation, such as u1, u2, u3, u4, etc., representing different object nodes; similarly, Arabic numerals are added after the resource node v to distinguish multiple resource nodes in the information flow scenario, such as v1, v2, v3, etc., representing different resource nodes.

[0073] The vertices in Figure 5 can be divided into two disjoint sets: an object node set and a resource node set. The object node set includes four object nodes: u1, u2, u3, and u4. The resource node set includes three resource nodes: v1, v2, and v3. If there is historical interaction data between an object node and a resource node, they can be connected by edges. For example, if u1 and v1 have historical interaction data, they can be connected; if u2 and v1 have historical interaction data, they can be connected, and so on.

[0074] After completing the connection operation based on the historical interaction data between object nodes and resource nodes, a bipartite graph corresponding to the information flow scenario can be obtained, which is also the scenario interaction graph. In the bipartite graph, the object nodes in the object node set are not connected to each other, and the resource nodes in the resource node set are also not connected to each other.

[0075] Because the scene interaction graphs involved in the embodiments of this application contain object nodes and resource nodes, the scene interaction graphs are heterogeneous graphs. Each scene interaction graph can be represented as G. i (u,v,e), where u represents an object node, v represents a resource node, and e represents historical interaction data between object nodes and resource nodes. Therefore, the scene interaction graph collected from multiple scenarios can be represented as a set:

[0076] G = {G1(u,v,e),G2(u,v,e)G} i (u,v,e)}, where i is a natural number greater than 2.

[0077] 102. Based on object nodes and resource nodes, connect the scene interaction graph to obtain the global scene interaction graph.

[0078] The specific steps are as follows: First, obtain the object description information corresponding to the object nodes and the resource description information corresponding to the resource nodes in the scene interaction graph. Object description information can include characteristics such as object attributes and behaviors, while resource description information can include characteristics such as resource type and content. Then, based on the object description information, construct at least one virtual object connection node in the scene interaction graph. The method for constructing virtual object connection nodes can be to cluster object nodes with similar object description information, using the cluster centers as virtual object connection nodes. Virtual object connection nodes are used to connect object nodes that have a relationship with them; the relationship can be determined by the similarity of the object description information. Similarly, based on the resource description information, construct at least one virtual resource connection node in the scene interaction graph, using a similar method to virtual object connection nodes. Finally, connect the scene interaction graph based on the virtual object connection nodes and virtual resource connection nodes. Specifically, connect virtual object connection nodes with the same virtual object description information and connect virtual resource connection nodes with the same virtual resource description information to obtain the global scene interaction graph.

[0079] Virtual object connection nodes are virtual nodes constructed based on the object description information of object nodes in the scene interaction graph. They are used to connect object nodes with similar object description information, establishing association channels between object nodes in different scenes during the multi-scene interaction graph connection process. Virtual resource connection nodes are virtual nodes constructed based on the resource description information of resource nodes in the scene interaction graph. They are used to connect resource nodes with similar resource description information, establishing association channels between resource nodes in different scenes during the multi-scene interaction graph connection process.

[0080] Figure 6 illustrates a schematic diagram of the processing method involved in the model training method provided in this application embodiment. As shown in Figure 6, after obtaining the scene interaction graphs corresponding to multiple information flow scenarios, the multiple scene interaction graphs can be integrated into a global scene interaction graph. Since this application embodiment involves multiple information flow scenarios, each information flow scenario corresponds to a scene interaction graph. As mentioned earlier, for a single information flow scenario, if a push algorithm is built within the scope of a single information flow scenario for item distribution, problems such as pattern bias and data bias will exist, affecting the accuracy of the push system. To solve the above problems, this application embodiment provides a multi-scenario joint modeling algorithm, which integrates multiple information flow scenarios, can fully utilize the commonalities and differences between multiple information flow scenarios, and then build a push algorithm based on the integrated global information, thereby enabling cross-scenario mining of user information flow preferences and eliminating pattern bias and data bias in single-scenario modeling.

[0081] This process involves connecting the scene interaction graphs corresponding to multiple information flow scenarios, based on the object nodes and resource nodes in each scene interaction graph, to obtain a global scene interaction graph. The global scene interaction graph is a comprehensive graph formed by constructing virtual connection nodes based on object and resource description information from the scene interaction graphs corresponding to multiple information flow scenarios, connecting virtual nodes with the same description information. It integrates the interaction information of multiple scenarios to uncover the relationships between different scenarios and eliminate data bias in single-scenario modeling.

[0082] A global scene interaction graph can be understood as a comprehensive global interaction graph encompassing various scene interaction graphs. It involves connecting scene interaction graphs based on object nodes and resource nodes to obtain the global scene interaction graph. This process includes: obtaining object description information for object nodes and resource description information for resource nodes in the scene interaction graphs; constructing at least one virtual object connection node in the scene interaction graph based on the object description information, whereby the virtual object connection node connects to object nodes that have a relationship with the virtual object connection node; constructing at least one virtual resource connection node in the scene interaction graph based on the resource description information, whereby the virtual resource connection node connects to resource nodes that have a relationship with the virtual object connection node; and connecting the scene interaction graphs based on the virtual object connection nodes and virtual resource connection nodes to obtain the global scene interaction graph.

[0083] In the process of connecting scene interaction graphs to obtain a global scene interaction graph in this embodiment, virtual object connection nodes and virtual resource connection nodes can be constructed in each scene interaction graph first. The virtual object connection nodes can be constructed based on the object description information corresponding to the object nodes in the scene interaction graph, and the virtual resource connection nodes can be constructed based on the resource description information corresponding to the resource nodes in the scene interaction graph. Then, using the virtual object connection nodes and virtual resource connection nodes as the connection basis, multiple scene interaction graphs are connected to form a global scene interaction graph.

[0084] Object description information can be understood as the node attribute characteristics of an object node. For example, object description information may reflect the basic profile information of the user corresponding to the object node, such as age, gender, education level, etc. Resource description information can be understood as the node attribute information of a resource node. For example, resource description information may reflect the category information, style information, etc., of the resource corresponding to the resource node.

[0085] Figure 7 illustrates a schematic diagram of a global scene interaction diagram provided in an embodiment of this application. As shown in Figure 7, the global scene interaction diagram is obtained by connecting scene interaction diagram A and scene interaction diagram B. To distinguish the scene interaction diagrams corresponding to different information flow scenarios, and the object nodes and resource nodes in different scene interaction diagrams, uppercase letters (such as A, B, C, D, etc.) can be added before the object nodes and resource nodes for differentiation. For example, A can be used to represent information flow scenario A, B can be used to represent information flow scenario B, and so on. For information flow scenario A, A-u1 can represent an object node in scene interaction diagram A corresponding to information flow scenario A, A-v1 can represent a resource node in scene interaction diagram A corresponding to information flow scenario A, and so on. For information flow scenario B, B-u1 can represent an object node in scene interaction diagram B corresponding to information flow scenario B, B-v1 can represent a resource node in scene interaction diagram B corresponding to information flow scenario B, and so on.

[0086] It should be noted that A-u1 and B-u1 belong to different scene interaction graphs. The object description information corresponding to the object nodes they represent may be the same or different. That is, A-u1 and B-u1 may be the same user or different users. Similarly, A-v1 and B-v1 belong to different scene interaction graphs. The resource description information corresponding to the resource nodes they represent may be the same or different, or they may be partially the same and partially different.

[0087] To easily distinguish between virtual object connection nodes and virtual resource connection nodes, a lowercase letter 'm' can be used to represent a virtual object connection node, and a lowercase letter 'n' can be used to represent a virtual resource connection node. For any scene interaction graph that has constructed virtual object connection nodes and / or virtual resource connection nodes, it can be represented as G. i (u,v,e,m,n), where u represents an object node, v represents a resource node, e represents historical interaction data between object nodes and resource nodes, m represents a virtual object connection node, and n represents a virtual resource connection node. Therefore, the scene interaction graph collected from multiple scenes can also be represented as a set:

[0088] G=[G1(u,v,e,p,q),G2(u,v,e,p,q)G i [u,v,e,p,q)], where i is a natural number greater than 2. In this case, G is the global scene interaction graph.

[0089] Building upon this, Arabic numerals can be added after the lowercase letter 'm' to represent different virtual object connection nodes, and after the lowercase letter 'n' to represent different virtual resource connection nodes. For example, different virtual object connection nodes can be represented as m1, m2, m3, m4, etc., and different virtual resource connection nodes can be represented as n1, n2, n3, n4, etc. For virtual object connection nodes or virtual resource connection nodes constructed in different scene interaction graphs, uppercase letters (such as A, B, C, D, etc.) can be added before the virtual object connection node or virtual resource connection node to distinguish them. For example, A-m1 can represent a virtual object connection node constructed in scene interaction graph A, A-n1 can represent a virtual resource connection node constructed in scene interaction graph A, and so on. Similarly, B-m1 can represent a virtual object connection node constructed in scene interaction graph B, B-n1 can represent a virtual resource connection node constructed in scene interaction graph B, and so on.

[0090] As shown in Figure 7, a virtual object connection node A-m1 and a virtual resource connection node A-n1 are constructed in scene interaction graph A. Virtual object connection node A-m1 is constructed based on object nodes A-u1 and A-u2, and virtual resource connection node A-n1 is constructed based on resource nodes A-v1 and A-v3. Similarly, a virtual object connection node B-m1 and a virtual resource connection node B-n1 are constructed in scene interaction graph B. Virtual object connection node B-m1 is constructed based on object nodes B-u1 and B-u2, and virtual resource connection node B-v1 and B-v2. After determining the corresponding virtual object connection nodes and virtual resource connection nodes in scene interaction graphs A and B, the scene interaction graphs can be connected based on these virtual object connection nodes to obtain a global scene interaction graph. During the connection process, the rules for connecting virtual object connection nodes and virtual resource connection nodes in different scene interaction graphs must be followed to complete the connection between different scene interaction graphs. For example, in Figure 7, virtual object connection node A-m1 is connected to virtual object connection node B-m1, and virtual resource connection node A-n1 is connected to virtual resource connection node B-n1, thereby achieving the effect of connecting scene interaction graph A and scene interaction graph B, resulting in a global scene interaction graph.

[0091] It should be noted that in Figure 7, connecting only virtual object connection node A-m1 and virtual object connection node B-m1 can also achieve the purpose of connecting scene interaction graph A and scene interaction graph B, thus obtaining a global scene interaction graph. Similarly, connecting only virtual resource connection node A-n1 and virtual resource connection node B-n1 can also achieve the purpose of connecting scene interaction graph A and scene interaction graph B, thus obtaining another global scene interaction graph.

[0092] It should be understood that the global scene interaction diagram shown in Figure 7 is merely an example for illustrating the technical solution of this application. The number of scene interaction diagrams, the specific correspondence and number of object nodes, resource nodes, virtual object connection nodes, and virtual resource connection nodes involved in each scene interaction diagram, etc., do not limit the actual scene interaction diagram, nor do they limit the global scene interaction diagram obtained based on the actual scene interaction diagram.

[0093] The following will explain in detail each step of constructing the global scene interaction graph. The object description information can include object description labels of at least one dimension, as well as the label values ​​of object nodes under those labels. Object description labels can include label types describing object nodes from different dimensions, such as age, gender, and education level. The label values ​​under object description labels can be understood as the specific attribute values ​​corresponding to each object node under that label. For example, if the object description label is age, the label value of an object node under that label can be 20 years old, 30 years old, 40 years old, or other values. If the object description label is gender, the label value of an object node under that label can be male or female. If the object description label is education level, the label value of an object node under that label can be primary school, middle school, high school, university, etc.

[0094] Based on this, according to the object description information, at least one virtual object connection node is constructed in the scene interaction graph, which may include: selecting at least one pushed object description tag from the object description tags, and adding an object tag value range to the pushed object description tag; based on the pushed object description tag and the object tag value range, at least one virtual object connection node is constructed in the scene interaction graph, and the virtual object description information corresponding to the virtual object connection node includes the pushed object description tag and the object tag value range.

[0095] Each scene interaction graph includes multiple object nodes, each with its own corresponding object description information. In this embodiment, at least one virtual object connection node can be constructed based on the object description tags included in the object description information and the tag values ​​of the object nodes under the object description tags. Specifically, at least one pushed object description tag can be selected from the object description tags, and the tag values ​​involved in the pushed object description tag can be obtained; an object tag value range can be added to the pushed object description tag according to the tag values; and then, a virtual object connection node can be constructed in the scene interaction graph based on the pushed object description tag and the object tag value range.

[0096] Figure 8 illustrates a schematic diagram of constructing virtual object connection nodes and virtual resource connection nodes in any scene interaction graph according to an embodiment of this application. As shown in Figure 8, the scene interaction graph includes 5 object nodes (represented as u1, u2, u3, u4, and u5, respectively) and 3 resource nodes (represented as v1, v2, and v3, respectively). For ease of explanation, the object description information of each object node in the scene interaction graph is listed in Table 1.

[0097] Table 1. Object description information for each object node in the scene interaction diagram of Figure 8.

[0098] Based on the data listed in Table 1, at least one of the three object description tags can be extracted as the target object description tag. For example, "Age," "Gender," and "Education Level" can all be used as target object description tags, resulting in three target object description tags. For each target object description tag, it is also necessary to determine the corresponding value range. For example, for the target object description tag "Age," its value range can be set to 22-26 years old; for the target object description tag "Gender," its value range can be set to female; and for the target object description tag "Education Level," its value range can be set to university.

[0099] Based on this, three virtual object connection nodes, m1, m2, and m3, can be constructed in the scene interaction diagram shown in Figure 8. For ease of distinction, m1 represents the virtual object connection node corresponding to the "age" description tag of the pushed object, m2 represents the virtual object connection node corresponding to the "gender" description tag of the pushed object, and m3 represents the virtual object connection node corresponding to the "education level" description tag of the pushed object. Since the object tag value range of m1 is 22-26 years old, the object node satisfying this condition is u2. m1 and u2 can be connected by a dashed line to indicate a relationship between them. Since the object tag value range of m2 is female, the object nodes satisfying this condition are u1, u3, and u4. Therefore, m2 can be connected to u1, u3, and u4 respectively by dashed lines to indicate a relationship between them. Similarly, since the object label value range of m3 is "university", and the object label value range that meets this condition is u3 and u5, m3 can be connected to u3 and u5 respectively with dashed lines to indicate that m2 is associated with u3 and u5 respectively.

[0100] It should be understood that the information on object nodes and virtual object connection nodes shown in Figure 8 and Table 1 is merely an example for the purpose of illustrating the technical solution of this application, and does not limit the specific information of actual object nodes and virtual object connection nodes.

[0101] The resource description information can include at least one dimension of resource description tags, as well as the tag values ​​of resource nodes under those tags. Resource description tags can include different categories such as technology, retro, humor, and lifestyle, describing resource nodes from various perspectives. The tag values ​​under resource description tags can be understood as the specific attribute values ​​corresponding to each resource node under that tag. For example, if the resource description tag is technology, the tag values ​​of resource nodes under that tag could be food technology, gene technology, computer technology, etc. If the resource description tag is lifestyle, the tag values ​​of resource nodes under that tag could be beverages, clothing, home appliances, flowers, etc.

[0102] Based on this, according to the resource description information, at least one virtual resource connection node is constructed in the scene interaction graph. This may include: selecting at least one push resource description tag from the resource description tags and adding a resource tag value range to the push resource description tag; based on the push resource description tag and the resource tag value range, at least one virtual resource connection node is constructed in the scene interaction graph, and the virtual resource description information corresponding to the virtual resource connection node includes the push resource description tag and the resource tag value range.

[0103] Each scene interaction graph includes multiple resource nodes, each with its own corresponding resource description information. In this embodiment, at least one virtual resource connection node can be constructed based on the resource description tags included in the resource description information and the tag values ​​of the resource nodes under the resource description tags. Specifically, at least one push resource description tag can be selected from the resource description tags, and the tag values ​​involved in the push resource description tag can be obtained; a resource tag value range can be added to the push resource description tag according to the tag values; and then, a virtual resource connection node can be constructed in the scene interaction graph based on the push resource description tag and the resource tag value range.

[0104] The scene interaction diagram shown in Figure 8 includes three resource nodes, namely v1, v2, and v3. For ease of explanation, the resource description information of each resource node in this scene interaction diagram is listed in Table 2.

[0105] Table 2. Resource description information for each resource node in the scene interaction diagram of Figure 8.

[0106] Based on the data listed in Table 2, at least one of the three resource description tags can be extracted as the push resource description tag. For example, "Technology" and "Lifestyle" can be used as push resource description tags, resulting in two push resource description tags. For each push resource description tag, it is also necessary to determine the range of resource tag values ​​corresponding to that tag. For example, for the push resource description tag "Technology", its resource tag value range can be set to computer technology; for the push resource description tag "Lifestyle", its resource tag value range can be set to "clothing".

[0107] Based on this, two virtual resource connection nodes, n1 and n2, can be constructed in the scene interaction diagram shown in Figure 8. For ease of distinction, n1 represents the virtual resource connection node corresponding to the push resource description tag "Technology," and n2 represents the virtual resource connection node corresponding to the push resource description tag "Lifestyle." Since the resource tag value range of n1 is "Computer Technology," the resource nodes satisfying this condition are v1, v2, and v3. n1 can be connected to v1, v2, and v3 respectively using dashed lines to indicate that n1 has an association with v1, v2, and v3. Similarly, since the resource tag value range of n2 is "Clothing," the resource nodes satisfying this condition are v2 and v3. Therefore, n2 can be connected to v2 and v3 respectively using dashed lines to indicate that n2 has an association with v2 and v3.

[0108] It should be understood that the information on resource nodes and virtual resource connection nodes shown in Figure 8 and Table 2 is merely an example for the purpose of illustrating the technical solution of this application, and does not limit the specific information of actual resource nodes and virtual resource connection nodes.

[0109] Specifically, based on virtual object connection nodes and virtual resource connection nodes, the scene interaction graph is connected to obtain a global scene interaction graph. This can include: obtaining the virtual object description information corresponding to the virtual object connection nodes and the virtual resource description information corresponding to the virtual resource connection nodes in each scene interaction graph; connecting virtual object connection nodes with the same virtual object description information and connecting virtual resource connection nodes with the same virtual resource description information to obtain a global scene interaction graph.

[0110] After determining the virtual object connection nodes and virtual resource connection nodes constructed in each scene interaction graph, multiple scene interaction graphs can be connected based on these virtual object connection nodes to obtain a global scene interaction graph. In this process, the virtual object description information corresponding to the virtual object connection nodes constructed in each scene interaction graph can be compared, as can the virtual resource description information corresponding to the virtual resource connection nodes constructed in each scene interaction graph. Based on the comparison results, virtual object connection nodes with the same virtual object description information are connected, and virtual resource connection nodes with the same virtual resource description information are connected, thereby achieving the goal of connecting the various scene interaction graphs to obtain the global scene interaction graph.

[0111] Taking Figure 7 as an example, the virtual object description information corresponding to virtual object connection node A-m1 in scene interaction graph A is the same as the virtual object description information corresponding to virtual object connection node B-m1 in scene interaction graph B. Therefore, virtual object connection node A-m1 and virtual object connection node B-m1 can be connected. Similarly, the virtual resource description information corresponding to virtual resource connection node A-n1 in scene interaction graph A is the same as the virtual resource description information corresponding to virtual resource connection node B-n1 in scene interaction graph B. Therefore, virtual resource connection node A-n1 and virtual resource connection node B-n1 can be connected. This results in a global scene interaction graph that covers scene interaction graphs A and B.

[0112] It should be noted that virtual object connection nodes are virtual nodes constructed based on actual existing object nodes, and virtual resource connection nodes are virtual nodes constructed based on actual existing resource nodes; that is, both virtual object connection nodes and virtual resource connection nodes are virtual connection nodes. In this embodiment, to distinguish them from the connection relationships between actual existing object nodes and resource nodes, dashed lines are used to represent the connection relationships between virtual object connection nodes and virtual resource connection nodes when connecting different scene interaction graphs using virtual object connection nodes and / or virtual resource connection nodes.

[0113] Furthermore, in the global scene interaction diagram shown in Figure 7, scene interaction diagram A and scene interaction diagram B can be connected via virtual object connection nodes A-m1 and B-m1. Virtual object connection node A-m1 is associated with object nodes A-u1 and A-u2 in scene interaction diagram A, and virtual object connection node B-m1 is associated with object nodes B-u1 and B-u2 in scene interaction diagram B. For object node A-u1, it can establish connections with B-u1 and B-u2 respectively through virtual object connection nodes A-m1 and B-m1. Since resource nodes that interact with B-u1 include B-v1 and B-v2, A-u1 can indirectly establish connections with resource nodes B-v1 and B-v2. Similarly, since resource nodes that interact with B-u2 include B-v2, A-u1 can also indirectly establish connections with resource node B-v2. The relationships between other object nodes in other scene interaction graphs can also be determined using the above method, and will not be elaborated here.

[0114] In the global scene interaction diagram shown in Figure 7, scene interaction diagram A and scene interaction diagram B can be connected through virtual resource connection nodes A-n1 and B-n1. Virtual resource connection node A-n1 is associated with resource nodes A-v1 and A-v3 in scene interaction diagram A, and virtual resource connection node B-n1 is associated with B-v1 and B-v2 in scene interaction diagram B. For resource node B-v1, it can establish connections with A-v1 and A-v3 respectively through virtual resource connection nodes A-n1 and B-n1. Since the object node interacting with A-v1 is A-u3, B-v1 can indirectly be associated with object node A-u3. Similarly, since the object node interacting with A-v3 is A-u2, B-v1 can also indirectly be associated with object node A-u2. The relationships of other resource nodes in other scene interaction diagrams can also be determined using the same method, and will not be elaborated upon here.

[0115] In this embodiment of the application, after obtaining the object description information corresponding to the object node and the resource description information corresponding to the resource node in the scene interaction graph, the method may further include: obtaining the number of object nodes corresponding to the object nodes and the number of resource nodes corresponding to the resource nodes in the scene interaction graph; determining the number of virtual object nodes to be constructed corresponding to the virtual object connection nodes in the scene interaction graph based on the number of object nodes, and determining the number of virtual resource nodes to be constructed corresponding to the virtual resource connection nodes in the scene interaction graph based on the number of resource nodes; constructing virtual object connection nodes in the scene interaction graph based on the number of virtual object nodes and the object description information, and constructing virtual resource connection nodes in the scene interaction graph based on the number of virtual resource nodes and the resource description information.

[0116] The number of object nodes and resource nodes contained in each scene interaction graph varies. Some scene interaction graphs contain more object nodes and / or resource nodes, and correspondingly, the number of object nodes and / or resource nodes in the scene interaction graph is larger. Other scene interaction graphs contain fewer object nodes and / or resource nodes, and correspondingly, the number of object nodes and / or resource nodes in the scene interaction graph is smaller.

[0117] In this embodiment, the number of virtual object connection nodes to be constructed in each scene interaction graph can be determined based on the number of object nodes in that graph, and the number of virtual resource connection nodes to be constructed in each scene interaction graph can be determined based on the number of resource nodes in that graph. For ease of description, the number of virtual object connection nodes to be constructed is referred to as the virtual object node construction quantity, and the number of virtual resource connection nodes to be constructed is referred to as the virtual resource node construction quantity. For example, for a scene interaction graph with a small number of object nodes, more virtual object connection nodes can be constructed, i.e., the virtual object node construction quantity in that scene interaction graph can be appropriately increased. In this way, the virtual object connection nodes can be refined, thereby increasing the connection probability between object nodes in that scene interaction graph and object nodes in other scene interaction graphs. Similarly, for a scene interaction graph with a small number of resource nodes, more virtual resource connection nodes can also be constructed, i.e., the virtual resource node construction quantity in that scene interaction graph can be appropriately increased. In this way, the virtual resource connection nodes can be refined, thereby increasing the connection probability between resource nodes in that scene interaction graph and resource nodes in other scene interaction graphs.

[0118] After determining the number of virtual object nodes and virtual resource nodes to be constructed in a scene interaction graph, the same number of virtual object connection nodes can be constructed based on the object description information in the scene interaction graph, and the same number of virtual resource connection nodes can be constructed based on the resource description information in the scene interaction graph. For example, if a scene interaction graph has 10 virtual object nodes and 6 virtual resource nodes, then 10 virtual object connection nodes and 6 virtual resource connection nodes need to be constructed in the scene interaction graph. The specific methods for constructing virtual object connection nodes and / or virtual resource connection nodes can refer to the methods described above, and will not be elaborated here.

[0119] It should be noted that, as mentioned earlier, the virtual object connection nodes and virtual resource connection nodes in the global scene interaction graph are all virtual connection nodes. That is to say, the virtual object connection nodes and virtual resource connection nodes serve to connect different scene interaction graphs, constructing information pathways between them. In the global scene interaction graph, neither the virtual object connection nodes nor the virtual resource connection nodes participate in actual graph neural network computation, and therefore, there is no feature update involved in updating the virtual object connection nodes and virtual resource connection nodes.

[0120] 103. Construct global object characteristics for each object node and global resource characteristics for each resource node in the global scene interaction graph.

[0121] For ease of distinction, object nodes in the global scene interaction graph can be called global object nodes, and resource nodes in the global scene interaction graph can be called global resource nodes.

[0122] Global object features are the feature vector representations of global object nodes in the global scene interaction graph within the graph neural network. They are obtained through steps such as data augmentation, sample pair determination, and graph network parameter updates to the global scene interaction graph. This integrates multi-scene interaction information and reflects the potential interaction characteristics of objects with resources in the global scene. Global object features can be understood as the feature vectors corresponding to global object nodes in the global scene interaction graph, that is, the embedding representation of global object nodes in the graph neural network corresponding to the global scene interaction graph. Global resource features are the feature vector representations of global resource nodes in the global scene interaction graph within the graph neural network. They are obtained through steps such as data augmentation, sample pair determination, and graph network parameter updates to the global scene interaction graph. This integrates multi-scene interaction information and reflects the potential interaction characteristics of resources with objects in the global scene. Global resource features can be understood as the feature vectors corresponding to global resource nodes in the global scene interaction graph, that is, the embedding representation of global resource nodes in the graph neural network corresponding to the global scene interaction graph.

[0123] In this embodiment of the application, constructing global object features for each object node and global resource features for each resource node in the global scene interaction graph may include: performing data augmentation on the global scene interaction graph to obtain a first augmented interaction graph and a second augmented interaction graph; determining positive sample pairs and negative sample pairs based on the first augmented interaction graph and the second augmented interaction graph, wherein positive sample pairs include node pairs composed of the same nodes in the first augmented interaction graph and the second augmented interaction graph, and negative sample pairs include node pairs composed of different nodes in the first augmented interaction graph and the second augmented interaction graph; and determining global object features for each object node and global resource features for each resource node based on the positive sample pairs and the negative sample pairs.

[0124] The first augmented interaction graph is obtained by performing data augmentation operations on the global scene interaction graph. It is generated by selecting a subset of object nodes and resource nodes and their interaction data from the global scene interaction graph, and is used to construct positive and negative sample pairs during comparative learning and training. The second augmented interaction graph is also obtained by performing data augmentation operations on the global scene interaction graph. It is generated by selecting a subset of object nodes and resource nodes and their interaction data from the global scene interaction graph, and may overlap with some nodes in the first augmented interaction graph. It is also used to construct positive and negative sample pairs during comparative learning and training.

[0125] Positive sample pairs are node pairs consisting of identical nodes in the first and second augmented interaction graphs. During contrastive learning training, they are used to help the graph network learn similar feature representations of the same nodes in different augmented graphs. Negative sample pairs are node pairs consisting of different nodes in the first and second augmented interaction graphs. During contrastive learning training, they are used to help the graph network learn feature differences between different nodes.

[0126] The specific methods for data augmentation are as follows: Node deletion and node addition can be used. For node deletion, object nodes and resource nodes in the global scene interaction graph are randomly deleted with a certain probability, generating a first part of the graph after deletion and a second part of the graph after deletion, respectively. For node addition, new object nodes and resource nodes are randomly added with a certain probability based on the deleted graphs, and the interaction relationships between the new nodes and existing nodes are established according to the historical interaction data in the global scene interaction graph. The graph after the first part of node deletion and addition is used as the first augmented interaction graph, and the graph after the second part of node deletion and addition is used as the second augmented interaction graph.

[0127] Specifically, we first define a loss function L, which considers the feature differences between positive and negative sample pairs. For positive sample pairs, we aim for the feature representations of their two nodes to be as similar as possible; for negative sample pairs, we aim for the feature representations of their two nodes to be as different as possible. The loss function L can be expressed as follows: Where P+ Let P represent the set of positive sample pairs. - Let represent the set of negative sample pairs, d represent the distance metric function between features (such as cosine distance), and α and β be the weight coefficients. Then, gradient descent is used to update the parameters of the initial first graph network and the initial second graph network to minimize the loss function L. After multiple iterations, the trained first graph network and second graph network are obtained. Based on these two graph networks, the global object features of each object node and the global resource features of each resource node are determined.

[0128] Figure 9 illustrates a flowchart of the training process starting from the global scene interaction graph provided in this embodiment of the application. As shown in Figure 9, after the global scene interaction graph is constructed, the graph neural network corresponding to the global scene interaction graph can be trained to obtain the global object features of each object node and the global resource features of each resource node in the global scene interaction graph.

[0129] In this embodiment, a contrastive learning paradigm can be used for self-supervised training to determine the global object features of each object node and the global resource features of each resource node in the global scene interaction graph. Specifically, based on the global scene interaction graph, two enhanced interaction graphs (a first enhanced interaction graph and a second enhanced interaction graph, respectively) can be constructed by controlling the random dropout rate of global object nodes and global resource nodes. The two enhanced interaction graphs are then compared to determine the positive and negative sample pairs used for training. The positive and negative sample pairs are then used as labels to train the two enhanced interaction graphs, resulting in trained enhanced interaction graphs. Finally, based on the trained enhanced interaction graphs, the global object features of each object node and the global resource features of each resource node are determined.

[0130] The global scene interaction graph can include multiple global object nodes, multiple global resource nodes, and historical interaction data between global object nodes and global resource nodes. In the global scene interaction graph, common graph neural network algorithms can be used to represent the features of global object nodes and global resource nodes, as follows:

[0131] Where i represents the index of different scene interaction graphs in the global scene interaction graph, u represents the object node of the scene interaction graph, v represents the resource node of the scene interaction graph, G represents the global scene interaction graph, and GNN is a graph neural network node embedding algorithm.

[0132] Based on this, data augmentation is performed on the global scene interaction graph to obtain a first augmented interaction graph and a second augmented interaction graph. This may include: extracting at least one global object node from multiple global object nodes as an augmented object node, and extracting at least one global resource node from multiple global resource nodes as an augmented resource node; filtering historical interaction data between augmented object nodes and augmented resource nodes from historical interaction data to obtain augmented node interaction data; and generating the first augmented interaction graph and the second augmented interaction graph based on the augmented object nodes, augmented resource nodes, and augmented node interaction data.

[0133] For example, the process of generating the first enhanced interaction graph may include the following steps: randomly selecting multiple global object nodes from the global scene interaction graph as enhanced object nodes of the first enhanced interaction graph, and randomly selecting multiple global resource nodes from the global scene interaction graph as enhanced resource nodes of the first enhanced interaction graph; after determining the enhanced object nodes and enhanced resource nodes corresponding to the first enhanced interaction graph, the historical interaction data between the enhanced object nodes and enhanced resource nodes can be filtered from the historical interaction data contained in the global scene interaction graph, and the filtered historical interaction data can be used as enhanced node interaction data; the first enhanced interaction graph can be generated based on the randomly selected enhanced object nodes, enhanced resource nodes, and enhanced node interaction data.

[0134] Similarly, the generation process of the second enhanced interaction graph can include the following steps: randomly selecting multiple global object nodes from the global scene interaction graph as enhanced object nodes of the second enhanced interaction graph, and randomly selecting multiple global resource nodes from the global scene interaction graph as enhanced resource nodes of the second enhanced interaction graph; after determining the enhanced object nodes and enhanced resource nodes corresponding to the second enhanced interaction graph, the historical interaction data between the enhanced object nodes and enhanced resource nodes can be filtered from the historical interaction data contained in the global scene interaction graph, and the filtered historical interaction data can be used as enhanced node interaction data; the second enhanced interaction graph can be generated based on the randomly selected enhanced object nodes, enhanced resource nodes, and enhanced node interaction data.

[0135] It should be understood that the number of enhanced object nodes in the first enhanced interaction graph is less than the number of global object nodes in the global scene interaction graph, and the number of enhanced resource nodes in the first enhanced interaction graph is also less than the number of global resource nodes in the global scene interaction graph. Similarly, the number of enhanced object nodes in the second enhanced interaction graph is less than the number of global object nodes in the global scene interaction graph, and the number of enhanced resource nodes in the second enhanced interaction graph is also less than the number of global resource nodes in the global scene interaction graph. Furthermore, the enhanced object nodes in the first enhanced interaction graph and the enhanced object nodes in the second enhanced interaction graph may partially overlap or may be completely different; similarly, the enhanced resource nodes in the first enhanced interaction graph and the enhanced resource nodes in the second enhanced interaction graph may partially overlap or may be completely different.

[0136] The process of generating a first enhanced interaction graph and a second enhanced interaction graph based on enhanced object nodes, enhanced resource nodes, and enhanced node interaction data may include: randomly extracting at least one enhanced object node from the enhanced object nodes as a first enhanced object node, and randomly extracting at least one enhanced object node from the enhanced object nodes as a second enhanced object node, with the first and second enhanced object nodes partially overlapping; randomly extracting at least one enhanced resource node from the enhanced resource nodes as a first enhanced resource node, and randomly extracting at least one enhanced resource node from the enhanced resource nodes as a second enhanced resource node, with the first and second enhanced resource nodes partially overlapping; filtering out first enhanced node interaction data between the first enhanced object node and the first enhanced resource node from the enhanced node interaction data, and filtering out second enhanced node interaction data between the second enhanced object node and the second enhanced resource node from the enhanced node interaction data; generating a first enhanced interaction graph based on the first enhanced object node, the first enhanced resource node, and the first enhanced node interaction data, and generating a second enhanced interaction graph based on the second enhanced object node, the second enhanced resource node, and the second enhanced node interaction data.

[0137] For ease of distinction, the enhanced object node used to generate the first enhanced interaction graph will be referred to as the first enhanced object node, and the enhanced object node used to generate the second enhanced interaction graph will be referred to as the second enhanced object node. Similarly, the enhanced resource node used to generate the first enhanced interaction graph will be referred to as the first enhanced resource node, and the enhanced resource node used to generate the second enhanced interaction graph will be referred to as the second enhanced resource node.

[0138] It should be noted that any given augmented object node may serve as the first augmented object node, the second augmented object node, or both. Similarly, any given augmented resource node may serve as the first augmented resource node, the second augmented resource node, or both.

[0139] Figure 10 illustrates the generation of the enhanced interaction graph provided in this embodiment. It should be understood that Figure 10 is merely an illustrative example of the process of generating the first and second enhanced interaction graphs based on the global scene interaction graph. The global scene interaction graph illustrated therein only shows global object nodes and global resource nodes, omitting virtual object connection nodes and / or virtual resource connection nodes. In practical applications, the global scene interaction graph may include virtual object connection nodes and / or virtual resource connection nodes. Therefore, the global scene interaction graph shown in Figure 10 does not limit the actual global scene interaction graph.

[0140] As shown in Figure 10, if the global scene interaction graph includes 4 global object nodes (denoted as u1, u2, u3, and u4) and 3 global resource nodes (denoted as v1, v2, and v3), then at least one of the 4 global object nodes can serve as an enhancement object node, and at least one of the 3 global resource nodes can serve as an enhancement resource node. The node interaction data between enhancement object nodes and enhancement resource nodes constitutes the enhancement node interaction data.

[0141] In Figure 10, u1, u3, and u4 are extracted from the augmented object nodes as first augmented object nodes. For ease of distinction, u1 as a first augmented object node is denoted as u1a, u3 as u3a, and u4 as u4a. Furthermore, u1, u2, u3, and u4 are also extracted from the augmented object nodes as second augmented object nodes. For ease of distinction, u1 as a second augmented object node is denoted as u1b, u2 as u2b, u3 as u3b, and u4 as u4b. The first augmented object nodes (u1a, u3a, u4a) and the second augmented object nodes (u1b, u2b, u3b, u4b) partially overlap.

[0142] Meanwhile, v1, v2, and v3 were extracted from the enhanced resource nodes as the first enhanced resource nodes. For ease of distinction, v1 as the first enhanced resource node is denoted as v1a, v2 as v2a, and v3 as v3a. Furthermore, v2 and v3 were also extracted from the enhanced resource nodes as the second enhanced resource nodes. For ease of distinction, v2 as the second enhanced resource node is denoted as v2b, and v3 as v3b. The first enhanced resource nodes (v1a, v2a, v3a) and the second enhanced resource nodes (v2b, v3b) partially overlap.

[0143] After determining the first augmented object node and the first augmented resource node, the first augmented node interaction data existing between the first augmented object node and the first augmented resource node can be filtered from the augmented node interaction data. Similarly, after determining the second augmented object node and the second augmented resource node, the second augmented node interaction data existing between the second augmented object node and the second augmented resource node can be filtered from the augmented node interaction data. Then, a first augmented interaction graph can be generated based on the first augmented object node, the first augmented resource node, and the first augmented node interaction data, and a second augmented interaction graph can be generated based on the second augmented object node, the second augmented resource node, and the second augmented node interaction data.

[0144] In this embodiment of the application, the enhanced object nodes and enhanced resource nodes in the two enhanced interaction graphs can be represented by node embedding, referring to the node sampling strategy and graph learning algorithm representation of global object nodes and global resource nodes in the global scene interaction graph:

[0145] Wherein, G1 represents the first enhanced interaction graph and G2 represents the second enhanced interaction graph.

[0146] The determination of positive and negative sample pairs based on the first and second enhanced interaction graphs may include: comparing the first enhanced object node in the first enhanced interaction graph with the second enhanced object node in the second enhanced interaction graph, and determining the object positive sample pair and the first negative sample pair corresponding to the enhanced object node based on the comparison result; comparing the first enhanced resource node in the first enhanced interaction graph with the second enhanced resource node in the second enhanced interaction graph, and determining the resource positive sample pair and the second negative sample pair corresponding to the enhanced resource node based on the comparison result; comparing the first enhanced object node in the first enhanced interaction graph, determining the third negative sample pair based on the comparison result, and comparing the first enhanced resource node in the first enhanced interaction graph, determining the fourth negative sample pair based on the comparison result; comparing the second enhanced object node in the second enhanced interaction graph, determining the fifth negative sample pair based on the comparison result, and comparing the second enhanced resource node in the second enhanced interaction graph, determining the sixth negative sample pair based on the comparison result; using the object positive sample pair and the resource positive sample pair as positive sample pairs, and using at least one of the first, second, third, fourth, fifth, and sixth negative sample pairs as negative sample pairs.

[0147] After constructing the first and second enhanced interaction graphs, positive and negative sample pairs can be built based on them. Specifically, the first and second enhanced interaction graphs can be compared, and positive and some negative sample pairs can be determined based on the comparison results; comparisons can also be made within the first enhanced interaction graph, and some negative sample pairs can be determined based on the comparison results; or comparisons can be made within the second enhanced interaction graph, and some negative sample pairs can be determined based on the comparison results. Table 3 shows the positive and negative sample pairs that can be obtained based on Figure 10.

[0148] Table 3. Positive and negative sample pairs corresponding to Figure 10

[0149] In Table 3, positive sample pairs include object positive sample pairs and resource positive sample pairs. The method for determining object positive sample pairs includes comparing the first enhanced object node in the first enhanced interaction graph with the second enhanced object node in the second enhanced interaction graph, and identifying identical enhanced object nodes as object positive sample pairs. For example, since the first enhanced object nodes in the first enhanced interaction graph include u1a, u3a, and u4a, and the second enhanced object nodes in the second enhanced interaction graph include u1b, u2b, u3b, and u4b, where u1a and u1b originate from the same enhanced object node u1, u3a and u3b originate from the same enhanced object node u3, and u4a and u4b originate from the same enhanced object node u4, it can be determined that object positive sample pairs include u1a-u1b, u3a-u3b, and u4a-u4b.

[0150] The method for determining positive resource pairs includes comparing the first enhanced resource node in the first enhanced interaction graph with the second enhanced resource node in the second enhanced interaction graph, and identifying identical enhanced resource nodes as positive resource pairs. For example, since the first enhanced resource nodes in the first enhanced interaction graph include v1a, v2a, and v3a, and the second enhanced resource nodes in the second enhanced interaction graph include v2b and v3b, where v2a and v2b originate from the same enhanced resource node v2, and v3a and v3b originate from the same enhanced resource node v3, positive resource pairs can be determined to include v2a-v2b and v3a-v3b.

[0151] There are several ways to determine negative sample pairs. For example, during the comparison of the first augmented object node in the first augmented interaction graph with the second augmented object node in the second augmented interaction graph, different augmented object nodes can be used as the first negative sample pairs. For instance, in Figure 10, the first augmented object node u1a in the first augmented interaction graph and the second augmented object nodes u2b, u3b, and u4b in the second augmented interaction graph originate from different augmented object nodes, thus determining that the first negative sample pairs include u1a-u2b, u1a-u3b, and u1a-u4b. Furthermore, the first augmented object node u3a in the first augmented interaction graph and the second augmented object nodes u1b, u2b, and u4b in the second augmented interaction graph originate from different augmented object nodes, thus determining that the first negative sample pairs also include u3a-u1b, u3a-u2b, and u3a-u4b. Furthermore, the first augmentation object node u4a in the first augmentation interaction graph and the second augmentation object nodes u1b, u2b, and u3b in the second augmentation interaction graph originate from different augmentation object nodes. Therefore, it can be determined that the first negative sample pair also includes u4a-u1b, u4a-u2b, and u4a-u3b.

[0152] The determination of negative sample pairs may further include: in the process of comparing the first enhanced resource node in the first enhanced interaction graph with the second enhanced resource node in the second enhanced interaction graph, different enhanced resource nodes are taken as the second negative sample pairs. For example, if the first enhanced resource node v1a in the first enhanced interaction graph and the second enhanced resource nodes v2b and v3b in the second enhanced interaction graph originate from different enhanced resource nodes, then the second negative sample pairs can be determined to include v1a-v2b and v1a-v3b. Furthermore, if the first enhanced resource node v2a in the first enhanced interaction graph and the second enhanced resource node v3b in the second enhanced interaction graph originate from different enhanced resource nodes, then the second negative sample pairs can also include v2a-v3b. Additionally, if the first enhanced resource node v3a in the first enhanced interaction graph and the second enhanced resource node v2b in the second enhanced interaction graph originate from different enhanced resource nodes, then the second negative sample pairs can also include v3a-v2b.

[0153] The determination of negative sample pairs can also include: comparing within the first augmentation interaction graph and determining a subset of negative sample pairs based on the comparison results. For example, the first augmentation object nodes in the first augmentation interaction graph can be compared, and different first augmentation object nodes can be used as the third negative sample pair; and the first augmentation resource nodes in the first augmentation interaction graph can be compared, and different first augmentation resource nodes can be used as the fourth negative sample pair. For example, the first augmentation object nodes in the first augmentation interaction graph include u1a, u3a, and u4a, where u1a and u3a are different first augmentation object nodes, u1a and u4a are different first augmentation object nodes, and u3a and u4a are different first augmentation object nodes. Therefore, the third negative sample pairs can be derived as u1a-u3a, u1a-u4a, and u3a-u4a. The first enhanced resource nodes in the first enhanced interaction graph include v1a, v2a and v3a, where v1a and v2a are different first enhanced resource nodes, v1a and v3a are different first enhanced resource nodes, and v2a and v3a are different first enhanced resource nodes. Therefore, the fourth negative sample pair includes v1a-v2a, v1a-v3a, and v2a-v3a.

[0154] The determination of negative sample pairs can also include: comparing within the second enhanced interaction graph and determining a subset of negative sample pairs based on the comparison results. For example, the second enhanced object nodes in the second enhanced interaction graph can be compared, and different second enhanced object nodes can be identified as the fifth negative sample pair; similarly, the second enhanced resource nodes in the second enhanced interaction graph can be compared, and different second enhanced resource nodes can be identified as the sixth negative sample pair. For instance, if the second enhanced object nodes in the second enhanced interaction graph include u1b, u2b, u3b, and u4b, the fifth negative sample pair can be obtained using the same determination method as the aforementioned third negative sample pair, including u1b-u2b, u1b-u3b, u1b-u4b, u2b-u3b, u2b-u4b, and u3b-u4b. Similarly, if the second enhanced resource nodes in the second enhanced interaction graph include v2b and v3b, the sixth negative sample pair can be obtained using the same determination method as the aforementioned fourth negative sample pair, including v2b-v3b.

[0155] As shown in Table 3, in the positive and negative sample pairs constructed using the first and second enhanced interaction graphs, the number of negative sample pairs is generally much higher than the number of positive sample pairs. Therefore, all positive sample pairs can be retained, and an appropriate number of negative sample pairs can be selected from the first to the sixth negative sample pairs as the final negative sample pairs used to train the graph neural network.

[0156] It should be understood that in the embodiments of this application, all positive and negative sample pairs may also be used as the final sample pairs for training the graph neural network.

[0157] The process of determining the global object features of each object node and the global resource features of each resource node based on positive and negative sample pairs can include: obtaining an initial first graph network corresponding to the first enhanced interaction graph and an initial second graph network corresponding to the second enhanced interaction graph; using the initial first graph network to predict the predicted first features of each enhanced node in the first enhanced interaction graph, where the predicted first features include the predicted first object features of the first enhanced object node and the predicted first resource features of the first enhanced resource node; using the initial second graph network to predict the predicted second features of each enhanced node in the second enhanced interaction graph, where the predicted second features include the predicted second object features of the second enhanced object node and the predicted second resource features of the second enhanced resource node; determining a first contrastive loss corresponding to the object node and a second contrastive loss corresponding to the resource node based on the positive and negative sample pairs, the predicted first features, and the predicted second features; converging the initial first graph network and the initial second graph network according to the first and second contrastive losses to obtain the first graph network and the second graph network; and determining the global object features of each object node and the global resource features of each resource node based on the first graph network and the second graph network.

[0158] In this embodiment of the application, the process of determining the global object features of each (global) object node and the global resource features of each (global) resource node based on positive sample pairs and negative sample pairs can be understood as the process of updating the parameters of the graph neural networks corresponding to the first and second enhanced interaction graphs based on positive sample pairs and negative sample pairs.

[0159] The training of a graph neural network model is similar to that of a traditional neural network, mainly including the following steps: (1) Initialize model parameters: The model parameters need to be initialized randomly, or fine-tuned using pre-trained model parameters; (2) Forward propagation: Input the data into the model and calculate the model's predicted value; (3) Calculate the loss function: Select the appropriate loss function according to the task type, such as cross-entropy, mean squared error loss function, etc.; (4) Backward propagation: Calculate the gradient of the loss function with respect to the model parameters and update the model parameters; (5) Repeated iteration: Repeat the forward propagation and backward propagation process until the model converges or reaches the specified number of iterations.

[0160] In this embodiment, for ease of distinction, the untrained graph neural network corresponding to the first augmented interaction graph is referred to as the initial first graph network, and the untrained graph neural network corresponding to the second augmented interaction graph is referred to as the initial second graph network. For the initial first graph network, the features of each first augmented object node and each first augmented resource node can be initialized. The initialized features of the first augmented object nodes are called initial first object features, and the initialized features of the first augmented resource nodes are called initial first resource features. Similarly, for the initial second graph network, the features of each second augmented object node and each second augmented resource node can be initialized. The initialized features of the second augmented object nodes are called initial second object features, and the initialized features of the second augmented resource nodes are called initial second resource features. Then, the initial first object features and initial first resource features are input into the initial first graph network, and the initial first graph network is used to predict the predicted first object features of each first augmented object node and the predicted first resource features of each first augmented resource node in the first augmented interaction graph. The predicted first object features and predicted first resource features can be collectively referred to as predicted first features. Simultaneously, the initial second object features and initial second resource features can be input into the initial second graph network. This initial second graph network is then used to predict the predicted second object features for each second augmented object node and the predicted second resource features for each second augmented resource node in the second augmented interaction graph. The predicted second object features and predicted second resource features can be collectively referred to as predicted second features.

[0161] After obtaining the predicted first features (including the predicted first object features of each first enhanced object node and the predicted first resource features of each first enhanced resource node) of each enhanced node in the first enhanced interaction graph, and the predicted second features (including the predicted second object features of each second enhanced object node and the predicted second resource features of each second enhanced resource node) of each enhanced node in the second enhanced interaction graph, the first contrast loss corresponding to the object node and the second contrast loss corresponding to the resource node can be determined based on positive sample pairs, negative sample pairs, predicted first features, and predicted second features. Specifically, the predicted first features and predicted second features can be combined into corresponding combined sample pairs according to the combination of positive and negative sample pairs. Then, the combined sample pairs are compared with the positive and negative sample pairs, and the first contrast loss corresponding to the object node and the second contrast loss corresponding to the resource node are determined based on the comparison results. For example, u1a-u1b is a positive object sample pair. The predicted first object feature corresponding to u1a and the predicted second object feature corresponding to u1b can be combined into a combined sample pair. Then, the combined sample pair is compared with u1a-u1b, and the comparison result is included in the first contrast loss corresponding to the object node. For example, if u1a-u2b is a negative sample pair, we can combine the predicted first object feature corresponding to u1a and the predicted second object feature corresponding to u2b into a combined sample pair. This combined sample pair is then compared with u1a-u2b, and the comparison result is included in the first contrastive loss corresponding to the object node. Similarly, if v2a-v2b is a positive resource sample pair, we can combine the predicted first resource feature corresponding to v2a and the predicted second resource feature corresponding to v2b into a combined sample pair. This combined sample pair is then compared with v2a-v2b, and the comparison result is included in the second contrastive loss corresponding to the resource node, and so on.

[0162] After determining the first and second contrastive losses, the initial first and second graph networks can be converged based on these losses to obtain the first and second graph networks. In this embodiment, the contrastive loss used in the self-supervised training process using the contrastive learning paradigm includes the first contrastive loss corresponding to the object node and the second contrastive loss corresponding to the resource node. The formula for calculating the contrastive loss is as follows: Contrast Loss

[0163] in, This represents the self-supervised contrastive loss of the object node, i.e., the first contrastive loss; G1 represents the self-supervised contrast loss of the resource node, i.e., the second contrast loss; G2 represents the first enhanced interaction graph and G1 represents the second enhanced interaction graph; U represents the object node and V represents the resource node; τ is a temperature coefficient commonly used in contrast loss to adjust the distribution of similarity scores in contrast loss; s is a similarity measurement function, which can be measured using cosine similarity in this embodiment.

[0164] As mentioned above, in the embodiments of this application, the first contrast loss and the second contrast loss can be added together, and the initial first graph network can be converged based on the summed contrast loss to obtain the first graph network. The initial second graph network can also be converged based on the summed contrast loss to obtain the second graph network.

[0165] The self-supervised training process using the contrastive learning paradigm described above can be understood as updating the parameters of the corresponding graph neural network to make the feature representations of two object nodes in the same positive sample pair in the first and second enhanced interaction graphs infinitely close, or even identical. Similarly, updating the parameters of the corresponding graph neural network to make the feature representations of two resource nodes in the same positive sample pair in the first and second enhanced interaction graphs infinitely close, or even identical. Simultaneously, updating the parameters of the corresponding graph neural network also needs to create differences in the feature representations of two object nodes in the same negative sample pair in the first and second enhanced interaction graphs, and also to create differences in the feature representations of two resource nodes in the same negative sample pair in the first and second enhanced interaction graphs.

[0166] After obtaining the trained first and second graph networks, the global object features of each object node and the global resource features of each resource node can be determined based on these networks. It should be understood that the first and second augmented interaction graphs may not completely cover every global object node and global resource node in the global scene interaction graph. However, the global object features of each first augmented object node and the global resource features of each first augmented resource node in the first augmented interaction graph can be directly determined based on the first graph network, and the global object features of each second augmented object node and the global resource features of each second augmented resource node in the second augmented interaction graph can be directly determined based on the second graph network. Based on this, the first augmented object nodes and second augmented resource nodes covered in the first augmented interaction graph, as well as the second augmented object nodes and second augmented resource nodes covered in the second augmented interaction graph, can be substituted into their corresponding positions in the global scene interaction graph. Then, based on the historical interaction data in the global scene interaction graph, the global object features corresponding to the remaining global object nodes and the global resource features corresponding to the remaining global resource nodes can be determined.

[0167] The global object features and global resource features obtained after self-supervised training using a contrastive learning paradigm integrate historical interaction data from multiple scene interaction graphs. This fully explores the high-order information between object nodes and resource nodes in multiple scene interaction graphs, and can well reflect the intrinsic relationship between object nodes and resource nodes in the global scene.

[0168] 104. Based on the global object features and global resource features, train the initial scene graph network corresponding to the scene interaction graph to obtain the scene graph network. The scene graph network is used to determine the push resources for the push object so as to push the push resources to the push object.

[0169] The Scene Graph Network is a graph neural network obtained by training an initial Scene Graph Network with global object features and global resource features built upon a global scene interaction graph. It is used to predict the interaction probability between a pushed object and resources, and to determine the resources to be pushed to that object.

[0170] Considering that the real-world situations of each information flow scenario are different, if similarity calculation is directly performed using global object features and global resource features for any information flow scenario, and resources are pushed based on the calculation results, it is inevitable that specific information of that information flow scenario will be ignored, resulting in a certain degree of deviation in the push results. Therefore, in this embodiment, based on the global object features and global resource features, an initial scene graph network corresponding to the scene interaction graph of each information flow scenario can be trained to obtain a trained scene graph network. This achieves the effect of fine-tuning the feature representations of object nodes and resource nodes in each scene interaction graph, enabling more accurate prediction of the interaction preferences between each object node and resource node.

[0171] The initial scene graph network is an untrained graph neural network model corresponding to the scene interaction graph. It serves as the basis for subsequent training based on global object features and global resource features to learn the interaction patterns between objects and resources within the scene. The initial scene graph network can be understood as an untrained graph neural network corresponding to the scene interaction graph.

[0172] Specifically, based on global object features and global resource features, an initial scene graph network corresponding to each scene interaction graph is trained to obtain a scene graph network. This process may include: selecting global object features corresponding to object nodes in each scene interaction graph from global object features to obtain scene global object features, and selecting global resource features corresponding to resource nodes in each scene interaction graph from global resource features to obtain scene global resource features; calculating predicted object features corresponding to scene global object features based on the initial scene graph network corresponding to the scene interaction graph, and calculating predicted resource features corresponding to scene global resource features based on the initial scene graph network; calculating the first similarity between predicted object features and predicted resource features to determine predicted interaction data between object nodes and resource nodes in the scene interaction graph; comparing the predicted interaction data with historical interaction data, and converging the initial scene graph network based on the comparison results to obtain the scene graph network.

[0173] The process of calculating predicted object features and predicted resource features using the initial scene graph network can be summarized as follows: The initial scene graph network contains multiple neural network layers, such as an input layer, hidden layers, and an output layer. First, the global object features of the scene are input into the input layer of the initial scene graph network, which then passes these features to the hidden layers. In the hidden layers, each neuron performs a weighted summation of the input features and applies a non-linear transformation using an activation function (such as the ReLU function) to obtain an intermediate feature representation. After processing by multiple hidden layers, the intermediate feature representation is finally passed to the output layer, which outputs the predicted object features. Similarly, the global resource features of the scene are processed in the same way through the input, hidden, and output layers to obtain the predicted resource features.

[0174] Specifically, a similarity threshold θ is preset. When the first similarity between the predicted object features and the predicted resource features is greater than θ, a predicted interaction relationship is considered to exist between the object node and the resource node, and the predicted interaction data between the object node and the resource node is marked as positive. When the first similarity is less than or equal to θ, no predicted interaction relationship is considered to exist between the object node and the resource node, and the predicted interaction data between the object node and the resource node is marked as negative. In this way, the predicted interaction data between object nodes and resource nodes in the scene interaction graph is determined based on the first similarity and the preset similarity threshold θ.

[0175] For example, for a scene interaction graph corresponding to an information flow scenario, the global object features corresponding to the object nodes in the scene interaction graph in the global interaction graph can be used as the scene global object features, and the global resource features corresponding to the resource nodes in the scene interaction graph in the global interaction graph can be used as the scene global resource features. This yields the scene global object features for each object node and the scene global resource features for each resource node in the scene interaction graph. Then, the scene global object features are input into the initial scene graph network to obtain predicted object features, and the scene global resource features are input into the initial scene graph network to obtain predicted resource features. In this process, the global object features corresponding to an object node can be considered as the initial object embedding representation of that object node in the initial scene graph network, and the global resource features corresponding to a resource node can be considered as the initial resource embedding representation of that resource node in the initial scene graph network.

[0176] Next, the first similarity between each predicted object feature and predicted resource feature is calculated. The first similarity can be cosine similarity. Based on the first similarity, the predicted interaction data between the object node and the resource node is determined. There are several ways to determine the predicted interaction data between the object node and the resource node based on the first similarity. For example, a similarity threshold can be preset. An interaction relationship can be established between object nodes and resource nodes whose first similarity is greater than the similarity threshold. That is, the predicted interaction data indicates that there may be an interaction relationship between a certain object node and a certain resource node.

[0177] The specific mechanism by which the calculation of the first similarity affects the determination of predicted interaction data is as follows: The first similarity reflects the degree of similarity of features between object nodes and resource nodes. When the first similarity is high, it indicates that the object nodes and resource nodes are quite similar in features, and there is a high probability that there is an interaction between them. In this case, the predicted interaction data can be marked as positive, indicating that there may be an interaction relationship between the object nodes and resource nodes. When the first similarity is low, it indicates that the object nodes and resource nodes are quite different in features, and there is a low probability that there is an interaction relationship between them. In this case, the predicted interaction data can be marked as negative, indicating that there may be no interaction relationship between the object nodes and resource nodes. A similarity threshold can be set; when the first similarity is greater than the threshold, it is marked as positive, and when it is less than or equal to the threshold, it is marked as negative.

[0178] Since each scene interaction graph already contains historical interaction data between object nodes and resource nodes, the predicted interaction data can be compared with the historical interaction data. The difference between the predicted and historical interaction data can be used as the loss information for the initial scene graph network, thereby enabling supervised training of the initial scene graph network and obtaining the trained scene graph network. Specifically, comparing the predicted and historical interaction data and converging the initial scene graph network based on the comparison results can include: calculating the difference information between the predicted and historical interaction data; determining the loss information of the initial scene graph network based on the difference information; and training the initial scene graph network based on the loss information to obtain the scene graph network.

[0179] It should be noted that historical interaction data can include both actual interactions between object nodes and resource nodes in the scene interaction graph and data where no interaction occurs. For example, in the scene interaction graph shown in Figure 5, u1 and v1 have a real interaction relationship, while u1 and v2 do not. The historical interaction data in Figure 5 can include both data showing a real interaction between u1 and v1 and data showing no interaction between u1 and v2. Therefore, the historical interaction data between u1 and v1 can be used as positive samples, and the historical interaction data between u1 and v2 as negative samples. Based on this, the predicted interaction data is compared with the positive and negative samples respectively, and the corresponding difference information is calculated. Then, based on the difference information, the loss information of the initial scene graph network is determined, and the initial scene network is converged according to the loss information to obtain the scene graph network.

[0180] In this embodiment of the application, for ease of distinction, the scene interaction graph that uses global object features to represent object nodes and global resource features to represent resource nodes is represented as follows: Based on this, the loss information during the supervised training process of the initial scene graph network using historical interaction data can be represented as follows:

[0181] Among them, E i express The actual "object-resource" interaction edges on the interaction graph; (u,v) represents positive samples of actual interactions; Negative samples represent samples where there is no interaction. σ represents the first similarity between the predicted object features and the predicted resource features, reflecting the interaction tendency of the object node to the resource node; σ is the activation function, and the sigmoid function can be used as the activation function in this embodiment.

[0182] As can be seen from the above, the embodiments of this application obtain scene interaction graphs corresponding to multiple information flow scenarios. The scene interaction graphs are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario. Then, based on object nodes and resource nodes, the scene interaction graphs are connected to obtain a global scene interaction graph. After that, global object features of each object node and global resource features of each resource node are constructed in the global scene interaction graph. Then, based on the global object features and global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain a scene graph network. The scene graph network is used to determine the push resources for the push object so as to push the push resources to the push object. This solution integrates the scene interaction graphs corresponding to multiple information flow scenarios to construct a global scene interaction graph that fully reflects the commonalities and differences between various information flow scenarios. This eliminates pattern bias and data bias in single-scenario modeling, more accurately mines user information flow preferences, and reduces operation and maintenance costs. Furthermore, for the global object features corresponding to each object node and the global resource features corresponding to each resource node in the global scene interaction graph, the global object features and global resource features are further used as the initial features of object nodes and resource nodes in a single scene interaction graph. This is then used to train the initial scene graph corresponding to the single scene interaction graph, resulting in a scene graph network that more accurately reflects the feature representation of object nodes and resource nodes in a single scene interaction graph. This improves the prediction accuracy of subsequent resource push using this scene graph network.

[0183] This application also provides a resource push method, apparatus, electronic device, and computer-readable storage medium. The resource push function can be integrated into an electronic device, which may be a server or a terminal, etc.

[0184] The server can be a standalone 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, content delivery network (CDN) acceleration services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these.

[0185] The application scenario diagram shown in Figure 2 is also applicable to the resource push method. As shown in Figure 2, taking the resource push device integrated into an electronic device, with the electronic device serving as a server, as an example, the user terminal (i.e., the client) where the user (including the push recipient) is located can establish a communication connection with the server. Based on this, when a resource push request is received for the push recipient, the server can detect the target information flow scene currently in which the push recipient is located; filter out the target scene graph network corresponding to the target information flow scene from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes; based on the target scene graph network, determine at least one push resource associated with the push recipient and push the push resource to the push recipient.

[0186] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0187] This embodiment will be described from the perspective of a resource push device, which can be integrated into an electronic device, such as a server.

[0188] A resource push method includes: when a resource push request is received for a target object, detecting the target information flow scene currently in which the target object is located; selecting a target scene graph network corresponding to the target information flow scene from a scene graph network, wherein the scene graph network includes global object features and global resource features constructed based on a global scene interaction graph, and a graph network trained on an initial scene graph network corresponding to each scene interaction graph, wherein the global scene interaction graph is obtained by connecting scene interaction graphs corresponding to multiple information flow scenes; and determining at least one push resource associated with the target object based on the target scene graph network, and pushing the push resource to the target object.

[0189] Figure 11 shows a flowchart of the resource push method provided in an embodiment of this application. As shown in Figure 11, the specific flow of the resource push method is as follows:

[0190] 201. When a resource push request is received for the target object, detect the target information flow scenario in which the target object is currently located.

[0191] The recipient of a resource push request is an object that receives resources pushed by the system. Based on its characteristics and context information, the system selects suitable resources from the context graph network for push. The target information flow context refers to the specific information flow context in which the recipient is currently located when a resource push request is received. The system will select the corresponding context graph network based on this context for resource push.

[0192] There are several ways to receive resource push requests for a target object. For example, the terminal can detect the current object's operation status. When it detects that the current object has launched an application software corresponding to a certain information flow scenario, it designates the current object as the target object and generates a resource push request corresponding to the target object. The terminal then sends this resource push request to the server corresponding to the resource push device. After receiving the resource push request for the target object, the server can send at least one push resource to the terminal where the target object is located based on the scenario graph network trained in the aforementioned steps, thereby completing the resource push to the target object.

[0193] The specific methods for detecting the current target information stream scenario of a push notification recipient are as follows: The target information stream scenario can be determined by analyzing the information carried in the resource push request. For example, the request may contain the identifier of the application currently being used by the recipient; based on the application identifier, the corresponding information stream scenario can be determined. Alternatively, the target information stream scenario can also be detected by monitoring the recipient's device status and behavioral data. For example, the current scenario of the recipient can be determined through device sensor data (such as screen display content, application running status, etc.). Furthermore, a comprehensive judgment can be made by combining the user's historical behavioral data and contextual information; for example, the target information stream scenario can be inferred based on the user's frequently used scenarios within a specific time period and current time and location.

[0194] 202. Select the target scene graph network corresponding to the target information flow scene from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes.

[0195] The target scene graph network is a graph neural network selected from scene graph networks that corresponds to the target information flow scene. It is trained based on features of the global scene interaction graph and is used to determine the push resources for the target object in the target scene.

[0196] Resource push requests can carry information about the target information flow scenario currently in which the target object is located. When a resource push request for a target object is received, the target information flow scenario can be identified from the resource push request. Since each information flow scenario corresponds to a scenario interaction graph, and each scenario interaction graph corresponds to a trained scenario graph network, once the target information flow scenario is determined, the scenario interaction graph corresponding to the target information flow scenario can be used as the target scenario interaction graph, and the scenario graph network corresponding to the target scenario interaction graph can be used as the target scenario graph network.

[0197] The specific implementation method for filtering target scene graph networks is as follows: Scene graph networks can be stored as a data structure, where each scene graph network is associated with a corresponding information flow scene. Once the target information flow scene is determined, this data structure can be traversed to find scene graph networks with the same identifier as the target information flow scene. Alternatively, an indexing mechanism can be used, creating an index for each information flow scene, and the corresponding target scene graph network can be directly located based on the index of the target information flow scene. Furthermore, if the scene graph networks are stored in a database, database query statements can be used to query based on the information of the target information flow scene to filter out the target scene graph networks.

[0198] 203. Based on the target scene graph network, determine at least one push resource associated with the push object, and push the push resource to the push object.

[0199] Push resources refer to resources that are associated with the target object and suitable for being pushed to that object, as determined by the prediction results of the scene graph network.

[0200] In this embodiment of the application, after supervised learning using historical interaction data in the scene interaction graph, the object nodes and resource nodes in each scene interaction graph can be represented as follows:

[0201] in, This represents the object node features corresponding to the object node after supervised fine-tuning in the i-th scene interaction graph. This represents the resource node feature corresponding to the resource node after supervised fine-tuning in the i-th scene interaction graph. and It can be directly used for resource pushing in the interaction diagram of this scene.

[0202] After determining the target scene graph network, the pushable object features of the target object can be determined based on the target scene graph network. These features are the characteristic representation of the target object in the target scene graph network, determined by comparing the target description information of the target object with the object description information of object nodes within the scene. They are used to calculate the similarity with scene resources to filter push resources. The pushable object features can be understood as the scene object features corresponding to the target object in the target scene graph network. Then, based on these scene object features, at least one push resource associated with the target object is calculated, and the push resource is pushed to the target object. Specifically, determining at least one push resource associated with the target object based on the target scene graph network can include: determining the pushable object features corresponding to the target object; determining the scene resource features corresponding to each resource node in the target information flow scene based on the target scene graph network; and calculating at least one push resource associated with the target object based on the pushable object features and scene resource features. Scene resource features are the feature representations of resource nodes calculated from the target scene graph network in the scene interaction graph corresponding to the target information flow scene. They are used to calculate the similarity with the pushable object features to filter suitable resources to be pushed to the target object.

[0203] In determining the characteristics of the target audience for a push notification, the following two scenarios may occur:

[0204] The first scenario: Determine the characteristics of the push object corresponding to the push object, which may include: obtaining the target description information of the push object, and the object description information corresponding to the object nodes contained in the target information stream scenario;

[0205] If the object description information is the same as the target description information, the object node corresponding to the pushed object is taken as the pushed object node; the scene object features corresponding to the pushed object node are determined according to the target scene graph network; the scene object features are taken as the pushed object features.

[0206] The first scenario concerns the process of determining the characteristics of the pushed object when the target object is already an object node in the target information flow scenario. To determine whether the pushed object is an existing object node in the target information flow scenario, we can first obtain the target description information of the pushed object, as well as the object description information corresponding to each object node in the target information flow scenario. The target description information can reflect the basic profile information of the user corresponding to the pushed object, such as age, gender, education, etc. Then, we compare the target description information with each object description information. If there is an object description information that matches the target description information, then the object node corresponding to that object description information is the pushed object node of the pushed object in the target information flow scenario. After determining the pushed object node, we can determine the scene object characteristics corresponding to the pushed object node based on the target scene graph network, and use these scene object characteristics as the pushed object characteristics corresponding to the pushed object.

[0207] The second scenario: Determine the characteristics of the push object corresponding to the push object, which may include: obtaining the target description information of the push object, and the object description information corresponding to the object nodes contained in the target information stream scenario;

[0208] If the object description information is different from the target description information, obtain the global object description information corresponding to each global object node in the global scene interaction graph; filter out the global object nodes corresponding to the global object description information that is the same as the target description information from the global scene interaction graph to obtain the target global object node; based on the scene graph network, determine the scene object features corresponding to the target global object node to obtain the push object features of the pushed object.

[0209] The second scenario addresses the process of determining the features of the pushed object when the target object is not an object node in the target information flow scenario. After obtaining the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information flow scenario, the target description information can be compared with each object description information. If no object description information matching the target description information is found, the comparison scope can be expanded to the global scene interaction graph. The global object description information corresponding to each global object node in the global scene interaction graph is compared with the target description information. The global object node corresponding to the global object description information matching the target description information is taken as the target global object node. Then, the scene object features corresponding to the target global object node are determined based on the scene graph network where the target global object node is located, and these scene object features are taken as the pushed object features of the pushed object.

[0210] It should be noted that in the second scenario, the scene interaction graph of the target global object node differs from the scene interaction graph of the target information flow scene. Consequently, the scene graph network of the target global object node also differs from the scene graph network of the target information flow scene. In other words, even if the pushed object is not an object node in its current target information flow scene, the characteristics of the pushed object can be determined from the global scene interaction graph. These characteristics can then be applied to the target scene graph network of the target information flow scene to calculate at least one push resource associated with the pushed object in the target information flow scene. This method enables object-side and resource-side cold start functionality in new scenarios.

[0211] Among them, scene resource features can be understood as the feature representations of each resource node in the scene interaction graph corresponding to the target information flow scene, calculated based on the target scene graph network.

[0212] The step of calculating at least one push resource associated with the push object based on the characteristics of the push object and the characteristics of the scene resources may include: calculating a second similarity between the characteristics of the push object and the characteristics of the scene resources; comparing the second similarity with a preset similarity threshold, and selecting push resource nodes from the resource nodes of the target information flow scene based on the comparison result; and using the resources corresponding to the push resource nodes as push resources.

[0213] The second similarity refers to the similarity between the characteristics of the target object and the characteristics of the scene resources calculated during the resource push process. It is used to filter out the push resource nodes from the resource nodes of the target information flow scene.

[0214] After determining the characteristics of the target object and the scene resource characteristics of each resource node in the target scene graph network, the similarity between the target object characteristics and each scene resource characteristic can be calculated (for ease of distinction, this similarity is called the second similarity), and at least one resource node can be selected as the push resource node based on the second similarity.

[0215] There are several ways to select at least one resource node as the push resource node based on the second similarity score. For example, the second similarity score can be compared with a preset similarity threshold, and the resource node with a second similarity score greater than the threshold can be selected as the push resource node. Alternatively, the calculated second similarities can be arranged in descending order, and the resource node corresponding to the highest-ranking second similarity score can be selected as the push resource node. After determining the push resource node, the resource corresponding to the push resource node can be used as the push resource and pushed to the target object.

[0216] It should be noted that, from the perspective of the organization of "user-item" interaction data, as the data scale increases, the traditional mainstream push algorithms that model user historical interaction data using paradigms such as point-based modeling (as shown in Figure 12A) and sequence modeling (as shown in Figure 12B) can no longer meet the needs of users. Therefore, it is necessary to continuously develop new push technologies. For example, the push system based on graph structure modeling used in the embodiments of this application can capture more comprehensive user preferences compared to point-based modeling and sequence modeling.

[0217] Based on graph structure modeling, researchers have developed several multi-scenario collaborative push systems, such as the auxiliary paradigm (Figure 13A), the multi-tower paradigm (Figure 13B), and the expert paradigm (Figure 13C). The auxiliary paradigm relies on a common module to learn the commonalities of multiple scenarios, and then uses scene features and a multilayer perceptron to learn scene differences, merging these two characteristics when predicting user preferences. Since the fusion of common feature information from multiple scenarios and difference information from a single scenario is unstable, the multi-tower paradigm improves upon this by using multiple sub-modules to learn single-scenario feature information on a shared underlying layer. The accuracy of the multi-tower paradigm modeling is easily affected by the sparsity of single-scenario data. The expert paradigm uses a lightweight expert network to characterize the differences between multiple scenarios, and then employs a gating mechanism for adaptive information fusion. However, biases in scene data can easily lead to uneven workloads during the learning process of multiple experts, resulting in suboptimal user preference characterization.

[0218] The technical solution of this application proposes virtual object connection nodes and virtual resource connection nodes to communicate scene interaction graphs corresponding to multiple information flow scenarios, based on the organization of user interaction data in a graph structure. Under the role of virtual object connection nodes and virtual resource connection nodes, the deviation of multi-scenario interaction data is solved by controlling the magnitude of cross-graph sampling. By establishing an information exchange mechanism between multiple scene interaction graphs, multi-scenario joint modeling is realized, and the problem of multi-scenario data deviation is solved through sampling strategies.

[0219] To comprehensively evaluate the effectiveness of the resource push method proposed in this application, evaluations were also conducted in multiple academic and industry scenarios. The recall effects of traditional single-scenario supervised modeling methods, multi-scenario joint construction push methods (such as auxiliary paradigms, multi-tower paradigms, and expert paradigms), and the resource push method proposed in this application were compared to demonstrate that the proposed resource push method has superior performance. Table 4 shows the evaluation results of the push effect in multiple academic scenarios, and Table 5 shows the evaluation results of the push effect in multiple industry scenarios.

[0220] In the comparative evaluation process, LightGCN (a push algorithm based on graph convolutional neural network GCN) was used uniformly as the graph structure node representation learning method; 2-hop (second-order) was used for both single-scene node sampling and multi-scene joint graph node sampling. The recall effect evaluation mainly focused on the gains of Recall, MRR (mean reciprocal ranking), NDCG (normalized discount cumulative gain), and Hit Ratio metrics when Top K=10.

[0221] Recall rate refers to the proportion of relevant items that are successfully pushed by the push system out of all relevant items. The higher the value, the better.

[0222] MRR (Mean Reciprocal Rank): A metric used to evaluate the ranking quality of a push notification system, with a particular focus on the ranking of the first relevant item. It is the average of the reciprocal rankings of all queries, and a higher value is better.

[0223] NDCG (Normalized Discounted Cumulative Gain): A metric that considers the relevance and ranking of push notifications. It evaluates the quality of push notifications by discounting the gain of relevant items in the push list; a higher value is better.

[0224] Hit Ratio: This refers to the percentage of items in the push notification list that contain at least one relevant item; the higher the value, the better.

[0225] Table 4. Evaluation Results of Push Notification Effectiveness in Multiple Academic Scenarios

[0226] Table 5. Evaluation Results of Push Notification Performance in Multiple Scenarios in the Industry

[0227] Based on the comprehensive recall metrics obtained in Tables 4 and 5, the resource push method proposed in this application has achieved significant benefits. Breaking down the data reveals that the benefits primarily stem from the unsupervised learning phase, which involves constructing two enhanced interaction graphs by controlling the random dropout rates of global object nodes and global resource nodes, based on the global scene interaction graph, and then performing comparative learning to determine the characteristics of global objects and global resources.

[0228] As can be seen from the above, the embodiments of this application obtain scene interaction graphs corresponding to multiple information flow scenarios. The scene interaction graphs are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario. Then, based on object nodes and resource nodes, the scene interaction graphs are connected to obtain a global scene interaction graph. After that, global object features of each object node and global resource features of each resource node are constructed in the global scene interaction graph. Then, based on the global object features and global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain a scene graph network. When a resource push request for a push object is received, at least one push resource is pushed to the push object based on the scene graph network. This solution integrates the scene interaction graphs corresponding to multiple information flow scenarios to construct a global scene interaction graph that fully reflects the commonalities and differences between various information flow scenarios. This eliminates pattern bias and data bias in single-scenario modeling, more accurately mines user information flow preferences, and reduces operation and maintenance costs. Furthermore, for the global object features corresponding to each object node and the global resource features corresponding to each resource node in the global scene interaction graph, the global object features and global resource features are further used as the initial features of object nodes and resource nodes in a single scene interaction graph. This is then used to train the initial scene graph corresponding to the single scene interaction graph, resulting in a scene graph network that more accurately reflects the feature representation of object nodes and resource nodes in a single scene interaction graph. Applying the scene graph network to resource push can improve the accuracy of resource push.

[0229] Based on the method described in the above embodiments, the following examples will provide further detailed explanations.

[0230] In this embodiment, the resource push device will be specifically integrated into an electronic device, with the electronic device being a server, as an example for explanation.

[0231] Figure 14 shows another flowchart of the resource push method provided in an embodiment of this application. As shown in Figure 14, the specific process of a resource push method is as follows:

[0232] 301. The server obtains the scene interaction diagrams corresponding to each of the multiple information flow scenarios. The scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario.

[0233] For example, the server can obtain the object nodes, resource nodes, and historical interaction data between object nodes and resource nodes corresponding to multiple information flow scenarios; and construct a bipartite graph based on the object nodes, resource nodes, and historical interaction data to obtain the scenario interaction graph.

[0234] 302. The server obtains the object description information corresponding to the object node and the resource description information corresponding to the resource node in the scene interaction graph.

[0235] For example, object description information can be understood as the node attribute characteristics of an object node. For instance, object description information could reflect the basic profile information of the user corresponding to that object node, such as age, gender, education level, etc. Resource description information can be understood as the node attribute information of a resource node. For instance, resource description information could reflect the category information, style information, etc., of the resource corresponding to that resource node.

[0236] 303. Based on the object description information, the server constructs at least one virtual object connection node in the scene interaction graph. The virtual object connection node is used to connect object nodes that have an association relationship with the virtual object connection node.

[0237] For example, the server can filter at least one pushed object description tag from the object description tags and add an object tag value range to the pushed object description tag; based on the pushed object description tag and the object tag value range, at least one virtual object connection node is constructed in the scene interaction graph, and the virtual object description information corresponding to the virtual object connection node includes the pushed object description tag and the object tag value range.

[0238] 304. Based on the resource description information, the server constructs at least one virtual resource connection node in the scene interaction graph. The virtual resource connection node is used to connect resource nodes that are associated with the virtual object connection node.

[0239] For example, the server can filter at least one push resource description tag from the resource description tags and add a resource tag value range to the push resource description tag; based on the push resource description tag and the resource tag value range, at least one virtual resource connection node is constructed in the scene interaction graph, and the virtual resource description information corresponding to the virtual resource connection node includes the push resource description tag and the resource tag value range.

[0240] 305. The server connects the scene interaction graph based on virtual object connection nodes and virtual resource connection nodes to obtain a global scene interaction graph.

[0241] For example, the server can obtain the virtual object description information corresponding to the virtual object connection node and the virtual resource description information corresponding to the virtual resource connection node in each scene interaction graph; connect the virtual object connection nodes with the same virtual object description information, and connect the virtual resource connection nodes with the same virtual resource description information to obtain the global scene interaction graph.

[0242] For example, the server can obtain the number of object nodes corresponding to object nodes and the number of resource nodes corresponding to resource nodes in the scene interaction graph; determine the number of virtual object nodes to be constructed corresponding to virtual object connection nodes in the scene interaction graph based on the number of object nodes, and determine the number of virtual resource nodes to be constructed corresponding to virtual resource connection nodes in the scene interaction graph based on the number of resource nodes; construct virtual object connection nodes in the scene interaction graph based on the number of virtual object nodes and object description information, and construct virtual resource connection nodes in the scene interaction graph based on the number of virtual resource nodes and resource description information.

[0243] 306. The server constructs global object characteristics for each object node and global resource characteristics for each resource node in the global scene interaction graph.

[0244] For example, the server can perform data augmentation on the global scene interaction graph to obtain a first augmented interaction graph and a second augmented interaction graph; based on the first augmented interaction graph and the second augmented interaction graph, positive sample pairs and negative sample pairs are determined; based on the positive sample pairs and negative sample pairs, the global object features of each object node and the global resource features of each resource node are determined.

[0245] Enhanced object nodes are selected from the global object nodes of the global scene interaction graph and are used to generate the first and second enhanced interaction graphs to increase the diversity and richness of the training data. Enhanced resource nodes are selected from the global resource nodes of the global scene interaction graph and are used to generate the first and second enhanced interaction graphs to increase the diversity and richness of the training data.

[0246] For example, the server can extract at least one global object node as an enhanced object node from multiple global object nodes, and at least one global resource node as an enhanced resource node from multiple global resource nodes; it can also filter historical interaction data between enhanced object nodes and enhanced resource nodes from historical interaction data to obtain enhanced node interaction data; based on the enhanced object nodes, enhanced resource nodes, and enhanced node interaction data, a first enhanced interaction graph and a second enhanced interaction graph are generated. The enhanced node interaction data, selected from the historical interaction data of the global scene interaction graph, represents the interaction data between enhanced object nodes and enhanced resource nodes, and is used to generate the first and second enhanced interaction graphs, providing data support for comparative learning.

[0247] For example, the server can randomly extract at least one enhanced object node from the enhanced object nodes as a first enhanced object node, and randomly extract at least one enhanced object node from the enhanced object nodes as a second enhanced object node, with the first and second enhanced object nodes partially overlapping; it can randomly extract at least one enhanced resource node from the enhanced resource nodes as a first enhanced resource node, and randomly extract at least one enhanced resource node from the enhanced resource nodes as a second enhanced resource node, with the first and second enhanced resource nodes partially overlapping; it can filter out the first enhanced node interaction data between the first enhanced object node and the first enhanced resource node from the enhanced node interaction data, and filter out the second enhanced node interaction data between the second enhanced object node and the second enhanced resource node from the enhanced node interaction data; it can generate a first enhanced interaction graph based on the first enhanced object node, the first enhanced resource node, and the first enhanced node interaction data, and generate a second enhanced interaction graph based on the second enhanced object node, the second enhanced resource node, and the second enhanced node interaction data.

[0248] For example, the server can compare the first enhanced object node in the first enhanced interaction graph with the second enhanced object node in the second enhanced interaction graph, and determine the object positive sample pair and the first negative sample pair corresponding to the enhanced object node based on the comparison result; compare the first enhanced resource node in the first enhanced interaction graph with the second enhanced resource node in the second enhanced interaction graph, and determine the resource positive sample pair and the second negative sample pair corresponding to the enhanced resource node based on the comparison result; compare the first enhanced object node in the first enhanced interaction graph, and determine the third negative sample pair based on the comparison result; compare the first enhanced resource node in the first enhanced interaction graph, and determine the fourth negative sample pair based on the comparison result; compare the second enhanced object node in the second enhanced interaction graph, and determine the fifth negative sample pair based on the comparison result; compare the second enhanced resource node in the second enhanced interaction graph, and determine the sixth negative sample pair based on the comparison result; use the object positive sample pair and the resource positive sample pair as positive sample pairs, and use at least one of the first negative sample pair, the second negative sample pair, the third negative sample pair, the fourth negative sample pair, the fifth negative sample pair, and the sixth negative sample pair as negative sample pairs.

[0249] For example, the server can obtain the initial first graph network corresponding to the first augmented interaction graph and the initial second graph network corresponding to the second augmented interaction graph; use the initial first graph network to predict the predicted first features of each augmented node in the first augmented interaction graph, the predicted first features including the predicted first object features of the first augmented object node and the predicted first resource features of the first augmented resource node; use the initial second graph network to predict the predicted second features of each augmented node in the second augmented interaction graph, the predicted second features including the predicted second object features of the second augmented object node and the predicted second resource features of the second augmented resource node; based on positive sample pairs, negative sample pairs, the predicted first features, and the predicted second features, determine the first contrastive loss corresponding to the object node and the second contrastive loss corresponding to the resource node; converge the initial first graph network and the initial second graph network according to the first contrastive loss and the second contrastive loss to obtain the first graph network and the second graph network; and determine the global object features of each object node and the global resource features of each resource node according to the first graph network and the second graph network. The first contrastive loss is the contrastive loss of the object node calculated during contrastive learning training based on positive sample pairs, negative sample pairs, the predicted features of nodes in the first augmented interaction graph, and the predicted features of nodes in the second augmented interaction graph, and is used to update the graph network parameters. The second contrastive loss is the contrastive loss of resource nodes calculated during contrastive learning training based on positive sample pairs, negative sample pairs, predicted features of nodes in the first augmented interaction graph, and predicted features of nodes in the second augmented interaction graph. This loss is used to update the graph network parameters.

[0250] 307. The server trains the initial scene graph network corresponding to the scene interaction graph based on the global object features and global resource features to obtain the scene graph network.

[0251] For example, the server can filter global object features corresponding to object nodes in each scene interaction graph from global object features to obtain scene global object features, and filter global resource features corresponding to resource nodes in each scene interaction graph from global resource features to obtain scene global resource features. Based on the initial scene graph network corresponding to the scene interaction graph, the server calculates the predicted object features corresponding to the scene global object features, and calculates the predicted resource features corresponding to the scene global resource features based on the initial scene graph network. The server calculates the first similarity between the predicted object features and the predicted resource features to determine the predicted interaction data between object nodes and resource nodes in the scene interaction graph. The server compares the predicted interaction data with historical interaction data, and converges the initial scene graph network based on the comparison results to obtain the scene graph network. The first similarity refers to the similarity between the predicted object features and the predicted resource features calculated during the training of the scene graph network, used to determine the predicted interaction data between object nodes and resource nodes in the scene interaction graph, assisting in model convergence.

[0252] For example, the server can calculate the difference between predicted interaction data and historical interaction data; based on the difference, determine the loss information of the initial scene graph network; and based on the loss information, train the initial scene graph network to obtain the scene graph network.

[0253] 308. When a resource push request is received for a target object, the server pushes at least one push resource to the target object based on the scene graph network.

[0254] For example, the server can detect the target information flow scene in which the push object is currently located, and filter out the target scene graph network corresponding to the target information flow scene from the scene graph network; based on the target scene graph network, calculate at least one push resource associated with the push object, and push the push resource to the push object.

[0255] For example, the server can determine the characteristics of the pushed object corresponding to the pushed object; determine the scene resource characteristics corresponding to each resource node in the target information flow scene according to the target scene graph network; and calculate at least one push resource associated with the pushed object based on the pushed object characteristics and scene resource characteristics.

[0256] For example, the server can obtain the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information flow scene; if the object description information is the same as the target description information, the object node corresponding to the pushed object is taken as the pushed object node; based on the target scene graph network, the scene object features corresponding to the pushed object node are determined; and the scene object features are taken as the pushed object features.

[0257] For example, the server can obtain the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information stream scene; if the object description information is different from the target description information, the server can obtain the global object description information corresponding to each global object node in the global scene interaction graph; the server can filter out the global object nodes corresponding to the global object description information that is the same as the target description information from the global scene interaction graph to obtain the target global object node; based on the scene graph network, the server can determine the scene object features corresponding to the target global object node to obtain the pushed object features of the pushed object.

[0258] For example, the server can calculate a second similarity between the characteristics of the pushed object and the characteristics of the scene resources;

[0259] The second similarity is compared with a preset similarity threshold, and the push resource nodes are selected from the resource nodes of the target information flow scene based on the comparison results; the resources corresponding to the push resource nodes are used as push resources.

[0260] As described above, the server obtains scene interaction graphs corresponding to multiple information flow scenarios. These scene interaction graphs indicate the historical interaction data between multiple object nodes and multiple resource nodes within the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario. Then, the server obtains object description information corresponding to the object nodes and resource description information corresponding to the resource nodes in the scene interaction graphs. Next, based on the object description information, the server constructs at least one virtual object connection node in the scene interaction graph. This virtual object connection node connects object nodes that have an association with it. Furthermore, based on the resource description information, the server... At least one virtual resource connection node is constructed, which is used to connect resource nodes that are associated with the virtual object connection node. Then, the server connects the scene interaction graph based on the virtual object connection node and the virtual resource connection node to obtain a global scene interaction graph. The server constructs global object features for each object node and global resource features for each resource node in the global scene interaction graph. After that, the server trains the initial scene graph network corresponding to the scene interaction graph based on the global object features and global resource features to obtain the scene graph network. Finally, when a resource push request is received for a push object, the server pushes at least one push resource to the push object based on the scene graph network. This solution integrates the scene interaction graphs corresponding to multiple information flow scenarios to construct a global scene interaction graph that fully reflects the commonalities and differences between various information flow scenarios. This eliminates pattern bias and data bias in single-scenario modeling, more accurately mines user information flow preferences, and reduces operation and maintenance costs. Furthermore, for the global object features corresponding to each object node and the global resource features corresponding to each resource node in the global scene interaction graph, the global object features and global resource features are further used as the initial features of object nodes and resource nodes in a single scene interaction graph. This is then used to continue training the initial scene graph corresponding to the single scene interaction graph, resulting in a scene graph network that more accurately reflects the feature representation of object nodes and resource nodes in a single scene interaction graph, thereby improving the accuracy of resource push.

[0261] To better implement the above model training methods, this application also provides a model training device that can be integrated into a network device, such as a server or terminal. The terminal may include a tablet computer, a laptop computer, and / or a personal computer.

[0262] Figure 15 shows a schematic diagram of the model training device provided in an embodiment of this application. As shown in Figure 15, the model training device may include an acquisition unit 401, a connection unit 402, a construction unit 403, and a training unit 404, as follows:

[0263] (1) Obtain unit 401;

[0264] The acquisition unit 401 is used to acquire scene interaction diagrams corresponding to multiple information flow scenarios. The scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario.

[0265] For example, the acquisition unit 401 can be used to acquire the object nodes, resource nodes, and historical interaction data between object nodes and resource nodes corresponding to multiple information flow scenarios; and construct a bipartite graph based on the object nodes, resource nodes, and historical interaction data to obtain the scenario interaction graph.

[0266] (2) Connecting unit 402;

[0267] The connection unit 402 is used to connect the scene interaction graph based on object nodes and resource nodes to obtain a global scene interaction graph.

[0268] For example, connection unit 402 can be specifically used to obtain object description information corresponding to object nodes and resource description information corresponding to resource nodes in the scene interaction graph; based on the object description information, construct at least one virtual object connection node in the scene interaction graph, the virtual object connection node being used to connect object nodes that have an association relationship with the virtual object connection node; based on the resource description information, construct at least one virtual resource connection node in the scene interaction graph, the virtual resource connection node being used to connect resource nodes that have an association relationship with the virtual object connection node; and connect the scene interaction graph based on the virtual object connection node and the virtual resource connection node to obtain a global scene interaction graph.

[0269] For example, the connection unit 402 can be used to filter at least one pushed object description tag from the object description tags, and add an object tag value range to the pushed object description tag; based on the pushed object description tag and the object tag value range, at least one virtual object connection node is constructed in the scene interaction graph, and the virtual object description information corresponding to the virtual object connection node includes the pushed object description tag and the object tag value range.

[0270] For example, the connection unit 402 can be used to filter at least one push resource description tag from the resource description tags and add a resource tag value range to the push resource description tag; based on the push resource description tag and the resource tag value range, at least one virtual resource connection node is constructed in the scene interaction graph, and the virtual resource description information corresponding to the virtual resource connection node includes the push resource description tag and the resource tag value range.

[0271] For example, the connection unit 402 can be used to obtain the virtual object description information corresponding to the virtual object connection node and the virtual resource description information corresponding to the virtual resource connection node in each scene interaction graph; connect the virtual object connection nodes with the same virtual object description information and connect the virtual resource connection nodes with the same virtual resource description information to obtain the global scene interaction graph.

[0272] (3) Constructing unit 403;

[0273] Construction unit 403 is used to construct the global object features of each object node and the global resource features of each resource node in the global scene interaction graph.

[0274] For example, the construction unit 403 can be used to perform data augmentation on the global scene interaction graph to obtain a first augmented interaction graph and a second augmented interaction graph; based on the first augmented interaction graph and the second augmented interaction graph, positive sample pairs and negative sample pairs are determined; based on the positive sample pairs and negative sample pairs, the global object features of each object node and the global resource features of each resource node are determined.

[0275] For example, the construction unit 403 can be used to extract at least one global object node as an enhanced object node from multiple global object nodes, and at least one global resource node as an enhanced resource node from multiple global resource nodes; to filter historical interaction data between enhanced object nodes and enhanced resource nodes from historical interaction data to obtain enhanced node interaction data; and to generate a first enhanced interaction graph and a second enhanced interaction graph based on enhanced object nodes, enhanced resource nodes and enhanced node interaction data.

[0276] For example, the construction unit 403 can be specifically used to randomly extract at least one enhanced object node from the enhanced object nodes as a first enhanced object node, and randomly extract at least one enhanced object node from the enhanced object nodes as a second enhanced object node, with the first enhanced object node and the second enhanced object node partially overlapping; randomly extract at least one enhanced resource node from the enhanced resource nodes as a first enhanced resource node, and randomly extract at least one enhanced resource node from the enhanced resource nodes as a second enhanced resource node, with the first enhanced resource node and the second enhanced resource node partially overlapping; filter out the first enhanced node interaction data between the first enhanced object node and the first enhanced resource node from the enhanced node interaction data, and filter out the second enhanced node interaction data between the second enhanced object node and the second enhanced resource node from the enhanced node interaction data; generate a first enhanced interaction graph based on the first enhanced object node, the first enhanced resource node, and the first enhanced node interaction data, and generate a second enhanced interaction graph based on the second enhanced object node, the second enhanced resource node, and the second enhanced node interaction data.

[0277] For example, the construction unit 403 can be specifically used to obtain the initial first graph network corresponding to the first enhanced interaction graph and the initial second graph network corresponding to the second enhanced interaction graph; use the initial first graph network to predict the predicted first features of each enhanced node in the first enhanced interaction graph, the predicted first features including the predicted first object features of the first enhanced object node and the predicted first resource features of the first enhanced resource node; use the initial second graph network to predict the predicted second features of each enhanced node in the second enhanced interaction graph, the predicted second features including the predicted second object features of the second enhanced object node and the predicted second resource features of the second enhanced resource node; based on positive sample pairs, negative sample pairs, predicted first features, and predicted second features, determine the first contrast loss corresponding to the object node and the second contrast loss corresponding to the resource node; converge the initial first graph network and the initial second graph network according to the first contrast loss and the second contrast loss to obtain the first graph network and the second graph network; and determine the global object features of each object node and the global resource features of each resource node according to the first graph network and the second graph network.

[0278] (4) Training Unit 404;

[0279] Training unit 404 is used to train the initial scene graph network corresponding to the scene interaction graph based on global object features and global resource features to obtain the scene graph network. The scene graph network is used to determine the push resources for the push object so as to push the push resources to the push object.

[0280] For example, training unit 404 can be used to filter out global object features corresponding to object nodes in each scene interaction graph from global object features to obtain scene global object features, and to filter out global resource features corresponding to resource nodes in each scene interaction graph from global resource features to obtain scene global resource features; based on the initial scene graph network corresponding to the scene interaction graph, calculate the predicted object features corresponding to the scene global object features, and based on the initial scene graph network, calculate the predicted resource features corresponding to the scene global resource features; calculate the first similarity between the predicted object features and the predicted resource features to determine the predicted interaction data between object nodes and resource nodes in the scene interaction graph; compare the predicted interaction data with historical interaction data, and converge the initial scene graph network based on the comparison results to obtain the scene graph network.

[0281] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0282] As can be seen from the above, in this embodiment of the application, the acquisition unit 401 acquires scene interaction graphs corresponding to multiple information flow scenarios. The scene interaction graphs are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario. Then, the connection unit 402 connects the scene interaction graphs based on the object nodes and resource nodes to obtain a global scene interaction graph. After that, the construction unit 403 constructs global object features of each object node and global resource features of each resource node in the global scene interaction graph. The training unit 404 then trains the initial scene graph network corresponding to the scene interaction graph based on the global object features and global resource features to obtain the scene graph network. This solution integrates the scene interaction graphs corresponding to multiple information flow scenarios to construct a global scene interaction graph that fully reflects the commonalities and differences between various information flow scenarios. This eliminates pattern bias and data bias in single-scenario modeling, more accurately mines user information flow preferences, and reduces operation and maintenance costs. Furthermore, for the global object features corresponding to each object node and the global resource features corresponding to each resource node in the global scene interaction graph, the global object features and global resource features are further used as the initial features of object nodes and resource nodes in a single scene interaction graph. This is then used to continue training the initial scene graph corresponding to the single scene interaction graph, resulting in a scene graph network that more accurately reflects the feature representation of object nodes and resource nodes in a single scene interaction graph. This improves the accuracy of subsequent resource push using this scene graph network.

[0283] To better implement the above resource push method, this application embodiment also provides a resource push device, which can be integrated into network devices, such as servers or terminals. The terminal may include tablet computers, laptop computers, and / or personal computers.

[0284] Figure 16 shows a schematic diagram of the resource push device provided in an embodiment of this application. As shown in Figure 16, the resource push device may include a receiving unit 501, a filtering unit 502, and a push unit 503, as follows:

[0285] (1) Receiving unit 501,

[0286] The receiving unit 501 is used to detect the target information flow scenario in which the push object is currently located when it receives a resource push request for the push object.

[0287] (2) Filtering unit 502,

[0288] The filtering unit 502 is used to filter out the target scene graph network corresponding to the target information flow scene from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. It is a graph network trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes.

[0289] (3) Push unit 503;

[0290] The push unit 503 is used to determine at least one push resource associated with the push object based on the target scene graph network, and push the push resource to the push object.

[0291] For example, push unit 503 can be used to determine the characteristics of the push object corresponding to the push object; determine the scene resource characteristics corresponding to each resource node in the target information flow scene according to the target scene graph network; and calculate at least one push resource associated with the push object based on the push object characteristics and scene resource characteristics.

[0292] For example, push unit 503 can be used to obtain the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information flow scene; if the object description information is the same as the target description information, the object node corresponding to the pushed object is taken as the pushed object node; the scene object features corresponding to the pushed object node are determined according to the target scene graph network; and the scene object features are taken as the pushed object features.

[0293] For example, push unit 503 can be used to obtain the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information flow scene; if the object description information is different from the target description information, obtain the global object description information corresponding to each global object node in the global scene interaction graph; filter out the global object nodes corresponding to the global object description information that is the same as the target description information from the global scene interaction graph to obtain the target global object node; and determine the scene object features corresponding to the target global object node based on the scene graph network to obtain the pushed object features of the pushed object.

[0294] For example, push unit 503 can be used to calculate the second similarity between the features of the push object and the features of the scene resources; compare the second similarity with a preset similarity threshold, and select push resource nodes from the resource nodes of the target information flow scene based on the comparison results; and use the resources corresponding to the push resource nodes as push resources.

[0295] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0296] This application also provides an electronic device, as shown in FIG17, which illustrates a schematic diagram of the structure of the electronic device involved in this application embodiment. Specifically:

[0297] The electronic device may include components such as a processor 601 with one or more processing cores, a memory 602 with one or more computer-readable storage media, a power supply 603, and an input unit 604. Those skilled in the art will understand that the electronic device structure shown in FIG17 does not constitute a limitation on the electronic device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0298] The processor 601 is the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes various functions and processes data by running or executing software programs and / or modules stored in the memory 602, and by calling data stored in the memory 602. Optionally, the processor 601 may include one or more processing cores; preferably, the processor 601 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 601.

[0299] The memory 602 can be used to store software programs and modules. The processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 602 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 602 may also include a memory controller to provide the processor 601 with access to the memory 602.

[0300] The electronic device also includes a power supply 603 that supplies power to the various components. Preferably, the power supply 603 can be logically connected to the processor 601 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 603 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0301] The electronic device may also include an input unit 604, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0302] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 601 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 602 according to the following instructions, and the processor 601 runs the applications stored in the memory 602 to realize various functions, as follows:

[0303] Obtain scene interaction graphs corresponding to multiple information flow scenarios. These graphs indicate historical interaction data between multiple object nodes and multiple resource nodes within the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects within the information flow scenario. Based on object nodes and resource nodes, connect the scene interaction graphs to obtain a global scene interaction graph. Construct global object features for each object node and global resource features for each resource node within the global scene interaction graph. Train the initial scene graph network corresponding to the scene interaction graph based on the global object features and global resource features to obtain a scene graph network. This scene graph network is used to determine push resources for the target object, so that the push resources can be pushed to the target object.

[0304] And / or,

[0305] When a resource push request is received for a target object, the target information flow scene currently in which the target object is located is detected; the target scene graph network corresponding to the target information flow scene is selected from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes; based on the target scene graph network, at least one push resource associated with the target object is determined, and the push resource is pushed to the target object.

[0306] Specifically, the electronic device can acquire scene interaction diagrams corresponding to multiple information flow scenarios. These scene interaction diagrams indicate historical interaction data between multiple object nodes and multiple resource nodes within the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects within the information flow scenario. Then, it acquires object description information corresponding to the object nodes and resource description information corresponding to the resource nodes in the scene interaction diagrams. Next, based on the object description information, it constructs at least one virtual object connection node in the scene interaction diagram. This virtual object connection node connects object nodes that have an association with it. Furthermore, based on the resource description information, it constructs... At least one virtual resource connection node is established, which is used to connect resource nodes that are associated with the virtual object connection node. Then, based on the virtual object connection node and the virtual resource connection node, the scene interaction graph is connected to obtain a global scene interaction graph. In the global scene interaction graph, global object features of each object node and global resource features of each resource node are constructed. Then, based on the global object features and global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain the scene graph network. Finally, when a resource push request is received for the push object, at least one push resource is pushed to the push object based on the scene graph network, and so on.

[0307] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0308] As can be seen from the above, the embodiments of this application obtain scene interaction graphs corresponding to multiple information flow scenarios. The scene interaction graphs are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects in the information flow scenario. Then, based on object nodes and resource nodes, the scene interaction graphs are connected to obtain a global scene interaction graph. After that, global object features of each object node and global resource features of each resource node are constructed in the global scene interaction graph. Then, based on the global object features and global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain a scene graph network. When a resource push request for a push object is received, at least one push resource is pushed to the push object based on the scene graph network. This solution integrates the scene interaction graphs corresponding to multiple information flow scenarios to construct a global scene interaction graph that fully reflects the commonalities and differences between various information flow scenarios. This eliminates pattern bias and data bias in single-scenario modeling, more accurately mines user information flow preferences, and reduces operation and maintenance costs. Furthermore, for the global object features corresponding to each object node and the global resource features corresponding to each resource node in the global scene interaction graph, the global object features and global resource features are further used as the initial features of object nodes and resource nodes in a single scene interaction graph. This is then used to continue training the initial scene graph corresponding to the single scene interaction graph, resulting in a scene graph network that more accurately reflects the feature representation of object nodes and resource nodes in a single scene interaction graph, thereby improving the accuracy of resource push.

[0309] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0310] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the model training methods and resource delivery methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0311] Obtain scene interaction graphs corresponding to multiple information flow scenarios. These graphs indicate historical interaction data between multiple object nodes and multiple resource nodes within the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects within the information flow scenario. Based on object nodes and resource nodes, connect the scene interaction graphs to obtain a global scene interaction graph. Construct global object features for each object node and global resource features for each resource node within the global scene interaction graph. Train the initial scene graph network corresponding to the scene interaction graph based on the global object features and global resource features to obtain a scene graph network. This scene graph network is used to determine push resources for the target object, so that the push resources can be pushed to the target object.

[0312] And / or,

[0313] When a resource push request is received for a target object, the target information flow scene currently in which the target object is located is detected; the target scene graph network corresponding to the target information flow scene is selected from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes; based on the target scene graph network, at least one push resource associated with the target object is determined, and the push resource is pushed to the target object.

[0314] Specifically, the process involves obtaining scene interaction diagrams corresponding to multiple information flow scenarios. These diagrams indicate historical interaction data between multiple object nodes and multiple resource nodes within the information flow scenario. Object nodes represent objects, and resource nodes represent interactive resources provided to objects within the information flow scenario. Then, object description information corresponding to the object nodes and resource description information corresponding to the resource nodes in the scene interaction diagrams are obtained. Next, based on the object description information, at least one virtual object connection node is constructed in the scene interaction diagram. This virtual object connection node connects object nodes that have an association with it. Finally, based on the resource description information, at least one virtual object connection node is constructed in the scene interaction diagram. A virtual resource connection node is used to connect resource nodes that are associated with the virtual object connection node. Then, based on the virtual object connection node and the virtual resource connection node, the scene interaction graph is connected to obtain a global scene interaction graph. Global object features for each object node and global resource features for each resource node are constructed in the global scene interaction graph. Then, based on the global object features and global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain the scene graph network. Finally, when a resource push request is received for the push object, at least one push resource is pushed to the push object based on the scene graph network, and so on.

[0315] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0316] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0317] Since the instructions stored in the computer-readable storage medium can execute the steps of any of the model training methods and resource push methods provided in the embodiments of this application, the beneficial effects that any of the model training methods and resource push methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0318] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the methods provided in the various alternative implementations of the data access aspect described above.

[0319] In summary, this application provides a model training method, resource push method, apparatus, device, computer-readable storage medium, and computer program product. The electronic device acquires scene interaction graphs corresponding to multiple information flow scenarios. These scene interaction graphs indicate historical interaction data between multiple object nodes and multiple resource nodes within the information flow scenario. A global scene interaction graph is obtained by connecting the scene interaction graphs based on object nodes and resource nodes. Global object features for each object node and global resource features for each resource node are constructed within the global scene interaction graph. An initial scene graph network corresponding to the scene interaction graph is trained based on the global object features and global resource features to obtain a scene graph network, which is used to determine and push resources to the target object. This solution integrates scene interaction graphs from multiple information flow scenarios to construct a global scene interaction graph, eliminating pattern and data biases present in single-scenario modeling. This allows for more comprehensive utilization of historical interaction data from multiple scenarios, uncovering more complex high-order information between object nodes and resource nodes, and thus more accurately reflecting the intrinsic connections between object nodes and resource nodes in the global scenario, improving the accuracy of user information flow preference mining. Meanwhile, by training the initial scene graph network with global object features and global resource features, the resulting scene graph network can more accurately represent the features of object nodes and resource nodes in a single scene interaction graph, thereby improving the prediction accuracy of subsequent resource pushes. Furthermore, multi-scene joint modeling reduces the number of push systems that need to be maintained, lowers system complexity, and improves resource utilization.

[0320] Furthermore, when connecting the scene interaction graphs to obtain the global scene interaction graph, the electronic device acquires the object description information corresponding to the object nodes and the resource description information corresponding to the resource nodes in the scene interaction graph. Based on the object description information, it constructs at least one virtual object connection node, and based on the resource description information, it constructs at least one virtual resource connection node. The scene interaction graph is then connected based on these virtual object connection nodes and virtual resource connection nodes. These virtual object connection nodes and virtual resource connection nodes establish association channels between object nodes and resource nodes in different scene interaction graphs, enabling information flow and sharing between different scenes. This helps break down the barriers of single-scene data, more effectively integrates multi-scene data, avoids information loss due to single-scene data bias, and improves the global scene interaction graph's comprehensive representation of multi-scene information, thus providing a better data foundation for subsequently constructing more accurate global object features and global resource features.

[0321] Furthermore, when the object description information includes at least one dimension of object description tags and the tag values ​​of object nodes under the object description tags, the electronic device filters at least one pushed object description tag from the object description tags, adds the object tag value range, and constructs virtual object connection nodes based on this. This method enables targeted node construction based on the specific characteristics of objects, allowing virtual object connection nodes to more accurately connect object nodes with similar characteristics. By refining the connection relationships between object nodes, the associations between object nodes can be captured more clearly, improving the connection tightness and information transmission efficiency of object nodes in the global scene interaction graph. This helps to more accurately mine the global characteristics of object nodes, thereby improving the accuracy of resource push.

[0322] Furthermore, if the resource description information includes at least one dimension of resource description tags and the tag values ​​of resource nodes under those tags, the electronic device filters at least one push resource description tag from the resource description tags, adds a range of resource tag values, and constructs a virtual resource connection node based on this. This associates the construction of virtual resource connection nodes with the specific characteristics of the resources, enabling more effective connection of resource nodes with similar characteristics. By strengthening the connections between resource nodes, resource information can be better integrated, improving the information flow of resource nodes in the global scene interaction graph, helping to uncover potential relationships between resource nodes, thereby more accurately determining the global characteristics of resource nodes and providing more precise resource information for resource push.

[0323] Furthermore, when obtaining the global scene interaction graph by connecting scene interaction graphs based on virtual object connection nodes and virtual resource connection nodes, the electronic device acquires the virtual object description information corresponding to the virtual object connection nodes and the virtual resource description information corresponding to the virtual resource connection nodes in each scene interaction graph. Virtual object connection nodes with the same virtual object description information are connected, and virtual resource connection nodes with the same virtual resource description information are connected. This description-based connection method ensures that nodes with the same characteristics in different scene interaction graphs can be accurately connected, enabling the global scene interaction graph to more accurately reflect the commonalities and differences between multiple scenes. Through this precise connection, multi-scene data can be integrated more effectively, avoiding errors and redundancy in the connection process, improving the accuracy and effectiveness of the global scene interaction graph, and providing a more reliable data foundation for subsequent feature construction and model training.

[0324] Furthermore, after acquiring the object description information corresponding to the object nodes and the resource description information corresponding to the resource nodes in the scene interaction graph, the electronic device acquires the number of object nodes corresponding to the object nodes and the number of resource nodes corresponding to the resource nodes in the scene interaction graph. Based on the number of object nodes, it determines the number of virtual object nodes to be constructed for virtual object connection nodes in the scene interaction graph, and based on the number of resource nodes, it determines the number of virtual resource nodes to be constructed for virtual resource connection nodes in the scene interaction graph. Virtual object connection nodes are constructed based on the number of virtual object nodes and the object description information, and virtual resource connection nodes are constructed based on the number of virtual resource nodes and the resource description information. This method of dynamically adjusting the number of virtual connection nodes based on the number of nodes allows for flexible processing according to the actual data situation of different scene interaction graphs. For scene interaction graphs with a small number of object nodes or resource nodes, increasing the number of virtual connection nodes can increase the connection probability between this scene and other scenes, enhance information exchange between different scenes, avoid information isolation due to insufficient data, ensure balanced integration of information from various scenes in the global scene interaction graph, and improve the model's adaptability and processing capabilities for different scene data.

[0325] Furthermore, when constructing global object features for each object node and global resource features for each resource node in the global scene interaction graph, the electronic device performs data augmentation on the global scene interaction graph, resulting in a first augmented interaction graph and a second augmented interaction graph. Positive and negative sample pairs are determined based on these two graphs, and the global object features for each object node and the global resource features for each resource node are determined according to these positive and negative sample pairs. Data augmentation increases the diversity and richness of the training data by generating different augmented interaction graphs, enabling the model to learn more diverse node features and interaction patterns. The use of positive and negative sample pairs provides the model with clear learning objectives, prompting it to learn the similar features of the same node in different augmented graphs and the feature differences between different nodes. Through this comparative learning, node features can be captured more accurately, improving the accuracy and representativeness of global object features and global resource features, thereby enhancing the model's understanding and prediction capabilities regarding user preferences.

[0326] Furthermore, when the global scene interaction graph includes multiple global object nodes, multiple global resource nodes, and historical interaction data between global object nodes and global resource nodes, the electronic device extracts at least one global object node as an enhanced object node from the multiple global object nodes and at least one global resource node as an enhanced resource node from the multiple global resource nodes. It also filters historical interaction data between enhanced object nodes and enhanced resource nodes from the historical interaction data to obtain enhanced node interaction data. Based on the enhanced object nodes, enhanced resource nodes, and enhanced node interaction data, a first enhanced interaction graph and a second enhanced interaction graph are generated. This targeted approach of selecting specific nodes and interaction data for data enhancement focuses on key information, reduces interference from irrelevant data, and improves the efficiency and effectiveness of data enhancement. By generating more representative enhanced interaction graphs, the node characteristics and interaction relationships in the global scene interaction graph can be more accurately reflected, providing better data support for determining global object characteristics and global resource characteristics.

[0327] Furthermore, when generating the first and second enhanced interaction graphs based on enhanced object nodes, enhanced resource nodes, and enhanced node interaction data, the electronic device randomly extracts at least one enhanced object node from the enhanced object nodes as the first enhanced object node, and at least one enhanced object node from the enhanced object nodes as the second enhanced object node, with partial overlap between the first and second enhanced object nodes. It also randomly extracts at least one enhanced resource node from the enhanced resource nodes as the first enhanced resource node, and at least one enhanced resource node from the enhanced resource nodes as the second enhanced resource node, with partial overlap between the first and second enhanced resource nodes. First enhanced node interaction data between the first enhanced object node and the first enhanced resource node is selected from the enhanced node interaction data, and second enhanced node interaction data between the second enhanced object node and the second enhanced resource node is selected from the enhanced node interaction data. The first enhanced interaction graph is generated based on the first enhanced object node, the first enhanced resource node, and the first enhanced node interaction data, and the second enhanced interaction graph is generated based on the second enhanced object node, the second enhanced resource node, and the second enhanced node interaction data. The random node extraction and partial overlap method increases the randomness and diversity of the enhanced interaction graphs, enabling the model to learn a wider range of node features and interaction patterns. This diversity helps avoid model overfitting, improves the model's generalization ability, and enables the model to more accurately predict node features when faced with different input data, thereby improving the quality of global object features and global resource features.

[0328] Furthermore, when determining the global object features of each object node and the global resource features of each resource node based on positive and negative sample pairs, the electronic device acquires the initial first graph network corresponding to the first enhanced interaction graph and the initial second graph network corresponding to the second enhanced interaction graph. The initial first graph network is used to predict the predicted first features of each enhanced node in the first enhanced interaction graph, and the initial second graph network is used to predict the predicted second features of each enhanced node in the second enhanced interaction graph. Based on the positive and negative sample pairs, the predicted first features, and the predicted second features, the first contrastive loss corresponding to the object node and the second contrastive loss corresponding to the resource node are determined. The initial first and second graph networks are converged based on the first and second contrastive losses to obtain the first and second graph networks. Based on the first and second graph networks, the global object features of each object node and the global resource features of each resource node are determined. Through the calculation of the contrastive loss and the convergence of the graph networks, the model can continuously adjust its parameters, making the feature representations of the same node more similar in different enhanced graphs and the feature differences between different nodes more obvious. This approach can learn node features more accurately, improve the accuracy and stability of global object features and global resource features, thereby enhancing the model's ability to capture user preferences and the accuracy of resource delivery.

[0329] Furthermore, when training the initial scene graph network corresponding to each scene interaction graph based on global object features and global resource features to obtain the scene graph network, the electronic device filters out the global object features corresponding to the object nodes of each scene interaction graph from the global object features to obtain the scene global object features, and filters out the global resource features corresponding to the resource nodes of each scene interaction graph from the global resource features to obtain the scene global resource features. Based on the initial scene graph network corresponding to the scene interaction graph, the predicted object features corresponding to the scene global object features are calculated, and based on the initial scene graph network, the predicted resource features corresponding to the scene global resource features are calculated. The first similarity between the predicted object features and the predicted resource features is calculated to determine the predicted interaction data between object nodes and resource nodes in the scene interaction graph. The predicted interaction data is compared with historical interaction data, and the initial scene graph network is converged based on the comparison results to obtain the scene graph network. This training method, which combines global features and historical scene interaction data, can make full use of global information to fine-tune a single scene. By comparing predicted interaction data with historical interaction data, biases in model predictions can be identified, allowing for targeted adjustments to the initial scene graph network. This enables the scene graph network to more accurately predict the interaction preferences between object nodes and resource nodes, improving its adaptability and accuracy to individual scenes and ultimately enhancing the effectiveness of resource delivery.

[0330] Furthermore, during the resource push process, when a resource push request is received for a target object, the electronic device detects the target information flow scene in which the target object is currently located, filters out the target scene graph network corresponding to the target information flow scene from the scene graph network, determines at least one push resource associated with the target object based on the target scene graph network, and pushes the push resource to the target object. This method of pushing resources based on the specific scene in which the target object is located takes into account the characteristics and needs of different scenes, improving the targeting of resource push. By using the target scene graph network for resource filtering, the needs of the target object can be matched more accurately, avoiding inaccurate pushes caused by using a general model in different scenes, improving the accuracy and effectiveness of resource push, and enhancing the user experience.

[0331] Furthermore, when determining at least one push resource associated with a target object based on the target scene graph network, the electronic device determines the target object characteristics corresponding to the target object. Based on the target scene graph network, it determines the scene resource characteristics corresponding to each resource node in the target information flow scene. Based on the target object characteristics and scene resource characteristics, it calculates at least one push resource associated with the target object. Accurately determining the target object characteristics and scene resource characteristics provides a more precise basis for resource calculation. By calculating the association between the two, resources matching the target object's needs can be more accurately selected, improving the accuracy and relevance of resource push notifications and making the pushed resources more aligned with the user's actual needs.

[0332] Furthermore, when determining the features of the pushed object, the electronic device acquires the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information stream scene. If the object description information is the same as the target description information, the object node corresponding to the pushed object is taken as the pushed object node. Based on the target scene graph network, the scene object features corresponding to the pushed object node are determined, and these scene object features are used as the pushed object features. This method of directly using the target scene graph network to obtain the pushed object features can make full use of the historical information and training results of the scene, improving the accuracy and efficiency of feature acquisition. By using scene object features as pushed object features, the characteristics of the pushed object in the current scene can be more accurately reflected, providing a more reliable basis for subsequent resource pushes.

[0333] Furthermore, after obtaining the target description information of the pushed object and the object description information corresponding to the object nodes contained in the target information flow scene, if the object description information differs from the target description information, the electronic device obtains the global object description information corresponding to each global object node in the global scene interaction graph. It then filters out the global object nodes corresponding to the global object description information that is identical to the target description information from the global scene interaction graph, obtaining the target global object node. Based on the scene graph network, it determines the scene object features corresponding to the target global object node, thus obtaining the pushed object features of the pushed object. This approach solves the problem of the pushed object not being in the target information flow scene. By expanding the search scope to the global scene interaction graph, it can accurately obtain the features of the pushed object in different scenarios. This enables the resource push system to adapt to new scenarios, achieve cold start on both the object and resource sides, improve the system's flexibility and applicability, and expand the coverage of resource push.

[0334] Furthermore, when calculating at least one push resource associated with the push object based on the characteristics of the push object and the characteristics of the scene resources, the electronic device calculates a second similarity between the characteristics of the push object and the characteristics of the scene resources, compares the second similarity with a preset similarity threshold, and filters push resource nodes from the resource nodes of the target information flow scene based on the comparison result, using the resources corresponding to the push resource nodes as push resources. By calculating similarity and comparing it with a threshold, resource nodes that match the characteristics of the push object can be objectively filtered out. This filtering method can improve the accuracy and relevance of resource push, avoid pushing irrelevant resources, improve resource utilization, and make the pushed resources more in line with the needs of the push object.

[0335] The foregoing has provided a detailed description of a model training method, resource push method, device, medium, and computer program product provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0336] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A model training method, performed by an electronic device, comprising: Obtain scene interaction diagrams corresponding to multiple information flow scenarios. The scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. The object node represents an object, and the resource node represents an interactive resource provided to the object in the information flow scenario; Based on the object nodes and the resource nodes, the scene interaction graph is connected to obtain a global scene interaction graph; In the global scene interaction graph, global object features of each object node and global resource features of each resource node are constructed. and Based on the global object features and the global resource features, the initial scene graph network corresponding to the scene interaction graph is trained to obtain the scene graph network. The scene graph network is used to determine the push resources for the push object, so as to push the push resources to the push object.

2. The model training method as described in claim 1, wherein connecting the scene interaction graph based on the object nodes and the resource nodes to obtain a global scene interaction graph includes: Obtain object description information corresponding to the object nodes and resource description information corresponding to the resource nodes in the scene interaction graph; Based on the object description information, at least one virtual object connection node is constructed in the scene interaction graph. The virtual object connection node is used to connect object nodes that have an association relationship with the virtual object connection node. Based on the resource description information, at least one virtual resource connection node is constructed in the scene interaction graph. The virtual resource connection node is used to connect resource nodes that are associated with the virtual object connection node. Based on the virtual object connection node and the virtual resource connection node, the scene interaction graph is connected to obtain the global scene interaction graph.

3. The model training method as described in claim 2, wherein the object description information includes at least one dimension of object description labels, and the label values ​​of the object nodes under the object description labels. The step of constructing at least one virtual object connection node in the scene interaction graph based on the object description information includes: Filter at least one pushed object description tag from the object description tags, and add an object tag value range to the pushed object description tag; Based on the description tag of the pushed object and the value range of the object tag, at least one virtual object connection node is constructed in the scene interaction graph. The virtual object description information corresponding to the virtual object connection node includes the description tag of the pushed object and the value range of the object tag.

4. The model training method as described in claim 2 or 3, wherein the resource description information includes at least one dimension of resource description labels, and the label value of the resource node under the resource description labels. The step of constructing at least one virtual resource connection node in the scene interaction graph based on the resource description information includes: Filter at least one push resource description tag from the resource description tags, and add a resource tag value range to the push resource description tag; Based on the push resource description tag and the value range of the resource tag, at least one virtual resource connection node is constructed in the scene interaction graph. The virtual resource description information corresponding to the virtual resource connection node includes the push resource description tag and the value range of the resource tag.

5. The model training method according to any one of claims 2 to 4, wherein connecting the scene interaction graph based on the virtual object connection node and the virtual resource connection node to obtain the global scene interaction graph includes: Obtain the virtual object description information corresponding to the virtual object connection node and the virtual resource description information corresponding to the virtual resource connection node in each scene interaction graph; Connect virtual object nodes with the same virtual object description information, and connect virtual resource nodes with the same virtual resource description information to obtain the global scene interaction graph.

6. The model training method according to any one of claims 2 to 5, further comprising, after obtaining the object description information corresponding to the object node and the resource description information corresponding to the resource node in the scene interaction graph: Obtain the number of object nodes corresponding to object nodes and the number of resource nodes corresponding to resource nodes in the scene interaction graph; The number of virtual object nodes corresponding to the virtual object connection nodes in the scene interaction graph is determined based on the number of object nodes, and the number of virtual resource nodes corresponding to the virtual resource connection nodes in the scene interaction graph is determined based on the number of resource nodes. Based on the number of virtual object nodes constructed and the object description information, virtual object connection nodes are constructed in the scene interaction graph, and based on the number of virtual resource nodes constructed and the resource description information, virtual resource connection nodes are constructed in the scene interaction graph.

7. The model training method according to any one of claims 1 to 6, wherein constructing the global object features of each object node and the global resource features of each resource node in the global scene interaction graph comprises: Data augmentation is performed on the global scene interaction graph to obtain a first augmented interaction graph and a second augmented interaction graph; Based on the first enhanced interaction graph and the second enhanced interaction graph, positive sample pairs and negative sample pairs are determined. The positive sample pairs include node pairs consisting of the same nodes in the first enhanced interaction graph and the second enhanced interaction graph, and the negative sample pairs include node pairs consisting of different nodes in the first enhanced interaction graph and the second enhanced interaction graph. Based on the positive sample pairs and the negative sample pairs, the global object features of each object node and the global resource features of each resource node are determined.

8. The model training method as described in claim 7, wherein the global scene interaction graph includes multiple global object nodes, multiple global resource nodes, and historical interaction data between the global object nodes and the global resource nodes. The step of performing data augmentation on the global scene interaction graph to obtain a first augmented interaction graph and a second augmented interaction graph includes: At least one global object node is extracted from the plurality of global object nodes as an enhanced object node, and at least one global resource node is extracted from the plurality of global resource nodes as an enhanced resource node; The historical interaction data between the enhanced object node and the enhanced resource node is filtered out from the historical interaction data to obtain the enhanced node interaction data; Based on the enhanced object node, the enhanced resource node, and the enhanced node interaction data, the first enhanced interaction graph and the second enhanced interaction graph are generated.

9. The model training method as described in claim 8, wherein generating the first augmented interaction graph and the second augmented interaction graph based on the augmented object node, the augmented resource node, and the augmented node interaction data comprises: At least one enhanced object node is randomly selected from the enhanced object nodes as a first enhanced object node, and at least one enhanced object node is randomly selected from the enhanced object nodes as a second enhanced object node, wherein the first enhanced object node and the second enhanced object node partially overlap. At least one enhanced resource node is randomly selected from the enhanced resource nodes as a first enhanced resource node, and at least one enhanced resource node is randomly selected from the enhanced resource nodes as a second enhanced resource node, wherein the first enhanced resource node and the second enhanced resource node partially overlap. The first enhanced node interaction data between the first enhanced object node and the first enhanced resource node is filtered out from the enhanced node interaction data, and the second enhanced node interaction data between the second enhanced object node and the second enhanced resource node is filtered out from the enhanced node interaction data. The first enhanced interaction graph is generated based on the first enhanced object node, the first enhanced resource node, and the interaction data of the first enhanced node, and the second enhanced interaction graph is generated based on the interaction data of the second enhanced object node, the second enhanced resource node, and the second enhanced node.

10. The model training method as described in claims 8 to 9, wherein determining the global object feature of each object node and the global resource feature of each resource node based on the positive sample pair and the negative sample pair comprises: Obtain the initial first graph network corresponding to the first enhanced interaction graph and the initial second graph network corresponding to the second enhanced interaction graph; The initial first graph network is used to predict the predicted first features of each augmented node in the first augmented interaction graph. The predicted first features include the predicted first object features of the first augmented object node and the predicted first resource features of the first augmented resource node. The initial second graph network is used to predict the predicted second features of each augmented node in the second augmented interaction graph. The predicted second features include the predicted second object features of the second augmented object node and the predicted second resource features of the second augmented resource node. Based on the positive sample pair, the negative sample pair, the predicted first feature, and the predicted second feature, determine the first contrast loss corresponding to the object node and the second contrast loss corresponding to the resource node; Based on the first contrast loss and the second contrast loss, the initial first graph network and the initial second graph network are converged to obtain the first graph network and the second graph network. Based on the first graph network and the second graph network, the global object characteristics of each object node and the global resource characteristics of each resource node are determined.

11. The model training method according to any one of claims 1 to 10, wherein training an initial scene graph network corresponding to each scene interaction graph based on the global object features and the global resource features to obtain a scene graph network includes: The global object features corresponding to each object node of the scene interaction graph are selected from the global object features to obtain the scene global object features, and the global resource features corresponding to each resource node of the scene interaction graph are selected from the global resource features to obtain the scene global resource features. Based on the initial scene graph network corresponding to the scene interaction graph, the predicted object features corresponding to the global object features of the scene are calculated, and based on the initial scene graph network, the predicted resource features corresponding to the global resource features of the scene are calculated. Calculate the first similarity between the predicted object features and the predicted resource features to determine the predicted interaction data between object nodes and resource nodes in the scene interaction graph; The predicted interaction data and the historical interaction data are compared, and the initial scene graph network is converged based on the comparison results to obtain the scene graph network.

12. A resource push method, executed by an electronic device, comprising: When a resource push request is received for a target object, the target information stream scenario in which the target object is currently located is detected. The target scene graph network corresponding to the target information flow scene is selected from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes. and Based on the target scene graph network, at least one push resource associated with the push recipient is determined, and the push resource is pushed to the push recipient.

13. The resource push method as described in claim 12, wherein determining at least one push resource associated with the push recipient based on the target scene graph network includes: Determine the characteristics of the push object corresponding to the push object; Based on the target scene graph network, determine the scene resource characteristics corresponding to each resource node in the target information flow scene; Based on the characteristics of the target object and the characteristics of the scene resources, at least one push resource associated with the target object is calculated.

14. The resource push method as described in claim 13, wherein determining the characteristics of the push recipient corresponding to the push recipient includes: Obtain the target description information of the pushed object, and the object description information corresponding to the object nodes contained in the target information stream scenario; If the object description information is the same as the target description information, the object node corresponding to the pushed object is taken as the pushed object node. Based on the target scene graph network, determine the scene object features corresponding to the pushed object node; The features of the scene object are used as the features of the object being pushed.

15. The resource push method as described in claim 14, further comprising, after obtaining the target description information of the pushed object and the object description information corresponding to the object nodes included in the target information stream scenario: If the object description information is different from the target description information, obtain the global object description information corresponding to each global object node in the global scene interaction graph; From the global scene interaction graph, global object nodes corresponding to global object description information that are the same as the target description information are selected to obtain the target global object node; Based on the scene graph network, the scene object features corresponding to the target global object node are determined, and the push object features of the push object are obtained.

16. The resource push method according to any one of claims 13 to 15, wherein calculating at least one push resource associated with the push object based on the characteristics of the push object and the scene resource characteristics includes: Calculate the second similarity between the features of the pushed object and the features of the scene resources; The second similarity is compared with a preset similarity threshold, and push resource nodes are selected from the resource nodes of the target information flow scenario based on the comparison results. The resources corresponding to the push resource node are used as the push resources.

17. A model training device, comprising: The acquisition unit is used to acquire scene interaction diagrams corresponding to multiple information flow scenarios, wherein the scene interaction diagrams are used to indicate the historical interaction data between multiple object nodes and multiple resource nodes in the information flow scenario. The object node represents an object, and the resource node represents an interactive resource provided to the object in the information flow scenario; A connection unit is used to connect the scene interaction graph based on the object node and the resource node to obtain a global scene interaction graph; A construction unit is used to construct the global object features of each object node and the global resource features of each resource node in the global scene interaction graph. The training unit is used to train the initial scene graph network corresponding to the scene interaction graph based on the global object features and the global resource features to obtain the scene graph network. The scene graph network is used to determine the push resources for the push object so as to push the push resources to the push object.

18. A resource delivery device, comprising: The receiving unit is used to detect the target information stream scenario in which the push object is currently located when a resource push request for the push object is received. A filtering unit is used to filter out the target scene graph network corresponding to the target information flow scene from the scene graph network. The scene graph network includes global object features and global resource features constructed based on the global scene interaction graph. The graph network is trained on the initial scene graph network corresponding to each scene interaction graph. The global scene interaction graph is obtained by connecting the scene interaction graphs corresponding to multiple information flow scenes. The push unit is used to determine at least one push resource associated with the push object based on the target scene graph network, and push the push resource to the push object.

19. An electronic device comprising a processor and a memory, the memory storing an application program, the processor being configured to run the application program in the memory to perform the steps of the model training method of any one of claims 1-11 and the resource push method of any one of claims 12-16.

20. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the model training method of any one of claims 1-11 and the resource push method of any one of claims 12-16.

21. A computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to perform the steps of the model training method of any one of claims 1-11 and the resource push method of any one of claims 12-16.

Images