Recommended content determination method and device, electronic equipment, computer readable storage medium and computer program product

Abstract

The invention provides a recommendation content determination method and device, electronic equipment, a computer readable storage medium and a computer program product, and the method comprises the steps: obtaining a first object having a direct association relationship with a target object, and a second object having an indirect association relationship with the target object, the first article has a direct association relationship, and the second article has an indirect association relationship; constructing a heterogeneous graph by taking the first object, the second object, the target object, the first article and the second article as nodes and taking the incidence relation between the nodes as edges; for each node in the heterogeneous graph, determining a neighbor node feature of a neighbor node in a hierarchy farthest from the node; and determining object features of the target object based on the neighbor node features of the neighbor nodes of the nodes, and determining recommended contents of the target object based on the object features. In this way, the accuracy and efficiency of determining the recommended content can be improved.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a method, device, electronic device, computer-readable storage medium, and computer program product for determining recommended content. Background Art

[0002] In related technologies, when determining recommended content for a user, the recommended content is mostly determined based on other users or items that are directly related to the user. This solution results in a low accuracy rate when determining the recommended content. Summary of the Invention

[0003] Embodiments of the present application provide a method, apparatus, electronic device, computer-readable storage medium, and computer program product for determining recommended content, which can improve the accuracy and efficiency of determining recommended content.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] This embodiment of the present application provides a method for determining recommended content, including:

[0006] Acquire a first object that is directly associated with a target object, a second object that is indirectly associated with the target object, and acquire a first item that is directly associated with the target object, and a second item that is indirectly associated with the target object;

[0007] Constructing a heterogeneous graph by using the first object, the second object, the target object, the first item, and the second item as nodes and the relationships between the nodes as edges;

[0008] For each node in the heterogeneous graph, determining a neighbor node in a layer farthest from the node, and obtaining neighbor node features of the neighbor node;

[0009] Based on the neighbor node features of the neighbor nodes of each node, the object features of the target object are determined, and based on the object features of the target object, the recommended content of the target object is determined.

[0010] This embodiment of the present application provides a device for determining recommended content, including:

[0011] an acquisition module, configured to acquire a first object directly associated with a target object, a second object indirectly associated with the target object, and a first item directly associated with the target object, and a second item indirectly associated with the target object;

[0012] a construction module, configured to construct a heterogeneous graph using the first object, the second object, the target object, the first item, and the second item as nodes and the relationships between the nodes as edges;

[0013] A first determining module is configured to determine, for each node in the heterogeneous graph, a neighboring node in a layer farthest from the node, and obtain neighboring node features of the neighboring node;

[0014] The second determining module is configured to determine object features of the target object based on neighbor node features of the neighbor nodes of each node, and determine recommended content of the target object based on the object features of the target object.

[0015] In the above scheme, the first determination module is also used to determine at least one level corresponding to the node based on multiple edges; select the level farthest from the node from the at least one level; determine at least one candidate neighbor node in the farthest level, and determine the neighbor node from the at least one candidate neighbor node.

[0016] In the above scheme, the first determination module is also used to determine at least one type of associated nodes corresponding to the node based on multiple edges; wherein, the number of edges connecting associated nodes of the same category to the node is the same, and the number of edges connecting associated nodes of different categories to the node is different; and the level of each type of associated node is determined as at least one level corresponding to the node.

[0017] In the above scheme, there is a one-to-one correspondence between the levels and the categories; the first determination module is further used to, for each level, use the number of edges connecting the associated nodes of the category corresponding to the level and the nodes as the number of edges corresponding to the level; based on the number of edges corresponding to each level, select the level with the largest number of corresponding edges from the at least one level as the level farthest from the node.

[0018] In the above scheme, there are multiple neighbor nodes, and the second determination module is also used to perform the following processing for each node: determine the neighbor node characteristics of multiple neighbor nodes of the node; aggregate the multiple neighbor node characteristics to obtain the node characteristics of the node.

[0019] In the above scheme, the number of levels is N, where N is a positive integer; the second determination module is further used to aggregate the neighbor node features of the neighbor nodes of the i-th level to obtain the neighbor node features of the i-1-th level; traverse the i to obtain the node features of the node; wherein the i is a positive integer less than or equal to N.

[0020] In the above scheme, the second determination module is further used to determine the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first item, and the item characteristics of the second item based on the heterogeneous graph; and select recommended content of the target object from the first object, the second object, the first item, and the second item in combination with the object characteristics of the target object, the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first item, and the item characteristics of the second item.

[0021] In the above scheme, the second determination module is also used to determine the inner product of the first object based on the object characteristics of the target object and the object characteristics of the first object, and to determine the inner product of the second object based on the object characteristics of the target object and the object characteristics of the second object, and to determine the inner product of the first item based on the object characteristics of the target object and the item characteristics of the first item, and to determine the inner product of the second item based on the object characteristics of the target object and the item characteristics of the second item; based on the determined inner products, select the object or item with the largest inner product from the first object, the second object, the first item and the second item as the target content, and determine the target content as the recommended content for the target object.

[0022] In the above scheme, the number of the first items is a first number, the number of the second items is a second number, the number of the first objects is a third number, and the number of the second objects is a fourth number. At least one of the first number, the second number, the third number, and the fourth number is multiple. The device also includes a third determination module, which is used to respectively determine the correlation value between each other content and the target content; wherein the other content is any object or item among the first object, the second object, the first item, and the second item except the target content; based on the size of each of the correlation values, the multiple other contents are sorted, and starting from the maximum correlation value, they are selected in sequence from the multiple other contents until the target number of other contents is selected; the second determination module is also used to determine the target content and the selected target number of other contents as recommended content for the target object.

[0023] In the above scheme, the third determination module is also used to obtain the content characteristics of each of the other contents and the content characteristics of the target content based on the heterogeneous graph; based on the content characteristics of each of the other contents and the content characteristics of the target content, respectively determine the correlation between each of the other contents and the target content; normalize each of the correlations to obtain the correlation value between each of the other contents and the target content.

[0024] In the above scheme, the first determination module is also used to obtain a target heterogeneous graph, which is associated with the heterogeneous graph; based on the target heterogeneous graph, a graph neural network model is trained to obtain a target graph neural network model; object information of the target object, object information of the first object, object information of the second object, item information of the first item, and item information of the second item are obtained; the object information of the target object, object information of the first object, object information of the second object, item information of the first item, and item information of the second item are input into the target graph neural network model to obtain neighbor node features of the neighbor nodes.

[0025] An embodiment of the present application provides an electronic device, including:

[0026] a memory for storing computer-executable instructions;

[0027] The processor is configured to implement the method for determining recommended content provided in the embodiment of the present application when executing the computer-executable instructions stored in the memory.

[0028] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for causing a processor to execute instructions to implement the method for determining recommended content provided in the embodiment of the present application.

[0029] Embodiments of the present application provide a computer program product comprising computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the method for determining recommended content provided in embodiments of the present application.

[0030] The embodiments of the present application have the following beneficial effects:

[0031] The target object is recommended based on the first object and the first item that are directly associated with the target object, and the second object and the second item that are indirectly associated with the target object. Compared with the scheme of recommending the target object based only on the first object and the first item that are directly associated with the target object, the accuracy of the determined recommended content is improved; at the same time, based on the heterogeneous graph, the object characteristics of the target object are determined according to the neighbor node characteristics of the neighbor nodes in the layer farthest from each node. Compared with the scheme in the related art that requires continuous sampling of the nodes corresponding to the target object and the nodes around the node, the possibility of overfitting is reduced and the efficiency of determining the recommended content is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the architecture of a system for determining recommended content provided in an embodiment of the present application;

[0033] Figure 2 is a structural diagram of an electronic device provided in an embodiment of the present application;

[0034] Figure 3 Schematic diagram of a flow chart of a method for determining recommended content provided in an embodiment of the present application;

[0035] Figure 4 is a schematic diagram of a heterogeneous graph provided in an embodiment of the present application;

[0036] Figure 5 1 is a flowchart of a process for determining a neighboring node in a layer farthest from a node, provided by an embodiment of the present application;

[0037] Figure 6 is a schematic diagram of a process for determining at least one layer corresponding to a node provided by an embodiment of the present application;

[0038] Figure 7 1 is a flowchart of a process for obtaining neighbor node features of a neighbor node provided in an embodiment of the present application;

[0039] Figure 8 It is a structural diagram of the graph neural network model provided in the embodiment of the present application;

[0040] Figure 9 It is a schematic diagram of two layers corresponding to nodes provided in an embodiment of the present application;

[0041] Figure 10 This is a flowchart of a process for determining recommended content for a target object provided by an embodiment of the present application;

[0042] Figure 11 This is a flowchart of a process for selecting recommended content for a target object provided by an embodiment of the present application;

[0043] Figure 12 This is a process diagram of the information flow high-order social relationship interest propagation and distribution method based on machine learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0045] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0046] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0048] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0049] 1) In response, it is used to indicate the conditions or states on which the executed operations depend. When the dependent conditions or states are met, one or more operations executed can be real-time or have a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations executed are executed.

[0050] 2) Client, also known as the user end, refers to the program corresponding to the server that provides local services to users. Except for some applications that can only run locally, it is generally installed on the terminal and needs to cooperate with the server to run. That is, there must be corresponding servers and service programs in the network to provide corresponding services. In this way, specific communication connections need to be established between the client and server to ensure the normal operation of the application.

[0051] 3) Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0052] 4) Content: The content recommended by the recommendation system to users may include text, pictures, or short videos. Videos are usually provided by professional content production institutions or organizations or UGC, and are finally provided in the form of short videos or small video content in the form of feeds.

[0053] 5) User Generated Content (UGC) refers to user-generated content.

[0054] 6) Machine Learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance.

[0055] 7) Deep learning. The concept of deep learning originated from the study of artificial neural networks. The multi-layer perceptron with multiple hidden layers is a deep learning structure. Deep learning combines low-level features to form more abstract high-level representation attribute categories or features to discover distributed feature representations of data.

[0056] 8) Higher-order relationships, also known as higher-order interactions, refer to interactions that act on the higher-order structure of the network and involve multiple entities. Compared with binary interactions, higher-order interactions can be used to describe the interactions involving multiple entities in the system.

[0057] 9) Contrastive learning, a discriminative representation learning framework (or method) based on contrastive thinking, is mainly used for unsupervised (self-supervised) representation learning. The specific idea adopted by contrastive learning is to compare the sample with examples that are semantically similar to it (positive examples) and examples that are semantically dissimilar to it (negative examples). It is hoped that by designing the model structure and contrastive loss, the representations corresponding to semantically similar examples will be closer in the representation space, and the representations corresponding to semantically dissimilar examples will be farther apart, so as to achieve a clustering-like effect.

[0058] See also Figure 1 , Figure 1 This is a schematic diagram of the architecture of the recommended content determination system 100 provided in an embodiment of the present application, including a terminal (terminal 400 is shown as an example), wherein the terminal 400 is connected to the server 200 via a network 300, wherein the network 300 may be a wide area network or a local area network, or a combination of the two, and data transmission is achieved using wireless or wired links.

[0059] The server 200 is configured to obtain a first object that is directly associated with a target object, a second object that is indirectly associated with the target object, and a first item that is directly associated with the target object, and a second item that is indirectly associated with the target object; construct a heterogeneous graph using the first object, the second object, the target object, the first item, and the second item as nodes, and the associations between the nodes as edges; for each node in the heterogeneous graph, determine a neighbor node in the farthest layer from the node, and obtain neighbor node features of the neighbor node; determine object features of the target object based on the neighbor node features of the neighbor nodes of each node, and determine recommended content for the target object based on the object features of the target object; and send the recommended content for the target object to the terminal 400.

[0060] The terminal 400 is configured to present the received recommended content of the target object.

[0061] In some embodiments, the server 200 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (CDNs), and big data and artificial intelligence platforms. The terminal 400 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a set-top box, an intelligent voice interaction device, a smart home appliance, a virtual reality device, a vehicle-mounted terminal, an aircraft, a portable music player, a personal digital assistant, a dedicated messaging device, a portable gaming device, an intelligent speaker, and a smart watch, etc., but is not limited thereto. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application.

[0062] Next, an electronic device that implements the method for determining recommended content provided by an embodiment of the present application is described. Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device can be a server or a terminal. Figure 1 Take the server shown in as an example, Figure 2The electronic device shown includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the terminal 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0063] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0064] The user interface 430 includes one or more output devices 431 that enable display of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0065] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0066] The memory 450 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0067] In some embodiments, the memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0068] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0069] A network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include Bluetooth, Wi-Fi, and Universal Serial Bus (USB).

[0070] a presentation module 453 for enabling display of information via one or more output devices 431 (e.g., a display screen, a speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0071] The input processing module 454 is configured to detect one or more user inputs or interactions from one of the one or more input devices 432 and to translate the detected inputs or interactions.

[0072] In some embodiments, the device for determining recommended content provided in the embodiments of the present application may be implemented in software. Figure 2 The device 455 for determining recommended content stored in the memory 450 is shown. This device can be software in the form of a program or plug-in, and includes the following software modules: an acquisition module 4551, a construction module 4552, a first determination module 4553, and a second determination module 4554. These modules are logical and can be arbitrarily combined or further separated according to the functions they implement. The functions of each module will be described below.

[0073] In other embodiments, the recommended content determination device provided in the embodiments of the present application can be implemented in hardware. As an example, the recommended content determination device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the recommended content determination method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0074] In some embodiments, the terminal or server can implement the method for determining the recommended content provided in the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; it can be a local (Native) application (APP, Application), that is, a local client, that is, a program that needs to be installed in the operating system to run, such as an instant messaging APP, a web browser APP; it can also be a small program, that is, a program that can be run only by downloading it into a browser environment; it can also be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of client, module or plug-in.

[0075] Based on the above description of the recommended content determination system and electronic device provided by the embodiment of the present application, the following describes the recommended content determination method provided by the embodiment of the present application. In actual implementation, the recommended content determination method provided by the embodiment of the present application can be implemented by the terminal or the server alone, or by the terminal and the server in collaboration, so that Figure 1 The method for determining the recommended content provided in the embodiment of the present application is described as an example in which the server 200 alone executes the method. Figure 3 , Figure 3 This is a flow chart of the method for determining recommended content provided by the embodiment of the present application. Next, Figure 3 The steps shown are explained.

[0076] In step 101 , the server obtains a first object directly associated with a target object, a second object indirectly associated with the target object, and a first item directly associated with the target object, and a second item indirectly associated with the target object.

[0077] In actual implementation, an application capable of determining recommended content is installed on the terminal, such as a playback client, a shopping client, or a social client. The objects (including the first object, the target object, and the second object) here can be user objects, that is, all users who use a certain application or a certain website. The target object is the user who needs to be recommended items. When the application is a shopping client, the objects (including the first object and the second object) can be users of the shopping client other than the target object, and the items (including the first item and the second item) can be item objects, that is, items purchased, collected, or browsed by the objects (including the target object, the first object, and the second object). The recommended content can be based on the items presented by the shopping client.

[0078] When the application is a playback client, the objects (including the first object and the second object) may be users other than the target object in the playback client, the items may be media information (such as videos, audio, etc.) played, commented on, or liked by the objects (including the target object, the first object, and the second object), and the recommended content may be based on the media information (such as videos, audio, etc.) presented by the playback client;

[0079] When the application is a social client, the objects (including the first object and the second object) may be users other than the target object in the social client, the items may be media information (such as video, text, image, audio, etc.) posted by the objects (including the target object, the first object, and the second object) in the social client, and the recommended content may be based on the user or media information presented by the playback client;

[0080] When the application is a travel client, the objects (including the first object and the second object) can be users in the travel client other than the target object, the items can be scenic spots and locations visited, liked, and collected by the objects (including the target object, the first object, and the second object) in the travel client, and the recommended content can be based on the scenic spots and locations recommended by the travel client.

[0081] When the user, ie, the target object, opens an application on the terminal and the terminal runs the application, the terminal sends a recommended content acquisition request carrying the object identifier of the target object to the server.

[0082] The server receives a recommended content acquisition request sent by the terminal and carrying the object identifier of the target object, analyzes the recommended content acquisition request, obtains the object identifier of the target object, and then determines the object information of the target object based on the object identifier of the target object, where the object information will be explained below, thereby obtaining a first object that is directly associated with the target object and a second object that is indirectly associated with the target object based on the object information of the target object, and obtaining a first item that is directly associated with the target object and a second item that is indirectly associated with the target object.

[0083] It should be noted that a direct association relationship is used to indicate a direct association with a target object, that is, a direct association relationship or a first object and a first item that are directly associated with the target object. The direct association relationship is determined based on an operation of the target object. For example, the first object that is directly associated with the target object may be a user that the target object follows, comments on, or likes; and the first item that is directly associated with the target object may be an item that the target object clicks, browses, comments on, likes, or purchases.

[0084] The indirect association relationship is used to indicate an indirect association with the target object, that is, an indirect association relationship or a second object and a second item that are indirectly associated with the target object, which are determined based on the first object and the first item that are directly associated with the target object. For example, the second object that is indirectly associated with the target object can be a user followed, commented on or liked by the first object, or a user who has followed, commented on or liked the same user as the target object, or a user who has clicked, browsed, commented, liked or purchased the same item as the target object; and the second item that is indirectly associated with the target object can be an item clicked, browsed, commented on, liked or purchased by the first object, or an item clicked, browsed, commented on, liked or purchased by the second object, etc., or an item similar to the first item, such as an item whose association with the first item reaches a association threshold. This is not limited in the embodiments of the present application.

[0085] It should be noted that acquiring the first object, the second object, the first item, and the second item means acquiring the object information of the first object, the object information of the second object, the item information of the first item, and the item information of the second item;

[0086] Among them, object information includes biological characteristics, sociological characteristics and behavioral characteristics of the corresponding object. The biological characteristics of the object include age, gender, height, weight, language, physical and mental health; the sociological characteristics of the user include education, kinship, living city, social status, profession, etc.; and the behavioral characteristics of the object include user registration behavior, such as the registration time of the user's current account for the corresponding client, user purchase behavior, such as the user's historical purchase behavior of goods, the user's historical collection behavior of goods and the user's historical comment behavior on goods, the user's online comment like behavior, the user's online attention behavior, etc.; and item information is used to indicate the interaction information with the corresponding item generated based on the object information, such as the purchase of the corresponding item or the collection of the corresponding item.

[0087] In step 102 , a heterogeneous graph is constructed with the first object, the second object, the target object, the first item, and the second item as nodes and the relationships between the nodes as edges.

[0088] It should be noted that the association relationship between nodes may refer to the interaction relationship between objects; for example, see Figure 4 , Figure 4 This is a schematic diagram of a heterogeneous graph provided in an embodiment of the present application, based on Figure 4 , the hollow points indicated by 401 are nodes in the heterogeneous graph, and the black solid lines indicated by 402 are edges in the heterogeneous graph.

[0089] As an example, when the application is a shopping client, the items (including the first item and the second item) are commodities presented based on the shopping client, such as commodities purchased, collected, and browsed by the object, and the objects (including the target object, the first object, and the second object) are customers of the shopping client. With customers and commodities as nodes, and the interaction relationships between customers and commodities and the interaction relationships between customers as edges, a heterogeneous graph corresponding to multiple customers and multiple commodities is constructed. By way of example, the interaction relationships between customers and commodities include customers' purchasing behavior, customers' evaluation of a certain commodity, and the frequency or duration of customers browsing a certain type of commodity, etc. The interaction relationships between customers include attention and comments between customers, etc.

[0090] As an example, when the application is a playback client, the items (including the first item and the second item) are media files presented based on the playback application, such as media information (such as video, audio, etc.) played, commented on, and liked by the object, and the objects (including the target object, the first object, and the second object) are viewers of the playback client. With the audience and media information as nodes, and the interaction relationship between the audience and media information and the interaction relationship between the audience as edges, a heterogeneous graph corresponding to multiple viewers and multiple media information is constructed. By way of example, the interaction relationship between the audience and the media information here includes the customer's viewing behavior such as replay, fast forward, etc., the customer's evaluation of a certain video, and the frequency of the customer watching a certain type of video, etc. The interaction relationship between the audience includes attention and comments between the audience.

[0091] As an example, when the application is a social client, the items are media information (such as video, text, pictures, audio, etc.) published by objects (including target objects, first objects, and second objects) in the social client, and the objects (including target objects, first objects, and second objects) are social users of the social client. With social users and media information as nodes, and the interaction relationships between social users and media information and the interaction relationships between social users as edges, a heterogeneous graph between multiple social users and multiple media information is constructed. By way of example, the interaction relationships between social users and media information include the social users' evaluation of a certain media information, and the length of time the social users watch a certain media information, while the interaction relationships between social users include the comment and like behaviors between social users, the follow-up relationships between social users, and the chat frequency between social users, etc.

[0092] Step 103 : for each node in the heterogeneous graph, determine the neighbor node in the layer farthest from the node, and obtain the neighbor node features of the neighbor node.

[0093] In actual implementation, after determining the heterogeneous graph, for each node in the heterogeneous graph, determine the neighbor node in the layer farthest from the node, see Figure 5 , Figure 5 This is a flowchart of a process for determining neighbor nodes in the layer farthest from a node provided by an embodiment of the present application, based on Figure 5 ,The process of determining the neighbor nodes in the layer farthest from the node can be achieved through the following steps.

[0094] Step 1031a: Determine at least one level corresponding to the node based on the multiple edges.

[0095] In actual implementation, after determining the heterogeneous graph, at least one level corresponding to the node is determined based on the multiple edges included in the heterogeneous graph. Specifically, at least one type of associated node corresponding to the node is determined based on the multiple edges; wherein the number of edges connecting associated nodes of the same category is the same, and the number of edges connecting associated nodes of different categories is different; the level at which each type of associated node is located is determined as at least one level corresponding to the node.

[0096] It should be noted that each node is connected by an edge. For the above nodes, the number of edges connected between each associated node and the above node in the heterogeneous graph is determined respectively, and associated nodes with the same number of connected edges are classified as associated nodes of the same category. For example, see Figure 6 , Figure 6 This is a schematic diagram of the process of determining at least one layer corresponding to a node provided by an embodiment of the present application, based on Figure 6 , nodes other than the node indicated by 601 are associated nodes, Figure 6 The number of edges between the middle gray node and the node indicated by 601 is 1, which means they are nodes of the same category. Figure 6 The number of edges connecting the black node and the node indicated by 601 is 2, and they are nodes of the same category.

[0097] Then, for each type of associated nodes, the level of the associated nodes of the corresponding category is determined based on the number of edges connecting the associated nodes of the corresponding category and the nodes. For example, when the number of edges connecting the associated nodes of category A and the nodes is 1, the associated nodes of category A are at the first level; when the number of edges connecting the associated nodes of category B and the nodes is 2, the associated nodes of category B are at the second level.

[0098] Finally, the levels at which each type of associated node is located are determined as at least one level corresponding to the node. Continuing with the above example, when there are two types of associated nodes, where the associated nodes of category A are at the first level and the associated nodes of category B are at the second level, the levels at which these two types of associated nodes are located are determined as the two levels corresponding to the node.

[0099] Step 1032a: Select a level farthest from the node from at least one level.

[0100] In actual implementation, as described above, the category determines the level of the associated node, and the level determines the hierarchy to which the node corresponds, that is, there is a one-to-one correspondence between the hierarchy and the category. Therefore, the process of selecting the hierarchy farthest from the node from at least one hierarchy can be, for each hierarchy, taking the number of edges connecting the associated nodes of the category corresponding to the hierarchy and the node as the number of edges corresponding to the hierarchy; based on the number of edges corresponding to each hierarchy, selecting the hierarchy with the largest number of corresponding edges from at least one hierarchy as the hierarchy farthest from the node. Continuing with the above example, when there are two hierarchies corresponding to the node, for each hierarchy, taking the number of edges connecting the associated nodes of the category corresponding to the hierarchy and the node as the number of edges corresponding to the hierarchy, that is, the number of edges corresponding to the first hierarchy is 1, and the number of edges corresponding to the second hierarchy is 2. Then, the hierarchy with the largest number of corresponding edges, i.e., the second hierarchy, is selected as the hierarchy farthest from the node.

[0101] Step 1033a: Determine at least one candidate neighbor node in the farthest layer, and determine a neighbor node from the at least one candidate neighbor node.

[0102] In actual implementation, after determining the level farthest from the node, at least one candidate neighbor node in the farthest level is determined, and then a target number of neighbor nodes is randomly determined from the at least one candidate neighbor node, where the target number can be pre-set, for example, it can be 1 or more.

[0103] In actual implementation, after determining the neighbor nodes, see Figure 7 , Figure 7 This is a flow chart of the process of obtaining neighbor node features of neighbor nodes provided by the embodiment of the present application, based on Figure 7 , the process of obtaining the neighbor node features of the neighbor node can be achieved through the following steps.

[0104] Step 1031b: Acquire a target heterogeneous graph, where the target heterogeneous graph is associated with the heterogeneous graph.

[0105] It should be noted that the target heterogeneous graph is constructed based on multiple training objects and multiple training items. The multiple training objects include the target object, the first object, and the second object, and the multiple training items include the first item and the second item. The association between the target heterogeneous graph and the heterogeneous graph indicates that the target heterogeneous graph includes the heterogeneous graph, i.e., portions of the target heterogeneous graph and the heterogeneous graph are identical.

[0106] Step 1032b: Based on the target heterogeneous graph, train the graph neural network model to obtain the target graph neural network model.

[0107] It should be noted that, see Figure 8 , Figure 8 This is a schematic diagram of the structure of the graph neural network model provided in the embodiment of the present application, based on Figure 8 The graph neural network model includes an embedding vector layer, a graph information propagation layer, and a prediction layer. Figure 8 The process of training a graph neural network model based on a target heterogeneous graph and obtaining a target graph neural network model can be as follows: through an embedding vector layer, vector encoding is performed on multiple training objects and multiple training items to obtain an initial feature vector corresponding to each training object and an initial feature vector corresponding to each training item; through a graph information propagation layer, based on the initial feature vector corresponding to each training object and the initial feature vector corresponding to each training item, a target feature vector corresponding to each node in the target heterogeneous graph is determined; wherein the target feature vector is a comprehensive feature vector formed by information propagation from the k-th order neighbor node of the node to the 1-th order neighbor node, and k is a positive integer; based on the target feature vector corresponding to each node in the target heterogeneous graph, the interaction possibility between each node object in the target heterogeneous graph is determined through a prediction layer; based on the determined interaction possibility, the model parameters of the graph neural network model are updated to obtain the target graph neural network model.

[0108] It should be noted that the feature vector here is also the feature, and the target feature vector corresponding to each node is also the node feature of each node; and the node object can be an item or an object, which refers to the item or object corresponding to the corresponding node.

[0109] In actual implementation, the process of determining the target feature vector corresponding to each node in the target heterogeneous graph based on the initial feature vector corresponding to each training object and the initial feature vector corresponding to each training item through the graph information propagation layer, that is, the process of determining the comprehensive feature vector formed by information propagation from the k-th order neighbor node to the 1st order neighbor node of the node, is similar to the process of determining the object features of the target object based on the neighbor node features of the neighbor nodes of each node as described later, which will be described in detail later and will not be repeated here in the embodiments of the present application.

[0110] Based on the target feature vector corresponding to each node in the target heterogeneous graph, the process of determining the possibility of interaction between node objects in the target heterogeneous graph through the prediction layer can be, for any two node objects (including a first node object and a second node object), multiplying the target feature vector of the first node object by the transposed vector of the target feature vector of the second node object to obtain the possibility of interaction between the first node object and the second node object. For example, when the first node object is a customer and the second node object is a product presented based on a shopping client, the possibility of interaction between the first node object and the second node object can represent the customer's desire score to purchase the product, that is, the score of whether the customer will purchase the product.

[0111] Among them, for the target feature vector of the node object, when the node object is an item, the target feature vector can be the item feature of the item corresponding to the corresponding node object, that is, the item feature vector; when the node object is an object, the target feature vector can be the object feature of the object corresponding to the corresponding node object, that is, the object feature vector.

[0112] In actual implementation, the process of updating the model parameters of the graph neural network model based on the determined interaction possibility to obtain the target graph neural network model can be, for each training object, obtaining the first interaction possibility between the training object and the first training item output by the prediction layer, and the second interaction possibility between the training object and the second training item; wherein, there is an interaction relationship between the first training item and the training object, and there is no interaction relationship between the second training item and the training object; based on the first interaction possibility and the second interaction possibility, constructing the loss function of the graph neural network model, and updating the model parameters of the graph neural network model based on the loss function.

[0113] It should be noted that the loss function here can be pre-set, and this is not limited in the embodiments of the present application. In this way, by training the graph neural network model with training items that have an interactive relationship with the training object and training items that do not have an interactive relationship with the training object, the training information of the model and the prediction accuracy of the target graph neural network model are improved.

[0114] Step 1033b: Acquire object information of the target object, object information of the first object, object information of the second object, item information of the first article, and item information of the second article.

[0115] In actual implementation, as mentioned above, object information includes the biological characteristics, sociological characteristics and behavioral characteristics of the corresponding object. The biological characteristics of the object include age, gender, height, weight, language, physical and mental health; the sociological characteristics of the user include education, marriage and family, kinship, city of residence, social status, profession, etc.; and the behavioral characteristics of the object include the user's purchasing behavior, such as the user's historical purchasing behavior of the product, the user's historical collection behavior of the product, the user's historical comment behavior on the product, the user's online comment like behavior, the user's online attention behavior, etc.; and the item information is used to indicate the interaction information with the corresponding item generated based on the object information, such as the purchase of the corresponding item or the collection of the corresponding item.

[0116] Step 1034b: Input the object information of the target object, the object information of the first object, the object information of the second object, the item information of the first item, and the item information of the second item into the target graph neural network model to obtain the neighbor node features of the neighbor nodes.

[0117] In actual implementation, after the object information of the target object, the object information of the first object, the object information of the second object, the item information of the first article and the item information of the second article are input into the target graph neural network model, the node features of each node in the heterogeneous graph composed of the target object, the first object, the second object, the first article and the second article can be determined, thereby also determining the neighbor node features of the neighbor nodes in the heterogeneous graph; wherein, the neighbor node features of the neighbor nodes here are also the target feature vectors of the neighbor nodes.

[0118] Step 104 : determining the object features of the target object based on the neighbor node features of the neighbor nodes of each node, and determining the recommended content of the target object based on the object features of the target object.

[0119] In actual implementation, the number of neighbor nodes can be one or more; in some embodiments, when the number of neighbor nodes is one, the neighbor node features of the neighbor node are directly used as the node features of the corresponding node, thereby determining the node features of each node in the heterogeneous graph. Since the target object is also a node in the heterogeneous graph, the object features of the target object are also determined.

[0120] In other embodiments, when there are multiple neighbor nodes, the process of determining the object characteristics of the target object based on the neighbor node characteristics of the neighbor nodes of each node may be to perform the following processing on each node: determine the neighbor node characteristics of multiple neighbor nodes of the node; aggregate the multiple neighbor node characteristics to obtain the node characteristics of the node.

[0121] In actual implementation, a node corresponds to at least one layer, and there are neighbor nodes of the node in each layer. The process of determining the neighbor node features of multiple neighbor nodes of a node can be to determine the neighbor node features of the neighbor node in the farthest layer in at least one layer corresponding to the node, and then aggregate the multiple neighbor node features to obtain the node features of the node.

[0122] For example, see Figure 9 , Figure 9 This is a schematic diagram of two levels of nodes corresponding to the embodiment of the present application, based on Figure 9 , the node is the node indicated by 901, A2-K2 are nodes of the second layer, among which A2, C2, D2, E2, F2, H2, and I2 are neighbor nodes in the second layer, and A1-E1 are nodes of the first layer, among which A1, B1, C1, and D1 are neighbor nodes in the first layer, thereby determining the neighbor node characteristics of multiple neighbor nodes of the node, that is, determining the neighbor node characteristics of multiple neighbor nodes of the node indicated by 901, namely A2, C2, D2, E2, F2, H2, I2, A1, B1, C1, and D1.

[0123] For the process of aggregating multiple neighbor node features to obtain the node features of a node, specifically, the number of levels is N, N is a positive integer, and the process of aggregating multiple neighbor node features to obtain the node features of a node can be, aggregating the neighbor node features of the neighbor nodes of the i-th level to obtain the neighbor node features of the i-1-th level; traversing i to obtain the node features of the node; wherein i is a positive integer less than or equal to N.

[0124] In actual implementation, the neighbor node features of the neighbor nodes at the i-th level are aggregated to obtain the neighbor node features at the i-1th level, that is,

[0125]

[0126] Among them, σ is the activation function, α ij is the weight between neighbor node i at level i-1 and neighbor node j at level i, such as the intimacy between neighbor node i and neighbor node j, such as the number of common friends, interaction frequency, common interests, similarities, etc. between the user corresponding to neighbor node i and the user corresponding to neighbor node j. Wh j Indicates the neighbor node features corresponding to the neighbor node at level i, N i Indicates the domain consisting of all neighbor nodes j at the i-th level that have a connection relationship with the neighbor node i at the i-1th level.

[0127] It should be noted that, first let i be N, that is, first aggregate the neighbor node features of the neighbor nodes at the Nth level to obtain the neighbor node features of the N-1th level, and then traverse i, that is, aggregate the neighbor node features of the neighbor nodes at the N-1th level to obtain the neighbor node features of the N-2th level, and repeat this until i is equal to 1, and obtain the neighbor node features of the 0th level, that is, the node features of the node.

[0128] Continuing with the above example, see Figure 9, the neighbor node features of multiple neighbor nodes of the node indicated by 901, namely A2, C2, D2, E2, F2, H2, I2, A1, B1, C1, and D1, are first aggregated to obtain the neighbor node features of A1, B1, C1, and D1, and then the neighbor node features of A1, B1, C1, and D1 are aggregated to obtain the node features of the node indicated by 901. Specifically, the neighbor node features of A2 and C2 are aggregated to obtain the neighbor node features of A1, the neighbor node features of D2 and E2 are aggregated to obtain the neighbor node features of B1, the neighbor node features of E2 and F2 are aggregated to obtain the neighbor node features of C1, and the neighbor node features of H2 and I2 are aggregated to obtain the neighbor node features of D1. Then, the neighbor node features of A1, B1, C1, and D1 are aggregated to obtain the node feature of the node indicated by 901.

[0129] In this way, the neighbor node features of the neighbor nodes at the i-th level are directly aggregated, which reduces the number of times the neighbor node features are acquired, improves the aggregation efficiency, and thus improves the efficiency of determining the recommended content.

[0130] It should be noted that the node characteristics of each node in the heterogeneous graph are determined, and since the target object is also a node in the heterogeneous graph, the object characteristics of the target object are also determined.

[0131] In actual implementation, after determining the object characteristics of the target object, see Figure 10 , Figure 10 This is a flow chart of the process of determining the recommended content for the target object provided by the embodiment of the present application, based on Figure 10 The process of determining the recommended content of the target object based on the object characteristics of the target object can be achieved through the following steps.

[0132] Step 1041 : determining, based on the heterogeneous graph, object features of the first object, object features of the second object, item features of the first article, and item features of the second article.

[0133] It should be noted that, as mentioned above, the node characteristics of each node in the heterogeneous graph are determined. Since the first object, the second object, the first item and the second item are also nodes in the heterogeneous graph, the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first item and the item characteristics of the second item can be determined directly based on the heterogeneous graph.

[0134] Step 1042 , combining the object features of the target object, the object features of the first object, the object features of the second object, the item features of the first article, and the item features of the second article, selects recommended content for the target object from the first object, the second object, the first article, and the second article.

[0135] In actual implementation, after determining the object characteristics of the target object, the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first article, and the item characteristics of the second article, the recommended content of the target object can be determined, wherein the recommended content can be an object, an article, or both an object and an article. This embodiment of the present application does not limit this.

[0136] In actual implementation, see Figure 11 , Figure 11 This is a flowchart of the process of selecting recommended content for a target object provided by an embodiment of the present application, based on Figure 11 The process of selecting recommended content for the target object from the first object, the second object, the first item, and the second item by combining the object characteristics of the target object, the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first item, and the item characteristics of the second item can be achieved through the following steps.

[0137] Step 10421: Determine the inner product of the first object based on the object characteristics of the target object and the object characteristics of the first object, determine the inner product of the second object based on the object characteristics of the target object and the object characteristics of the second object, determine the inner product of the first item based on the object characteristics of the target object and the item characteristics of the first item, and determine the inner product of the second item based on the object characteristics of the target object and the item characteristics of the second item.

[0138] It should be noted that, as mentioned above, the object feature here is also a feature vector, so the process of determining the inner product of the first object based on the object feature of the target object and the object feature of the first object is to multiply the feature vector of the target object by the transposed vector of the feature vector of the first object to obtain the inner product of the first object;

[0139] Accordingly, the process of determining the inner product of the second object based on the object feature of the target object and the object feature of the second object may be to multiply the feature vector of the target object by the transposed vector of the feature vector of the second object to obtain the inner product of the second object;

[0140] Accordingly, the process of determining the inner product of the first item based on the object feature of the target object and the item feature of the first item may be to multiply the feature vector of the target object by the transposed vector of the feature vector of the first item to obtain the inner product of the first item;

[0141] Accordingly, the process of determining the inner product of the second item based on the object characteristics of the target object and the item characteristics of the second item may be to multiply the feature vector of the target object by the transposed vector of the feature vector of the second item to obtain the inner product of the second item.

[0142] Step 10422: Based on the determined inner product, select the object or item with the largest inner product from the first object, the second object, the first item, and the second item as the target content, and determine the target content as the recommended content for the target object.

[0143] In actual implementation, the size of the inner product is used to indicate the size of the correlation between the corresponding object or item and the target object. Based on the determined inner product, by comparing the sizes of each inner product, the object or item with the largest inner product is selected as the target content of interest to the target object, and the target content is determined as the recommended content for the target object.

[0144] In some embodiments, the number of first items is a first number, the number of second items is a second number, the number of first objects is a third number, and the number of second objects is a fourth number. At least one of the first number, the second number, the third number, and the fourth number is multiple. After selecting the object or objects with the largest inner product as the target content, the correlation values ​​between each other content and the target content can be determined respectively; wherein the other content is any object or item other than the target content among the first object, the second object, the first item, and the second item; based on the size of each correlation value, the multiple other contents are sorted, and starting from the maximum correlation value, they are selected in sequence from the multiple other contents until the target number of other contents are selected. Thus, the process of determining the target content as the recommended content of the target object can be to determine the target content and the selected target number of other contents as the recommended content of the target object.

[0145] It should be noted that the target number can be pre-set, for example, 5, or 10.

[0146] In actual implementation, the process of separately determining the correlation values ​​between each other content and the target content may be: based on the heterogeneous graph, obtaining the content features of each other content and the content features of the target content; based on the content features of each other content and the content features of the target content, separately determining the correlation between each other content and the target content; normalizing each correlation to obtain the correlation value between each other content and the target content.

[0147] It should be noted that, as described above, the node features of each node in the heterogeneous graph are determined. Since the other contents and the target content are also nodes in the heterogeneous graph, the content features of the other contents and the target content can be directly obtained based on the heterogeneous graph.

[0148] Among them, as for content features, as mentioned above, content refers to items or objects. Therefore, when the content corresponding to the content feature is an item, the content feature refers to the item feature of the item; when the content corresponding to the content feature is an object, the content feature refers to the object feature of the object.

[0149] In actual implementation, for the process of determining the relevance of each other content to the target content based on the content features of each other content and the content features of the target content, in some embodiments, an attention mechanism may be used to determine the relevance of each other content to the target content based on the content features of each other content and the content features of the target content, that is,

[0150] e ij =Attention(Wh i ,Wh j )...Formula (2);

[0151] Among them, W is a weight matrix of size F×F, h i is the content feature of the target content, h j For the content features of other content, attention indicates that the attention mechanism is used for processing, e ij Indicates the relevance of the target content to other content.

[0152] In other embodiments, as described above, content features are also feature vectors. Therefore, the process of determining the correlation between each other content and the target content based on the object features of each other content and the content features of the target content can be to obtain the distance between the feature vector of the target content and the feature vectors of each other content; wherein the size of the distance is used to indicate the size of the difference between the target content and other content; and convert the distance between the feature vector of the target content and the feature vectors of each other content to obtain the correlation between each other content and the target content.

[0153] It should be noted that the process of converting the distance between the feature vector of the target content and the feature vector of other contents may be to obtain a mapping relationship between a pre-set distance and correlation; based on the mapping relationship and the distance between the feature vector of the target content and the feature vector of each other content, determine the correlation between each other content and the target content.

[0154] It should be noted that the process of determining the relevance of each other content to the target content based on the content characteristics of each other content and the content characteristics of the target content can also be implemented in other ways, including but not limited to the above two methods, and the embodiments of this application do not limit this.

[0155] In actual implementation, after determining each correlation, each correlation is normalized, for example, by performing softmax processing, to obtain the correlation value between each other content and the target content.

[0156] In this way, after determining the target content, the recommended content of the target object can be determined again based on other content that is relevant to the target content. This not only improves the diversity of the recommended content, but also improves the efficiency of the recommended content determination process and reduces the consumption of computing resources compared to the solution that needs to determine the recommended content of the target object based on the characteristics of the target object again.

[0157] In some embodiments, after determining the recommended content for the target object, the determined recommended content is recommended to the terminal. Specifically, the server sends the determined recommended content to the terminal corresponding to the target object so that the terminal presents the received recommended content, thereby displaying the recommended content to the target object.

[0158] By applying the above-mentioned embodiments of the present application, the target object is recommended based on the first object and the first item that are directly associated with the target object, and the second object and the second item that are indirectly associated with the target object. Compared with the scheme of recommending the target object based only on the first object and the first item that are directly associated with the target object, the accuracy of the determined recommended content is improved; at the same time, based on the heterogeneous graph, the object characteristics of the target object are determined according to the neighbor node characteristics of the neighbor nodes in the layer farthest from each node. Compared with the scheme in the related art that requires continuous sampling of the nodes corresponding to the target object and the nodes around the node, the possibility of overfitting is reduced and the efficiency of determining the recommended content is improved.

[0159] Below, an exemplary application of the embodiment of the present application in a practical application scenario will be described.

[0160] The recommendation systems of related technologies have the following problems: First, different friends have different influences on the user's decision-making and judgment. For example, user B has two friends. Both B and A like playing ball games and singing. B considers A's suggestions more when doing outdoor activities and C's suggestions more when doing indoor activities. Therefore, singing and ball games have different degrees of influence on users. Related technologies mostly make recommendations based on the importance of friends, but ignore the multi-faceted influence of different characteristics of specific interest areas on users, and the results are incomplete. Second, the recommendation system mainly uses paired user relationships in the modeling process, but ignores the complex high-order relationships between users. The actual situation is that the relationship between users is a complex network structure. In addition to being influenced by friends, a user also has friends of friends. The relationships formed by the complex user social activities in the entire social network will have an impact. The recommendation schemes of related technologies mainly focus on the joint modeling of explicit and implicit relationships in social networks, but ignore high-order implicit relationships. There is a clear difference in the importance of high-order implicit relationships to a user with enough neighbors and a user with only a few neighbors. Third, in the field of content recommendation on social networks, content producers and The consumer relationship of content and the social friend relationship of users are actually two sets of relationship networks. When a user consumes or reads a piece of content, he or she will usually pay close attention to the author of the content, but this author is not necessarily the user's social friend. If a user likes an author, he or she will usually take the initiative to follow and subscribe to the author's content, forming a follow-up relationship between the user and the self-media author (here it refers to broad attention, including users' attention to the author, likes, shares, clicks on the author's homepage, and other behaviors that reflect the user's preference for the author). Similarly, although another user is not a friend of this user, he or she also follows the same self-media author. It can be considered that these two users actually have common hobbies. Their common follow-up behavior is also an important feature for characterizing the user itself, but such features are ignored in the recommendation process of related technologies.

[0161] Based on this, the embodiment of the present application provides a method for propagating and distributing high-order social relationship interests in information flows based on machine learning. First, the social relationships between users in social networks, including multi-level social relationships and attention relationships between users and self-media content creators, are combined with the description information of the content to construct high-order social relationships. The high-order relationships are transmitted from the multi-level relationships between users, and the relationship between users and authors and related attribute information are expanded and introduced into a larger heterogeneous network. Since the author, as the creator of the content, is the source of content creation, it directly determines the quality of the content itself. In essence, the content transmitted and distributed in social networks is a small number of top self-media authors who occupy most of the content traffic. In addition, high-order relationships are directly modeled, and deep learning networks are used to obtain implicit representations of high-order relationships. The advantages of all parties are fully utilized to combine the user's interest network, social network, and attention network.

[0162] Second, we make full use of the characteristics of high-order relationships, from the first order to the Nth order, and use the attention mechanism to aggregate the embeddings of heterogeneous graphs to accurately learn the representation of nodes in the graph. At the same time, for graphs of different orders, we use the graph convolution algorithm of Graph Neural Networks (GNN) to aggregate the embedded representations of user nodes and content nodes of the current order, capture neighbors of different orders, and construct high-order implicit relationships between users, thereby predicting the relevance between the content to be recommended (such as videos or products) and the current user.

[0163] Among them, GNN establishes a deep learning representation of graph structure. Compared with the network representation learning method, it can simultaneously utilize graph structure information and node feature information to build a more complex and deeper neural network for representation learning. By aggregating neighbor nodes and the information of the previous layer and the aggregation and splicing of multi-layer propagation results, it can capture social relationships and behavioral relationships of different depths, realize hierarchical expression of nodes, and then realize high-order implicit relationship modeling. At the same time, for the item embedding in the propagation process, it means integrating the information of the propagation and diffusion of the item in the user's high-order relationship graph. In this way, the attention network integrates the influence between users of different orders, and finally obtains the embedding of high-order users and content through N-order propagation fusion, which is the expression of the prediction and final correlation between the two.

[0164] In this way, first, by modeling the high-order relationship chains in social networks, we can better understand the various different relationships in social networks, attention networks, and interest networks, make full use of the diffusion and dissemination in interest networks and social networks, achieve the goal of targeted content distribution, increase the channels for more content recall sources, and provide better recommendation effects, thereby improving the overall recommendation performance and user satisfaction; second, we can make full use of the high-order relationships between friends to transfer related behaviors and interests, and use the attention mechanism to improve the existing related social recommendation algorithms to accurately capture their joint impact on user-content interaction and better simulate the dynamic characteristics of user and content changes; third, we can effectively strengthen the role of the attention relationship between users and self-media authors in the final expression results, and incorporate the relevant information of self-media authors into the final embedded expression network, which can reflect the characteristics of the creators themselves and the user's preferences for the authors, so that the content of good authors has more opportunities to be disseminated and distributed, and provide help for the construction of the entire content ecosystem.

[0165] Next, the technical solution of this application is explained from the product side.

[0166] In actual implementation, the social relationships between users, including multi-level social relationships and the attention relationships between users and self-media content creators, are portrayed to assist in the distribution and dissemination of content interests in multi-level networks, help break out of the information cocoon of the recommendation process, and increase the diversity of recommended content results.

[0167] Next, the technical solution of this application is explained from a technical perspective.

[0168] When introducing social recommendation models, recommendation systems in related technologies leverage direct social relationships between users (such as friends, followers, interactions, and user relevance) as auxiliary information to improve recommendation performance. Due to the random nature of social relationship formation, explicit relationships are not always available. Related social recommendation systems often directly use these relationships as regularizers to constrain the user's final representation or as input to enhance the user's original embedding. This only considers the influence of a user's first-order neighbors and ignores the recursive diffusion of higher-order influences within social networks. In real life, users may be influenced by their friends' friends. Furthermore, social networks present many complex content diffusion phenomena, such as the spread of hot content, which cannot be effectively modeled using binary interactions alone. High-order networks and high-order interactions can be used to address these issues. High-order networks are important tools for characterizing complex communication. High-order interactions refer to interactions that occur on high-order networks and involve multiple entities. Compared to binary interactions, high-order interactions can be used to describe interactions involving multiple entities within a system.

[0169] In actual implementation, this application constructs high-order social relationships from the social relationships between users in social networks, including multi-level social relationships and attention relationships between users and self-media content creators, and content description information. The high-order relationships are transmitted from the multi-order relationships between users, and the relationship between users and authors and related attribute information are expanded to a larger heterogeneous network. Since the author, as the creator of the content, is the source of content creation, it directly determines the quality of the content itself. In essence, the content transmitted and distributed in social networks is a small number of top self-media authors who occupy most of the content traffic; directly model the high-order relationships, use deep learning networks to obtain implicit representations of high-order relationships, make full use of the advantages of all parties' capabilities, combine the user's interest network, social network, and attention network, and obtain multiple user embedding vectors and item embedding vectors through the aggregation and splicing of multi-layer and multi-order propagation results, and continuously repeat the propagation process. They are added as the final user embedding vector, which can effectively capture social relationships, behavioral relationships, and attention relationships of different depths, and achieve efficient feature expression of user nodes and content nodes.

[0170] This approach fully leverages the characteristics of high-order relationships, using an attention mechanism to aggregate the embeddings of heterogeneous graphs from the first to the Nth order to accurately learn the representations of nodes in the graph. At the same time, for graphs of different orders, the GNN graph convolution algorithm is used to aggregate the embedded representations of user nodes and content nodes at the current order, capturing neighbors of different orders and building high-order implicit relationships between users. GNNs are also used to establish the graph structure of the overall communication network using multi-scenario data from social networks. Using graph structure information and node feature information, a more complex and deeper neural network is constructed for representation learning. This aggregates information from neighboring nodes and the previous layer, and by aggregating and splicing multi-layer communication results, it can capture social relationships, behavioral relationships, and attention relationships at different depths, achieving hierarchical expression of nodes and, in turn, modeling high-order implicit relationships.

[0171] See also Figure 12 , Figure 12 This is a process diagram of the information flow high-order social relationship interest propagation and distribution method based on machine learning provided by the embodiment of the present application, based on Figure 12 In order to fully utilize the characterization of users and content in the process of high-order social relationship propagation, we first need to obtain the pre-built basic GNN multi-level graph, the specific construction process and processing of this graph, and then based on this constructed GNN graph, hierarchically integrate other relationships at different levels, such as attention relationships and items as initialization input, and then hierarchically propagate the characterization to obtain the final high-order embedded expression.

[0172] Typical graph nodes include users, authors, items (including items and friends), and attribute descriptions of each node. Friends and their items in multi-level networks can also be used to determine a user's potential preferences, so they need to be taken into account when characterizing users for a more comprehensive approach. The following details the implementation process and specific principles of the key technical modules in the diagram:

[0173] like Figure 12 As shown in the figure, the first level is used as the starting point for propagation and characterization. The core focus at this time is the current user node and its first-order domain node, as well as the currently focused author node and its attribute information. The content node is the item node and its attribute information triggered by the actual observed user operation behavior. This is also a GNN graph composed of nodes, edges, and their neighbors. By following the diffusion process, we can see that by the time the propagation reaches the Nth level, the vast majority of nodes in the entire network will be spread. At this point, the final node embedding expression can include information from the previous multiple layers, and the characterization ability of the entire node embedding expression will be much stronger.

[0174] There are two main propagation mechanisms for multi-level relationship network graphs: average pooling and attention. Average pooling treats the social influence of friends equally, while the graph attention network performs neighbor aggregation on the user social graph, interaction graph, and user-author attention graph, respectively. This is referred to as graph convolution (GNN) and node convolution in the graph. The former represents the entire layer, while the latter represents a specific node. When the next layer is reached, the results of the previous layer are brought in. It uses the attention mechanism to distinguish the influence of different friends, which is more in line with objective reality. Alternatively, instead of directly learning the node embedding vector, an aggregation operation can be learned to aggregate the node's neighbor information to obtain the node embedding vector, which can reduce some computational complexity.

[0175] In the GNN algorithm, for any node u on a graph, the information carried by each of its neighbors is first calculated. Second, the information from all neighbors is aggregated to node u. Finally, the representation of node u is updated. To enable GNN application to larger graphs, a node's neighbors are sampled and the sampled neighbor information is aggregated. Specifically, neighbor sampling is performed first, followed by aggregation, and finally, the node is updated based on the aggregated results. This is illustrated by the propagation of vectors and author profiles of self-media authors along social network relationships, author relationships, and the relationships between nodes and their friends. This fully exploits the rich information and relationships of various node types in heterogeneous graph networks.

[0176] The core principle of graph attention networks is to use the attention mechanism to aggregate neighbor information for each node in the graph network. Assuming that node i has j neighbors, the first-order neighbor topological aggregation features of node i are shown in the figure above. Here, the topological features of node i are obtained by aggregating the features of its neighbors and then undergoing a nonlinear transformation. The essence and core idea of ​​aggregation can be understood as dimensionality reduction. The node's own information is represented by merging and fusing domain node information, as shown in formula (1). After a nonlinear transformation, the values ​​are normalized through a softmax layer.

[0177] The process of aggregating node neighbor information based on an aggregation operation to obtain the node embedding vector is as follows for each batch: (1) For each node in this batch, sample the node's 1st-order neighbors, 2nd-order neighbors, and so on to the Kth-order neighbors; (2) Starting from the node sampled from the Kth-order neighbor, perform the aggregation operation and aggregate K times, and finally obtain the embedding vector information of each node in this batch; (3) Calculate the loss of the obtained embedding vector using an unsupervised or supervised loss function, and apply the gradient descent algorithm to update the parameters. The sampling operation and aggregation operation of this model are performed in opposite directions, which ultimately ensures that the embedding vectors of all nodes required for this batch can be generated, while also avoiding the huge amount of computation brought by aggregating all neighbors.

[0178] To capture neighbors of different orders and construct high-order implicit relationships between users, if dealing with heterogeneous graphs containing multiple types of vertices and edges, we can first generate a walk sequence based on math-path, called Meta-Path-Based Random Walks. When sampling, we consider the node type and meta-path information, and then learn the node representation.

[0179] It should be noted that if Figure 12 As shown in the figure, during the information propagation process, social recommendation is reformulated as a heterogeneous graph with multi-order social networks, interest networks, and attention networks as inputs. The embedded expressions of each user and content are iteratively aggregated from three aspects: including the embedding between users, authors, and content, the influence aggregation of social neighbors from the social network, the interest aggregation of neighbors from the user-content interest network, and the feature aggregation from friends following authors. During the modeling process, user nodes are propagated simultaneously in the social network, behavior network, and attention network, and item nodes are propagated in the behavior network. The expression of nodes is updated between layers, and a high-order propagation process is learned. By aggregating and splicing the multi-layer propagation results, social relationships and behavioral relationships of different depths can be captured, and hierarchical expression of nodes can be achieved.

[0180] In actual implementation, this application separates graph storage and processing from the basic operations of the GNN algorithm, forming a two-layer system: (1) The graph storage layer is used to store graph topology, node attribute information, a fast sampling index mechanism, and a cache mechanism. The operator operation layer integrates the basic operation operators of the GNN algorithm into the existing machine learning framework, allowing algorithm developers to simply call these operators to complete the development of the GNN algorithm.

[0181] It should be noted that the high-level relationships mentioned in this application are not just multi-level social relationships between users, but also another network formed by users and authors, which also becomes a part of this high-level social relationship. Integrating this relationship will also bring direct help to the distribution of content and the dissemination of interests. In addition, through the supplement of high-order relationships, unlike the Internet media in related technologies, the biggest advantage of this social media in the distribution of social network information flow content is that it introduces non-equivalent user relationships for following authors, which can improve the current recommendation effect, increase additional recall, improve the efficiency of the distribution algorithm of the recommendation system content, and improve the diversity of recommendation results.

[0182] By applying the above-mentioned embodiments of the present application, the target object is recommended based on the first object and the first item that are directly associated with the target object, and the second object and the second item that are indirectly associated with the target object. Compared with the scheme of recommending the target object based only on the first object and the first item that are directly associated with the target object, the accuracy of the determined recommended content is improved; at the same time, based on the heterogeneous graph, the object characteristics of the target object are determined according to the neighbor node characteristics of the neighbor nodes in the layer farthest from each node. Compared with the scheme in the related art that requires continuous sampling of the nodes corresponding to the target object and the nodes around the node, the possibility of overfitting is reduced and the efficiency of determining the recommended content is improved.

[0183] The following continues to describe the exemplary structure of the recommended content determination device 455 provided in the embodiment of the present application implemented as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the recommended content determination device 455 of the memory 450 may include:

[0184] Acquisition module 4551, configured to acquire a first object directly associated with a target object, a second object indirectly associated with the target object, and a first item directly associated with the target object, and a second item indirectly associated with the target object;

[0185] A construction module 4552 is configured to construct a heterogeneous graph using the first object, the second object, the target object, the first item, and the second item as nodes and the relationships between the nodes as edges;

[0186] A first determining module 4553 is configured to determine, for each node in the heterogeneous graph, a neighboring node in a layer farthest from the node, and obtain neighboring node features of the neighboring node;

[0187] The second determination module 4554 is configured to determine the object features of the target object based on the neighbor node features of the neighbor nodes of each node, and determine the recommended content of the target object based on the object features of the target object.

[0188] In some embodiments, the first determination module 4553 is further used to determine at least one level corresponding to the node based on multiple edges; select the level farthest from the node from the at least one level; determine at least one candidate neighbor node in the farthest level, and determine the neighbor node from the at least one candidate neighbor node.

[0189] In some embodiments, the first determination module 4553 is further used to determine at least one type of associated nodes corresponding to the node based on multiple edges; wherein the number of edges connecting associated nodes of the same type to the node is the same, and the number of edges connecting associated nodes of different types to the node is different; and the level at which each type of associated node is located is determined as at least one level corresponding to the node.

[0190] In some embodiments, there is a one-to-one correspondence between the hierarchy and the category; the first determination module 4553 is also used to, for each hierarchy, use the number of edges connecting the associated nodes of the category corresponding to the hierarchy and the nodes as the number of edges corresponding to the hierarchy; based on the number of edges corresponding to each hierarchy, select the hierarchy with the largest number of corresponding edges from the at least one hierarchy as the hierarchy farthest from the node.

[0191] In some embodiments, there are multiple neighbor nodes, and the second determination module 4554 is also used to perform the following processing for each node: determine the neighbor node characteristics of multiple neighbor nodes of the node; aggregate the multiple neighbor node characteristics to obtain the node characteristics of the node.

[0192] In some embodiments, the number of levels is N, where N is a positive integer; the second determination module 4554 is further used to aggregate the neighbor node features of the neighbor nodes of the i-th level to obtain the neighbor node features of the i-1-th level; traverse the i to obtain the node features of the node; wherein the i is a positive integer less than or equal to N.

[0193] In some embodiments, the second determination module 4554 is further used to determine the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first item, and the item characteristics of the second item based on the heterogeneous graph; and select recommended content for the target object from the first object, the second object, the first item, and the second item in combination with the object characteristics of the target object, the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first item, and the item characteristics of the second item.

[0194] In some embodiments, the second determination module 4554 is further used to determine the inner product of the first object based on the object characteristics of the target object and the object characteristics of the first object, and to determine the inner product of the second object based on the object characteristics of the target object and the object characteristics of the second object, and to determine the inner product of the first item based on the object characteristics of the target object and the item characteristics of the first item, and to determine the inner product of the second item based on the object characteristics of the target object and the item characteristics of the second item; based on the determined inner products, select the object or item with the largest inner product from the first object, the second object, the first item and the second item as the target content, and determine the target content as the recommended content for the target object.

[0195] In some embodiments, the quantity of the first items is a first quantity, the quantity of the second items is a second quantity, the quantity of the first objects is a third quantity, and the quantity of the second objects is a fourth quantity. At least one of the first quantity, the second quantity, the third quantity, and the fourth quantity is multiple. The device also includes a third determination module, which is used to respectively determine the correlation value between each other content and the target content; wherein the other content is any object or item other than the target object among the first object, the second object, the first item, and the second item; based on the size of each of the correlation values, the multiple other contents are sorted, and starting from the maximum correlation value, they are selected in sequence from the multiple other contents until the target number of other contents is selected; the second determination module 4554 is also used to determine the target content and the selected target number of other contents as recommended content for the target object.

[0196] In some embodiments, the third determination module is further used to obtain content features of each of the other contents and content features of the target content based on the heterogeneous graph; determine the correlation between each of the other contents and the target content based on the content features of each of the other contents and the content features of the target content; and normalize each of the correlations to obtain a correlation value between each of the other contents and the target content.

[0197] In some embodiments, the first determination module 4553 is also used to obtain a target heterogeneous graph, which is associated with the heterogeneous graph; based on the target heterogeneous graph, train a graph neural network model to obtain a target graph neural network model; obtain the object information of the target object, the object information of the first object, the object information of the second object, the item information of the first item, and the item information of the second item; input the object information of the target object, the object information of the first object, the object information of the second object, the item information of the first item, and the item information of the second item into the target graph neural network model to obtain the neighbor node features of the neighbor nodes.

[0198] An embodiment of the present application provides a computer program product comprising computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the instructions, causing the electronic device to perform the method for determining recommended content described in the embodiment of the present application.

[0199] The embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions. The computer-executable instructions are stored in the medium. When the computer-executable instructions are executed by a processor, the processor will execute the method for determining recommended content provided by the embodiment of the present application, for example, Figure 3 The method for determining recommended content is shown.

[0200] In some embodiments, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a magnetic surface memory, an optical disk, or a CD-ROM; or various devices including one or any combination of the above memories.

[0201] In some embodiments, computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0202] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0203] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0204] In summary, the embodiments of the present application have the following beneficial effects:

[0205] (1) Recommending the target object based on the first object and the first item that are directly associated with the target object, and the second object and the second item that are indirectly associated with the target object improves the accuracy of the determined recommended content compared to a solution in which the target object is recommended only based on the first object and the first item that are directly associated with the target object.

[0206] (2) Based on the heterogeneous graph, the object features of the target object are determined according to the neighbor node features of the neighbor nodes in the layer farthest from each node. Compared with the solution in the related art that requires continuous sampling of the nodes corresponding to the target object and the nodes around the node, the possibility of overfitting is reduced and the efficiency of determining the recommended content is improved.

[0207] It should be noted that in the embodiments of the present application, when it comes to obtaining data related to behavioral characteristics, object information, etc., when the embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0208] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. A method for determining recommended content, characterized in that: The method comprises: Acquire a first object that is directly associated with a target object, a second object that is indirectly associated with the target object, and acquire a first item that is directly associated with the target object, and a second item that is indirectly associated with the target object; Constructing a heterogeneous graph by using the first object, the second object, the target object, the first item, and the second item as nodes and the relationships between the nodes as edges; For each node in the heterogeneous graph, determining a neighbor node in a layer farthest from the node, and obtaining neighbor node features of the neighbor node; Based on the neighbor node features of the neighbor nodes of each of the nodes, object features of the target object are determined, and based on the object features of the target object, recommended content of the target object is determined.

2. The method according to claim 1, wherein The determining of the neighbor node in the layer farthest from the node includes: Determining at least one hierarchy corresponding to the node based on the plurality of edges; Selecting, from the at least one hierarchy, a hierarchy farthest from the node; At least one candidate neighbor node in the farthest layer is determined, and the neighbor node is determined from the at least one candidate neighbor node.

3. The method according to claim 2, wherein The determining, based on the plurality of edges, at least one hierarchy corresponding to the node includes: Determining at least one type of associated nodes corresponding to the node based on the plurality of edges; The number of edges connected between the associated nodes of the same category and the nodes is the same, and the number of edges connected between the associated nodes of different categories and the nodes is different; The level at which each type of associated node is located is determined as at least one layer corresponding to the node.

4. The method according to claim 3, wherein There is a one-to-one correspondence between the hierarchy and the category; and selecting the hierarchy farthest from the node from the at least one hierarchy includes: For each of the layers, the number of edges connecting the associated nodes of the category corresponding to the layer and the nodes is used as the number of edges corresponding to the layer; Based on the number of edges corresponding to each of the levels, a level corresponding to the largest number of edges is selected from the at least one level as the level farthest from the node.

5. The method according to claim 1, wherein There are multiple neighboring nodes, and determining the object feature of the target object based on the neighboring node features of the neighboring nodes of each node includes: The following processing is performed for each of the nodes: Determining neighbor node characteristics of a plurality of neighbor nodes of the node; Aggregate the features of multiple neighbor nodes to obtain the node feature of the node.

6. The method according to claim 5, wherein The number of layers is N, where N is a positive integer. Aggregating the features of the plurality of neighboring nodes to obtain the node features of the node includes: Aggregate the neighbor node features of the neighbor nodes at the i-th level to obtain the neighbor node features at the i-1th level; Traverse the i to obtain the node features of the node; Wherein, i is a positive integer less than or equal to N.

7. The method according to claim 1, wherein The determining the recommended content for the target object based on the object feature of the target object includes: determining, based on the heterogeneous graph, an object feature of the first object, an object feature of the second object, an item feature of the first article, and an item feature of the second article; In combination with the object characteristics of the target object, the object characteristics of the first object, the object characteristics of the second object, the item characteristics of the first article, and the item characteristics of the second article, recommended content for the target object is selected from the first object, the second object, the first article, and the second article.

8. The method according to claim 7, wherein The selecting, from the first object, the second object, the first item, and the second item, recommended content for the target object based on the object feature of the target object, the object feature of the first object, the object feature of the second object, the item feature of the first item, and the item feature of the second item, includes: Based on the object feature of the target object and the object feature of the first object, an inner product of the first object is determined, and Based on the object feature of the target object and the object feature of the second object, an inner product of the second object is determined, and Based on the object feature of the target object and the item feature of the first item, an inner product of the first item is determined, and determining an inner product of the second item based on an object feature of the target object and an item feature of the second item; Based on the determined inner product, an object or item with the largest inner product is selected from the first object, the second object, the first item, and the second item as target content, and the target content is determined as recommended content for the target object.

9. The method according to claim 8, wherein The quantity of the first items is a first quantity, the quantity of the second items is a second quantity, the quantity of the first objects is a third quantity, and the quantity of the second objects is a fourth quantity, and at least one of the first quantity, the second quantity, the third quantity, and the fourth quantity is a plurality. After selecting the object or item with the largest inner product as the target content, the method further includes: respectively determining a correlation value between each other content and the target object; The other content is any one of the first object, the second object, the first item, and the second item except the target content; sorting the plurality of other objects based on the magnitude of the respective correlation values, and selecting from the plurality of other objects in sequence starting from the largest correlation value until a target number of other contents are selected; The step of determining the target content as the recommended content for the target object includes: The target content and the target number of other selected contents are determined as recommended content for the target object.

10. The method according to claim 9, wherein The respectively determining the correlation value between each other content and the target content includes: Based on the heterogeneous graph, obtaining content features of each of the other contents and content features of the target content; Determining the relevance of each of the other contents to the target content based on content features of each of the other contents and content features of the target content; Normalization is performed on each of the correlations to obtain a correlation value between each of the other contents and the target content.

11. The method according to claim 1, wherein The obtaining of neighbor node features of the neighbor node includes: Acquire a target heterogeneous graph, where the target heterogeneous graph is associated with the heterogeneous graph; Based on the target heterogeneous graph, training a graph neural network model to obtain a target graph neural network model; Acquire object information of the target object, object information of the first object, object information of the second object, item information of the first article, and item information of the second article; The object information of the target object, the object information of the first object, the object information of the second object, the item information of the first article, and the item information of the second article are input into the target graph neural network model to obtain the neighbor node features of the neighbor nodes.

12. A device for determining recommended content, characterized in that: The device comprises: an acquisition module, configured to acquire a first object directly associated with a target object, a second object indirectly associated with the target object, and a first item directly associated with the target object, and a second item indirectly associated with the target object; a construction module, configured to construct a heterogeneous graph using the first object, the second object, the target object, the first item, and the second item as nodes and the relationships between the nodes as edges; A first determining module is configured to determine, for each node in the heterogeneous graph, a neighboring node in a layer farthest from the node, and obtain neighboring node features of the neighboring node; The second determining module is configured to determine object features of the target object based on neighbor node features of the neighbor nodes of each node, and determine recommended content of the target object based on the object features of the target object.

13. An electronic device, characterized in that: include: a memory for storing computer-executable instructions; The processor is configured to implement the method for determining recommended content according to any one of claims 1 to 11 when executing the computer-executable instructions stored in the memory.

14. A computer-readable storage medium, characterized in that Computer executable instructions are stored, which are used to cause a processor to execute and implement the method for determining recommended content according to any one of claims 1 to 11.

15. A computer program product comprising computer executable instructions, characterized in that When the computer-executable instructions are executed by a processor, the method for determining recommended content according to any one of claims 1 to 11 is implemented.

Images