CN111651671B

CN111651671B - User object recommendation method, device, computer equipment and storage medium

Info

Publication number: CN111651671B
Application number: CN202010461415.4A
Authority: CN
Inventors: 陈昊
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-05-27
Filing date: 2020-05-27
Publication date: 2023-11-21
Anticipated expiration: 2040-05-27
Also published as: CN111651671A

Abstract

The application relates to a user object recommendation method, a user object recommendation device, computer equipment and a storage medium based on artificial intelligence. The method comprises the following steps: acquiring user characteristic information of a target user; obtaining local graph network structure information associated with the target user in a user graph network; the user graph network is constructed based on the user information and the user interaction behavior information of all user objects in the user set; performing information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model to obtain user graph representation information corresponding to the target user; determining a matching degree predicted value between the target user and each user object in the user graph network based on the user graph representation information; and determining a user object meeting recommendation conditions according to the matching degree predicted value, and recommending the user object to the target user. By adopting the method, the recommendation efficiency and recommendation accuracy of the user object can be effectively improved.

Description

User object recommendation method, device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method, an apparatus, a computer device, and a storage medium for recommending a user object.

Background

With the rapid development of internet technology, a variety of social networking sites and social networking applications have emerged, through which users can conduct social activities. With the rapid development of artificial intelligence (Artificial Intelligence, AI) technology, user objects can be intelligently recommended to users based on cloud computing, distributed storage, big data processing and other technologies. For example, some friend making software may recommend some crowd objects to the user that the user may like based on the user's interests.

Existing user recommendation methods typically make recommendations based on geographic location matches or user feature rules. However, the geographic position matching mode is seriously affected by the geographic position, and the user is required to actively screen; in the mode of matching based on the user characteristic rules, the considered characteristic factors are limited and single, so that the efficiency and accuracy of recommending the user object are not high.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a user object recommendation method, apparatus, computer device, and storage medium that can effectively improve the recommendation efficiency and recommendation accuracy of user objects.

A user object recommendation method, the method comprising:

acquiring user characteristic information of a target user;

obtaining local graph network structure information associated with the target user in a user graph network; the user graph network is constructed based on the user information and the user interaction behavior information of all user objects in the user set;

performing information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model to obtain user graph representation information corresponding to the target user;

determining a matching degree predicted value between the target user and each user object in the user graph network based on the user graph representation information;

and determining a user object meeting recommendation conditions according to the matching degree predicted value, and recommending the user object to the target user.

A user object recommendation device, the device comprising:

the information acquisition module is used for acquiring user characteristic information of the target user; obtaining local graph network structure information associated with the target user in a user graph network; the user graph network is constructed based on the user information and the user interaction behavior information of all user objects in the user set;

The information transfer module is used for carrying out information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model to obtain user graph representation information corresponding to the target user;

the matching degree prediction module is used for determining a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information;

and the user object recommending module is used for determining the user object meeting the recommending condition according to the matching degree predicted value and recommending the user object to the target user.

In one embodiment, the device further includes a graph network construction module, configured to obtain user information and user interaction behavior information of each user object in the user set; determining user nodes corresponding to the user objects according to the user information; determining the connection relation among all user nodes according to the user interaction behavior information; and constructing a user graph network based on the user nodes and the connection relation.

In one embodiment, the graph network construction module is further configured to determine intimacy between user nodes according to the user interaction behavior information; determining the connection weight among the user nodes according to the intimacy; and determining the connection relation among the user nodes based on the connection weight.

In one embodiment, the device further includes a graph embedding processing module, configured to extract a user feature vector corresponding to the target user according to the user feature information; node migration is carried out on each user node in the user graph network, and a migration track is obtained; obtaining local graph network structure information associated with the target user according to the migration track; performing graph embedding based on the migration track to obtain a graph embedded feature vector corresponding to the target user; and connecting the user feature vector and the graph embedded feature vector to obtain a node feature vector.

In one embodiment, the user characteristic information comprises user portrait characteristic information and user social dynamic information; the information acquisition module is also used for extracting the user feature vector corresponding to the target user according to the user portrait feature information and the user social dynamic information.

In one embodiment, the graph embedding processing module is further configured to obtain a connection weight between each user node in the user graph network; and carrying out weighted random walk on each user node in the user graph network based on the connection weight to obtain a walk track.

In one embodiment, the information transfer module is further configured to perform information transfer processing on node feature vectors of each local user node based on the local graph network structure information through a graph neural network layer included in the matching degree prediction model, so as to obtain user graph representation information corresponding to the target user.

In one embodiment, the information transfer module is further configured to input the local graph network structure information and node feature vectors of each local user node in the local graph network structure information to the graph neural network layer; performing error back propagation on node feature vectors of all local user nodes through the graph neural network layer according to the local graph network structure information to obtain local transfer feature vectors; and obtaining user graph representation information corresponding to the target user based on the local transfer feature vector and the node feature vector of the target user.

In one embodiment, the matching degree prediction module is further configured to determine, through a matching degree prediction layer included in the matching degree prediction model, a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information of the target user and the user graph representation information of each user object in the user graph network.

In one embodiment, the matching degree prediction module is further configured to screen candidate user objects from the user graph network according to the user characteristic information; and determining a matching degree prediction value between the target user and each candidate user object based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model.

In one embodiment, the device further comprises a model training module, configured to obtain a user information sample, a training tag of the user information sample, and the user graph network; the user information sample comprises user information and historical interaction behavior information; the training label is a user matching score of the user information sample; and training a matching degree prediction model based on the user information sample and the training label and the user graph network.

In one embodiment, the model training module is further configured to extract sample user feature information of each sample user according to the user information sample; inputting the sample user characteristic information into the user graph network to obtain sample node characteristic vectors of the sample users; the information transfer processing is carried out on the basis of the sample graph network structure information and the sample node feature vectors associated with each sample user through a graph neural network layer included in the matching degree prediction model, so that sample user graph representation information of each sample user is obtained; determining sample matching degree prediction values among the sample users based on the sample user graph representation information through a matching degree prediction layer included in the matching degree prediction model; and adjusting parameters of the matching degree prediction model based on the difference between the sample matching degree prediction value and the training label, and continuing training until the training condition is met.

A computer device comprising a memory storing a computer program and a processor which when executing the computer program performs the steps of:

acquiring user characteristic information of a target user;

A computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

acquiring user characteristic information of a target user;

The user object recommending method, the user object recommending device, the computer equipment and the storage medium acquire the user characteristic information of the target user, and acquire the local graph network structure information associated with the target user in the user graph network; because the user graph network is constructed based on the user information and the user interaction behavior information of each user object in the user set, the social network graph structure information corresponding to the friend circle of the target user can be effectively obtained. And carrying out information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model, so that user graph representation information containing the user characteristics of a target user and the comprehensive characteristics of a friend circle can be effectively obtained. Determining a matching degree predicted value between a target user and each user object in a user graph network based on user graph representation information; and determining the user object meeting the recommendation condition according to the matching degree predicted value, and recommending the user object to the target user. The user graph representation information obtained through information transmission can reflect social interests of users more accurately, and the matching degree prediction value among the users is determined according to the user graph representation information, so that the matching degree among the users can be predicted more accurately, and the recommending efficiency and recommending accuracy of the user objects are improved effectively.

Drawings

FIG. 1 is an application environment diagram of a user object recommendation method in one embodiment;

FIG. 2 is a flow chart of a user object recommendation method in one embodiment;

FIG. 3 is a flow chart illustrating the steps of constructing a user graph network in one embodiment;

FIG. 4 is a diagram of a local network relationship for a user object in one embodiment;

FIG. 5 is a flowchart of a user object recommendation method according to another embodiment;

FIG. 6 is a schematic diagram of a model structure of a matching prediction layer in one embodiment;

FIG. 7 is a flowchart illustrating a training process of the matching degree prediction model in one embodiment;

FIG. 8 is a flowchart of a user object recommendation method in one embodiment;

FIG. 9 is a flow diagram of a process for predicting a degree of matching between user objects in one embodiment;

FIG. 10 is a block diagram of a user object recommendation device in one embodiment;

FIG. 11 is a block diagram illustrating a user object recommendation apparatus according to another embodiment;

FIG. 12 is a block diagram of a user object recommendation device in yet another embodiment;

fig. 13 is an internal structural view of a computer device in one embodiment.

Detailed Description

The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.

The scheme provided by the embodiment of the application relates to artificial intelligence (ML) technology and other technologies. Artificial intelligence is a theory, technology and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, sense environment, acquire knowledge and use knowledge to obtain optimal results, so that the machine has the functions of sensing, reasoning and decision. Machine learning involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc., and studies on how a computer simulates or implements learning behavior of a human being to obtain new knowledge or skills, and reorganizes existing knowledge structures to continuously improve their own performance. The graph network structure data can also be processed through artificial intelligence in an attempt to build an artificial intelligence system that can obtain information from the graph network structure data or multidimensional data. By performing processing based on artificial intelligence and machine learning=technology and the like on the user information, matching user objects can be effectively recommended to the user.

The user object recommendation method provided by the application can be applied to an application environment shown in figure 1. Wherein the user terminal 102 communicates with the server 104 via a network. After obtaining the user characteristic information of the target user, the server 104 obtains the local graph network structure information associated with the target user in the user graph network. The server 104 performs information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model, and obtains user graph representation information corresponding to the target user. The server 104 further determines a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information; and determining the user object meeting the recommendation condition according to the matching degree predicted value, and recommending the user object to the user terminal 102 corresponding to the target user. The user terminal 102 may be, but not limited to, various personal computers, notebook computers, smartphones, tablet computers, and portable wearable devices, and the server 104 may be implemented by a stand-alone server or a server cluster composed of a plurality of servers.

In one embodiment, as shown in fig. 2, a user object recommendation method is provided, and the method is applied to the server in fig. 1 for illustration, and includes the following steps:

s202, user characteristic information of a target user is obtained.

With the popularization of social networks, recommendation systems based on social application platforms such as social applications and social software can recommend user objects which are possibly interested to users according to user characteristic information of each user in the social networks.

The user characteristic information may be information extracted from user portrait information and used for characterizing the user characteristic. The user characteristic information is key information for analyzing social demands of users. The user portrait information is tagged user information abstracted for describing information such as attributes, user preferences, living habits, user behaviors and the like of the target user. The user portrait information may be obtained based on user information such as user profile, personal introduction page, etc. For example, the user profile may include information about the user's age, gender, location, hobbies, occupation, and the like. This information can be used to extract user features.

Specifically, the server corresponding to the recommendation system may acquire the user characteristic information of the target user, and perform friend recommendation based on the user characteristic information of the target user.

S204, obtaining local graph network structure information associated with a target user in a user graph network; the user graph network is constructed based on user information and user interaction behavior information of each user object in the user set.

The user information refers to attribute information of the user, such as personal information of the user. The user interaction behavior information refers to user social behavior information, and can include behavior information of a user interacting with other users in a social platform, for example.

The user graph network refers to a user-based social network graph, i.e., a social graph. The social relationship between the individual users is presented in the form of a graph with the user network. For example, a social network may be defined by using graphs G (V, E). Where V is a set of user nodes, each node representing a user, E is a set of edges, and if users Va and Vb have a social network relationship, then there is an edge E connecting the two user nodes.

Before the server recommends the user object, a user graph network is constructed by utilizing a user set comprising all users in the social application platform. Specifically, the server acquires user information and user interaction behavior information of each user object in the user set, determines user nodes corresponding to each user object according to the user information, determines connection relations among the user nodes according to the user interaction behavior information, and further constructs a global user graph network corresponding to all users in the social network platform according to the user nodes and the connection relations of each user object in the user set.

The local graph network structure information associated with the target user refers to local user graph network structures corresponding to target user nodes and user nodes adjacent to the target user nodes in the user graph network. Wherein the user nodes adjacent to the target user node may be a preset adjacent threshold or a preset number of neighbor nodes. The local graph network structure information represents graph structure information corresponding to the local user graph network, and can reflect social network structure relations of local user nodes, for example, the local graph network structure information can be information in the form of matrix arrays, matrix vectors and the like.

The neighbor nodes of the target user are usually friends of the target user, the local graph network structure information associated with the target user can reflect the network structure information corresponding to the current social circle of the target user,

the server acquires local graph network structure information associated with the target user from a pre-constructed user graph network in the process of recommending the user object to the target user. Specifically, the local graph network structure information may be local graph representation information associated with the target user, which is obtained after the target user node performs graph embedding processing on the user graph network.

S206, carrying out information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model, and obtaining user graph representation information corresponding to the target user.

The matching degree prediction model is a model with matching degree prediction capability after training, and can be specifically a neural network model based on machine learning. The matching degree prediction model may include a graph neural network and a matching degree prediction neural network. The graph neural network may be a network structure based on a graph neural network model, for example, a metamodel of the graph convolutional neural network. The graph neural network is used for extracting user graph representation information of a target user, and the matching degree prediction neural network is used for calculating the matching degree between all user objects according to the user graph representation information.

Wherein the user graph representation information refers to information representing user characteristics obtained based on graph network structure information. For example, the user graph representation information may be information in the form of a matrix or vector in particular.

After the server acquires the user characteristic information and the local graph network structure information of the target user, information transfer processing is carried out based on the user characteristic information and the local graph network structure information through a matching degree prediction model, and after information transfer, the whole local information representation of the local graph network associated with the target user can be obtained, so that the user graph representation information of the target user can be obtained based on the user characteristic information and combined with the local information representation. That is, the obtained user graph representation information not only contains the user characteristics of the target user, but also contains the comprehensive characteristics of the friend circle of the whole friend circle of the target user. Therefore, the obtained user graph representation information can reflect the social interests of the user more accurately.

S208, determining a matching degree predicted value between the target user and each user object in the user graph network based on the user graph representation information.

The matching degree refers to the similarity degree of social interests of individual user objects in a user graph network based on the social network. The higher the matching, the more consistent the social interests between the user objects, and the higher the degree of interest between the user objects.

The server obtains user graph representation information corresponding to the target user based on the user characteristic information and the local graph network structure information through the matching degree prediction model, and the server can obtain user graph representations of all user objects in the user graph network at the same time. The server further predicts a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation of each user object in the user graph network through a matching degree prediction model. The matching degree of the target user and each user object can be accurately predicted according to the extracted user graph identification information.

S210, determining the user object meeting the recommendation condition according to the matching degree predicted value, and recommending the user object to the target user.

The pushing condition refers to a condition that a user object can be pushed to a target user, and may specifically be a threshold of a certain index, for example, a probability threshold for predicting that the target user will add the recommended user object as a friend.

Specifically, the server determines a matching degree prediction value between a target user and each user object in the user graph network based on the user graph representation information through a matching degree prediction model, and judges whether the user object meets the recommendation condition according to the matching degree prediction value. And recommending the user object corresponding to the matching degree predicted value meeting the recommendation condition to the target user only when the matching degree predicted value is met.

In another embodiment, after constructing a user graph network based on user information and user interaction behavior information of each user object in the user set, the server performs graph embedding processing on the user graph network according to user feature information of each user object to obtain local graph network structure information and node feature vectors of each user object. The local graph network structure information and the node feature vector of each user object are input into a pre-trained matching degree prediction model, and information transfer processing is carried out on the basis of the local graph network structure information and the node feature vector of each user object through a graph neural network in the matching degree prediction model, so that user graph representation information of each user object can be effectively obtained.

When the user object needs to be recommended to the target user, the server can calculate the matching degree prediction value of the target user and each user object through the matching degree prediction model directly based on the obtained user graph representation information of each user object. The server further determines the user object meeting the recommendation condition according to the matching degree predicted value, and recommends the user object to the target user. Therefore, the recommendation accuracy of the user object is ensured, and the recommendation efficiency of the user object can be improved more effectively.

In one embodiment, social application software is installed in a user terminal corresponding to a user. When the target user is online on the social application platform through social application software deployed in the user terminal or reaches a period of a preset periodic frequency, the server can automatically acquire user graph representation information of the target user, and a matching degree prediction value between the target user and each user object in the user graph network is determined according to the user graph representation information through a matching degree prediction model. The server further determines the user object meeting the recommendation condition according to the matching degree predicted value, and recommends the user object to the target user, so that the user object which is matched with the target user is accurately and efficiently recommended to the target user.

In one embodiment, when a user initiates a friend matching request on a social application platform through a corresponding user terminal, the server may acquire user graph representation information of the target user based on the friend matching request initiated by the target user, and determine a matching degree prediction value between the target user and each user object in the user graph network according to the user graph representation information through a matching degree prediction model. The server further determines the user object meeting the recommendation condition according to the matching degree predicted value, and recommends the user object to the target user, so that the matched user object is accurately and efficiently recommended to the target user.

In the user object recommendation method, after the server acquires the user characteristic information of the target user, the server acquires the local graph network structure information associated with the target user in the user graph network; because the user graph network is constructed based on the user information and the user interaction behavior information of each user object in the user set, the social network graph structure information corresponding to the friend circle of the target user can be effectively obtained. The server performs information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model, so that user graph representation information containing the user characteristics of the target user and the comprehensive characteristics of the friend circle can be effectively obtained. The server further determines a matching degree predicted value between the target user and each user object in the user graph network based on the user graph representation information; and determining the user object meeting the recommendation condition according to the matching degree predicted value, and recommending the user object to the target user. The user graph representation information obtained through information transmission can reflect social interests of users more accurately, and the matching degree prediction value among the users is determined according to the user graph representation information, so that the matching degree among the users can be predicted more accurately, and the recommending efficiency and recommending accuracy of the user objects are improved effectively.

In one embodiment, before obtaining the local graph network structure information associated with the target user in the user graph network, the user object recommendation method further includes: acquiring user information and user interaction behavior information of each user object in a user set; determining user nodes corresponding to all user objects according to the user information; determining the connection relation between the user nodes according to the user interaction behavior information; and constructing a user graph network based on the user nodes and the connection relation.

Wherein the graph network (GraphNetwork, GN) is a collection of functions organized by graph structure in topological space to perform relational reasoning. In deep learning theory is the generalization of the graph neural network (Graph neural network, GNN) and the probability graph model (Probabilistic Graphical Model, PGM). The graph network is made up of interconnected graph network blocks, also referred to as "nodes" in a neural network implementation. The connections between nodes are called "edges" and represent the dependencies between the nodes. Each node of the graph network has an internal state and a system state, referred to as node attributes.

The user graph network comprises user nodes and edges, wherein the user nodes are composed of user objects in a user set of the social network platform, and the edges between the user nodes are composed of user interaction behaviors. Namely, the user connection of the whole network is constructed into a corresponding user graph network, and the connection edges between user nodes are determined by the intimacy between users.

Specifically, before recommending the user objects, the server acquires user information and user interaction behavior information of each user object in the user set. And determining user nodes corresponding to the user objects according to the user information, namely, each user object corresponds to one user node. Wherein the user object may further comprise a corresponding associated user node. For example, some users may register two associated user accounts with the social networking platform, and thus the two accounts may correspond to one user object. The server can determine the user node corresponding to the user object according to the user information of the user account, and can further analyze the associated account information of the user to determine the associated user node of the user object. The server may combine the user nodes of the user object with the associated nodes to form corresponding user nodes. In another embodiment, the server may further associate the user node of the user object with the association node and add the association identifier.

The server further determines the connection relation between the user nodes according to the user interaction behavior information, namely if interaction behavior exists between the two users, a connection line exists between the two user nodes; if there is no interaction between the two users, there is no connection between the two user nodes. The server builds the user graph network based on the user nodes and the connection relationship, so that the user graph network based on the social network relationship can be effectively built.

In one embodiment, as shown in fig. 3, a flowchart illustrating steps for constructing a user graph network in one embodiment specifically includes the following:

s302, user information and user interaction behavior information of all user objects in a user set are obtained.

S304, determining user nodes corresponding to the user objects according to the user information.

S306, determining the intimacy between the user nodes according to the user interaction behavior information.

S308, determining the connection weight among all user nodes according to the intimacy; connection relationships between the user nodes are determined based on the connection weights.

S310, constructing a user graph network based on the user nodes and the connection relation.

The user interaction behavior information also comprises information such as user interaction behavior times, interaction behavior accumulated time length, interaction discussion quantity, interaction frequency and the like.

In the process that the server constructs the user graph network according to the user information and the user interaction behavior information of each user object in the user set, the connection relation of each user node is determined according to the user interaction behavior information, and the connection weight of the connection relation can be further determined according to the user interaction behavior information. Specifically, the server determines the affinity between the user nodes according to the user interaction behavior information, and specifically, the affinity can be determined according to the interaction frequency and the interaction duration in the user interaction behavior information, and the higher the interaction frequency and the interaction duration, the higher the affinity. The server further determines a connection weight between the user nodes according to the affinity, and the higher the affinity is, the higher the connection weight is. For example, a social network may be defined by using graphs G (V, E, W). Where V is a set of user nodes, each node representing a user, E is a set of edges, if users Va and Vb have a social network relationship, there is an edge E (Va, vb) connecting the two users, and W (Va, vb) is used to define the weight of the edge.

The connection edge between the user nodes is determined by the affinity of the users, if more interaction behaviors exist between the two users, the connection weight of the two users is higher, and if less interaction behaviors exist between the two users, the connection weight of the two users is lower. As shown in fig. 4, fig. 4 is a partial network relationship diagram of a user object in one embodiment. Fig. 4 shows the connection relationship between the user nodes corresponding to the users u1, u2, u 3. The connection weight between the u node 1 and the u2 node is higher, the interaction number between the u1 node and the u3 node is smaller, and the connection weight between the u1 node and the u3 node is lower. And if the user u1 has no interaction with other users, the u1 node has no connection with other user nodes.

The server further determines connection relations among all the user nodes based on the connection weights, and the user nodes and the connection relations construct a user graph network, so that the user graph network based on the social network relations, which reflects social behaviors of the users, can be effectively constructed.

In one embodiment, as shown in fig. 5, another user object recommendation method is provided, which includes the following steps:

s502, obtaining user characteristic information of a target user.

S504, obtaining a user graph network constructed based on the user information and the user interaction behavior information of each user object in the user set.

S506, extracting the user characteristic vector corresponding to the target user according to the user characteristic information.

S508, carrying out node migration on each user node in the user graph network to obtain a migration track; and obtaining the local graph network structure information associated with the target user according to the migration track.

S510, performing graph embedding based on the walk track to obtain a graph embedded feature vector corresponding to the target user.

S512, connecting the user feature vector and the graph embedded feature vector to obtain a node feature vector.

S514, information transfer processing is carried out based on the node characteristic vector and the local graph network structure information through a pre-trained matching degree prediction model, and user graph representation information corresponding to the target user is obtained.

S516, based on the user graph representation information, a matching degree predicted value between the target user and each user object in the user graph network is determined.

And S518, determining the user object meeting the recommendation condition according to the matching degree predicted value, and recommending the user object to the target user.

The graph embedding is a process of mapping graph data (generally, a high-dimensional dense matrix) into low-micro dense vectors, and is used for representing the whole graph structure or the partial graph structure as one vector for embedded representation, so that the problem that the graph data is difficult to input into a machine learning model efficiently can be effectively solved. Graph embedding requires capturing the topology of the graph, vertex-to-vertex relationships, and other information (e.g., subgraphs, edges, etc.). If more information is presented, the downstream tasks will get better performance. During graph embedding, nodes in vector space that remain connected are close to each other.

The server builds a user graph network based on the user information and the user interaction behavior information of each user object in the user set, and then further performs graph embedding processing on the user graph network. Specifically, the server starts from each user node in the graph and randomly walks a plurality of tracks based on the user graph network to obtain a walk track corresponding to each user node. Typically, the trace corresponding to each user node is the user node associated with the user node and the connection relationship, so that the local graph network structure information associated with the target user can be determined according to the trace.

After the server obtains the walk track, the graph embedding processing is further performed based on the walk track. Specifically, the server trains the obtained walk tracks as a walk corpus by using a preset word vector embedding algorithm, and learns a representation of a network, so that a vector representation corresponding to each user node can be obtained, and the obtained vector representation is a graph embedded feature vector. For example, node walk can be performed by deep walk, and graph embedding processing can be performed by using algorithms such as Node2Vec Node embedding or Word2Vec Word vector embedding. Local context information of the user node in the user graph network can be obtained through node wandering, and therefore the learned representation vector reflects the local structure of the user node in the user graph network.

After the server obtains the user characteristic information of the target user, extracting the user characteristic vector corresponding to the target user according to the user characteristic information. After obtaining the user feature vector of the target user and the graph embedded feature vector, the server connects the user feature vector and the graph embedded feature vector to obtain the node feature vector. And the same is true, so that node characteristic vectors corresponding to all user nodes in the user graph network can be obtained. And determining the node characteristic vector as the node attribute characteristic of each user node in the user graph network by the server, thereby effectively generating the user graph network comprising the node attribute characteristic.

In one embodiment, the server may update the user graph network based on the user characteristic information and the user interaction behavior information of each user object in the user set according to the preset frequency, and perform graph embedding processing on the user graph network, so as to update the network structure information of the user graph network. For example, the preset frequency may be updated once a week, once two weeks, once a month, etc., without limitation. And updating the network information of the user graph according to the preset frequency, so that the accuracy of the network structure information of the user graph network is ensured.

In this embodiment, by performing graph embedding processing on the user graph network and generating the node feature vector based on the user feature vector and the graph embedding feature vector obtained by graph embedding, the node attribute feature of each user node can be effectively obtained, and thus the user graph network can be effectively represented by a vector. By performing the graph embedding processing on the user graph network, the information based on the user graph network can be processed more efficiently.

In one embodiment, the user characteristic information includes user portrait characteristic information and user social dynamic information; obtaining a user feature vector corresponding to the user feature information, including: and extracting the user feature vector corresponding to the target user according to the user portrait feature information and the user social dynamic information.

The user characteristic information comprises user portrait characteristic information of the user and user social dynamic information of the user. The user social dynamic information comprises information such as user introduction information, user sharing content, user release content and the like. For example, taking a user account in which a user object is registered with a social networking platform as an example, the user account may include profile information, as well as social dynamic information such as user avatars, user signatures, user circle content, and the like.

The server can generate user portrait feature information according to the personal information of the user account, and can also acquire user social dynamic information to generate corresponding user social feature information. Specifically, the user social dynamic information may include information such as text, pictures and videos, and the server may convert the user dynamic information into a vector describing the social characteristics of the user by adopting an image processing technology, an NLP (Natural Langunge Possns, natural language processing) technology, a user tag technology, and the like. And the server extracts the user feature vector corresponding to the target user by combining the user portrait feature information and the user social feature information. The user portrait characteristic information is combined with the user social characteristic information to extract the user characteristic vector, so that the user characteristic capable of reflecting the user property, social style and hobbies more accurately can be obtained effectively, and matched user objects can be recommended to the user more accurately.

In one embodiment, performing a walk on each user node in the user graph network to obtain a walk track includes: acquiring a connection weight among user nodes in a user graph network; and carrying out weighted random walk on each user node in the user graph network based on the connection weight to obtain a walk track.

The user graph network comprises user nodes and connection relations among the user nodes, and the connection relations also comprise connection weights among the user nodes. The connection weight reflects the affinity between the user objects.

Because the connection weights before the user nodes are different, the server acquires the connection weights among the user nodes in the user graph network when performing graph embedding processing on the user graph network; and carrying out weighted random walk on each user node in the user graph network based on the connection weight, so that the walk track corresponding to each user node which is relatively close and is associated with the target user node can be obtained.

In this embodiment, local context information of a user node in the user graph network can be obtained through weighted random walk, and the more neighboring points (or higher-order neighboring points) shared by two points in the graph, the shorter the distance between the corresponding two vectors. Therefore, the learned expression vector reflects the local structure of the user node in the user graph network, so that the local graph network structure information reflecting the social interests of the user can be accurately and effectively obtained.

In one embodiment, the information transfer processing is performed based on the user characteristic information and the local graph network structure information by using a pre-trained matching degree prediction model, so as to obtain user graph representation information corresponding to the target user, which includes: and carrying out information transfer processing on node feature vectors of each local user node based on local graph network structure information through a graph neural network layer included in the matching degree prediction model, and obtaining user graph representation information corresponding to the target user.

The matching degree prediction model comprises a graph neural network layer, and the graph neural network layer can be a network structure based on the graph neural network model. The Graph neural network layer may adopt a Graph neural network model based on Graph convolution network (Graph ConvolutionNetworks, GCN), graph attention network (Graph attention networks), graph self-encoder (Graph automatic encoders), graph generation network (Graph Generative Networks), graph space-time network (Graph Spatial-temporal Networks), or the like, and is not limited herein. The graph neural network layer may be a meta model of the graph neural network model. Meta-models are elements, relationships between elements, and representations that describe the model in which they are included. Taking the neural network model as an example, the meta-model may be considered as a part of the neural network structure of the model, for extracting a specific feature representation. The graph neural network layer in this embodiment is used to extract user graph representation information of the target user.

Specifically, after the server obtains the user characteristic information and the user graph network of the target user, the server obtains the local graph network structure information of the target user and the node characteristic vector of each local user node based on the user characteristic information and the user graph network. The server further inputs the local graph network structure information of the target user and the node feature vectors of the local user nodes into a graph neural network layer included in the matching degree prediction model, the graph neural network layer carries out information transfer processing on the node feature vectors of the local user nodes based on the local graph network structure information, learns the information of the local user nodes in the local graph network structure, and can obtain user graph representation information corresponding to the target user according to the learned information after the information transfer. Therefore, the obtained user graph representation information not only comprises the user characteristic vector of the user node, but also comprises the comprehensive characteristic information of the local graph network, so that the user graph representation information reflecting the social interests of the user can be accurately and effectively extracted.

In one embodiment, the information transfer processing is performed on the node feature vector of the local user node based on the local graph network structure information through the graph neural network layer included in the matching degree prediction model, to obtain user graph representation information corresponding to the target user, including: inputting the local graph network structure information and node characteristic vectors of each local user node in the local graph network structure information to a graph neural network layer; carrying out error back propagation on node feature vectors of all local user nodes through a graph neural network layer according to local graph network structure information to obtain local transfer feature vectors; and obtaining user graph representation information corresponding to the target user based on the local transfer feature vector and the node feature vector of the target user.

The information transfer process may be processed using an error back propagation algorithm (ErrorBackPropagation, BP). The error back propagation algorithm is an algorithm for training a multi-layer neural network, and has the advantages of firm theoretical basis, strict deduction process, clear physical concept, strong universality and the like. The learning process consists of two processes, forward propagation of information and reverse propagation of errors. The output error is transmitted reversely, and the error is distributed to all units of each layer, so that the error information of the units of each layer is obtained, and the weight of each unit is corrected, and the process is a weight adjustment process.

The graph neural network layer may be a BP network-based graph neural network model. For the model structure of the graphic neural network layer, the graphic neural network layer specifically further comprises an input layer, a hidden layer and an output layer. The hidden layer may further include an information transfer layer, where the information transfer layer is a network structure based on an error back propagation network.

After the server acquires the user characteristic information and the user graph network of the target user, local graph network structure information and node characteristic vectors of nodes of each local user of the target user are acquired based on the user characteristic information and the user graph network. The server then inputs the local graph network structure information of the target user and the node characteristic vectors of the local user nodes to the graph neural network layer. And the graph neural network layer performs error back propagation on node characteristic vectors of all local user nodes according to the local graph network structure information. Specifically, the node characteristic vector of each local user node is subjected to error back propagation at least twice based on the local graph network structure information. Preferably, the graph neural network layer performs error back propagation on each node feature vector twice based on the local graph network structure information.

In the process of carrying out error back propagation on node characteristic vectors of each local user node by the graph neural network layer, forward transmitting input information until an error is generated by output, and updating a parameter matrix by back propagation error information. Parameters are optimized in the flow between the nodes. Specifically, the output of each node in each layer is weighted and summed, and error parameters of each layer are calculated based on the result of the weighted summation. The error is further counter-propagated according to the error parameters, the error parameters are guided to be adjusted in a better direction in the propagation process, the parameter optimization can be directly carried out through the error parameters, and the error can be reduced to the minimum after a plurality of iterations.

The node characteristic vectors of all local user nodes are subjected to error back propagation processing according to the local graph network structure information through the graph neural network layer, and after information transmission, the overall comprehensive local information corresponding to all local user nodes in the local graph network structure can be effectively learned, so that the local transmission characteristic vectors corresponding to the local graph network nodes of the target user are obtained, and the user graph representation information corresponding to the target user can be obtained.

In this embodiment, error back propagation processing is performed on the node feature vectors of each local user node according to the local graph network structure information through the graph neural network layer, so that the comprehensive feature of the friend circle of the whole friend circle of the target user can be effectively captured. Therefore, the obtained user graph representation information not only comprises the user characteristic vector of the user node, but also comprises the comprehensive characteristic information of the local graph network, so that the user graph representation information reflecting the social interests of the user can be accurately and effectively extracted, and the obtained user graph representation information can reflect the social interests of the user more accurately.

In one embodiment, determining a match prediction value between a target user and each user object in a user graph network based on user graph representation information includes: and determining a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information of the target user and the user graph representation information of each user object in the user graph network through a matching degree prediction layer included in the matching degree prediction model.

The matching degree prediction model comprises a matching degree prediction layer. The matching degree prediction layer may be a network structure based on a prediction model. The matching degree prediction layer may adopt CNN (ConvolutionalNeural Network ), DNN model (Deep Neural Network, deep neural network), LSTM (Long Short Term Memory, long-term memory neural network), etc. The matching degree prediction layer may specifically further include an input layer, a hidden layer, and an output layer.

After obtaining the user characteristic information and the user graph network of the target user, the server obtains node characteristic vectors of all local user nodes based on the user characteristic information and the user graph network. The server inputs the local graph network structure information of the target user and the node characteristic vectors of the local user nodes into a graph neural network layer in the matching degree prediction model, and the local graph network structure information and the node characteristic vectors of the local user nodes are subjected to information transfer processing through the graph neural network layer to obtain user graph representation information of the target user. In the same way, the user graph representation information of each user node in the user graph network can be obtained after information transfer processing is performed on the basis of the user graph network according to the node characteristic vector of each user node through the graph neural network layer in the matching degree prediction model.

After obtaining the user graph representation information of the target user, the server takes the user graph representation information of the target user and other user objects in the user graph network as input of a matching degree prediction layer, and a hidden layer of the matching degree prediction layer processes the user graph representation information and calculates a matching degree prediction value between the target user and each user object in the user graph network according to the user graph representation information. And in the process of calculating the matching degree, specifically calculating a matching degree predicted value between the target user node and other user nodes which are not adjacent to the target user node. The server then outputs the result obtained by the hidden layer through the output layer, so that the matching degree predicted value between the target user and each user object is effectively obtained.

For example, as shown in fig. 6, a model structure diagram of the matching degree prediction layer is shown. The matching degree prediction layer specifically comprises an input layer, a hidden layer and an output layer. The input of the matching degree prediction layer is user graph representation information of a target user and other user objects in the user graph network. For example, if the matching degree between the user 1 and the user 2, the user 3 and the user 4 is to be predicted, the graph representation information of the user 1 and the user 2 is sequentially input, the user graph representation information of the user 1 and the user 2 is processed by the hidden layer, and then the matching degree prediction value between the target user 1 and the user 2 object is output by the output layer. This operation is then repeated, and the matching degree prediction values between the user 1 and the user 2, the user 3, and the user 4 can be obtained, respectively. Therefore, the matching degree between the target user and each user object can be accurately and effectively predicted.

In one embodiment, the output layer may further determine the calculated matching degree predicted value, and only output the user node whose matching degree predicted value meets the recommendation condition. Specifically, the result output by the matching degree prediction layer may be a user identifier corresponding to the user object. And after the server determines the user object meeting the recommendation condition according to the matching degree predicted value, recommending the user object meeting the recommendation condition to the target user.

In this embodiment, the matching degree prediction layer of the matching degree prediction model is used to efficiently predict the matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation of each user object in the user graph network. Because the user graph representation information contains the user characteristics of the target user and the comprehensive characteristics of the friend circle, the matching degree among the users can be predicted more accurately, and the user object recommendation efficiency and recommendation accuracy are improved effectively.

In one embodiment, determining, by a matching degree prediction layer included in the matching degree prediction model, a matching degree prediction value between a target user and each user object in the user graph network based on user graph representation information of the target user and user graph representation information of each user object in the user graph network includes: screening candidate user objects from a user graph network according to the user characteristic information; and determining a matching degree prediction value between the target user and each candidate user object based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model.

When the server recommends user objects to the target user, some completely unmatched user objects can be filtered according to the user characteristic information of the target user, and then the user objects matched with the target user are predicted based on the user graph information representation, so that computing resources in the recommendation process are reduced.

Specifically, the server acquires user characteristic information and a user graph network of a target user and local graph network structure information associated with the target user in the user graph network, and performs information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model, so as to acquire user graph representation information corresponding to the target user. User graph representation information of individual user objects in the user graph network can thus also be obtained.

After obtaining user graph representation information of each user object in the user graph network, the server screens candidate user objects from the user graph network according to the user characteristic information. Specifically, the server may filter out some completely unmatched user objects according to the user feature information according to a preset filtering rule. The user object which is already friends with the target user can be filtered according to the user account information. Filtering user objects in the user graph network of the whole network to obtain candidate user objects.

When the server recommends the user object to the target user, the matching degree prediction value between the target user and each candidate user object is calculated based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model, so that the matching degree between each user can be predicted more efficiently and accurately, and the recommending efficiency and recommending accuracy of the user object are improved effectively.

In one embodiment, the fitness prediction model is obtained by training through a training step comprising: acquiring a user information sample and a training label of the user information sample, and a user graph network; the user information sample comprises user information and historical interaction behavior information; training labels are user matching scores for user information samples; a matching prediction model is trained based on the user information samples and training labels and the user graph network.

The matching degree prediction model is trained by using user information sample data. Before user object recommendation processing is performed on a target user through the matching degree prediction model, the matching degree prediction model needs to be trained in advance.

The user information sample can be user information and real user historical interaction behavior information in a historical time period. I.e. the actual interactive behavior information of the user over a period of time. The user information sample comprises a real positive sample and a real negative sample, wherein the real positive sample can be user history interaction behavior information with more user interaction behaviors; the true negative sample may be user historical interaction behavior information that has interactions but has low interaction behavior. Specifically, the number of positive samples can be increased or the weight of the positive samples can be increased in an oversampling manner, so that training sample data with good training effect can be obtained. The user information samples also include user matching scores for the user information samples. The user matching score can be obtained according to the data of the user object actually added by the user, and can also be obtained in advance through manual annotation.

In the process of training the matching degree prediction model, a user information sample is used as training sample data for training, and a user matching score is used as a training label. The training label is used for carrying out parameter adjustment and other processing on the training result of each time so as to further train and optimize the matching degree prediction model.

Specifically, after the server acquires the user information sample, the user information sample is input into a preset matching degree prediction model for training, and the matching degree prediction model is subjected to parameter adjustment and optimization by using a training label so as to train the matching degree prediction model meeting the conditions. By using the real user history interaction behavior information as a training sample and using the user matching score as a training label, the matching degree prediction model with higher accuracy can be effectively trained.

In a specific embodiment, as shown in fig. 7, the training step of the matching degree prediction model includes:

s702, acquiring a user information sample and a training label of the user information sample and a user graph network; the user information sample comprises user information and historical interaction behavior information; training labels are user match scores for user information samples.

And S704, extracting sample user characteristic information of each sample user according to the user information samples.

S706, the sample user characteristic information is input into a user graph network, and sample node characteristic vectors of each sample user are obtained.

S708, through a graph neural network layer included in the matching degree prediction model, information transfer processing is carried out based on sample graph network structure information and sample node feature vectors associated with each sample user, and sample user graph representation information of each sample user is obtained.

S710, determining sample matching degree predicted values among sample users based on sample user graph representation information through a matching degree prediction layer included in the matching degree prediction model.

S712, based on the difference between the sample matching degree predicted value and the training label, adjusting parameters of the matching degree predicted model and continuing training until the training condition is met.

The matching degree prediction model comprises a graph neural network layer and a matching degree prediction layer. The graph neural network layer can be a neural network model based on the graph neural network, and the matching degree prediction layer can be a prediction model based on the neural network. Before training the matching degree prediction model, the server firstly builds a user graph network based on the user information and the user interaction behavior information of all user objects in the user set.

In the process of training the matching degree prediction model, a server acquires a user information sample, training labels of the user information sample and a user graph network, extracts sample user characteristic information of each sample user according to the user information sample, and then inputs the sample user characteristic information into the user graph network to obtain sample node characteristic vectors of each sample user.

The server further inputs the user graph network and the sample node feature vectors of the sample users to a graph neural network layer included in the matching degree prediction model, performs information transfer processing based on sample graph network structure information and sample node feature vectors associated with the sample users through the graph neural network layer, and outputs sample user graph representation information of the sample users through the graph neural network layer. The server further inputs sample user graph representation information of each sample user to a matching degree prediction layer included in the matching degree prediction model, determines sample matching degree prediction values among the sample users based on the sample user graph representation information through the matching degree prediction layer, and outputs the sample matching degree prediction values. The server can train the matching degree prediction model based on the error back propagation graph neural network layer and train the matching degree prediction layer based on the output of the graph neural network layer. The server further adjusts parameters of the matching degree prediction model based on the difference between the sample matching degree prediction value and the training label, and continues training until the training condition is met.

The difference between the matching degree and the efficiency label between the sample user objects can be measured by a loss function, for example, a function such as a mean absolute value loss function (MAE), a smoothed mean absolute error (Huber loss), a cross entropy loss function, and the like can be selected as the loss function. The training conditions are conditions for ending model training. The training stopping condition may be that the preset iteration number is reached, or that the predicted performance index of the matching degree prediction model after the parameter adjustment reaches a preset index.

In one embodiment, the user information samples may also be generated into a training sample set and a test sample set when training the fitness prediction module. And the server inputs the sample user characteristic information of each sample user in the training sample set to the user graph network to obtain sample node characteristic vectors of each sample user. The server also inputs the characteristic information of the test users in the test sample set to the user graph network to obtain the characteristic vector of the test nodes of the users. The training sample set and the testing sample set are divided into user nodes, so that each sample user and each testing user are not overlapped. The server further trains the matching degree prediction model by using the user information sample and the training label in the training sample set and the corresponding user graph network, after the initial matching degree prediction model is obtained, the matching degree prediction model is verified by using the user information sample and the training label in the testing sample set and the corresponding user graph network, and when the training condition is met, training is stopped, so that the required matching degree prediction model can be effectively trained.

In the embodiment, the matching degree prediction model is trained by utilizing the real user history interaction behavior information, and the training results of each time are subjected to parameter adjustment by utilizing the training labels, so that the matching degree prediction model is further trained and optimized, and the matching degree prediction model with higher accuracy can be effectively trained and obtained.

In a specific embodiment, as shown in fig. 8, a specific user object recommendation method is provided, which includes the following steps:

s802, obtaining user characteristic information of a target user.

S804, a user graph network constructed based on the user information and the user interaction behavior information of each user object in the user set is obtained.

And S806, extracting a user characteristic vector corresponding to the target user according to the user characteristic information.

S808, node migration is carried out on each user node in the user graph network, and a migration track is obtained; and obtaining the local graph network structure information associated with the target user according to the migration track.

And S810, performing graph embedding based on the walk track to obtain a graph embedded feature vector corresponding to the target user.

And S812, connecting the user feature vector and the graph embedded feature vector to obtain a node feature vector.

S814, the local graph network structure information and the node feature vectors of the local user nodes in the local graph network structure information are input to the graph neural network layer.

And S816, carrying out error back propagation on the node characteristic vectors of each local user node through the graph neural network layer according to the local graph network structure information to obtain local transfer characteristic vectors.

S818, obtaining user graph representation information corresponding to the target user based on the local transfer feature vector and the node feature vector of the target user.

S820, screening candidate user objects from the user graph network according to the user characteristic information.

S822, determining a matching degree prediction value between the target user and each candidate user object based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model.

S824, determining the user object meeting the recommendation condition according to the matching degree predicted value, and recommending the user object to the target user.

In this embodiment, by performing graph embedding processing on the user graph network, the user graph network can be effectively represented by vectors, and node attribute features of each user node can be effectively obtained. And carrying out information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model, so that user graph representation information containing the user characteristics of a target user and the comprehensive characteristics of a friend circle can be effectively obtained. The user graph representation information obtained through information transmission can reflect social interests of users more accurately, and the matching degree prediction value among the users is determined according to the user graph representation information, so that the matching degree among the users can be predicted more accurately, and the recommending efficiency and recommending accuracy of the user objects are improved effectively.

In a specific embodiment, as shown in fig. 9, fig. 9 is a schematic flow chart of a process of predicting a matching degree between user objects in a user graph network. The server first describes the user through user information such as user portrait and the like, and obtains user characteristic information. And further constructing a user graph network according to the user information and the actual interaction behavior of the user, and further performing graph embedding processing on the user graph network to obtain node feature vectors corresponding to all the user nodes and associated local graph network structure information of all the user nodes. The constructed user graph network corresponding to the whole network user in the social network platform can be constructed. The server further inputs the user graph network and the node feature vectors corresponding to the user nodes to the matching degree prediction model, and error back propagation is performed on the node feature vectors corresponding to the user nodes based on the user graph network through an information transfer layer of the matching degree prediction model. After the information is transferred twice, user graph information representation corresponding to each user node in the user graph network can be obtained. The obtained user graph information represents user information comprising the user characteristics of each user node and direct or indirect interaction behaviors with the user. The server further inputs the user graph information representation of each user node to a matching degree prediction layer in the matching degree prediction model, and the matching degree prediction layer represents the matching degree between the prediction target user and each user object based on the user graph information. For example, when a user object is recommended to a target user i, the user i and each user object to be recommended form a user pair, and user graph representation information of a plurality of user pairs is sequentially input into the matching degree prediction layer. Specifically, when the matching degree between the target user i and the user j is predicted, user graph representation information of the user i and the user j is input to a matching degree prediction layer, and a matching degree prediction value between the user i and the user j is output through the matching degree prediction layer. Thus, the matching degree prediction value between the target user and each user object can be effectively predicted. The user graph representation information containing the user characteristics of the target user and the associated friend characteristics is obtained based on information transfer through a pre-trained matching degree prediction model, so that the matching degree between the users can be accurately predicted based on the user graph representation information.

The application also provides an application scene, which applies the user object recommendation method. Specifically, the application of the user object recommendation method in the application scene is as follows:

a recommendation system is deployed in the social application platform for recommending user objects to registered users in the social application platform. The recommended objects may specifically be strangers for each user in the social application platform. The social application platform may correspond to a respective social application or social application software. Social application software is installed in a user terminal corresponding to the user.

Specifically, when a target user is online on a social application platform through social application software installed in a user terminal, and a friend matching request is initiated on the social application platform, a server acquires user characteristic information of the target user, and a user graph network constructed based on user information and user interaction behavior information of each user object in a user set is acquired. And the server performs information transfer processing based on the user characteristic information and the local graph network structure information through a pre-trained matching degree prediction model to obtain user graph representation information of the target user. And the server further determines a matching degree prediction value between the target user and each user object in the user graph network according to the user graph representation information through the matching degree prediction model. The server further determines the user object meeting the recommendation condition according to the matching degree predicted value and recommends the user object to the target user, so that the user object which is matched with the target user is accurately and efficiently recommended to the target user.

It should be understood that, although the steps in the flowcharts of fig. 2, 3, 5, 7, 8 are shown in order as indicated by the arrows, these steps are not necessarily performed in order as indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps of fig. 2, 3, 5, 7, 8 may comprise a plurality of steps or stages, which are not necessarily performed at the same time, but may be performed at different times, nor does the order in which the steps or stages are performed necessarily follow one another, but may be performed alternately or alternately with at least some of the other steps or stages.

In one embodiment, as shown in fig. 10, there is provided a user object recommendation apparatus 100, which may employ a software module or a hardware module, or a combination of both, as a part of a computer device, and specifically includes: an information acquisition module 1002, an information delivery module 1004, a matching degree prediction module 1006, and a user object recommendation module 1008, wherein:

An information obtaining module 1002, configured to obtain user feature information of a target user; acquiring local graph network structure information associated with a target user in a user graph network; the user graph network is constructed based on the user information and the user interaction behavior information of all user objects in the user set;

the information transfer module 1004 is configured to perform information transfer processing based on the user feature information and the local graph network structure information through a pre-trained matching degree prediction model, so as to obtain user graph representation information corresponding to the target user;

a matching degree prediction module 1006, configured to determine a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information;

and the user object recommending module 1008 is used for determining the user object meeting the recommending condition according to the matching degree predicted value and recommending the user object to the target user.

In one embodiment, as shown in fig. 11, the user object recommendation apparatus 100 further includes a graph network construction module 1000, configured to obtain user information and user interaction behavior information of each user object in the user set; determining user nodes corresponding to all user objects according to the user information; determining the connection relation between the user nodes according to the user interaction behavior information; and constructing a user graph network based on the user nodes and the connection relation.

In one embodiment, the graph network construction module 1000 is further configured to determine intimacy between user nodes according to user interaction behavior information; determining the connection weight between the user nodes according to the intimacy; connection relationships between the user nodes are determined based on the connection weights.

In one embodiment, the apparatus for recommending a user object 100 further includes a graph embedding processing module, configured to extract a user feature vector corresponding to the target user according to the user feature information; carrying out node migration on each user node in the user graph network to obtain a migration track; obtaining local graph network structure information associated with a target user according to the migration track; graph embedding is carried out based on the walk track, and graph embedded feature vectors corresponding to the target user are obtained; and connecting the user feature vector with the graph embedded feature vector to obtain a node feature vector.

In one embodiment, the user characteristic information includes user portrait characteristic information and user social dynamic information; the information acquisition module is also used for extracting user feature vectors corresponding to the target users according to the user portrait feature information and the user social dynamic information.

In one embodiment, the information transfer module 1004 is further configured to perform, through a graph neural network layer included in the matching degree prediction model, information transfer processing on node feature vectors of each local user node based on the local graph network structure information, to obtain user graph representation information corresponding to the target user.

In one embodiment, the information transfer module 1004 is further configured to input the local graph network structure information and node feature vectors of each local user node in the local graph network structure information to the graph neural network layer; carrying out error back propagation on node feature vectors of all local user nodes through a graph neural network layer according to local graph network structure information to obtain local transfer feature vectors; and obtaining user graph representation information corresponding to the target user based on the local transfer feature vector and the node feature vector of the target user.

In one embodiment, the matching degree prediction module 1006 is further configured to determine, through a matching degree prediction layer included in the matching degree prediction model, a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information of the target user and the user graph representation information of each user object in the user graph network.

In one embodiment, the matching degree prediction module 1006 is further configured to filter candidate user objects from the user graph network according to the user characteristic information; and determining a matching degree prediction value between the target user and each candidate user object based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model.

In one embodiment, as shown in fig. 12, the user object recommendation apparatus 100 further includes a model training module 1001, configured to obtain a user information sample and a training label of the user information sample, and a user graph network; the user information sample comprises user information and historical interaction behavior information; training labels are user matching scores for user information samples; a matching prediction model is trained based on the user information samples and training labels and the user graph network.

In one embodiment, the model training module 1001 is further configured to extract sample user feature information of each sample user according to the user information samples; inputting the sample user characteristic information into a user graph network to obtain sample node characteristic vectors of each sample user; carrying out information transfer processing based on sample graph network structure information and sample node feature vectors associated with each sample user through a graph neural network layer included in the matching degree prediction model, and obtaining sample user graph representation information of each sample user; determining a sample matching degree predicted value among sample users based on sample user graph representation information through a matching degree prediction layer included in the matching degree prediction model; based on the difference between the sample matching degree predicted value and the training label, parameters of the matching degree predicted model are adjusted, training is continued until training conditions are met, and training is stopped.

For specific limitations of the user object recommendation device, reference may be made to the above limitation of the user object recommendation method, and no further description is given here. The respective modules in the above-described user object recommendation apparatus may be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a server, and the internal structure of which may be as shown in fig. 13. The computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer equipment is used for storing data such as user characteristic information, user interaction behavior information, user graph network and the like. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a user object recommendation method.

It will be appreciated by those skilled in the art that the structure shown in FIG. 13 is merely a block diagram of some of the structures associated with the present inventive arrangements and is not limiting of the computer device to which the present inventive arrangements may be applied, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.

In an embodiment, there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the method embodiments described above when the computer program is executed.

In one embodiment, a computer-readable storage medium is provided, storing a computer program which, when executed by a processor, implements the steps of the method embodiments described above.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, or the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic RandomAccess Memory, DRAM), and the like.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The above examples illustrate only a few embodiments of the application, which are described in detail and are not to be construed as limiting the scope of the application. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of protection of the present application is to be determined by the appended claims.

Claims

1. A method for recommending user objects, the method comprising:

acquiring user characteristic information of a target user;

extracting a user characteristic vector corresponding to the target user according to the user characteristic information;

carrying out node migration on each user node in the user graph network to obtain a migration track;

performing graph embedding based on the migration track to obtain a graph embedded feature vector corresponding to the target user;

Connecting the user feature vector and the graph embedded feature vector to obtain a node feature vector;

performing information transfer processing based on the node feature vector and the local graph network structure information through a pre-trained matching degree prediction model to obtain user graph representation information corresponding to the target user;

2. The method of claim 1, wherein prior to the obtaining local graph network structure information associated with the target user in the user graph network, the method further comprises:

acquiring user information and user interaction behavior information of each user object in the user set;

determining user nodes corresponding to the user objects according to the user information;

Determining the connection relation among all user nodes according to the user interaction behavior information;

and constructing a user graph network based on the user nodes and the connection relation.

3. The method according to claim 2, wherein determining the connection relationship between the user nodes according to the user interaction behavior information comprises:

determining the intimacy between the user nodes according to the user interaction behavior information;

determining the connection weight among the user nodes according to the intimacy;

and determining the connection relation among the user nodes based on the connection weight.

4. The method of claim 1, wherein the obtaining local graph network structure information associated with the target user in the user graph network comprises:

and obtaining local graph network structure information associated with the target user according to the migration track.

5. The method of claim 1, wherein the user characteristic information comprises user portrait characteristic information and user social dynamic information;

the extracting the user feature vector corresponding to the target user according to the user feature information comprises the following steps:

and extracting the user feature vector corresponding to the target user according to the user portrait feature information and the user social dynamic information.

6. The method of claim 1, wherein the performing node migration on each user node in the user graph network to obtain a migration track includes:

acquiring the connection weight among all user nodes in the user graph network;

and carrying out weighted random walk on each user node in the user graph network based on the connection weight to obtain a walk track.

7. The method according to claim 1, wherein the obtaining, by the pretrained matching degree prediction model, user graph representation information corresponding to the target user based on the node feature vector and the local graph network structure information through information transfer processing includes:

and carrying out information transfer processing on node feature vectors of each local user node based on the local graph network structure information through a graph neural network layer included in the matching degree prediction model, and obtaining user graph representation information corresponding to the target user.

8. The method according to claim 7, wherein the obtaining, by the graph neural network layer included in the matching degree prediction model, the user graph representation information corresponding to the target user by performing information transfer processing on node feature vectors of each local user node based on the local graph network structure information includes:

Inputting the local graph network structure information and node feature vectors of each local user node in the local graph network structure information to the graph neural network layer;

performing error back propagation on node feature vectors of all local user nodes through the graph neural network layer according to the local graph network structure information to obtain local transfer feature vectors;

and obtaining user graph representation information corresponding to the target user based on the local transfer feature vector and the node feature vector of the target user.

9. The method of claim 1, wherein determining a match prediction value between the target user and each user object in the user graph network based on the user graph representation information comprises:

and determining a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information of the target user and the user graph representation information of each user object in the user graph network through a matching degree prediction layer included in the matching degree prediction model.

10. The method according to claim 9, wherein the determining, by the matching degree prediction layer included in the matching degree prediction model, a matching degree prediction value between the target user and each user object in the user graph network based on the user graph representation information of the target user and the user graph representation information of each user object in the user graph network, includes:

Screening candidate user objects from the user graph network according to the user characteristic information;

and determining a matching degree prediction value between the target user and each candidate user object based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model.

11. The method according to any one of claims 1 to 10, wherein the matching degree prediction model is obtained by training through a training step comprising:

acquiring a user information sample and a training label of the user information sample and the user graph network; the user information sample comprises user information and historical interaction behavior information; the training label is a user matching score of the user information sample;

and training a matching degree prediction model based on the user information sample and the training label and the user graph network.

12. The method of claim 11, wherein the training a matching prediction model based on the user information samples and the training labels and the user graph network comprises:

Extracting sample user characteristic information of each sample user according to the user information sample;

inputting the sample user characteristic information into the user graph network to obtain sample node characteristic vectors of the sample users;

the information transfer processing is carried out on the basis of the sample graph network structure information and the sample node feature vectors associated with each sample user through a graph neural network layer included in the matching degree prediction model, so that sample user graph representation information of each sample user is obtained;

determining sample matching degree prediction values among the sample users based on the sample user graph representation information through a matching degree prediction layer included in the matching degree prediction model;

and adjusting parameters of the matching degree prediction model based on the difference between the sample matching degree prediction value and the training label, and continuing training until the training condition is met.

13. A user object recommendation device, the device comprising:

the information acquisition module is used for acquiring user characteristic information of the target user;

the figure embedding processing module is used for extracting the user characteristic vector corresponding to the target user according to the user characteristic information; carrying out node migration on each user node in the user graph network to obtain a migration track; performing graph embedding based on the migration track to obtain a graph embedded feature vector corresponding to the target user; connecting the user feature vector and the graph embedded feature vector to obtain a node feature vector;

The information acquisition module is also used for acquiring local graph network structure information associated with the target user in the user graph network; the user graph network is constructed based on the user information and the user interaction behavior information of all user objects in the user set;

the information transfer module is used for carrying out information transfer processing based on the node characteristic vector and the local graph network structure information through a pre-trained matching degree prediction model to obtain user graph representation information corresponding to the target user;

14. The apparatus of claim 13, wherein the apparatus further comprises:

the graph network construction module is used for acquiring the user information and the user interaction behavior information of each user object in the user set; determining user nodes corresponding to the user objects according to the user information; determining the connection relation among all user nodes according to the user interaction behavior information; and constructing a user graph network based on the user nodes and the connection relation.

15. The apparatus of claim 14, wherein the graph network construction module is further configured to determine a degree of affinity between the user nodes based on the user interaction behavior information; determining the connection weight among the user nodes according to the intimacy; and determining the connection relation among the user nodes based on the connection weight.

16. The apparatus of claim 13, wherein the information acquisition module is further configured to obtain local graph network structure information associated with the target user based on the travel track.

17. The apparatus of claim 13, wherein the user characteristic information comprises user portrait characteristic information and user social dynamic information; the information acquisition module is also used for extracting the user feature vector corresponding to the target user according to the user portrait feature information and the user social dynamic information.

18. The apparatus of claim 13, wherein the graph embedding processing module is further configured to obtain a connection weight between user nodes in the user graph network; and carrying out weighted random walk on each user node in the user graph network based on the connection weight to obtain a walk track.

19. The apparatus of claim 13, wherein the information delivery module is further configured to perform information delivery processing on node feature vectors of each local user node based on the local graph network structure information through a graph neural network layer included in the matching degree prediction model, to obtain user graph representation information corresponding to the target user.

20. The apparatus of claim 19, wherein the information delivery module is further configured to input the local graph network structure information and node feature vectors of each local user node in the local graph network structure information to the graph neural network layer; performing error back propagation on node feature vectors of all local user nodes through the graph neural network layer according to the local graph network structure information to obtain local transfer feature vectors; and obtaining user graph representation information corresponding to the target user based on the local transfer feature vector and the node feature vector of the target user.

21. The apparatus of claim 13, wherein the matching prediction module is further configured to determine, by a matching prediction layer included in the matching prediction model, a matching prediction value between the target user and each user object in the user graph network based on user graph representation information of the target user and user graph representation information of each user object in the user graph network.

22. The apparatus of claim 21, wherein the matching prediction module is further configured to filter candidate user objects from the user graph network based on the user characteristic information; and determining a matching degree prediction value between the target user and each candidate user object based on the user graph representation information of the target user and the user graph representation information of each candidate user object through a matching degree prediction layer included in the matching degree prediction model.

23. The apparatus according to any one of claims 13 to 22, further comprising:

the model training module is used for acquiring a user information sample, a training label of the user information sample and the user graph network; the user information sample comprises user information and historical interaction behavior information; the training label is a user matching score of the user information sample; and training a matching degree prediction model based on the user information sample and the training label and the user graph network.

24. The apparatus of claim 23, wherein the model training module is further configured to extract sample user characteristic information for each sample user from the user information samples; inputting the sample user characteristic information into the user graph network to obtain sample node characteristic vectors of the sample users; the information transfer processing is carried out on the basis of the sample graph network structure information and the sample node feature vectors associated with each sample user through a graph neural network layer included in the matching degree prediction model, so that sample user graph representation information of each sample user is obtained; determining sample matching degree prediction values among the sample users based on the sample user graph representation information through a matching degree prediction layer included in the matching degree prediction model; and adjusting parameters of the matching degree prediction model based on the difference between the sample matching degree prediction value and the training label, and continuing training until the training condition is met.

25. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 12 when the computer program is executed.

26. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method of any one of claims 1 to 12.