Target user identification method and device
By constructing a relational knowledge graph and graph neural network model, we can identify whether the user is a competitive group, and solve the problem of judging whether the user is a competitive group when the credit report data cannot be obtained, and accurately identify and market the competitive group is achieved. cover.
Patent Information
- Application Number
- CN202111045661.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-07
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-09-07
AI Technical Summary
In the loan marketing pricing scenario, when the user's credit data cannot be obtained, it becomes difficult to determine whether the user is a competitive group, resulting in the competitive user being ignored and not being marketed.
By obtaining the user's basic features and associated users and associated applications, building a relational knowledge graph, and inputting it into the graph neural network model, obtaining the user's vector representation, and finally inputting the user's basic features and vector representation to the recognition model to identify whether the user is For the target user.
This method can accurately identify competitors without credit reporting data, make full use of user relationship data and application data, and improve the accuracy and coverage of marketing pricing.
Smart Images

Figure CN113569059B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present specification relate to the field of computer technology, and in particular to a method for identifying a target user. Background Art
[0002] In the marketing and pricing scenario of loans, combined with expert experience and experimental data analysis, it is found that the external competitors who have loan behavior have strong capital needs themselves, so marketing to this group of people can bring greater benefits. However, to determine whether a user is a competitor, the only way to obtain relevant information is through the user's active authorization to check the platform's credit, but the amount of such credit data is very small compared to the overall population. Competitors in the population who are not authorized to check their credit may be ignored and not marketed. Therefore, how to determine whether a user is a competitor when the user's credit data cannot be obtained has become a technical problem that urgently needs to be solved. Summary of the invention
[0003] In view of this, the present specification provides a method for identifying a target user. One or more embodiments of the present specification also relate to a target user identification device, a computing device, and a computer-readable storage medium to solve the technical defects existing in the prior art.
[0004] According to a first aspect of an embodiment of this specification, a target user identification method is provided, including:
[0005] Obtaining basic user characteristics of a user, associated users of the user, and associated applications of the user;
[0006] Constructing a relational knowledge graph according to the relation between the user and the associated user, and the relation between the user and the associated application;
[0007] Inputting the relational knowledge graph into a graph neural network model to obtain a vector representation of the user;
[0008] The user basic features of the user and the target vector representation of the user are input into a recognition model to obtain a recognition result of whether the user is a target user.
[0009] According to a second aspect of an embodiment of this specification, a graph neural network model training method is provided, including:
[0010] Obtaining an associated user of each sample user among a plurality of sample users and an associated application of each sample user;
[0011] According to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, a sample relationship knowledge graph is constructed, wherein the sample relationship knowledge graph includes user nodes and program nodes;
[0012] Determine a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and interaction features between the user node and the program node;
[0013] The graph neural network model is trained according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
[0014] According to a third aspect of the embodiments of this specification, a target user identification device is provided, including:
[0015] A first acquisition module is configured to acquire basic user characteristics of a user, associated users of the user, and associated applications of the user;
[0016] A first graph construction module is configured to construct a relational knowledge graph according to the association relationship between the user and the associated user, and the association relationship between the user and the associated application;
[0017] A vector representation acquisition module is configured to input the relational knowledge graph into a graph neural network model to obtain a vector representation of the user;
[0018] The recognition module is configured to input the user basic features of the user and the target vector representation of the user into the recognition model to obtain a recognition result of whether the user is a target user.
[0019] According to a fourth aspect of an embodiment of this specification, a graph neural network model training device is provided, including:
[0020] A fourth acquisition module is configured to acquire associated users of each sample user among a plurality of sample users and associated applications of each sample user;
[0021] A fourth graph construction module is configured to construct a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, wherein the sample relationship knowledge graph includes user nodes and program nodes;
[0022] A feature acquisition module, configured to determine a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and an interaction feature between the user node and the program node;
[0023] The third training module is configured to train the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
[0024] According to a fifth aspect of an embodiment of this specification, a computing device is provided, including:
[0025] Memory and processor;
[0026] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the target user identification method are implemented.
[0027] According to the sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the above-mentioned target user identification method are implemented.
[0028] An embodiment of the present specification implements a target user identification method and device, wherein the target object processing method includes obtaining the user's basic features, the user's associated users, and the user's associated applications; constructing a relationship knowledge graph based on the association relationship between the user and the associated users, and the association relationship between the user and the associated applications; inputting the relationship knowledge graph into a graph neural network model to obtain a vector representation of the user; inputting the user's basic features and the user's target vector representation into a recognition model to obtain a recognition result of whether the user is a target user.
[0029] Specifically, the target user identification method is a target user (i.e., competitor) identification scheme combined with a graph, which makes full use of the relational knowledge graph constructed based on relational data, and uses the user's features from associated users and associated applications as a supplement to the missing features of the target user, thereby accurately identifying the competitor. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 is a flow chart of a target user identification method provided by an embodiment of this specification;
[0031] Figure 2 It is a schematic diagram of the structure of a relational knowledge graph in a target user identification method provided by an embodiment of this specification;
[0032] Figure 3 It is a structural schematic diagram of an RCF-GCN graph neural network model in a target user identification method provided by an embodiment of this specification;
[0033] Figure 4 It is a schematic diagram of a specific recognition process of a recognition model in a target user recognition method provided by an embodiment of this specification;
[0034] Figure 5 This is a schematic diagram of a specific processing process of a target user identification method provided by an embodiment of this specification applied in a competitor user identification scenario;
[0035] Figure 6 It is a flowchart of a graph neural network model training method provided by one embodiment of this specification;
[0036] Figure 7 It is a structural schematic diagram of a target user identification device provided by an embodiment of this specification;
[0037] Figure 8 It is a structural schematic diagram of a graph neural network model training device provided by an embodiment of this specification;
[0038] Fig. 9 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION
[0039] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.
[0040] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.
[0041] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0042] First, the terms involved in one or more embodiments of this specification are explained.
[0043] Basic user characteristics: basic user attributes (such as name, age, etc.), behavioral characteristics such as access, transfer, expenditure, withdrawal, etc. on the fund trading platform, and access, transfer, expenditure, credit, etc. at financial institutions.
[0044] Heterogeneous graph: refers to a graph with more than one node type or relationship type. In real-world scenarios, the graph data objects that are usually studied are of multiple types, and the interactions between objects are also diverse. Therefore, heterogeneous graphs can better fit reality.
[0045] Bipartite graph: It is a special type of heterogeneous graph. We split the vertex set V in the graph G into two subsets A and B. If for any edge ij in the graph, vi∈A, vj∈B or vi∈B, vj∈A, then the graph G is called a bipartite graph. Bipartite graph is a very common graph data object that describes the interaction between two types of objects, such as users and products, authors and papers, etc.
[0046] Multi-relationship graph: The edges in the graph can be represented as (u, v, r), which means there is an edge of type r from node u to node v, that is, there are multiple types of edges between two nodes in a graph. For example, there are different types of relationships between two users, such as transfer relationship, address book relationship, friend relationship, etc. The modeling of multi-relationship graph is more practical and has higher application value.
[0047] Collaborative Filtering (CF): It includes two operations: collaboration and filtering. Collaboration is to analyze the similarity between users or things based on the information left by the interaction between users and things (such as ratings); filtering is to filter out the things that users are interested in from all candidate things based on the analysis and recommend these things to users.
[0048] Graph collaborative filtering network: Using a neural network framework to learn a nonlinear collaborative filtering model in the connection graph of user-item interactions can obtain more collaborative filtering signals.
[0049] Competitors: users who are eligible for loans.
[0050] APP:Application, application.
[0051] AGG: aggregation;
[0052] GCN: Graph Convolutional Neural Network.
[0053] In the marketing and pricing scenario of loans, combined with expert experience and experimental data analysis, it is found that the competitors who have loan behavior outside have strong capital needs, are active in visiting and spending on the loan platform, and are highly sensitive to rights and interests. Marketing to this group of people can bring greater benefits. To determine whether a user is a competitor, the only way to obtain relevant information is through the user's active authorization to check the credit of the platform. However, the amount of such credit data is very small compared to the overall population. However, at present, there is a serious lack of new customers on various loan platforms, and new customers with credit data for nearly one year only account for about 4% of the total new customers. For this large group of people who have not authorized credit checks, using common marketing methods, the competitors are likely to be ignored and not marketed. Among them, the definition of new customers is users who have not spent money on the loan platform in the past. Obviously, for new customers of the loan platform, there is a lack of data on their spending, repayment and other behaviors on the loan platform. These data are precisely the manifestation of the activeness of the competitors, which plays an important role in identifying competitors.
[0054] In the embodiments of this specification, in order to solve the problem of lack of new customer data, two aspects are mainly considered for supplementation: 1. User relationship data, such as transfer relationships, address book relationships, friend relationships, etc. It is believed that the transfer and consumption situations between related users may be similar, which can supplement the new customer information; 2. External data, such as APP data. Users' use of different APPs reflects the users' characteristics and needs to a certain extent, which can help characterize the behavioral characteristics of new customers.
[0055] In summary, there are three challenges in identifying competitive users:
[0056] 1. New customers lack historical behavior information on the loan platform. At the same time, the proportion of new customers who authorize credit checks is very small, making them the group that competitors focus on identifying.
[0057] 2. There is a large amount of relationship data, but it has not been fully utilized, such as friend relationships, transfer relationships, bank flow relationships, APPs, etc.
[0058] 3. There are only a small number of users who are authorized to check credit reports. That is to say, only a small number of samples have labels, and most samples are unlabeled.
[0059] Studies have shown that introducing graphs as auxiliary information into scenarios where information is sparse or missing can enrich the description of users and items and enhance the mining ability of algorithms, thereby effectively compensating for the sparseness or lack of interactive information. Typically, graphs are used to represent the relationship between objects. Any system containing relationships can be described by graphs. In the scenario of the embodiment of this specification, the interactive relationship between users and between users and APPs can be well constructed through graphs. However, the user-user graph and the user-APP graph are two different heterogeneous graphs. The user-user graph belongs to a multi-relationship graph with the same type of nodes and different types of edges. There are many different types of edges such as transfer relationships, address book relationships, and friend relationships between two users. Traditional graph neural networks (GNNs) such as GNN, GCN, and GAT have no specific meaning in the composition of the edge. They are only used as link relationships between nodes. Different types of edges are represented by the features of the edge, and the differences between different types of edges are not well reflected in the process of graph learning. For different types of edges between two nodes, their impact on the nodes is also different. Therefore, in the embodiment of this specification, the type of its edge is defined when composing the graph, and the role played by different relationships in user interaction is learned. The user-APP relationship graph is quite special, being a bipartite graph with different node types and the same edge type. Bipartite graphs are often used to construct relationship graphs between users and products in recommendation systems. Their characteristics are that only edge relationships exist between users and products, and there are no relationships between users or between products. Similarly, in the embodiments of this specification, the collaborative filtering concept in the recommendation system is also applied to the user-APP graph learning.
[0060] Therefore, the solution of the embodiment of this specification proposes a RCF-GCN graph neural network (GraphConvolution Network based on multi Relation learning and CollaborativeFiltering). Based on multi-relation learning and collaborative filtering, a relationship graph between users and between users and APPs is constructed. Different weights are used for different types of edges in the graph neural network to distinguish different connections between two nodes. At the same time, with the help of the idea of collaborative filtering, user-APP interaction information is added to graph learning to purposefully learn the degree of user preference for different APPs. In addition, this solution also improves the loss function of the network, effectively combining unlabeled samples with labeled samples for training. Finally, the obtained vector representation of the user is combined with the basic features of the user and input into the competitor crowd recognition model to obtain the competitor recognition result.
[0061] In the following, a method for identifying a target user is provided. One or more embodiments of the present specification also relate to a device for identifying a target user, a computing device, and a computer-readable storage medium. The implementation of the above-mentioned scheme is described in detail one by one in the following embodiments.
[0062] See also Figure 1 , Figure 1 A flow chart of a target user identification method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.
[0063] Step 102: Obtain basic user characteristics of the user, associated users of the user, and associated applications of the user.
[0064] Among them, the basic characteristics of users include but are not limited to the basic attributes of users, such as name, age, occupation, income level, etc.; the behavioral attributes of users, such as the behavioral characteristics of users' visits, transfers, expenditures, withdrawals, etc. on the capital trading platform and the behavioral characteristics of users' visits, transfers, expenditures and credit granting, etc. on the loan platform.
[0065] Associated users are users who have an associated relationship with the user, such as the user's friends, users who have a transfer relationship with the user, users in the user's address book, etc.; associated applications are mobile phone apps that have an associated relationship with the user, such as shopping apps, loan apps, fund transaction apps, etc. in the user's mobile phone.
[0066] Specifically, the target user identification method is applied in the scenario of identifying competing users. The basic user characteristics of the user to be identified, associated users with whom the user has an associated relationship, and associated applications with whom the user has an associated relationship can be obtained through a third-party platform (such as some credit reporting platforms, mobile terminals, and applications on mobile terminals where the user has registered information or used, etc.).
[0067] Step 104: construct a relational knowledge graph based on the association relationship between the user and the associated users, and the association relationship between the user and the associated applications.
[0068] Specifically, after obtaining the user's associated users and associated applications, a relationship knowledge graph is constructed based on the user, the associated users, the associated applications, the association relationship between the user and the associated users, and the association relationship between the user and the associated applications. The specific implementation method is as follows:
[0069] The user, the user's associated users, and the user's associated applications are taken as nodes, the association relationship between the user and the associated users, and the association relationship between the user and the associated applications are taken as edges, and a relational knowledge graph is constructed based on the nodes and the edges.
[0070] Among them, the relational knowledge graph is composed of nodes and edges. In specific implementation, the user, each associated user that has an associated relationship with the user, and each associated application that has an associated relationship with the user are taken as each node of the relational knowledge graph; then each association relationship between the user and the associated users, and each association relationship between the user and the associated application are taken as each edge of the relational knowledge graph; finally, the relational knowledge graph is constructed based on each node and each edge.
[0071] See also Figure 2 , Figure 2 A schematic diagram of the structure of a relational knowledge graph in a target user identification method provided according to an embodiment of the present specification is shown.
[0072] In practical applications, the relational knowledge graph is composed of several triples (h, r, t), where h and t represent the head node and tail node of a relationship, and r represents the relationship between h and t.
[0073] Take the nodes as users and applications, and the relationships as transfer relationships, address book relationships, and friend relationships as examples to build Figure 2 The relational knowledge graph in .
[0074] Depend on Figure 2 From the relational knowledge graph, we can see that U1, U2, U3, U4, U5, U6, A1, and A2 are nodes. The transfer relationship between U1 and U3 is an edge, the friend relationship between U1 and U2 is an edge, the transfer relationship between U1 and U2 is an edge, the address book relationship between U2 and U5 is an edge, the transfer relationship between U2 and U5 is an edge, the friend relationship between U2 and U4 is an edge, the friend relationship between U5 and U6 is an edge, the interaction relationship between U1 and A1 is an edge, the interaction relationship between U2 and A1 is an edge, and the interaction relationship between U5 and A2 is an edge. Based on the above nodes and edges, we construct Figure 2 The relational knowledge graph in the figure; among them, nodes U1, U2, U3, U4, U5, U6, etc. are users, and A1 and A2 are applications.
[0075] When U1 is the user to be identified, the associated users of U1 are U2 and U3, and the associated application is A1.
[0076] pass Figure 2 From the relational knowledge graph in the figure, we can see that the relational knowledge graph has two characteristics: first, users and users are connected by various types of relationships, such as U1 and U2 can have both transfer and friend relationships; second, users and APPs are connected by whether they interact with the APPs. There are directly connected edges between users and APPs, but there are no directly connected edges between APPs.
[0077] In the embodiments of this specification, a relational knowledge graph is constructed through users, associated users, associated applications, the associations between users and associated users, and the associations between users and associated applications. When the user is subsequently identified as a competitive user based on the relational knowledge graph, the relational knowledge graph can be referenced to enrich the description of the user to be identified and enhance the mining capability of the algorithm, thereby obtaining an accurate vector representation of the user through a graph neural network model.
[0078] Step 106: Input the relational knowledge graph into the graph neural network model to obtain the vector representation of the user.
[0079] Specifically, after constructing the user's relational knowledge graph, the relational knowledge graph can be input into a graph neural network model to obtain a vector representation of the user. The graph neural network model is a pre-trained graph neural network.
[0080] In practical applications, graph neural network models can exchange / obtain information about surrounding nodes by aggregating (AGG) neighbor information. This aggregation operation is repeated T times, which is equivalent to each node taking its own neighbor nodes that can be reached by T hops to propagate and exchange. Finally, after the information is aggregated to the central node, the downstream will design the corresponding loss function according to different tasks, and optimize the parameters through back propagation to perform node classification or link prediction tasks. Graph convolutional neural network GCN uses convolution to perform AGG on neighbor nodes.
[0081] Assume that there are N nodes in a graph, each node has its own initial features, and the initial features of all nodes form an N*D-dimensional matrix X. The relationship between the nodes also forms an N*N-dimensional matrix A, also known as the adjacency matrix. Among them, X and A are the inputs of the graph convolutional neural network model, so the aggregation method of the graph convolutional neural network GCN for the i-th node is described in formula 1:
[0082]
[0083] Among them, c i is the regularization constant, which is calculated by matrix A; represents the vector representation of the i-th node in the (l+1)-th layer; is the initial feature of the i-th node in the l-th layer; W l is the parameter matrix of the lth layer; N i Indicates the number of neighbor nodes connected to the i-th node.
[0084] The basic loss function of the graph convolutional neural network GCN is calculated using cross entropy, as described in Formula 2:
[0085]
[0086] Among them, |Y L | represents the number of samples; K represents the number of classifications; I(y l =i) is the indicator function, which means that when y l =i, the output is 1, y l ≠i, the output is 0; Z li Represents the softmax output of the final vector.
[0087] Finally, the graph convolutional neural network GCN is trained by the vector representation of each node obtained by formula 1 and the loss function obtained by formula 2. However, in the current scenario of competitive user identification, the graph convolutional neural network GCN has the following shortcomings: As can be seen from formula 1, W l It is shared by all the edges in layer l. In other words, although there are different types of edges in the graph, the types of edges are not distinguished during the learning process, and the differences in edges cannot be learned; the nodes are encoded to obtain embeddings, and then the loss function is constructed and decoded based on the definition of the relationship between the nodes. That is, in the process of constructing the embeddings, the interactive information between the nodes is not well utilized, and the learned embeddings lack important collaborative information between nodes of different types; and the supervised learning model can only use a small number of samples with labels (sample labels), which will result in a large amount of missing information, resulting in sample bias, making the trained graph convolutional neural network inaccurate.
[0088] In order to solve this technical problem, the graph convolutional neural network model can be improved by using the RCF-GCN graph neural network. The RCF-GCN graph neural network uses the GCN computing framework and designs the network structure accordingly for the application scenario of the competitive user identification in the embodiment of this specification. The specific implementation method is as follows:
[0089] The training steps of the graph neural network model are as follows:
[0090] Obtaining an associated user of each sample user among a plurality of sample users and an associated application of each sample user;
[0091] According to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, a sample relationship knowledge graph is constructed, wherein the sample relationship knowledge graph includes user nodes and program nodes;
[0092] Determine a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and interaction features between the user node and the program node;
[0093] The graph neural network model is trained according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
[0094] Among them, since there are currently only a small number of users authorized to check credit reports, that is, among multiple sample users, only a small number of sample users are labeled (such as whether they can be determined to be competing users), and some sample users are unlabeled (it is impossible to determine whether they are competing users); therefore, multiple sample users can be understood as labeled sample users and unlabeled sample users.
[0095] Specifically, first obtain multiple labeled and unlabeled sample users, and determine the associated users of each sample user and the associated applications of each sample user; then take each sample user, the associated users of each sample user, and the associated applications of each sample user as nodes, and take the association relationship between each sample user and the associated users of each sample user, and the association relationship between each sample user and the associated applications of each sample user as edges, and build a sample relationship knowledge graph based on the nodes and edges; then, based on the sample relationship knowledge graph, determine the feature matrix of each node in the sample relationship knowledge graph, the relationship matrix of each edge of the same node type, the weight matrix of each relationship matrix, and the interaction features of nodes of different node types; finally, train the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix, and the interaction features to obtain the graph neural network model (i.e., the RCF-GCN graph neural network).
[0096] In the embodiments of this specification, the graph neural network model is trained using the feature matrix of each node, the relationship matrix of different types of edges, the weight matrix of the relationship matrix, and the interaction features, so that the graph neural network model initializes different parameter matrices for different types of edges to learn, distinguishes different connection types between two nodes, and can obtain more abundant information. The attention mechanism is used to give each parameter matrix different weights to reflect the degree of influence of different edge relationships on users. At the same time, by adding the interaction features between users and APPs in the embedding layer, the user's preference information for different APPs is learned, which solves the problem of lack of important collaborative information between users and APPs in the vector representation learned by GCN, and improves the accuracy of the RCF-GCN graph neural network.
[0097] Specifically, the specific implementation methods of the feature matrix, relationship matrix, weight matrix and interaction features are as follows:
[0098] Taking each sample user and the associated users of each sample user as the user nodes of the sample relationship knowledge graph, and taking the associated application of each sample user as the program node of the sample relationship knowledge graph, and determining the feature matrix of the user node and the program node in the sample relationship knowledge graph;
[0099] Using different types of edges formed between each sample user and the associated users of each sample user as edges of the user node, and configuring a different relationship matrix for each edge of the user node;
[0100] Determine the weight matrix of each relationship matrix through the attention mechanism; and
[0101] The interaction feature between the user node and the program node is determined according to the feature matrix of the user node, the relationship matrix, the weight matrix and the feature matrix of the program node.
[0102] Among them, the feature matrix is the feature matrix of each user node (sample user, sample user's associated user) and the feature matrix of each program node in the sample relationship knowledge graph, that is, the feature matrix generated after vectorization of each sample user, each sample user's associated user, and each associated application. The relationship matrix is the relationship matrix for user nodes and different types of edge configurations (that is, the above-mentioned parameter matrix); at the same time, different weight matrices are assigned to each relationship matrix through the attention mechanism; and the interaction features are the interaction features between different types of user nodes and APP nodes.
[0103] In the embodiments of the present specification, before training the graph neural network model, the feature matrix of each user node, the relationship matrix of different types of edges of the user nodes, the weight matrix of each relationship matrix, and the interaction features between different types of nodes are obtained; the above features enable the graph neural network model to define the parameter matrix of different relationships between nodes of the same type when training the graph neural network model, and use the attention mechanism to learn the importance of different relationships; at the same time, the synergy between the user and the APP is considered, and the vector representation with user-APP interaction information is learned.
[0104] In a specific implementation, determining the interaction feature between the user node and the program node according to the feature matrix of the user node, the relationship matrix, the weight matrix, and the feature matrix of the program node includes:
[0105] Calculating the feature matrix of the user node, the relationship matrix, and the weight matrix through a first aggregation algorithm to obtain a vector representation of the user node;
[0106] Calculating the feature matrix of the program node using a second aggregation algorithm to obtain a vector representation of the program node;
[0107] The vector representation of the user node and the vector representation of the program node are interactively processed to obtain an interactive feature of the user node and the program node.
[0108] The first aggregation algorithm may be understood as Formula 4 in the embodiment of this specification, and the second aggregation algorithm may be understood as Formula 1 in the embodiment of this specification.
[0109] Specifically, in the RCF-GCN graph neural network model, by using different parameter matrices for different types of edges of user nodes, the effects of different relationships on nodes can be learned in the process of node aggregation information. For example, if there are three relationships: transfer, address book, and friend, then three weight matrices W are needed in the network. r (r=0,1,2) to learn, and finally we will get 3 representation vectors.
[0110] Considering that different relationships have different effects on different users, different weights should be assigned to the importance of the relationship based on the network's own learning. Therefore, in the embodiments of this specification, the attention mechanism can be used to calculate the importance of different relationships of the same user, so as to integrate the relationships.
[0111] Among them, the attention mechanism is implemented using a neural network, as described in Formula 3:
[0112]
[0113] in, is the weight matrix of different relationships, are the parameters learned by the network, A parameter matrix representing the different relations at layer l.
[0114] After determining the weight matrices of different relationships, the vector representation of each node (ie, user node) can be obtained through the feature matrix, the relationship matrix, and the weight matrix, wherein the aggregation method of the vector representation of each user node is as described in Formula 4:
[0115]
[0116] Among them, c i is the regularization constant; represents the vector representation of the i-th node in the (l+1)-th layer; is the initial feature of the i-th node in the l-th layer; N is the parameter matrix representing different relations at layer l;r,i Represents the number of neighbor nodes that have the rth relationship with the i-th node.
[0117] At the same time, according to the feature matrix of each program node, the vector representation of each program node is calculated by the above formula 1.
[0118] After obtaining the vector representation of each user node and the vector representation of each program node, the vector representation of the user node is multiplied by the vector representation of the program node element by element, and then input into the network layer for learning. This allows the interaction features of the user and the APP to be included before embedding learning, and important collaborative filtering information between the user and the APP can be learned through the network.
[0119] That is, after obtaining the vector representation of each user node and the vector representation of each program node, the vector representation of each user node with associated applications is interactively processed with the vector representation of the program node to obtain the interactive features of the user node and the program node, so that when the graph neural network model is subsequently trained, important collaborative filtering information between users and apps can be learned. That is, the vector representation of the user node and the interactive features are input into the graph neural network model, and after multi-layer vector aggregation learning, the final target vector representation of each node is obtained.
[0120] In specific implementation, the specific training process of the graph neural network model is as follows:
[0121] The step of training the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix, and the interaction features to obtain the graph neural network model includes:
[0122] Based on the vector representation of the user node and the interaction characteristics between the user node and the program node, obtaining the target vector representation of each sample user;
[0123] Determine labeled sample users and unlabeled sample users among the multiple sample users, train a graph neural network model according to vector representations of the labeled sample users, vector representations of the unlabeled sample users, and sample labels of the labeled sample users, to obtain the graph neural network model.
[0124] Specifically, the vector representation of each sample user is obtained based on the vector representation of each user node and the interaction characteristics between the user node and the program node.
[0125] In practical applications, when training a graph neural network model, we first design a loss function to effectively combine unlabeled sample users and labeled sample users, and then perform training on the full sample data.
[0126] Drawing on the design idea of pairwise loss, for a labeled sample user, the vector representation of the unlabeled sample user connected to it should be similar to the vector representation of the labeled sample user. Therefore, in specific implementation, the following loss function can be used to supervise the training of the unlabeled sample user. The specific calculation method of the loss function is shown in the following formula 5:
[0127]
[0128] Among them, O={(u,i,j)|(u,i)∈R + , (u, j)∈R-, R + represents the set of unlabeled sample users connected to labeled sample users, R - represents the set of unlabeled sample users that are not connected to labeled sample users, z u Represents the vector representation of the sample user with a label, z i The vector representation of the unlabeled sample users that are connected to the labeled sample users; z j The vector representation of the unlabeled sample users that do not have a connection relationship with the labeled sample users.
[0129] Specifically, firstly, the vector representation of the user node obtained after aggregation and the interaction features between the user node and the program node are learned through the embedding layer to obtain the final target vector representation of each sample user. Then, the reference loss function for supervised training of unlabeled sample users during model training is calculated based on the target vector representation of labeled sample users among multiple sample users, the target vector representation of unlabeled sample users, and the sample labels of labeled sample users.
[0130] In the embodiments of this specification, in order to address the problem that the number of labeled sample users in supervised learning modeling is relatively small, resulting in biased model training, the reference loss function for supervised training of unlabeled sample users is designed to effectively combine the unlabeled sample users with the labeled sample users, thereby making full use of the full amount of sample data for model training, thereby achieving better results for the graph neural network model obtained through training.
[0131] In a specific implementation, the graph neural network model is trained according to the target vector representation of the labeled sample user, the target vector representation of the unlabeled sample user, and the sample label of the labeled sample user, including:
[0132] Obtaining an initial loss function of the labeled sample according to the target vector representation of the labeled sample user and the sample label of the labeled sample user;
[0133] Determining unlabeled sample users associated with the labeled sample users and unlabeled sample users not associated with the labeled sample users;
[0134] Obtaining an initial loss function of the unlabeled sample according to the target vector representation of the unlabeled sample user associated with the labeled sample user, the target vector representation of the unlabeled sample user not associated with the labeled sample user, and the sample label of the labeled sample user;
[0135] According to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained, and the graph neural network model is trained based on the target vector representation of each sample user and the target loss function.
[0136] Specifically, first, according to the target vector representation of the labeled sample users and the sample labels of the labeled sample users, the above formula 2 is used to calculate the reference loss function for supervised training of labeled samples; then, according to the target vector representation of the unlabeled sample users associated with the labeled sample users, the target vector representation of the unlabeled sample users not associated with the labeled sample users, and the sample labels of the labeled sample users, the above formula 5 is used to calculate the initial loss function of the unlabeled samples, that is, the reference loss function for supervised training of unlabeled samples; finally, according to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained.
[0137] In practical applications, the target loss function of the graph neural network model obtained in the above manner is expressed by the following formula 6:
[0138]
[0139] Among them, Loss1 represents the initial loss function of labeled samples, and Loss2 represents the initial loss function of unlabeled samples. Represents adjustable parameters, is the regularization term.
[0140] After determining the target loss function of the graph neural network model, the graph neural network model can be trained according to the target vector representation of each sample user, the sample label of the target vector representation of each sample user, and the target loss function; that is, the graph neural network model is trained according to the target vector representation of each sample user and the sample label of the target vector representation of each sample user until the loss function obtained by training the graph neural network model satisfies the target loss function. Among them, the graph neural network model outputs the target vector representation of each sample user and the sample label of each target vector representation, and the sample label can be understood as yes or no, that is, the sample label is whether the sample user is a competing user or not a competing user.
[0141] See also Figure 3 , Figure 3 A schematic diagram of the structure of an RCF-GCN graph neural network model in a target user identification method provided according to an embodiment of the present specification is shown.
[0142] Figure 3 In the figure, U1, U2, U3, U4, and A1 are five nodes in the relational knowledge graph, among which U1, U2, U3, and U4 are user nodes, and A1 is a program node.
[0143] In practical applications, the feature matrix of each node of U1, U2, U3, and U4 is first obtained, and then the relationship matrix of each different type of edge is obtained. At the same time, different weight matrices are set for the relationship matrices of edges with the same node type but different edge types through the attention mechanism. For example, weight W1 is set for the edge of U1's transfer relationship, weight W2 is set for the edge of U1's friend relationship, and weight W3 is set for the edge of U3's address book. In addition, the aggregated vector representation of U2 interacts with the vector of A1 to obtain the interaction feature, and the interaction feature and the vector representation of each user node are input into the embedding layer for learning to obtain the final target vector representation of each user node.
[0144] The target user identification method of the embodiment of this specification proposes an RCF-GCN graph neural network, which distinguishes different types of edges between two nodes by defining parameter matrices of different relationships, and uses the attention mechanism to learn the importance of different relationships; at the same time, user-APP interaction information is added to the network to purposefully learn the user's preference for different APPs. In addition, the loss function of the network is designed to effectively combine unlabeled samples with labeled samples for training in the full amount of samples, thereby improving the accuracy of the graph neural network model obtained through training.
[0145] Step 108: Input the basic user features of the user and the target vector representation of the user into a recognition model to obtain a recognition result of whether the user is a target user.
[0146] Specifically, after determining the basic features of the user and the target vector representation of the user, the basic features of the user and the target vector representation of the user are input into the recognition model to obtain a recognition result of whether the user is a target user.
[0147] The training steps of the recognition model are as follows:
[0148] Obtaining user basic features of each sample user among multiple sample users, associated users of each sample user, and associated applications of each sample user;
[0149] Constructing a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user;
[0150] Inputting the sample relationship knowledge graph into the graph neural network model to obtain a target vector representation of each sample user;
[0151] The recognition model is trained according to the basic user features of each sample user and the target vector representation of each sample user.
[0152] Specifically, the process of constructing the sample relationship knowledge graph can be referred to the above embodiment and will not be repeated here.
[0153] The sample relationship knowledge graph is input into the graph neural network model to obtain the target vector representation of each sample user. Finally, the recognition model is trained based on the user basic characteristics of each sample user and the target vector representation of each sample user.
[0154] See also Figure 4 , Figure 4 A schematic diagram of a specific recognition process of a recognition model in a target user recognition method provided according to an embodiment of the present specification is shown.
[0155] Figure 4 It includes the user's basic features, the user's target vector representation obtained through the graph neural network model, and the recognition model.
[0156] Among them, the recognition model includes an activation function layer, a sigmoid function layer and a normal connection layer.
[0157] In practical applications, the user's basic features and the user's target vector representation are input into the recognition model. Through the activation function, sigmoid function and normal connection of the recognition model, the recognition result of whether the user is the target user is obtained, that is, whether the user is a competing user.
[0158] The target user identification method provided in the embodiment of this specification is a target user (i.e., competitor) identification solution combined with a graph, which makes full use of the relational knowledge graph constructed based on relational data, and uses the user's features from associated users and associated applications as a supplement to the missing features of the target user, thereby accurately identifying the competitor.
[0159] See also Figure 5 , Figure 5 A schematic diagram showing a specific processing process of a target user identification method provided according to an embodiment of the present specification applied in a competitor user identification scenario.
[0160] Step 1: Obtain basic user characteristics of the user to be identified, associated users who have an associated relationship with the user to be identified, and associated applications that have an associated relationship with the user to be identified.
[0161] Step 2: Construct a relational knowledge graph based on the user to be identified, associated users with whom the user to be identified has an associated relationship, associated applications with whom the user to be identified has an associated relationship, the relationship between the user to be identified and associated users with whom the user to be identified has an associated relationship, and the relationship between the user to be identified and associated applications with whom the user to be identified has an associated relationship.
[0162] The association relationship between the user to be identified and the associated user with which the user to be identified has an association relationship includes but is not limited to a transfer relationship, a contact book relationship, and a friend relationship, etc. The association relationship between the user to be identified and the associated application (APP) with which the user to be identified has an association relationship is only an interactive relationship.
[0163] Step 3: Input the relational knowledge graph into the pre-trained RCF-GCN graph neural network model to obtain the target vector representation of the user to be identified.
[0164] Step 4: Input the basic user features and target vector representation of the user to be identified into the competitor identification model to obtain the user label corresponding to the user to be identified. Through the user label, it can be determined whether the user to be identified is a competitor user.
[0165] The target user identification method provided in the embodiments of this specification provides a set of competitive crowd identification solutions combined with graphs. Making full use of relationship data to construct a graph can well supplement the missing features of new customers from the data of other users; and through the RCF-GCN graph neural network, by defining the parameter matrix of different relationships to distinguish the different types of edges between two nodes, and using the attention mechanism to learn the importance of different relationships; adding user-APP interaction information to the network, purposefully learning the user's preference for different APPs, and designing the network's loss function to effectively combine unlabeled samples with labeled samples for training in the full sample, so that the identification of competitive crowds is faster and more accurate.
[0166] See also Figure 6 , Figure 6 A flowchart of a graph neural network model training method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.
[0167] Step 602: Obtain the associated users of each sample user among a plurality of sample users and the associated applications of each sample user.
[0168] Step 604: Construct a sample relationship knowledge graph based on the relationship between each sample user and the associated users of each sample user, and the relationship between each sample user and the associated applications of each sample user, wherein the sample relationship knowledge graph includes user nodes and program nodes.
[0169] Step 606: Determine the feature matrix of the user node in the sample relational knowledge graph, the relation matrix of each edge of the user node, the weight matrix of each relation matrix, and the interaction features between the user node and the program node.
[0170] Step 608: Train the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
[0171] Optionally, determining a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and an interaction feature between the user node and the program node includes:
[0172] Taking each sample user and the associated users of each sample user as the user nodes of the sample relationship knowledge graph, and taking the associated application of each sample user as the program node of the sample relationship knowledge graph, and determining the feature matrix of the user node and the program node in the sample relationship knowledge graph;
[0173] Using different types of edges formed between each sample user and the associated users of each sample user as edges of the user node, and configuring a different relationship matrix for each edge of the user node;
[0174] Determine the weight matrix of each relationship matrix through the attention mechanism; and
[0175] The interaction feature between the user node and the program node is determined according to the feature matrix of the user node, the relationship matrix, the weight matrix and the feature matrix of the program node.
[0176] Optionally, determining the interaction feature between the user node and the program node according to the feature matrix of the user node, the relationship matrix, the weight matrix, and the feature matrix of the program node includes:
[0177] Calculating the feature matrix of the user node, the relationship matrix, and the weight matrix through a first aggregation algorithm to obtain a vector representation of the user node;
[0178] Calculating the feature matrix of the program node using a second aggregation algorithm to obtain a vector representation of the program node;
[0179] The vector representation of the user node and the vector representation of the program node are interactively processed to obtain an interactive feature of the user node and the program node.
[0180] Optionally, the training a graph neural network model according to the feature matrix, the relationship matrix, the weight matrix, and the interaction features to obtain the graph neural network model includes:
[0181] Based on the vector representation of the user node and the interaction characteristics between the user node and the program node, obtaining the target vector representation of each sample user;
[0182] Determine labeled sample users and unlabeled sample users among the multiple sample users, train a graph neural network model according to the target vector representations of the labeled sample users, the target vector representations of the unlabeled sample users, and the sample labels of the labeled sample users, to obtain the graph neural network model.
[0183] Optionally, the training the graph neural network model according to the target vector representation of the labeled sample user, the target vector representation of the unlabeled sample user, and the sample labels of the labeled sample user includes:
[0184] Obtaining an initial loss function of the labeled sample according to the target vector representation of the labeled sample user and the sample label of the labeled sample user;
[0185] Determining unlabeled sample users associated with the labeled sample users and unlabeled sample users not associated with the labeled sample users;
[0186] Obtaining an initial loss function of the unlabeled sample according to the target vector representation of the unlabeled sample user associated with the labeled sample user, the target vector representation of the unlabeled sample user not associated with the labeled sample user, and the sample label of the labeled sample user;
[0187] According to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained, and the graph neural network model is trained based on the target vector representation of each sample user and the target loss function.
[0188] Specifically, the graph neural network model training method provided in the embodiments of this specification is consistent with the graph neural network model training method in the target user identification method in the above embodiments. The specific details can be referred to the above embodiments and will not be repeated here.
[0189] The graph neural network model training method provided in the embodiments of this specification distinguishes different types of edges between two nodes by defining parameter matrices of different relationships, and uses the attention mechanism to learn the importance of different relationships; adds user-APP interaction information to the network, purposefully learns the user's preference for different APPs, and designs the network's loss function to effectively combine unlabeled samples with labeled samples for training in the full sample set, thereby obtaining a more accurate graph neural network model.
[0190] Corresponding to the above method embodiment, this specification also provides a target user identification device embodiment, Figure 7 FIG. 1 shows a schematic diagram of the structure of a target user identification device provided by an embodiment of the present specification. Figure 7 As shown, the device comprises:
[0191] A first acquisition module 702 is configured to acquire basic user characteristics of a user, associated users of the user, and associated applications of the user;
[0192] A first graph construction module 704 is configured to construct a relational knowledge graph according to the association relationship between the user and the associated user, and the association relationship between the user and the associated application;
[0193] A vector representation obtaining module 706 is configured to input the relational knowledge graph into a graph neural network model to obtain a vector representation of the user;
[0194] The recognition module 708 is configured to input the basic user features of the user and the target vector representation of the user into a recognition model to obtain a recognition result of whether the user is a target user.
[0195] Optionally, the graph neural network model is trained through the following modules:
[0196] A second acquisition module is configured to acquire associated users of each sample user among a plurality of sample users and associated applications of each sample user;
[0197] A second graph construction module is configured to construct a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, wherein the sample relationship knowledge graph includes user nodes and program nodes;
[0198] A first interaction feature determination module is configured to determine a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and an interaction feature between the user node and the program node;
[0199] The first training module is configured to train the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
[0200] Optionally, the first interaction feature determination module is further configured to:
[0201] Taking each sample user and the associated users of each sample user as the user nodes of the sample relationship knowledge graph, and taking the associated application of each sample user as the program node of the sample relationship knowledge graph, and determining the feature matrix of the user node and the program node in the sample relationship knowledge graph;
[0202] Using different types of edges formed between each sample user and the associated users of each sample user as edges of the user node, and configuring a different relationship matrix for each edge of the user node;
[0203] Determine the weight matrix of each relationship matrix through the attention mechanism; and
[0204] The interaction feature between the user node and the program node is determined according to the feature matrix of the user node, the relationship matrix, the weight matrix and the feature matrix of the program node.
[0205] Optionally, the first interaction feature determination module is further configured to include:
[0206] Calculating the feature matrix of the user node, the relationship matrix, and the weight matrix through a first aggregation algorithm to obtain a vector representation of the user node;
[0207] Calculating the feature matrix of the program node using a second aggregation algorithm to obtain a vector representation of the program node;
[0208] The vector representation of the user node and the vector representation of the program node are interactively processed to obtain an interactive feature of the user node and the program node.
[0209] Optionally, the first training module is further configured to:
[0210] Based on the vector representation of the user node and the interaction characteristics between the user node and the program node, obtaining the target vector representation of each sample user;
[0211] Determine labeled sample users and unlabeled sample users among the multiple sample users, train a graph neural network model according to the target vector representations of the labeled sample users, the target vector representations of the unlabeled sample users, and the sample labels of the labeled sample users, to obtain the graph neural network model.
[0212] Optionally, the first training module is further configured to:
[0213] Obtaining an initial loss function of the labeled sample according to the target vector representation of the labeled sample user and the sample label of the labeled sample user;
[0214] Determining unlabeled sample users associated with the labeled sample users and unlabeled sample users not associated with the labeled sample users;
[0215] Obtaining an initial loss function of the unlabeled sample according to the target vector representation of the unlabeled sample user associated with the labeled sample user, the target vector representation of the unlabeled sample user not associated with the labeled sample user, and the sample label of the labeled sample user;
[0216] According to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained, and the graph neural network model is trained based on the target vector representation of each sample user and the target loss function.
[0217] Optionally, the recognition model is trained by the following modules:
[0218] A third acquisition module is configured to acquire user basic features of each sample user among a plurality of sample users, associated users of each sample user, and associated applications of each sample user;
[0219] A third graph construction module is configured to construct a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user;
[0220] A first vector representation acquisition module is configured to input the sample relationship knowledge graph into the graph neural network model to obtain a target vector representation of each sample user;
[0221] The second training module is configured to train the recognition model according to the basic user features of each sample user and the vector representation of each sample user.
[0222] Optionally, the first graph construction module 704 is further configured to:
[0223] The user, the user's associated users, and the user's associated applications are taken as nodes, the association relationship between the user and the associated users, and the association relationship between the user and the associated applications are taken as edges, and a relational knowledge graph is constructed based on the nodes and the edges.
[0224] The target user identification device provided in the embodiment of this specification is a target user (i.e., competitor) identification solution combined with a graph, which makes full use of the relational knowledge graph constructed by relational data, and uses the user's features from associated users and associated applications as a supplement to the missing features of the target user, thereby accurately identifying the competitor.
[0225] The above is a schematic scheme of a target user identification device of this embodiment. It should be noted that the technical scheme of the target user identification device and the technical scheme of the target user identification method described above belong to the same concept, and the details not described in detail in the technical scheme of the target user identification device can be referred to the description of the technical scheme of the target user identification method described above.
[0226] Corresponding to the above method embodiment, this specification also provides a graph neural network model training device embodiment, Figure 8 FIG. 1 is a schematic diagram showing the structure of a graph neural network model training device provided by an embodiment of the present specification. Figure 8 As shown, the device comprises:
[0227] A fourth acquisition module 802 is configured to acquire associated users of each sample user among a plurality of sample users and associated applications of each sample user;
[0228] The fourth graph construction module 804 is configured to construct a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, wherein the sample relationship knowledge graph includes user nodes and program nodes;
[0229] A feature acquisition module 806 is configured to determine a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and an interaction feature between the user node and the program node;
[0230] The third training module 808 is configured to train the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
[0231] Optionally, the feature acquisition module 806 is further configured to:
[0232] Taking each sample user and the associated users of each sample user as the user nodes of the sample relationship knowledge graph, and taking the associated application of each sample user as the program node of the sample relationship knowledge graph, and determining the feature matrix of the user node and the program node in the sample relationship knowledge graph;
[0233] Using different types of edges formed between each sample user and the associated users of each sample user as edges of the user node, and configuring a different relationship matrix for each edge of the user node;
[0234] Determine the weight matrix of each relationship matrix through the attention mechanism; and
[0235] The interaction feature between the user node and the program node is determined according to the feature matrix of the user node, the relationship matrix, the weight matrix and the feature matrix of the program node.
[0236] Optionally, the feature acquisition module 806 is further configured to:
[0237] Calculating the feature matrix of the user node, the relationship matrix, and the weight matrix through a first aggregation algorithm to obtain a vector representation of the user node;
[0238] Calculating the feature matrix of the program node using a second aggregation algorithm to obtain a vector representation of the program node;
[0239] The vector representation of the user node and the vector representation of the program node are interactively processed to obtain an interactive feature of the user node and the program node.
[0240] Optionally, the third training module 808 is further configured to:
[0241] Based on the vector representation of the user node and the interaction characteristics between the user node and the program node, obtaining the target vector representation of each sample user;
[0242] Determine labeled sample users and unlabeled sample users among the multiple sample users, train a graph neural network model according to the target vector representations of the labeled sample users, the target vector representations of the unlabeled sample users, and the sample labels of the labeled sample users, to obtain the graph neural network model.
[0243] Optionally, the third training module 808 is further configured to:
[0244] Obtaining an initial loss function of the labeled sample according to the target vector representation of the labeled sample user and the sample label of the labeled sample user;
[0245] Determining unlabeled sample users associated with the labeled sample users and unlabeled sample users not associated with the labeled sample users;
[0246] Obtaining an initial loss function of the unlabeled sample according to the target vector representation of the unlabeled sample user associated with the labeled sample user, the target vector representation of the unlabeled sample user not associated with the labeled sample user, and the sample label of the labeled sample user;
[0247] According to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained, and the graph neural network model is trained based on the target vector representation of each sample user and the target loss function.
[0248] The graph neural network model training device provided in the embodiments of this specification distinguishes different types of edges between two nodes by defining parameter matrices of different relationships, and uses the attention mechanism to learn the importance of different relationships; adds user-APP interaction information to the network, purposefully learns the user's preference for different APPs, and designs the network's loss function to effectively combine unlabeled samples with labeled samples for training in the full sample set, thereby obtaining a more accurate graph neural network model.
[0249] The above is a schematic scheme of a graph neural network model training device of this embodiment. It should be noted that the technical scheme of the graph neural network model training device and the technical scheme of the graph neural network model training method mentioned above belong to the same concept. For details not described in detail in the technical scheme of the graph neural network model training device, please refer to the description of the technical scheme of the graph neural network model training method mentioned above.
[0250] Fig. 9 The block diagram of a computing device 900 according to an embodiment of the present specification is shown. The components of the computing device 900 include but are not limited to a memory 910 and a processor 920. The processor 920 is connected to the memory 910 via a bus 930, and a database 950 is used to store data.
[0251] The computing device 900 also includes an access device 940 that enables the computing device 900 to communicate via one or more networks 960. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 940 may include one or more of any type of network interface (e.g., a network interface card (NIC)) that is wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.
[0252] In one embodiment of the present specification, the above components of the computing device 900 and Fig. 9 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Fig. 9 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.
[0253] The computing device 900 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 900 may also be a mobile or stationary server.
[0254] Among them, the processor 920 is used to execute the following computer-executable instructions, which, when executed by the processor, can implement the steps of the above-mentioned target user identification method or graph neural network model training method.
[0255] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the target user identification method or the graph neural network model training method described above belong to the same concept. For details not described in detail in the technical scheme of the computing device, please refer to the description of the technical scheme of the target user identification method or the graph neural network model training method described above.
[0256] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned target user identification method or graph neural network model training method.
[0257] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium belongs to the same concept as the technical scheme of the target user identification method or the graph neural network model training method mentioned above. For details not described in detail in the technical scheme of the storage medium, please refer to the description of the technical scheme of the target user identification method or the graph neural network model training method mentioned above.
[0258] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0259] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.
[0260] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.
[0261] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0262] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.
Claims
1. A target user identification method, comprising: Obtaining basic user characteristics of a user, associated users of the user, and associated applications of the user; Constructing a relational knowledge graph according to the relation between the user and the associated user, and the relation between the user and the associated application; The relational knowledge graph is input into a graph neural network model to obtain a vector representation of the user, wherein the training steps of the graph neural network model are as follows: Obtaining an associated user of each sample user among a plurality of sample users and an associated application of each sample user; According to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, a sample relationship knowledge graph is constructed, wherein the sample relationship knowledge graph includes user nodes and program nodes; Determine the feature matrix of the user node in the sample relational knowledge graph, the relation matrix of each edge of the user node, the weight matrix of each relation matrix, and the interaction features between the user node and the program node, wherein the feature matrix is the feature matrix of each user node and the feature matrix of each program node in the sample relational knowledge graph, the relation matrix is the relation matrix of different types of edge configurations for the user node, the weight matrix is the matrix assigned to each relation matrix through the attention mechanism, and the interaction features are the interaction features between different types of the user node and the program node; Training a graph neural network model according to the feature matrix, the relationship matrix, the weight matrix, and the interaction features to obtain the graph neural network model; The user basic features of the user and the target vector representation of the user are input into a recognition model to obtain a recognition result of whether the user is a target user.
2. According to the target user identification method of claim 1, the determining of the feature matrix of the user node in the sample relationship knowledge graph, the relationship matrix of each edge of the user node, the weight matrix of each relationship matrix, and the interaction features between the user node and the program node comprises: Taking each sample user and the associated users of each sample user as the user nodes of the sample relationship knowledge graph, and taking the associated application of each sample user as the program node of the sample relationship knowledge graph, and determining the feature matrix of the user node and the program node in the sample relationship knowledge graph; Using different types of edges formed between each sample user and the associated users of each sample user as edges of the user node, and configuring a different relationship matrix for each edge of the user node; Determine the weight matrix of each relationship matrix through the attention mechanism; as well as The interaction feature between the user node and the program node is determined according to the feature matrix of the user node, the relationship matrix, the weight matrix and the feature matrix of the program node.
3. The target user identification method according to claim 2, wherein determining the interaction features between the user node and the program node according to the feature matrix of the user node, the relationship matrix, the weight matrix, and the feature matrix of the program node comprises: Calculating the feature matrix of the user node, the relationship matrix, and the weight matrix through a first aggregation algorithm to obtain a vector representation of the user node; Calculating the feature matrix of the program node using a second aggregation algorithm to obtain a vector representation of the program node; The vector representation of the user node and the vector representation of the program node are interactively processed to obtain an interactive feature of the user node and the program node.
4. According to the target user identification method of claim 3, the training of the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix and the interaction feature to obtain the graph neural network model comprises: Based on the vector representation of the user node and the interaction characteristics between the user node and the program node, obtaining the target vector representation of each sample user; Determine labeled sample users and unlabeled sample users among the multiple sample users, train a graph neural network model according to the target vector representations of the labeled sample users, the target vector representations of the unlabeled sample users, and the sample labels of the labeled sample users, to obtain the graph neural network model.
5. According to the target user identification method of claim 4, the training of the graph neural network model according to the target vector representation of the labeled sample user, the target vector representation of the unlabeled sample user, and the sample label of the labeled sample user comprises: Obtaining an initial loss function of the labeled sample according to the target vector representation of the labeled sample user and the sample label of the labeled sample user; Determining unlabeled sample users associated with the labeled sample users and unlabeled sample users not associated with the labeled sample users; Obtaining an initial loss function of the unlabeled sample according to the target vector representation of the unlabeled sample user associated with the labeled sample user, the target vector representation of the unlabeled sample user not associated with the labeled sample user, and the sample label of the labeled sample user; According to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained, and the graph neural network model is trained based on the target vector representation of each sample user and the target loss function.
6. According to the target user identification method of claim 1, the training steps of the identification model are as follows: Obtaining user basic features of each sample user among multiple sample users, associated users of each sample user, and associated applications of each sample user; Constructing a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user; Inputting the sample relationship knowledge graph into the graph neural network model to obtain a target vector representation of each sample user; The recognition model is trained according to the basic user features of each sample user and the vector representation of each sample user.
7. The target user identification method according to claim 1, wherein the step of constructing a relational knowledge graph based on the relation between the user and the associated user, and the relation between the user and the associated application, comprises: The user, the user's associated users, and the user's associated applications are taken as nodes, the association relationship between the user and the associated users, and the association relationship between the user and the associated applications are taken as edges, and a relational knowledge graph is constructed based on the nodes and the edges.
8. A graph neural network model training method, comprising: Obtaining an associated user of each sample user among a plurality of sample users and an associated application of each sample user; According to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, a sample relationship knowledge graph is constructed, wherein the sample relationship knowledge graph includes user nodes and program nodes; Determine the feature matrix of the user node in the sample relational knowledge graph, the relation matrix of each edge of the user node, the weight matrix of each relation matrix, and the interaction features between the user node and the program node, wherein the feature matrix is the feature matrix of each user node and the feature matrix of each program node in the sample relational knowledge graph, the relation matrix is the relation matrix of different types of edge configurations for the user node, the weight matrix is the matrix assigned to each relation matrix through the attention mechanism, and the interaction features are the interaction features between different types of the user node and the program node; The graph neural network model is trained according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
9. The method for training a graph neural network model according to claim 8, wherein determining the feature matrix of the user node in the sample relationship knowledge graph, the relationship matrix of each edge of the user node, the weight matrix of each relationship matrix, and the interaction features between the user node and the program node comprises: Taking each sample user and the associated users of each sample user as the user nodes of the sample relationship knowledge graph, and taking the associated application of each sample user as the program node of the sample relationship knowledge graph, and determining the feature matrix of the user node and the program node in the sample relationship knowledge graph; Using different types of edges formed between each sample user and the associated users of each sample user as edges of the user node, and configuring a different relationship matrix for each edge of the user node; Determine the weight matrix of each relationship matrix through the attention mechanism; as well as The interaction feature between the user node and the program node is determined according to the feature matrix of the user node, the relationship matrix, the weight matrix and the feature matrix of the program node.
10. The graph neural network model training method according to claim 9, wherein determining the interaction features between the user node and the program node according to the feature matrix of the user node, the relationship matrix, the weight matrix, and the feature matrix of the program node comprises: Calculating the feature matrix of the user node, the relationship matrix, and the weight matrix through a first aggregation algorithm to obtain a vector representation of the user node; Calculating the feature matrix of the program node using a second aggregation algorithm to obtain a vector representation of the program node; The vector representation of the user node and the vector representation of the program node are interactively processed to obtain an interactive feature of the user node and the program node.
11. The graph neural network model training method according to claim 10, wherein the graph neural network model is trained according to the feature matrix, the relationship matrix, the weight matrix, and the interaction feature to obtain the graph neural network model, comprising: Based on the vector representation of the user node and the interaction characteristics between the user node and the program node, obtaining the target vector representation of each sample user; Determine labeled sample users and unlabeled sample users among the multiple sample users, train a graph neural network model according to the target vector representations of the labeled sample users, the target vector representations of the unlabeled sample users, and the sample labels of the labeled sample users, to obtain the graph neural network model.
12. The graph neural network model training method according to claim 11, wherein the graph neural network model is trained according to the target vector representation of the labeled sample users, the target vector representation of the unlabeled sample users, and the sample labels of the labeled sample users, comprising: Obtaining an initial loss function of the labeled sample according to the target vector representation of the labeled sample user and the sample label of the labeled sample user; Determining unlabeled sample users associated with the labeled sample users and unlabeled sample users not associated with the labeled sample users; Obtaining an initial loss function of the unlabeled sample according to the target vector representation of the unlabeled sample user associated with the labeled sample user, the target vector representation of the unlabeled sample user not associated with the labeled sample user, and the sample label of the labeled sample user; According to the initial loss function of the labeled samples and the initial loss function of the unlabeled samples, the target loss function of the graph neural network model is obtained, and the graph neural network model is trained based on the target vector representation of each sample user and the target loss function.
13. A target user identification device, comprising: A first acquisition module is configured to acquire basic user characteristics of a user, associated users of the user, and associated applications of the user; A first graph construction module is configured to construct a relational knowledge graph according to the association relationship between the user and the associated user, and the association relationship between the user and the associated application; The vector representation acquisition module is configured to input the relational knowledge graph into a graph neural network model to obtain the vector representation of the user, wherein the training steps of the graph neural network model are as follows: Obtaining an associated user of each sample user among a plurality of sample users and an associated application of each sample user; According to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, a sample relationship knowledge graph is constructed, wherein the sample relationship knowledge graph includes user nodes and program nodes; Determine the feature matrix of the user node in the sample relational knowledge graph, the relation matrix of each edge of the user node, the weight matrix of each relation matrix, and the interaction features between the user node and the program node, wherein the feature matrix is the feature matrix of each user node and the feature matrix of each program node in the sample relational knowledge graph, the relation matrix is the relation matrix of different types of edge configurations for the user node, the weight matrix is the matrix assigned to each relation matrix through the attention mechanism, and the interaction features are the interaction features between different types of the user node and the program node; Training a graph neural network model according to the feature matrix, the relationship matrix, the weight matrix, and the interaction features to obtain the graph neural network model; The recognition module is configured to input the user basic features of the user and the target vector representation of the user into the recognition model to obtain a recognition result of whether the user is a target user.
14. A graph neural network model training device, comprising: A fourth acquisition module is configured to acquire associated users of each sample user among a plurality of sample users and associated applications of each sample user; A fourth graph construction module is configured to construct a sample relationship knowledge graph according to the association relationship between each sample user and the associated user of each sample user, and the association relationship between each sample user and the associated application of each sample user, wherein the sample relationship knowledge graph includes user nodes and program nodes; A feature acquisition module is configured to determine a feature matrix of the user node in the sample relationship knowledge graph, a relationship matrix of each edge of the user node, a weight matrix of each relationship matrix, and an interaction feature between the user node and the program node, wherein the feature matrix is a feature matrix of each user node and a feature matrix of each program node in the sample relationship knowledge graph, the relationship matrix is a relationship matrix configured for different types of edges of the user node, the weight matrix is a matrix assigned to each relationship matrix through an attention mechanism, and the interaction feature is an interaction feature between different types of user nodes and program nodes; The third training module is configured to train the graph neural network model according to the feature matrix, the relationship matrix, the weight matrix and the interaction features to obtain the graph neural network model.
15. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the target user identification method described in any one of claims 1 to 7 or the graph neural network model training method described in any one of claims 8 to 12 are implemented.
16. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the target user identification method described in any one of claims 1 to 7 or the graph neural network model training method described in any one of claims 8 to 12.
