Group key person identification method, electronic equipment and readable storage medium

Abstract

The invention discloses a group key person identification method, electronic equipment and a readable storage medium, and belongs to the technical field of mobile communication. According to the method, the target sub-graph is sent to the terminal equipment corresponding to the target entity, so that the terminal equipment carries out fusion based on the received target sub-graph, and a global heterogeneous network graph corresponding to the target entity after multiple knowledge graphs are linked is obtained; training a local business model according to the global heterogeneous network atlas to obtain local training gradient data, and sending the local training gradient data to a service party; the service party fuses the local training gradient data provided by each service party to obtain global gradient data corresponding to the graph federal learning model, and feeds back the global gradient data to each service party; and respectively updating model parameters of a local service model by each service party according to the fused global gradient data. According to the invention, the group key person identification precision and efficiency can be improved.

Description

Technical Field

[0001] The present application relates to the field of mobile communication technology, and in particular to a method for identifying key persons in a group, an electronic device, and a readable storage medium. Background Art

[0002] As competition in the individual market intensifies, the family market is emerging as a new competitive arena. Family services are becoming a key revenue driver. Therefore, it's imperative to fully tap the value of user data for individuals and families, strengthen data operations capabilities, and achieve refined user operations in segmented scenarios by identifying family groups and key individuals.

[0003] At present, the federated learning model for group key person identification in family groups is mainly based on user feature data such as user location and call behavior, and family groups are formed through aggregation and distance calculation. This method is still rule-based in user feature processing. For example, it is based on simple relationships such as short number services and location overlap for identification. It does not fully utilize the user's call relationship and the characteristics of the base station time and spatial structure for mining. As a result, the model cannot fully utilize the content information characteristics of the node, and performs poorly on heterogeneous graphs with rich node information, which leads to low accuracy and slow speed in group key person identification. It may even cause the model to fail to converge, making it difficult to produce meaningful results. Summary of the Invention

[0004] The main purpose of this application is to provide a method for identifying key persons in a group, an electronic device, and a readable storage medium, aiming to improve the accuracy and efficiency of identifying key persons in a group.

[0005] To achieve the above objectives, the present application provides a method for identifying key individuals in a group. The method is based on a graph federated learning model. The graph federated learning model is jointly trained by multiple business parties based on local training gradient data of a knowledge graph. A single knowledge graph is formed by the business party to which it belongs based on the private data of each entity it owns, and includes the entity identifiers of each entity, the private data of each entity, and the associations between the entities. The method is executed by any one of the multiple business parties and includes:

[0006] Based on the knowledge graph owned by the party, any one entity is selected from the entities as the target entity, and a target subgraph corresponding to the target entity is generated using a neighborhood sampling mechanism;

[0007] Sending the target subgraph to a terminal device corresponding to the target entity, so that the terminal device fuses the target subgraphs received from the multiple business parties to obtain a global heterogeneous network graph corresponding to the target entity after linking multiple knowledge graphs;

[0008] Training a local service model according to the global heterogeneous network graph to obtain local training gradient data, and sending the local training gradient data to a service provider;

[0009] In response to the graph federated learning model meeting the gradient update conditions, the service provider integrates the local training gradient data provided by each business party to obtain the global gradient data corresponding to the graph federated learning model, and feeds the global gradient data back to each business party;

[0010] Each business party updates the model parameters of the local business model based on the fused global gradient data to iteratively train the local business model until the gradient converges and training is stopped to obtain the target business model. The target business model identifies key people in the group based on the input knowledge graph.

[0011] Optionally, the step of generating a target subgraph corresponding to the target entity by using a neighborhood sampling mechanism includes:

[0012] Randomly sampling different types of neighbor nodes of the target entity to obtain randomly sampled neighbor nodes;

[0013] According to the weight from high to low, fixed-length sampling is performed on each randomly sampled neighbor node to obtain each fixed-length sampled neighbor node;

[0014] The neighbor nodes sampled with fixed length are aggregated and updated to obtain the target subgraph corresponding to the target entity.

[0015] Optionally, the step of aggregating and updating the fixed-length sampled neighbor nodes to obtain a target subgraph corresponding to the target entity includes:

[0016] Aggregate and update the neighbor nodes of the fixed-length sampling to obtain the aggregated node embedding corresponding to the target entity;

[0017] Variable screening is performed based on the aggregation node embedding to obtain key dimension data;

[0018] A target subgraph corresponding to the target entity is generated according to the key dimension data.

[0019] Optionally, the key dimension data includes user call data, user subscription service data, and user base station data, and the step of generating a target subgraph corresponding to the target entity according to the key dimension data includes:

[0020] Constructing a call relationship network map between different users in a target time period based on the user call data;

[0021] Constructing a subscription network graph between users and services based on the user subscription service data;

[0022] Constructing a location relationship network map of base station locations where calls occur between different users based on the user base station data;

[0023] Based on the call relationship network graph, the ordering network graph and the location relationship network graph, a spatiotemporal heterogeneous network graph is constructed, and the spatiotemporal heterogeneous network graph is used as the target subgraph corresponding to the target entity.

[0024] Optionally, before the step of selecting any one entity from the entities as the target entity based on the knowledge graph owned by the party, the method further includes:

[0025] Obtaining sample data, the sample data including key data within a preset time period in a user's call details, wherein the key data includes family network data, group user information, enterprise payment and reimbursement information, and enterprise virtual network information;

[0026] Performing group division and identification based on the sample data to obtain multiple social groups, and determining a positive sample relationship pair set and a negative sample relationship pair set based on each of the social groups;

[0027] Based on the positive sample relationship pair set and the negative sample relationship pair set, group data cascade is performed to determine the knowledge graph possessed by the party, wherein the knowledge graph possessed by the party includes call behavior information, business processing information and geographic location information of the base station.

[0028] Optionally, the step of dividing and identifying groups according to the sample data to obtain multiple social groups includes:

[0029] The Fast Unfolding algorithm is used to perform group division and identification in combination with the sample data to obtain multiple social groups.

[0030] Optionally, the step of sending the target subgraph to a terminal device corresponding to the target entity includes:

[0031] Encrypt the private data of each entity in the target subgraph using the fully homomorphic encryption method negotiated with other business parties;

[0032] The encrypted target subgraph is sent to a terminal device corresponding to the target entity.

[0033] Optionally, the target subgraph includes multiple nodes and edges for connecting the nodes, the node attributes of the nodes include user basic information, the edge attributes of the edges include user call details information, service subscription information, and base station location information, wherein the user basic information includes user age information and user gender information.

[0034] In addition, to achieve the above-mentioned purpose, the present application also provides an electronic device, which includes: a memory, a processor, and a group key person identification program stored in the memory and runnable on the processor. When the group key person identification program is executed by the processor, the group key person identification method as described above is implemented.

[0035] In addition, to achieve the above-mentioned purpose, the present application also provides a readable storage medium, which is a computer-readable storage medium. A group key person identification program is stored on the computer-readable storage medium. When the group key person identification program is executed by the processor, the group key person identification method as described above is implemented.

[0036] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a group key person identification program, and when the group key person identification program is executed by a processor, it implements the steps of the group key person identification method as described above.

[0037] The present application proposes a group key person identification method, electronic device and readable storage medium, wherein the group key person identification method is based on a graph federated learning model, the graph federated learning model is jointly trained by multiple business parties based on local training gradient data of the knowledge graph, a single knowledge graph is formed by the business party to which it belongs based on the privacy data of each entity it owns, and includes the entity identification of each entity, the privacy data of each entity and the association relationship between the entities, the method is executed by any business party among the multiple business parties, the method includes: based on the knowledge graph owned by the party, selecting any one entity from each entity as the target entity, using the neighborhood sampling mechanism to generate a target subgraph corresponding to the target entity, and sending the target subgraph to the terminal device corresponding to the target entity, so that the terminal device can access the data from multiple business parties. The service party fuses the target subgraphs received separately to obtain the global heterogeneous network graph corresponding to the target entity after linking multiple knowledge graphs. Then, based on the global heterogeneous network graph, the local business model is trained to obtain local training gradient data, and the local training gradient data is sent to the service party. Then, in response to the graph federated learning model meeting the gradient update conditions, the service party fuses the local training gradient data provided by each business party to obtain the global gradient data corresponding to the graph federated learning model, and feeds the global gradient data back to each business party. Each business party updates the model parameters of the local business model according to the fused global gradient data to iteratively train the local business model until the gradient converges and the training is stopped to obtain the target business model. The target business model performs group key person identification based on the input knowledge graph.

[0038] This application adopts a graph federated learning method to train a graph convolutional neural network model through collaboration among multiple participants without sharing their private data, in order to solve the data isolation problem. By optimizing the GraphSAGE algorithm, a group and key person identification method based on ultra-large heterogeneous graph federated learning is implemented, which fully utilizes the characteristics of users' call relationships and base station time and spatial structure for mining, so that the model can fully utilize the content information characteristics of the nodes and perform better on heterogeneous graphs with rich node information, thereby improving the accuracy and efficiency of group key person identification.

[0039] It is worth mentioning that this application constructs an ultra-large virtual heterogeneous graph network through horizontal federated learning, and constructs a virtual social network by combining multi-party data, which expands the graph data space and makes the relationship graph more complete. It solves the problem of data silos and isolation while ensuring data privacy and security. In addition, this application also optimizes the GraphSAGE algorithm and solves the problem that GraphSAGE is only applicable to homogeneous graphs by adopting a heterogeneous network aggregator, thereby realizing a group and key person identification method based on ultra-large heterogeneous graph federated learning. The R-GraphSAGE algorithm uses a sampling mechanism to solve the memory and video memory limitations during full-graph training, and can obtain embedding representations even for unknown new nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0041] Figure 1 This is a flow chart of the first embodiment of the method for identifying key persons in the present application group;

[0042] Figure 2 This is a flow chart of the second embodiment of the method for identifying key persons in the present application group;

[0043] Figure 3 A schematic diagram of the process scenario of the key person identification method for this application group;

[0044] Figure 4 A schematic diagram of a construction scenario of a heterogeneous network graph according to an embodiment of the present application;

[0045] Figure 5 This is a schematic diagram of the hardware structure of the electronic device involved in the embodiment of the present application.

[0046] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0047] It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0048] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0049] It should be noted that all directional indications in the embodiments of the present application (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0050] In this application, unless otherwise specified or limited, the terms "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can mean fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two elements or interaction between two elements, unless otherwise specified. For those skilled in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0051] In addition, the descriptions of "first", "second", etc. in this application are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0052] At present, the federated learning model for group key person identification in family groups is mainly based on user feature data such as user location and call behavior, and family groups are formed through aggregation and distance calculation. This method is still rule-based in user feature processing. For example, it is based on simple relationships such as short number services and location overlap for identification. It does not fully utilize the user's call relationship and the characteristics of the base station time and spatial structure for mining. As a result, the model cannot fully utilize the content information characteristics of the node, and performs poorly on heterogeneous graphs with rich node information, which leads to low accuracy and slow speed in group key person identification. It may even cause the model to fail to converge, making it difficult to produce meaningful results.

[0053] Based on this, the embodiment of the present application provides a method for identifying key persons in a group, referring to Figure 1 , Figure 1 This is a flow chart of an embodiment of a group key person identification method of the present application. In this embodiment, the group key person identification method is based on a graph federated learning model, which is jointly trained by multiple business parties based on local training gradient data of a knowledge graph. A single knowledge graph is formed by the business party based on the private data of each entity it owns, and includes the entity identifiers of each entity, the private data of each entity, and the relationships between the entities. The method is executed by any one of the multiple business parties and includes:

[0054] Step S10: Based on the knowledge graph owned by the party, any one entity is selected from the entities as the target entity, and a target subgraph corresponding to the target entity is generated using a neighborhood sampling mechanism;

[0055] Those skilled in the art will know that neighborhood sampling is also called neighbor sampling. Those skilled in the art have already conducted in-depth research on this, which will not be elaborated here.

[0056] Understandably, knowledge graphs are typically constructed from nodes that correspond one-to-one to entities, and edges connecting nodes (corresponding to the relationships between entities), resulting in potentially complex structures. For different business entities, if their business data contains different entities or different relationships, the graph data structure will differ significantly. For example, if one business entity is a payment platform, in the graph data constructed using its business data, the entities corresponding to the nodes may be users, and the relationships between nodes may include address book friends, transfer frequency, and so on. Another business entity is a shopping platform, and the entities corresponding to the nodes may include both consumers and merchants. The relationships between nodes may include sharing product links between consumers or between merchants and consumers, and consumers purchasing products from merchants, and so on. More business entities can also include graph data with more diverse structures. In other words, each business entity can construct one or more types of graph data based on its own business data, forming a "heterogeneous graph." In this case, it is difficult to process the local graph data of each business entity using a unified graph neural network.

[0057] Under the technical concept of this embodiment, the concept of a model basis is proposed. A model basis can be considered the basic parameter unit that constitutes a business model (such as a graph neural network). A model basis can contain several baseline model parameters (or baseline parameters). The model basis is shared by all business parties, and business parties locally build their own business models, such as graph neural networks, based on the model basis. The basic architecture of the model basis can be agreed upon by all business parties, for example, the number of model bases, the dimensions of each model base, and the meaning of each dimension. The activation functions, optimization methods, and task types to be predicted in the business model can also be agreed upon by each business party. When a single business party builds a business model, its model parameters can be defined by a combination of model bases in a predetermined manner. For example, the model parameters can be a linear combination of model bases, or they can be determined by combining other combinations of model bases (such as network structure search (NAS)). When building a graph neural network, the parameters of each layer of the neural network can be a linear combination of each model basis. The model basis can be in the form of a vector or matrix composed of baseline parameters. Model parameters can be represented in vector or matrix form. The following example uses the linear combination of model bases to construct model parameters.

[0058] Exemplarily, the target subgraph includes multiple nodes and edges for connecting the nodes, the node attributes of the nodes include user basic information, the edge attributes of the edges include user call details information, service subscription information, and base station location information, wherein the user basic information includes user age information and user gender information.

[0059] Step S20: sending the target subgraph to a terminal device corresponding to the target entity, so that the terminal device fuses the target subgraphs received from the multiple business parties to obtain a global heterogeneous network graph corresponding to the target entity after linking multiple knowledge graphs;

[0060] Exemplarily, in step S20, the step of sending the target subgraph to a terminal device corresponding to the target entity includes:

[0061] Step A10: Encrypt the private data of each entity in the target subgraph using the fully homomorphic encryption method negotiated with other business parties;

[0062] Step A20: sending the encrypted target subgraph to a terminal device corresponding to the target entity.

[0063] In this embodiment, those skilled in the art will appreciate that, since fully homomorphic encryption supports arbitrary calculations on ciphertext without decryption, it can immediately solve data privacy and security issues and has great application demand. For example, in a cloud environment, users encrypt data and store it in the cloud. Due to data encryption, the cloud cannot obtain the content of the data, thereby ensuring the privacy of the data. In addition, due to fully homomorphic encryption, the cloud can perform arbitrary calculations on the ciphertext data. In short, fully homomorphic encryption not only protects data through encryption, but also does not lose computational performance.

[0064] After step S20, step S30 is executed to train a local business model according to the global heterogeneous network graph, obtain local training gradient data, and send the local training gradient data to the service provider;

[0065] In step S40, in response to the graph federated learning model satisfying the gradient update condition, the service party integrates the local training gradient data provided by each business party to obtain the global gradient data corresponding to the graph federated learning model, and feeds the global gradient data back to each business party;

[0066] In step S50, each business party updates the model parameters of the local business model based on the fused global gradient data to iteratively train the local business model until the gradient converges and the training is stopped to obtain the target business model. The target business model is based on the input knowledge graph to identify key people in the group.

[0067] In this embodiment, the global model is stored on the central server, and the local models are stored on all the clients communicating with the central server.

[0068] Wherein, a local subgraph of the global graph distribution is stored on all clients communicating with the central server, and the graph structure data nodes of the local subgraph on each client do not overlap with each other.

[0069] In each iteration, a subset of clients is randomly selected from all clients to participate in the iteration. After each iteration, the local model parameters of each client in the subset are collected. The client samples a corresponding subset of nodes based on the sampling probability of each graph structure data node on the local subgraph. The local model on the client is trained based on this subset of nodes to obtain local training gradient data.

[0070] Specifically, all clients calculate the sampling probability of each graph structure data node on their local subgraph. At the beginning of each iteration, the central server randomly selects a subset of clients from all clients to participate in the iteration. Based on the sampling probability of each client in the client subset, the corresponding node subset is collected. Each node subset is used to train the local model on the client corresponding to the client subset, and the local training gradient data of the trained local model is obtained.

[0071] After each round of iteration, the local model parameters of each client in the client subset are averaged and aggregated to obtain updated global gradient data, and the updated global gradient data is used as the initial model parameters of the local model in the next round of iteration until the node classification accuracy of the global model corresponding to the updated global gradient data reaches a preset threshold on the global graph distribution test set that conforms to the local subgraphs of all clients.

[0072] Specifically, after each round of iteration, the central server collects the local model parameters of each client in the client subset selected in the iteration round and averages and aggregates them to obtain updated global gradient data, and sends the updated global gradient data to all clients as the initial model parameters of the local model for the next round of iteration, until the node classification accuracy of the global model corresponding to the updated global gradient data on the central server reaches a preset threshold on the global graph distribution test set that conforms to the local subgraphs of all clients. This can solve the problems of low efficiency and low precision of federated learning in the scenario of graph structured data in related technologies, effectively suppress the degree of deviation of each local model in graph federated learning, and at the same time reduce the storage and computing costs of local model training, and improve the accuracy of node classification of graph structured data.

[0073] In an embodiment of the present application, federated learning based on homomorphic encryption and the R-GraphSAGE graph learning algorithm are used to enrich the graph data space by combining multi-party data to construct virtual call graph data, making the user call graph relationship network more complete. Under the premise of ensuring data privacy and security, the data isolation problem is solved, and the groups and key people among the members of different networks are effectively identified. The embodiment of the present application solves the problem that GraphSAGE is only applicable to isomorphic graphs by using a heterogeneous network aggregator, expands the use and popularization of the algorithm in multi-relationship graph structures, and uses a sampling mechanism to solve the limitation that traditional graph convolutional neural networks cannot handle very large heterogeneous graphs, and can also obtain representation embeddings for unknown new nodes. While improving the GraphSAGE algorithm to solve the problem of heterogeneous graphs, it uses a sampling mechanism to solve the limitations of memory and video memory during full graph information training, so that even unknown new nodes can be represented.

[0074] It is worth mentioning that this embodiment can be widely used in scenarios such as family groups, key family members, telecommunications anti-harassment, anti-fraud, family marketing and financial risk control, and can provide financial risk control and group identification services to improve risk control capabilities.

[0075] This embodiment can learn node information and predict heterogeneous graph node information by sampling current neighbor nodes in a heterogeneous graph structure network. It has broad application value in large-scale heterogeneous graph networks, can save model training time, and solve the problem of model inability to update. At this stage, it is widely used in group key person identification and other aspects.

[0076] Through the solution provided by this embodiment, multiple business factors are used to identify family relationship pairs, and multiple types of business that can indicate family relationships are comprehensively considered. Businesses are classified according to the importance of the connection between business and family relationships. Modeling is carried out based on this starting point, making the identification of family relationships more in-depth. Family relationship pairs are identified using location distance, thereby improving the accuracy of relationship identification.

[0077] In addition, in order to address the problem that each business party provides graph data of different structures (heterogeneous graphs) in the federated learning process and cannot be modeled uniformly, a technical concept of consistent local model parameters is proposed. Through the setting of the model base, a model base can include several benchmark parameters. Each business party uses the combination of the model base in a predetermined manner to construct a local model parameter vector or matrix, thereby constructing a local business model. In the federated learning process, the business party uploads local gradient data in units of model bases, the service party updates global gradient data in units of model bases, and the business party updates local model parameters in units of model bases. This technical solution of locally updating model parameters can effectively break the barriers of federated learning under heterogeneous graphs, provide a new idea for machine learning, and can be extended to various business models, not limited to machine learning models that process graph data.

[0078] During federated learning, each business entity can maintain local sample data, which is private and not intended for other parties to access. During federated learning, each business entity trains its business model using local sample data to obtain the local gradients of its model parameters.

[0079] The service provider can provide assistance to the federated learning of each business party, for example, assisting in nonlinear calculations, global model parameter gradient calculations, etc. The service provider can be in the form of other parties independently set up from each business party, such as a trusted third party. In practice, the service provider can also be distributed among various business parties, or composed of various business parties, and the various business parties can use secure computing protocols (such as secret sharing, etc.) to complete joint auxiliary computing. This specification does not limit this.

[0080] Each business party can also obtain the gradient data of the global model parameters from the service party to update the local model parameters.

[0081] When business data is described using a graph structure, forming graph data, the business model can be a machine learning model such as graph neural networks, RDF2Vec, or Weisfeiler-Lehman kernels (WL). During federated learning, each business entity can combine pre-built graph data and process relevant training samples using the corresponding business model to generate output results.

[0082] For example, graph neural networks are generalized neural networks based on graph structures (i.e., graph data). They use the underlying graph data as a computational graph and learn neural network primitives to generate node embedding vectors by transferring, transforming, and aggregating node feature information across the entire graph. These node embedding vectors can be used as input to any differentiable prediction layer for node classification or predicting connections between nodes. Complete graph neural networks can be trained in an end-to-end manner.

[0083] The embodiment of the present application proposes a group key person identification method, electronic device and readable storage medium, wherein the group key person identification method is based on a graph federated learning model, the graph federated learning model is jointly trained by multiple business parties based on local training gradient data of the knowledge graph, a single knowledge graph is formed by the business party to which it belongs based on the privacy data of each entity it owns, and includes the entity identification of each entity, the privacy data of each entity and the association relationship between each entity, the method is executed by any business party among the multiple business parties, the method includes: based on the knowledge graph owned by the party, selecting any one entity from each entity as the target entity, using the neighborhood sampling mechanism to generate a target subgraph corresponding to the target entity, and sending the target subgraph to the terminal device corresponding to the target entity, so that the terminal device can access the data from multiple entities. The business parties fuse the target subgraphs received separately to obtain a global heterogeneous network graph corresponding to the target entity after linking multiple knowledge graphs. Then, based on the global heterogeneous network graph, the local business model is trained to obtain local training gradient data, and the local training gradient data is sent to the service party. Then, in response to the graph federated learning model meeting the gradient update conditions, the service party fuses the local training gradient data provided by each business party to obtain the global gradient data corresponding to the graph federated learning model, and feeds the global gradient data back to each business party. Each business party updates the model parameters of the local business model based on the fused global gradient data to iteratively train the local business model until the gradient converges and the training is stopped to obtain the target business model. The target business model performs group key person identification based on the input knowledge graph.

[0084] The embodiment of the present application adopts a graph federated learning method to train a graph convolutional neural network model through collaboration among multiple participants without sharing their private data, so as to solve the data isolation problem, and optimizes the GraphSAGE algorithm to realize a group and key person identification method based on ultra-large heterogeneous graph federated learning, which makes full use of the characteristics of users' call relationships and base station time and spatial structure for mining, so that the model can make full use of the content information characteristics of the nodes, and perform better on heterogeneous graphs with rich node information, thereby improving the accuracy and efficiency of group key person identification.

[0085] It is worth mentioning that the embodiment of the present application constructs an ultra-large virtual heterogeneous graph network through horizontal federated learning, and constructs a virtual social network by combining multi-party data, thereby expanding the graph data space and making the relationship graph more complete. Under the premise of ensuring data privacy and security, it solves the problem of data silos and isolation. In addition, the embodiment of the present application also optimizes the GraphSAGE algorithm and solves the problem that GraphSAGE is only applicable to homogeneous graphs by adopting a heterogeneous network aggregator, thereby realizing a group and key person identification method based on ultra-large heterogeneous graph federated learning. The R-GraphSAGE algorithm uses a sampling mechanism to solve the memory and video memory limitations during full graph training, and can obtain embedding representations even for unknown new nodes.

[0086] In one possible implementation, please refer to Figure 2 In step S10, the step of generating a target subgraph corresponding to the target entity by using a neighborhood sampling mechanism includes:

[0087] Step S11, randomly sampling different types of neighbor nodes of the target entity to obtain randomly sampled neighbor nodes;

[0088] Step S12, performing fixed-length sampling on each randomly sampled neighbor node according to weight from high to low, to obtain each fixed-length sampled neighbor node;

[0089] Step S13: aggregate and update the neighboring nodes sampled with fixed length to obtain a target subgraph corresponding to the target entity.

[0090] The embodiment of the present application randomly samples different types of neighbor nodes of the target entity to obtain randomly sampled neighbor nodes, and performs fixed-length sampling on each randomly sampled neighbor node according to the weight size from high to low to obtain each fixed-length sampled neighbor node, and then aggregates and updates each fixed-length sampled neighbor node to obtain a target subgraph corresponding to the target entity, so that the embodiment randomly samples neighbors of different edge types through R-GraphSAGE, and then performs fixed-length sampling from high to low according to the edge weight size, and uses a heterogeneous network aggregator to aggregate and update node embedding of the sampled different types of neighbor nodes, thereby finally achieving the processing of heterogeneous data.

[0091] Based on the first embodiment of the present application, a second embodiment of the group key person identification method of the present application is proposed. In step S13, the step of aggregating and updating the fixed-length sampled neighbor nodes to obtain the target subgraph corresponding to the target entity includes:

[0092] Step C10: aggregate and update the neighbor nodes of the fixed-length sampling to obtain the aggregated node embedding corresponding to the target entity;

[0093] Step C20: Filter variables based on the aggregation node embedding to obtain key dimension data;

[0094] Step C30: Generate a target subgraph corresponding to the target entity based on the key dimension data.

[0095] This embodiment aggregates and updates the neighboring nodes sampled from a fixed length to obtain the aggregated node embedding corresponding to the target entity, and performs variable screening based on the aggregated node embedding to obtain key dimension data. Then, based on the key dimension data, a target subgraph corresponding to the target entity is generated, thereby laying a solid foundation for subsequent heterogeneous network data processing. Since the network data constructed in actual business is basically heterogeneous networks, traditional graph convolutional neural networks and GraphSAGE algorithms are only suitable for processing homogeneous networks and cannot process heterogeneous networks. Based on this, this embodiment introduces a heterogeneous network processing algorithm by improving traditional algorithms to effectively identify group members and key people in heterogeneous network members. Using the optimized R-GraphSAGE algorithm to learn the feature information and structural information of nodes can deeply extract the hidden and complex patterns of graph features.

[0096] In a possible implementation, the key dimension data includes user call data, user subscription service data, and user base station data. Step C30, generating a target subgraph corresponding to the target entity based on the key dimension data, includes:

[0097] Step D10: constructing a call relationship network map between different users in a target time period based on the user call data;

[0098] Step D20: constructing a subscription network graph between the user and the service based on the user subscription service data;

[0099] Step D30, constructing a location relationship network map of base station locations where calls occur between different users based on the user base station data;

[0100] Step D40: constructing a spatiotemporal heterogeneous network graph based on the call relationship network graph, the ordering network graph, and the location relationship network graph, and using the spatiotemporal heterogeneous network graph as a target subgraph corresponding to the target entity.

[0101] The embodiment of the present application constructs a call relationship network graph between different users in a target time period based on user call data, constructs a subscription network graph between users and services based on user service subscription data, and constructs a location relationship network graph of base station locations where calls occur between different users based on user base station data. Based on the call relationship network graph, the subscription network graph, and the location relationship network graph, a spatiotemporal heterogeneous network graph is constructed, and the spatiotemporal heterogeneous network graph is used as the target subgraph corresponding to the target entity, so that this embodiment constructs a call relationship network graph between users at a certain moment from user call data, constructs a subscription network graph between users and services from user service subscription data, and constructs a location relationship network graph between users and base stations, and between base stations from user base station data, and finally constructs a complete spatiotemporal heterogeneous network graph, making the relationship graph more complete. On the premise of ensuring data privacy and security, the data isolation problem is solved, and groups and key people among members of different networks are effectively identified.

[0102] Through the solution provided by this embodiment, the family structure can be further analyzed, making it easier for users to further handle services that meet their family needs. The family users who are mainly responsible for handling services in the family structure can be further identified, the identification accuracy can be improved, and it is easier for family users to further customize the operator's services. The business handling, business usage and business affordability of family users can be judged, and the main key people in the family group can be further accurately identified. The roles of family members in the family group can be accurately identified, making it easier for family users to customize services that meet their needs. The contact frequency of family users' contacts can be analyzed to identify their terminal usage frequency, the terminal model of family users can be analyzed to identify their age, and the application software usage information of family users can be analyzed to identify their interests, thereby accurately identifying the age, gender and role of family users and achieving accurate classification. Avoid role identification errors in family groups and improve the accuracy of classification.

[0103] In a possible implementation, in step S10, before the step of selecting any one entity from the entities as the target entity based on the knowledge graph owned by the party, the method further includes:

[0104] Step E10: Acquire sample data, wherein the sample data includes key data within a preset time period in the user's call details, wherein the key data includes family network data, group user information, enterprise payment and reimbursement information, and enterprise virtual network information;

[0105] Step E20: performing group division and identification based on the sample data to obtain multiple social groups, and determining a positive sample relationship pair set and a negative sample relationship pair set based on each of the social groups;

[0106] Exemplarily, in step E20, the step of dividing and identifying groups based on the sample data to obtain multiple social groups includes:

[0107] In step F10 , the Fast Unfolding algorithm is used to perform group division and identification in combination with the sample data to obtain multiple social groups.

[0108] Those skilled in the art will appreciate that the Fast Unfolding algorithm is an iterative algorithm whose primary goal is to continuously divide communities so that the modularity of the entire network increases. The Fast Unfolding algorithm's algorithmic process is mainly divided into two stages:

[0109] The first stage is called Modularity Optimization, which mainly divides each node into the community of its adjacent nodes so that the modularity value continues to increase;

[0110] The second phase, called Community Aggregation, involves aggregating the communities identified in the first step into a single point, essentially reconstructing the network based on the community structure generated in the previous step. This process is repeated until the network structure no longer changes.

[0111] The specific algorithm process of the Fast Unfolding algorithm can be as follows:

[0112] 1. Initialization: divide each point into different communities;

[0113] 2. For each node, try to divide it into the community where its adjacent points are located, calculate the modularity at this time, and determine whether the difference in modularity before and after the division is positive. If it is positive, accept this division; if not, abandon this division;

[0114] 3. Repeat the above process until the modularity cannot be increased;

[0115] 4. Construct a new graph. Each point in the new graph represents each community drawn in step 3. Continue to perform steps 2 and 3 until the structure of the community no longer changes.

[0116] After step E20, execute step E30 to perform group data cascade based on the positive sample relationship pair set and the negative sample relationship pair set to determine the knowledge graph possessed by the party, wherein the knowledge graph possessed by the party includes call behavior information, business processing information and geographic location information of the base station.

[0117] This embodiment obtains sample data, which includes key data within a preset time period in the user's call details, wherein the key data includes family network data, group user information, enterprise payment information and enterprise virtual network information, and divides and identifies groups based on the sample data to obtain multiple social groups, and determines a positive sample relationship pair set and a negative sample relationship pair set based on each social group, and then cascades the group data based on the positive sample relationship pair set and the negative sample relationship pair set to determine the knowledge graph possessed by the party, wherein the knowledge graph possessed by the party includes call behavior information, business processing information and base station geographic location information, thereby filtering the call data for a period of time in the user's call details. By completing a series of key data such as user call rejection records to ensure the accuracy of the model, and based on family network data, group user information, corporate payment and reimbursement, corporate virtual network and other information, user pairs with call relationships are screened as a set of users with known group relationships. Among them, user pairs with the same corporate business ID or family network ID form a positive sample relationship pair set, and user pairs with different corporate business IDs or family network IDs form a negative relationship pair set (among them, users who have not handled the corresponding business but have call relationships and users whose group relationships cannot be determined form a set of user pairs with unknown group relationships). This facilitates the final formation of a family group with precise target users and accurate business positioning, which is conducive to the promotion of subsequent businesses.

[0118] Furthermore, in order to help understand the technical concept of the embodiments of the present application, a specific embodiment is listed, including:

[0119] Considering the shortcomings of related technical solutions, this embodiment employs a graph federated learning approach to address data isolation by enabling multiple participants to collaboratively train a graph convolutional neural network model without sharing their private data. The GraphSAGE algorithm is optimized to implement a group and key person identification method based on federated learning on very large heterogeneous graphs. Key technical highlights include: improving the GraphSAGE algorithm, namely the R-GraphSAGE algorithm, to enable it to process heterogeneous network data, and utilizing a sampling mechanism to address the limitation of traditional graph convolutional neural networks that cannot handle very large heterogeneous graphs.

[0120] (1) Heterogeneous network data processing

[0121] In practice, network data constructed is essentially heterogeneous. Traditional graph convolutional neural networks and the GraphSAGE algorithm are only suitable for processing homogeneous networks and cannot handle heterogeneous networks. Based on this, this embodiment improves traditional algorithms and introduces a heterogeneous network processing algorithm to effectively identify group members and key individuals in heterogeneous networks. The optimized R-GraphSAGE algorithm learns node feature information and structural information, enabling deep extraction of hidden and complex patterns in graph features.

[0122] (2) Processing of very large heterogeneous graphs

[0123] Traditional GCNs require storing the entire graph in memory and video memory, making them incapable of handling very large heterogeneous graphs. Furthermore, training requires knowledge of the entire graph's structure, including the nodes to be predicted, which is impossible in real-world scenarios. The R-GraphSAGE algorithm utilizes a sampling mechanism to address memory and video memory limitations during full-graph training, enabling representations even for unknown new nodes.

[0124] This embodiment constructs a large virtual heterogeneous graph network through horizontal federated learning, and constructs a virtual social network by combining multi-party data, which expands the graph data space and makes the relationship graph more complete. It solves the data island and isolation problems while ensuring data privacy and security. By optimizing the GraphSAGE algorithm and using a heterogeneous network aggregator to solve the problem that GraphSAGE is only applicable to homogeneous graphs, a group and key person identification method based on large heterogeneous graph federated learning is implemented. The R-GraphSAGE algorithm uses a sampling mechanism to solve the memory and video memory limitations during full graph training, and can obtain embedding representations even for unknown new nodes. The main processes and steps are as follows: Figure 3 shown.

[0125] Step 1: Secret calculation

[0126] Homomorphic encryption refers to an encryption algorithm that satisfies the property of homomorphic operation on ciphertext. That is, after the data is homomorphically encrypted, certain specific calculations are performed on the ciphertext. The resulting ciphertext calculation result is equivalent to the same calculation performed directly on the plaintext data after the corresponding homomorphic decryption, achieving "calculable but invisible" data. Existing homomorphic encryption includes partially homomorphic encryption (partially homomorphic), somewhat homomorphic encryption (somewhat homomorphic), and fully homomorphic encryption (fully homomorphic). Partially homomorphic encryption algorithms allow a certain operation to be performed an unlimited number of times, somewhat homomorphic encryption allows a limited number of arbitrary operations on ciphertext, and fully homomorphic encryption allows an unlimited number of arbitrary homomorphic operations on ciphertext. The principle of fully homomorphic encryption is as follows:

[0127] Encryption method:

[0128] Encryption(m):c=m+2r+pq;

[0129] Decryption method:

[0130] Decode(c):(c mod p)mod 2=(cp*「c / p」)mod 2=m;

[0131] Among them, m represents the plaintext to be encrypted, r is the encrypted random number, p and q are two large prime numbers for encryption, and c is the encrypted ciphertext.

[0132] This embodiment uses a method based on fully homomorphic encryption to perform encrypted computations. The method of using homomorphic encryption in federated learning is as follows: (1) The participant first encrypts the local model, then uses local data to perform biased parameter estimation and uploads the encrypted model gradient; (2) The server aggregates the local model gradient and performs ciphertext operations; (3) The server sends the encrypted model parameters, and the participant decrypts the new round of global model parameters and continues the operation in (1).

[0133] Step 2: Data aggregation and preprocessing

[0134] 2.1 Sample Data Preparation

[0135] We screened a period of call data from user call details and ensured model accuracy by supplementing key data, such as user call rejection records. Based on family network data, group user information, enterprise payment and reimbursement, and enterprise virtual network information, we selected user pairs with call relationships as the known group relationship user set. User pairs with the same enterprise business ID or family network ID formed the positive sample relationship pair set, while user pairs with different enterprise business IDs or family network IDs formed the negative relationship pair set. Users who had not yet completed the corresponding service but had call relationships and whose group relationships could not be determined formed the unknown group relationship user pair set.

[0136] 2.2 Variable screening

[0137] For the number itself, we selected basic information such as the region where the number was registered, the registration time, and the type of package, as well as graph statistical features such as the geographic location of the base station located by the number through recent signaling, the number of recent call records, and the number of call numbers as node features. For call records, we collected information such as the call time and duration, the geographic location of the base station where the call occurred, and the number of recent calls between the two parties. Specifically, the dimensions of each variable include but are not limited to those shown in the following Table (1):

[0138]

[0139]

[0140] (one)

[0142] Step 3: Social Group Identification

[0143] The Fast Unfolding algorithm, a community discovery algorithm, is used to divide the family social circle. Referring to the modularity gain concept proposed by Newman et al., the formula is as follows:

[0144]

[0145] Among them, k i is the sum of the weights of the links associated with node i, k i,in is the sum of links from node i to nodes in community C, Represents the sum of all weights in the network.

[0146] Based on user intimacy, groups are merged and members are expanded. When the modularity Q value no longer increases, the division of family social circles is completed, and each community becomes a family. The basic steps are as follows:

[0147] (1) Initialization, divide each point into different communities.

[0148] (2) Select each node one by one and calculate the modularity gain of its neighboring community according to the modularity gain formula. If the maximum modularity gain is greater than 0, it is assigned to the corresponding neighboring community; otherwise, it remains in the original community.

[0149] (3) Repeat step (2) until the community of the node no longer changes.

[0150] (4) Construct a new graph. The points in the new graph represent the different communities generated in the previous stage, and the edge weight is the sum of the edge weights of all node pairs in the two communities. Repeat step ② until the maximum modularity value is obtained.

[0151] Step (5): Identify key people in the group based on the R-GraphSAGE algorithm

[0152] The embodiment of the present application constructs a heterogeneous network map based on time and space by fusing multi-dimensional feature data, which has stronger information representation capabilities than conventional map construction. The embodiment of the present application constructs a call relationship network map between users at a certain moment from user call data, a subscription network map between users and services from user subscription service data, and a location relationship network map between users and base stations, and between base stations and base stations from user base station data, and finally constructs a complete spatiotemporal heterogeneous network map. Figure 4 As shown in , the weight calculation algorithm for the edge of a user at time t within a period of time can be defined as:

[0153] w el =w i,k *c i,r *w t ;

[0154] Among them, w t represents the weight of the edge at time t, represents the set of neighbor nodes of node i in relation r, w i,k Represents edge e i,k The weight of the edge.

[0155] The steps for constructing a heterogeneous network graph are as follows: (1) Preprocessing of user call details, user information data, and call base station information data, including cleaning of data such as garbled characters, missing fields that cannot be completed, etc. in the details, user information, base station, and service subscription information, to ensure that the established user call spatiotemporal graph data is more accurate. (2) Construction of a call and service subscription graph model. The construction steps include: constructing user call graph model nodes based on the user information table and service table, with user basic information including age, gender, etc. as node attributes; constructing edges between user nodes based on user call details information, service subscription information, and base station location information, with call start time, number of calls, network access time, service subscription time, and call duration as edge attributes. The user call heterogeneous graph model is constructed through the relationships between user nodes, service nodes, and edges.

[0156] In this example, R-GraphSAGE is an inductive learning framework that includes two major steps: sampling and aggregation. Sampling is used to sample the number of neighbors, obtain the embeddings of neighboring nodes, and then aggregate these embeddings to update the embedding information of the node itself. The learning process of R-GraphSAGE is as follows:

[0157] (1) Randomly sample neighbors of different edge types, and then sample them with fixed length from high to low according to the edge weight;

[0158] (2) The sampled neighbor embeddings are passed to the node, and an aggregation function is used to aggregate these neighbor information to update the node embedding;

[0159] (3) Predict the node label based on the updated embedding.

[0160] The mean aggregator used by the GraphSAGE algorithm for sampling is as follows:

[0161]

[0162] Since heterogeneous networks need to be processed, this embodiment uses a heterogeneous network aggregator as follows:

[0163]

[0164] in, represents the k-th layer neighbor aggregation representation of node v, w el It represents the weight matrix of different network graphs in heterogeneous networks, and M represents the edge type.

[0165] The R-GraphSAGE proposed in this embodiment randomly samples neighbors of different edge types, and then samples them in a fixed length from high to low according to the edge weight. A heterogeneous network aggregator is used to aggregate the sampled neighbor nodes of different types to update the node embedding, and finally achieve the processing of heterogeneous data, solving the problem that GCN and GraphSAGE can only process isomorphic graphs. R-GraphSAGE processes the principle of ultra-large heterogeneous graphs through sampling, mainly adopting fixed-length sampling of S neighbor nodes, and achieving S through methods such as resampling with replacement and negative sampling. Ensure that the number of neighbors of each node is consistent after sampling, and splice multiple nodes and their neighbors as Tensor input into the GPU for batch training, so that the model converges quickly. GCN needs to put the entire graph into memory and video memory, which will consume a lot of memory and video memory, and cannot process ultra-large-scale graphs.

[0166] It is worth mentioning that in the current graph federated learning methods, communication and memory consumption are common performance bottlenecks. R-GraphSAGE's sampling solves this problem better. By extracting a fixed number of nodes S through resampling with replacement and negative sampling, the amount of data in the communication between the models and the memory consumption of the server are reduced. In graph federated learning, each data owner holds a subgraph data A. k It can be seen as a partition from the potential entire graph A, that is, A=>{A k}. Then the R-GraphSAGE global model is:

[0167]

[0168] in, Indicates output, X (k) and A (k) Represents the features and adjacency matrix of the k-th data owner dataset.

[0169] After using FedAvg, the objective function is:

[0170]

[0171]

[0172] In summary, the technical problems and solutions addressed by the embodiments of this application can be summarized as follows: optimizing the GraphSAGE algorithm to process heterogeneous network data, addressing the issue of GraphSAGE's limited applicability to homogeneous graphs. Subsequently, the R-GraphSAGE model is trained through federated learning to predict whether a node is a key figure in a community. This entire solution, based on heterogeneous networks and federated learning theory, addresses the diversity and complexity of social relationships, enabling the rational construction of social networks and accurate analysis of key figures.

[0173] It should be noted that what is disclosed above is only a preferred embodiment of the present application, and it is certainly not intended to limit the scope of protection of the present application. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope covered by the present application.

[0174] That is to say, the above specific embodiments are only used to help understand the technical concept of the embodiments of this application, and do not constitute a limitation of this application. More simple transformations based on the technical concept should all be within the scope of protection of this application.

[0175] In addition, the present invention also provides an electronic device, referring to Figure 5 , Figure 5 This is a hardware structure diagram of an electronic device provided in an embodiment of the present application. Figure 5 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (WIreless-FIdelity, WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM) memory, or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0176] Those skilled in the art will understand that Figure 5 The structure shown in the figure does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange components differently. Figure 5 As shown, the memory 1005 as a storage medium may include an operating system, a data storage module, a network communication module, a user interface module, and a group key person recognition program.

[0177] exist Figure 5 In the electronic device shown, the network interface 1004 is mainly used for data communication with other devices; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this embodiment can be set in the communication device, and the communication device calls the group key person identification program stored in the memory 1005 through the processor 1001, and executes the group key person identification method provided in any of the above embodiments.

[0178] The terminal proposed in this embodiment and the method for group key person identification proposed in the above embodiment belong to the same inventive concept. Technical details not described in detail in this embodiment can be referred to any of the above embodiments, and this embodiment has the same beneficial effects as executing the group key person identification method.

[0179] In addition, an embodiment of the present application also proposes a storage medium, which is a computer storage medium. The computer-readable storage medium can be a non-volatile computer-readable storage medium. A group key person identification program is stored on the computer-readable storage medium. When the group key person identification program is executed by the processor, the group key person identification method of the present application as described above is implemented.

[0180] The various embodiments of the electronic device and computer-readable storage medium of the present application can refer to the various embodiments of the group key person identification method of the present application, and will not be repeated here.

[0181] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0182] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0183] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling an electronic device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0184] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for identifying key persons in a group, characterized in that: The group key person identification method is based on a graph federated learning model, which is jointly trained by multiple business parties based on local training gradient data of a knowledge graph. A single knowledge graph is formed by the business party based on the private data of each entity it owns, and includes the entity identifiers of each entity, the private data of each entity, and the relationships between the entities. The method is executed by any one of the multiple business parties and includes: Based on the knowledge graph owned by the party, any one entity is selected from the entities as the target entity, and a target subgraph corresponding to the target entity is generated using a neighborhood sampling mechanism; Sending the target subgraph to a terminal device corresponding to the target entity, so that the terminal device fuses the target subgraphs received from the multiple business parties to obtain a global heterogeneous network graph corresponding to the target entity after linking multiple knowledge graphs; Training a local service model according to the global heterogeneous network graph to obtain local training gradient data, and sending the local training gradient data to a service provider; In response to the graph federated learning model meeting the gradient update conditions, the service provider integrates the local training gradient data provided by each business party to obtain the global gradient data corresponding to the graph federated learning model, and feeds the global gradient data back to each business party; Each business party updates the model parameters of the local business model based on the fused global gradient data to iteratively train the local business model until the gradient converges and training is stopped to obtain the target business model. The target business model identifies key people in the group based on the input knowledge graph.

2. The method for identifying key persons in a group according to claim 1, wherein: The step of generating a target subgraph corresponding to the target entity by using a neighborhood sampling mechanism includes: Randomly sampling different types of neighbor nodes of the target entity to obtain randomly sampled neighbor nodes; According to the weight from high to low, fixed-length sampling is performed on each randomly sampled neighbor node to obtain each fixed-length sampled neighbor node; The neighbor nodes sampled with fixed length are aggregated and updated to obtain the target subgraph corresponding to the target entity.

3. The method for identifying key persons in a group according to claim 2, wherein: The step of aggregating and updating the neighboring nodes sampled by fixed length to obtain the target subgraph corresponding to the target entity includes: Aggregate and update the neighbor nodes of the fixed-length sampling to obtain the aggregated node embedding corresponding to the target entity; Variable screening is performed based on the aggregation node embedding to obtain key dimension data; A target subgraph corresponding to the target entity is generated according to the key dimension data.

4. The method for identifying key persons in a group according to claim 3, wherein: The key dimension data includes user call data, user subscription service data, and user base station data. The step of generating a target subgraph corresponding to the target entity based on the key dimension data includes: Constructing a call relationship network map between different users in a target time period based on the user call data; Constructing a subscription network graph between users and services based on the user subscription service data; Constructing a location relationship network map of base station locations where calls occur between different users based on the user base station data; Based on the call relationship network graph, the ordering network graph and the location relationship network graph, a spatiotemporal heterogeneous network graph is constructed, and the spatiotemporal heterogeneous network graph is used as the target subgraph corresponding to the target entity.

5. The method for identifying key persons in a group according to claim 1, wherein: Before the step of selecting any one entity from the entities as a target entity based on the knowledge graph owned by the party, the method further includes: Obtaining sample data, the sample data including key data within a preset time period in a user's call details, wherein the key data includes family network data, group user information, enterprise payment and reimbursement information, and enterprise virtual network information; Performing group division and identification based on the sample data to obtain multiple social groups, and determining a positive sample relationship pair set and a negative sample relationship pair set based on each of the social groups; Based on the positive sample relationship pair set and the negative sample relationship pair set, group data cascade is performed to determine the knowledge graph possessed by the party, wherein the knowledge graph possessed by the party includes call behavior information, business processing information and geographic location information of the base station.

6. The method for identifying key persons in a group according to claim 5, wherein: The step of dividing and identifying groups according to the sample data to obtain multiple social groups includes: The Fast Unfolding algorithm is used to perform group division and identification in combination with the sample data to obtain multiple social groups.

7. The group key person identification method according to any one of claims 1 to 6, characterized in that: The step of sending the target subgraph to a terminal device corresponding to the target entity includes: Encrypt the private data of each entity in the target subgraph using the fully homomorphic encryption method negotiated with other business parties; The encrypted target subgraph is sent to a terminal device corresponding to the target entity.

8. The group key person identification method according to any one of claims 1 to 6, characterized in that: The target subgraph includes multiple nodes and edges for connecting the nodes. The node attributes of the nodes include user basic information, and the edge attributes of the edges include user call details information, service subscription information, and base station location information. The user basic information includes user age information and user gender information.

9. An electronic device, characterized in that: include: A memory, a processor, and a group key person identification program stored in the memory and executable on the processor, wherein the group key person identification program implements the group key person identification method according to any one of claims 1 to 8 when executed by the processor.

10. A readable storage medium, characterized in that: The readable storage medium is a computer-readable storage medium, which stores a group key person identification program. When the group key person identification program is executed by a processor, the group key person identification method according to any one of claims 1 to 8 is implemented.

11. A computer program product, characterized in that The computer program product comprises a group key person identification program, which implements the steps of the group key person identification method according to any one of claims 1 to 8 when executed by a processor.

Images