An entity identity correlation mapping method and system based on a graph neural network

Abstract

The application discloses an entity identity correlation mapping method and system based on a graph neural network, wherein the method comprises: feature extraction and representation based on graph embedding; relevant information and extracted features in different information systems are taken as nodes for graph representation, edges between the nodes and users are constructed, feature nodes in the graph are mapped to the same feature space, and normalized representation of the features is obtained; identity correlation mapping based on a multi-layer attention network; structural noise and attribute noise existing in a real network are simulated; hidden features of each graph embedding structure are learned through a multi-layer graph attention network, embedding of each graph attention layer is aggregated, and embedding representation results are obtained; correlation matching scores are obtained by aggregating embedding representation results of each layer of graph attention layers, and are represented by using a hierarchical correlation matrix, and finally, a correlation matrix is obtained through a greedy algorithm. The application can accurately realize network entity identity correlation mapping.

Description

Technical Field

[0001] This invention relates to the field of network information security technology, and in particular to an entity identity association mapping method and system based on graph neural networks. Background Technology

[0002] With the rapid development of the Internet and information technology, various information systems in different industries often have multi-source identity data due to the time span of construction and the impact of technological iteration. This provides opportunities for criminals to exploit the inconsistency of identity data to carry out illegal activities such as privacy theft and phishing. If it is possible to establish a correlation mapping between network entity identities, it will help to carry out research on trust mechanism propagation, abnormal user detection and intervention, public opinion analysis and control, user privacy protection, behavior auditing and tracing, etc., and greatly reduce the security risks in this regard.

[0003] Entity identity association mapping is the foundation for the applications described above. However, different information systems have different identity management systems and lack an identity alliance management system, making it impossible to directly associate virtual identities in different information systems. Furthermore, characteristics such as single-source attribute similarity and multi-source behavioral inconsistency pose significant challenges to identity feature extraction and association mapping. Therefore, research on entity identity association mapping methods has become a new trend. However, traditional methods based on feature combination weights and machine learning methods relying on expert knowledge design have not yielded ideal association results due to weak identity feature representation capabilities and unclear potential relationships. Research is gradually shifting from traditional identity association methods and machine learning-based methods to deep learning-based methods.

[0004] The information of network entities in various information systems mainly includes: identity attribute information, identity relationship topology, and entity behavior details. Faced with numerous challenges, effectively integrating and utilizing this information is key to accurately extracting entity identity features and achieving identity association mapping. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an entity identity association mapping method and system based on graph neural networks. To address the problem of weak representation capabilities, it utilizes the propagation mechanism of graph neural networks to achieve full coverage of identity feature information extraction, thereby improving the node feature representation capability. To address the issue of unclear potential relationships, it employs an attention mechanism for self-learning, improving adaptability to noise and refining the representation of node neighborhoods, thus accurately achieving network entity identity association mapping.

[0006] The technical solution adopted in this invention is as follows:

[0007] An entity identity association mapping method based on graph neural networks includes:

[0008] Feature extraction and representation based on graph embedding: relevant information and extracted features from different information systems are used as nodes to form a graph representation. The edges between these nodes and users are constructed, and then the feature nodes in the graph are mapped to the same feature space to obtain a normalized feature representation. The relevant information includes identity attribute information, identity relationship topology, and entity behavior details.

[0009] Identity association mapping based on multi-layer attention network: Simulates structural noise and attribute noise existing in real networks; learns the hidden features of each graph embedding structure through multi-layer graph attention network, aggregates the embeddings of each graph attention layer, balances local and global information in the embedding, and obtains the embedding representation result; aggregates the embedding representation results of each graph attention layer to obtain the association matching score, and uses a hierarchical association matrix for representation, and then obtains the final association matrix through a greedy algorithm, thereby obtaining the association matching candidate users for each entity identity node.

[0010] Furthermore, the feature extraction and representation based on graph embedding includes: tree structure determination, graph embedding structure construction, identity attribute information feature extraction, entity behavior detail feature extraction, identity relationship topology feature extraction, and node vectorization representation.

[0011] Furthermore, the identity association mapping based on the multi-layer attention network includes: attribute noise simulation, introduction of adaptive noise loss function, attention coefficient normalization, node embedding fusion, multi-layer feature representation, and association identity recognition calculation.

[0012] Furthermore, the feature extraction and representation based on graph embedding includes the following steps:

[0013] S101. Tree structure determination: The first layer is the expression layer, which includes identity attribute information nodes, identity relationship topology nodes, and entity behavior detail nodes; the second layer is the type layer, which includes category nodes composed of attribute classification and relationship topology classification; the third layer is the feature layer, which includes the specific results of feature extraction. Feature layer nodes have uniqueness and universality.

[0014] S102. Graph Embedding Structure Construction: Represent the graph embedding structure as η = {υ, ε}, where υ = {v i Let |i=1,...,N} be the set of nodes, and ε={e ij |(i,j)=1,...,N} is the set of edges; the set of nodes υ consists of entity nodes υ U Identity attribute information node υ P Identity relationship topology node υ G Entity behavior details S Category node υ C and feature node υ F It consists of six types of nodes, namely υ={υU ,υ P ,υ G ,υ S ,υ C ,υ F};

[0015] S103. Identity Attribute Information Feature Extraction: For simple-dimensional identity attribute information, it is directly used as feature nodes to construct the graph embedding structure; for short text information, characters are directly extracted as features, and the position of the characters is recorded.

[0016] S104. Entity Behavior Detail Feature Extraction: The entity behavior details obtained from the fusion analysis are spliced ​​and combined, and clustering is used to classify the entity behavior details. Then, the TF-IDF method is used to calculate the word frequency and extract the behavior detail keywords as entity behavior detail nodes. For features with a sequence, the sequence information is recorded.

[0017] S105. Identity Relationship Topological Feature Extraction: Utilizing the first-order similarity in LINE, each identity relationship is embedded and represented, and relationship embedding edges e are constructed. ij Obtain the vector representation z of identity in the latent space. i ,z j And add a mapping function to obtain the feature vector h' of the relation embedding edges in the latent space;

[0018] S106. Vectorized Representation of Nodes: Using the Glorot Normal Distribution for the Node Set υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F The extracted representation is vectorized, the attention weights of the extracted representation are calculated, the internal correlations between features are captured, and the embedding vectors in the latent space are dynamically updated during subsequent training.

[0019] Furthermore, the identity association mapping based on a multi-layer attention network includes the following steps:

[0020] S201. Attribute noise simulation: Establish an adjacency matrix for reinforcement learning and set a zero mask matrix. Calculate the Hadamard product of the zero mask matrix and the adjacency matrix to enable the source entity identity node set υ to randomly delete or add edges e with probability p, thereby simulating the structural noise present in real information systems.

[0021] S202. Introduction of Adaptive Noise Loss Function: An adaptive noise loss function is introduced to minimize the difference in multi-level features between nodes before and after reinforcement learning, and to control whether the reinforcement learning process goes out of bounds based on a confidence mechanism;

[0022] S203. Normalize Attention Coefficients: Perform a shared linear transformation on the node attribute matrix F using the parameterized matrix W, and use the softmax function to normalize the attention coefficients μ. ab =α(WF υ WF υ' The attention coefficient is normalized to make it comparable across different identity nodes; the attention coefficient is μ. ab =α(WF υ WF υ' This is used to reflect the degree of importance between identity nodes a and b;

[0023] S204. Node Embedding Fusion: Construct k graph attention layers, each learning different attribute features in different neighborhood structures; the 0th layer embeds H... (0) The initial attribute matrix F is used, and H is embedded in each subsequent layer. (x) The attribute features are obtained by learning the structural information of its neighborhood, and the identity attribute information nodes of the entity node set υ are combined using a multi-layer graph attention mechanism. P Identity relationship topology node υ G Entity behavior details S Fusion is performed at different network layers;

[0024] S205. Multi-layer Feature Representation: Using the embedding representation results of each layer as the feature representation of the entity identity node {H (1) H (2) ,...,H (k)}, where k is a hyperparameter and also the number of graph attention layers;

[0025] S206. Association Identity Recognition Calculation: Constructing a Hierarchical Association Matrix Where m is the number of the attention layer in the current graph. For the feature representation of the entity identity node in information system s, The feature representation of the entity identity node of information system t; obtain the associated matching candidate user for each entity identity node based on all structural and attribute information of the current graph attention layer.

[0026] An entity identity association mapping system based on graph neural networks includes:

[0027] The feature extraction and representation module is configured to use relevant information from different information systems and extracted features as nodes to represent a graph, construct the connections between these nodes and the user, and then map the feature nodes in the graph to the same feature space to obtain a normalized feature representation; the relevant information includes identity attribute information, identity relationship topology, and entity behavior details;

[0028] The identity association mapping module is configured to simulate structural noise and attribute noise present in real networks. It learns the hidden features of each graph embedding structure through a multi-layer graph attention network, aggregates the embeddings of each graph attention layer, balances local and global information in the embedding, and obtains the embedding representation result. The association matching score is obtained by aggregating the embedding representation results of each graph attention layer and representing it using a hierarchical association matrix. Then, a greedy algorithm is used to obtain the final association matrix, thereby obtaining the association matching candidate users for each entity identity node.

[0029] Furthermore, the feature extraction and representation module includes a tree structure determination unit, a graph embedding structure construction unit, an identity attribute information feature extraction unit, an entity behavior detail feature extraction unit, an identity relationship topology feature extraction unit, and a node vectorization representation unit connected in sequence.

[0030] Furthermore, the identity association mapping based on the multi-layer attention network includes, in sequence, an attribute noise simulation unit, an adaptive noise loss function introduction unit, an attention coefficient normalization unit, a node embedding fusion unit, a multi-layer feature expression unit, and an associated identity recognition calculation unit.

[0031] Furthermore, the feature extraction and representation based on graph embedding includes:

[0032] The tree-structured unit consists of three layers: the first layer is the expression layer, which includes identity attribute information nodes, identity relationship topology nodes, and entity behavior detail nodes; the second layer is the type layer, which includes category nodes composed of attribute classification and relationship topology classification; and the third layer is the feature layer, which includes the specific results of feature extraction. The feature layer nodes are unique and universal.

[0033] The graph embedding structure building unit is configured to represent the graph embedding structure as η = {υ, ε}, where υ = {v i Let |i=1,...,N} be the set of nodes, and ε={e ij |(i,j)=1,...,N} is the set of edges; the set of nodes υ consists of entity nodes υ U Identity attribute information node υ P Identity relationship topology node υ G Entity behavior details S Category node υ C and feature node υ F It consists of six types of nodes, namely υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F};

[0034] The identity attribute information feature extraction unit is configured to directly construct the graph embedding structure as feature nodes for simple-dimensional identity attribute information; for short text information, it directly extracts characters as features and records the position of the characters.

[0035] The entity behavior detail feature extraction unit is configured to concatenate and combine the fused and parsed entity behavior details, classify the entity behavior details using clustering, calculate word frequency using the TF-IDF method, extract behavior detail keywords as entity behavior detail nodes; for features with a sequence, the sequence information is recorded.

[0036] The identity relation topology feature extraction unit is configured to use the first-order similarity in LINE to embed each identity relation into a representation and construct relation embedding edges e. ij Obtain the vector representation z of identity in the latent space. i ,z j And add a mapping function to obtain the feature vector h' of the relation embedding edges in the latent space;

[0037] The node vectorized representation unit is configured to use a Glorot normal distribution for the node set υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F The extracted representation is vectorized, the attention weights of the extracted representation are calculated, the internal correlations between features are captured, and the embedding vectors in the latent space are dynamically updated during subsequent training.

[0038] Furthermore, the identity association mapping based on a multi-layer attention network includes:

[0039] The attribute noise simulation unit is configured to build an adjacency matrix for reinforcement learning and set a zero mask matrix. By calculating the Hadamard product of the zero mask matrix and the adjacency matrix, the source entity identity node set υ can randomly delete or add edges e with probability p, thereby simulating the structural noise that exists in real information systems.

[0040] The adaptive noise loss function introduction unit is configured to introduce an adaptive noise loss function to minimize the difference in multi-level features between nodes before and after reinforcement learning, and to control whether the reinforcement learning process goes out of bounds based on a confidence mechanism.

[0041] The attention coefficient normalization unit is configured to perform a shared linear transformation on the node attribute matrix F through the parameterization matrix W, and use the softmax function to normalize the attention coefficients μ. ab =α(WF υ WFυ' The attention coefficient is normalized to make it comparable across different identity nodes; the attention coefficient is μ. ab =α(WF υ WF υ' This is used to reflect the degree of importance between identity nodes a and b;

[0042] Node embedding fusion units are configured to construct k graph attention layers, each learning different attribute features in a different neighborhood structure; the 0th layer embeds H... (0) The initial attribute matrix F is used, and H is embedded in each subsequent layer. (x) The attribute features are obtained by learning the structural information of its neighborhood, and the identity attribute information nodes of the entity node set υ are combined using a multi-layer graph attention mechanism. P Identity relationship topology node υ G Entity behavior details S Fusion is performed at different network layers;

[0043] Multi-layer feature representation units are configured to use the embedding representation results of each layer as the feature representation {H} of entity identity nodes. (1) H (2) ,...,H (k)}, where k is a hyperparameter and also the number of graph attention layers;

[0044] The associated identity recognition calculation unit is configured to construct a hierarchical association matrix. Where m is the number of the attention layer in the current graph. For the feature representation of the entity identity node in information system s, The feature representation of the entity identity node of information system t; obtain the associated matching candidate user for each entity identity node based on all structural and attribute information of the current graph attention layer.

[0045] The beneficial effects of this invention are as follows:

[0046] 1) To address the problem of weak representation ability, a feature extraction and representation method based on graph embedding is proposed. The identity attribute information, identity relationship topology, entity behavior details and their extracted features from different information systems are represented in the graph embedding structure, and the edges between these nodes and users are constructed. The feature nodes in the graph are mapped to the latent feature space to achieve feature normalization representation, which facilitates subsequent association mapping.

[0047] 2) To address the issue of unclear potential relationships, a multi-layer attention network-based identity association mapping model is proposed. This model can be trained without prior knowledge and employs adaptive reinforcement learning to address attribute noise present in real-world information systems, thus improving its adaptability to noise. Furthermore, by learning the weights of neighboring nodes through an attention mechanism, the model provides a refined representation of the node's neighborhood, thereby accurately achieving the identification of associated users. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the entity identity association mapping method according to an embodiment of the present invention. Detailed Implementation

[0049] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment provides an entity identity association mapping method based on graph neural networks, including feature extraction and representation based on graph embedding and identity association mapping based on multi-layer attention networks, as detailed below.

[0052] Graph embedding is a method for network representation learning, which facilitates the representation of nodes or edges in a network in a low-order vector space while preserving the structure and properties of the nodes or edges, thus possessing strong relational representation capabilities. In this embodiment, relevant information and extracted features from different information systems are used as nodes for graph representation, and edges between these nodes and users are constructed. Then, the feature nodes in the graph are mapped to the same feature space to obtain a normalized feature representation, which facilitates subsequent association mapping. Preferably, the relevant information from different information systems can be identity attribute information, identity relationship topology, and entity behavior details.

[0053] Preferably, feature extraction and representation based on graph embedding includes the following steps:

[0054] S101. Tree structure determination: The first layer is the expression layer, which includes identity attribute information nodes, identity relationship topology nodes, and entity behavior detail nodes; the second layer is the type layer, which includes category nodes composed of attribute classification and relationship topology classification; the third layer is the feature layer, which includes the specific results of feature extraction. The feature layer nodes have uniqueness and universality.

[0055] S102. Graph Embedding Structure Construction: Represent the graph embedding structure as η = {υ, ε}, where υ = {v i Let |i=1,...,N} be the set of nodes, and ε={e ij |(i,j)=1,...,N} is the set of edges; the set of nodes υ consists of entity nodes υ U Identity attribute information node υ P Identity relationship topology node υ G Entity behavior details S Category node υ C and feature node υ F It consists of six types of nodes, namely υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F}

[0056] S103. Identity Attribute Information Feature Extraction: For simple-dimensional identity attribute information, it is directly used as feature nodes to construct the graph embedding structure; for short text information, characters are directly extracted as features, and the position of the characters is recorded.

[0057] S104. Entity Behavior Detail Feature Extraction: The entity behavior details obtained from the fusion analysis are spliced ​​and combined, and clustering is used to classify the entity behavior details. Then, the TF-IDF method is used to calculate the word frequency and extract the behavior detail keywords as entity behavior detail nodes. For features with a sequence, the sequence information is recorded.

[0058] S105. Identity Relationship Topological Feature Extraction: Utilizing the first-order similarity in LINE, each identity relationship is embedded and represented, and relationship embedding edges e are constructed. ij Obtain the vector representation z of identity in the latent space. i ,z j And add a mapping function to obtain the feature vector h' of the relation embedding edges in the latent space.

[0059] S106. Vectorized Representation of Nodes: Using the Glorot Normal Distribution for the Node Set υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F The extracted representation is vectorized, the attention weights of the extracted representation are calculated, the internal correlations between features are captured, and the embedding vectors in the latent space are dynamically updated during subsequent training.

[0060] To identify related entities in two information systems, this embodiment proposes an identity association mapping model based on a multi-layer attention network. The entire model consists of three parts:

[0061] (1) Enhanced learning process: By simulating structural noise and attribute noise that exist in reality, the model's adaptability to real network conditions is improved;

[0062] (2) Multi-layer graph attention network embedding: The hidden features of each graph embedding structure are learned through a multi-layer graph attention network, the embeddings of each graph attention layer are aggregated, and the local and global information in the embedding are balanced to obtain the final embedding representation.

[0063] (3) Association recognition calculation: The association matching score is represented by a hierarchical association matrix. These scores are aggregated from the embedding results of each graph attention layer. The final association matrix is ​​obtained through a greedy algorithm.

[0064] Preferably, the identity association mapping based on a multi-layer attention network includes the following steps:

[0065] S201. Attribute Noise Simulation: Establish an adjacency matrix for reinforcement learning and set a zero mask matrix. Calculate the Hadamard product of the zero mask matrix and the adjacency matrix to enable the source entity identity node set υ to randomly delete or add edges e with probability p, thereby simulating the structural noise present in real information systems.

[0066] S202. Introduction of Adaptive Noise Loss Function: Due to the consistency constraint of noise, the source entity identity should be associated with the target entity identity and the enhanced target entity identity to enhance the robustness of the model to structural noise and attribute noise. Therefore, this embodiment introduces an adaptive noise loss function to minimize the difference in multi-level features between nodes before and after reinforcement learning, and controls whether the reinforcement learning process goes out of bounds based on a confidence mechanism.

[0067] S203. Normalize Attention Coefficients: Perform a shared linear transformation on the node attribute matrix F using the parameterization matrix W, and define the attention coefficients μ. ab =α(WF υ WF υ' The attention coefficient reflects the importance of identity nodes a and b. Then, the softmax function is used to normalize the attention coefficient to make it comparable between different identity nodes.

[0068] S204. Node Embedding Fusion: Construct k graph attention layers, each learning different attribute features in different neighborhood structures; the 0th layer embeds H... (0) The initial attribute matrix F is used, and H is embedded in each subsequent layer. (x)The attribute features are obtained by learning the structural information of its neighborhood, and the identity attribute information nodes of the entity node set υ are combined using a multi-layer graph attention mechanism. P Identity relationship topology node υ G Entity behavior details S Fusion is performed at different network layers.

[0069] S205. Multi-layer Feature Representation: Using the embedding representation results of each layer as the feature representation of the entity identity node {H (1) H (2) ,...,H (k)}, where k is a hyperparameter, which is also the number of layers in the graph attention layer. The optimal value of k can be selected through experimentation.

[0070] S206. Association Identity Recognition Calculation: Multi-layer graph attention models are shared in information systems, so their mapping embedding spaces are the same. Association matching results can be determined by comparing the similarity of their embedding representations. Therefore, a hierarchical association matrix is ​​utilized. Based on all structural and attribute information of the current layer, we can obtain the associated matching candidate users for each entity identity node. Here, m represents the number of the current graph attention layer. For the feature representation of the entity identity node in information system s, This represents the feature representation of the entity identity node in information system t.

[0071] Example 2

[0072] This embodiment provides an entity identity association mapping system based on graph neural networks, including:

[0073] The feature extraction and representation module is configured to use relevant information from different information systems and extracted features as nodes to represent a graph, construct the connections between these nodes and the user, and then map the feature nodes in the graph to the same feature space to obtain a normalized feature representation; the relevant information includes identity attribute information, identity relationship topology, and entity behavior details;

[0074] The identity association mapping module is configured to simulate structural noise and attribute noise present in real networks. It learns the hidden features of each graph embedding structure through a multi-layer graph attention network, aggregates the embeddings of each graph attention layer, balances local and global information in the embedding, and obtains the embedding representation result. The association matching score is obtained by aggregating the embedding representation results of each graph attention layer and representing it using a hierarchical association matrix. Then, a greedy algorithm is used to obtain the final association matrix, thereby obtaining the association matching candidate users for each entity identity node.

[0075] Preferably, the feature extraction and representation module includes a tree structure determination unit, a graph embedding structure construction unit, an identity attribute information feature extraction unit, an entity behavior detail feature extraction unit, an identity relationship topology feature extraction unit, and a node vectorization representation unit connected in sequence.

[0076] More preferably, the graph embedding-based feature extraction and representation includes:

[0077] The tree-structured unit consists of three layers: the first layer is the expression layer, which includes identity attribute information nodes, identity relationship topology nodes, and entity behavior detail nodes; the second layer is the type layer, which includes category nodes composed of attribute classification and relationship topology classification; and the third layer is the feature layer, which includes the specific results of feature extraction. The feature layer nodes are unique and universal.

[0078] The graph embedding structure building unit is configured to represent the graph embedding structure as η = {υ, ε}, where υ = {v i Let |i=1,...,N} be the set of nodes, and ε={e ij |(i,j)=1,...,N} is the set of edges; the set of nodes υ consists of entity nodes υ U Identity attribute information node υ P Identity relationship topology node υ G Entity behavior details S Category node υ C and feature node υ F It consists of six types of nodes, namely υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F};

[0079] The identity attribute information feature extraction unit is configured to directly construct the graph embedding structure as feature nodes for simple-dimensional identity attribute information; for short text information, it directly extracts characters as features and records the position of the characters.

[0080] The entity behavior detail feature extraction unit is configured to concatenate and combine the fused and parsed entity behavior details, classify the entity behavior details using clustering, calculate word frequency using the TF-IDF method, extract behavior detail keywords as entity behavior detail nodes; for features with a sequence, the sequence information is recorded.

[0081] The identity relation topology feature extraction unit is configured to use the first-order similarity in LINE to embed each identity relation into a representation and construct relation embedding edges e. ij Obtain the vector representation z of identity in the latent space.i ,z j And add a mapping function to obtain the feature vector h' of the relation embedding edges in the latent space;

[0082] The node vectorized representation unit is configured to use a Glorot normal distribution for the node set υ={υ U ,υ P ,υ G ,υ S ,υ C ,υ F The extracted representation is vectorized, the attention weights are calculated, the internal correlations between features are captured, and the embedding vectors in the latent space are dynamically updated during subsequent training.

[0083] Preferably, the identity association mapping based on a multi-layer attention network includes, in sequence, an attribute noise simulation unit, an adaptive noise loss function introduction unit, an attention coefficient normalization unit, a node embedding fusion unit, a multi-layer feature expression unit, and an associated identity recognition calculation unit.

[0084] More preferably, the identity association mapping based on a multi-layer attention network includes:

[0085] The attribute noise simulation unit is configured to build an adjacency matrix for reinforcement learning and set a zero mask matrix. By calculating the Hadamard product of the zero mask matrix and the adjacency matrix, the source entity identity node set υ can randomly delete or add edges e with probability p, thereby simulating the structural noise that exists in real information systems.

[0086] The adaptive noise loss function introduction unit is configured to introduce an adaptive noise loss function to minimize the difference in multi-level features between nodes before and after reinforcement learning, and to control whether the reinforcement learning process goes out of bounds based on a confidence mechanism.

[0087] The attention coefficient normalization unit is configured to perform a shared linear transformation on the node attribute matrix F through the parameterization matrix W, and to use the softmax function to normalize the attention coefficients μ. ab =α(WF υ WF υ' The attention coefficient is normalized to make it comparable across different identity nodes; the attention coefficient is μ. ab =α(WF υ WF υ' This is used to reflect the degree of importance between identity nodes a and b;

[0088] Node embedding fusion units are configured to construct k graph attention layers, each learning different attribute features in a different neighborhood structure; the 0th layer embeds H... (0)The initial attribute matrix F is used, and H is embedded in each subsequent layer. (x) The attribute features are obtained by learning the structural information of its neighborhood, and the identity attribute information nodes of the entity node set υ are combined using a multi-layer graph attention mechanism. P Identity relationship topology node υ G Entity behavior details S Fusion is performed at different network layers;

[0089] Multi-layer feature representation units are configured to use the embedding representation results of each layer as the feature representation {H} of entity identity nodes. (1) H (2) ,...,H (k)}, where k is a hyperparameter, which is also the number of graph attention layers.

[0090] The associated identity recognition calculation unit is configured to construct a hierarchical association matrix. Where m is the number of the attention layer in the current graph. For the feature representation of the entity identity node in information system s, The feature representation of the entity identity node of information system t; obtain the associated matching candidate user for each entity identity node based on all structural and attribute information of the current graph attention layer.

[0091] Example 3

[0092] This embodiment is based on embodiment 1:

[0093] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the entity identity association mapping method based on graph neural networks in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0094] Example 4

[0095] This embodiment is based on embodiment 1:

[0096] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the entity identity association mapping method based on graph neural networks in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes: any entity or device capable of carrying computer program code, recording media, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the storage medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electrical carrier signals and telecommunication signals.

[0097] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for entity identity association mapping based on graph neural networks, characterized in that, include: Feature extraction and representation based on graph embedding: relevant information and extracted features from different information systems are used as nodes to form a graph representation. The edges between these nodes and users are constructed, and then the feature nodes in the graph are mapped to the same feature space to obtain a normalized feature representation. The relevant information includes identity attribute information, identity relationship topology, and entity behavior details. Identity association mapping based on multi-layer attention network: Simulates structural noise and attribute noise existing in real networks; learns hidden features of each graph embedding structure through multi-layer graph attention network, aggregates the embeddings of each graph attention layer, balances local and global information in the embedding, and obtains the embedding representation result; aggregates the embedding representation results of each graph attention layer to obtain the association matching score, and uses a hierarchical association matrix to represent it, and then obtains the final association matrix through a greedy algorithm, thereby obtaining the association matching candidate users for each entity identity node; The graph embedding-based feature extraction and representation includes the following steps: S101. Tree structure determination: The first layer is the expression layer, which includes identity attribute information nodes, identity relationship topology nodes, and entity behavior detail nodes; the second layer is the type layer, which includes category nodes composed of attribute classification and relationship topology classification; the third layer is the feature layer, which includes the specific results of feature extraction. Feature layer nodes have uniqueness and universality. S102. Graph Embedding Structure Construction: Represent the graph embedding structure as follows ,in For a set of nodes, The set of edges; the set of nodes From entity nodes Identity attribute information node Identity Relationship Topology Nodes Detailed nodes of entity behavior Category Node and feature nodes It consists of six types of nodes, namely ; S103. Identity Attribute Information Feature Extraction: For simple-dimensional identity attribute information, it is directly used as feature nodes to construct the graph embedding structure; for short text information, characters are directly extracted as features, and the position of the characters is recorded. S104. Entity Behavior Detail Feature Extraction: The entity behavior details obtained from the fusion analysis are concatenated and combined, and clustering is used to classify the entity behavior details. Then, the TF-IDF method is used to calculate the word frequency and extract the behavior detail keywords as entity behavior detail nodes; for features with a sequence, the sequence information is recorded. S105. Identity Relationship Topological Feature Extraction: Utilizing the first-order similarity in LINE, embedding representations are performed for each identity relationship, and relationship embedding edges are constructed. Obtain the vector representation of identity in the latent space. And add a mapping function to obtain the feature vector of the relation embedding edges in the latent space. ; S106. Vectorized representation of nodes: Using the Glorot normal distribution for the node set Vectorization is performed, attention weights of the extracted representations are calculated, internal correlations between features are captured, and embedding vectors in the latent space are dynamically updated during subsequent training.

2. The entity identity association mapping method based on graph neural networks according to claim 1, characterized in that, The graph embedding-based feature extraction and representation includes: tree structure determination, graph embedding structure construction, identity attribute information feature extraction, entity behavior detail feature extraction, identity relationship topology feature extraction, and node vectorization representation.

3. The entity identity association mapping method based on graph neural networks according to claim 1, characterized in that, The identity association mapping based on multi-layer attention network includes: attribute noise simulation, introduction of adaptive noise loss function, attention coefficient normalization, node embedding fusion, multi-layer feature representation and associated identity recognition calculation.

4. The entity identity association mapping method based on graph neural networks according to claim 1, characterized in that, The identity association mapping based on a multi-layer attention network includes the following steps: S201. Attribute Noise Simulation: Establish an adjacency matrix for reinforcement learning and set a zero-mask matrix. Calculate the Hadamard product of the zero-mask matrix and the adjacency matrix to make the source entity identity node set... Able to use probability To randomly delete or add edges This simulates the structural noise present in real information systems. S202. Introduction of Adaptive Noise Loss Function: An adaptive noise loss function is introduced to minimize the difference in multi-level features between nodes before and after reinforcement learning, and to control whether the reinforcement learning process goes out of bounds based on a confidence mechanism; S203. Normalize attention coefficients: through parameterization matrix node attribute matrix Perform a shared linear transformation and use the softmax function to adjust the attention coefficients. Normalization is performed to make it comparable across different identity nodes; the attention coefficient is... Used to reflect the degree of importance between identity nodes a and b; S204. Node Embedding and Fusion: Building Each graph attention layer learns different attribute features in different neighborhood structures; layer 0 embedding. For the initial attribute matrix Embedded in each subsequent layer The attribute features are obtained by learning the structural information of its neighborhood, and the entity node set is integrated using a multi-layer graph attention mechanism. Identity attribute information node Identity Relationship Topology Nodes Detailed nodes of entity behavior Fusion is performed at different network layers; S205. Multi-layer feature representation: Use the embedding representation results of each layer as the feature representation of the entity identity node. ,in This is a hyperparameter, and also the number of layers in the graph attention layer; S206. Association Identity Recognition Calculation: Constructing a Hierarchical Association Matrix ,in m This represents the current attention layer number. For information systems s The feature representation of entity identity nodes, For information systems t The feature representation of the entity identity node; based on all structural and attribute information of the current graph attention layer, obtain the associated matching candidate users for each entity identity node.

5. An entity identity association mapping system based on graph neural networks, characterized in that, include: The feature extraction and representation module is configured to use relevant information from different information systems and extracted features as nodes to represent a graph, construct the connections between these nodes and the user, and then map the feature nodes in the graph to the same feature space to obtain a normalized feature representation; the relevant information includes identity attribute information, identity relationship topology, and entity behavior details; The identity association mapping module is configured to simulate structural noise and attribute noise present in real networks. It learns the hidden features of each graph embedding structure through a multi-layer graph attention network, aggregates the embeddings of each graph attention layer, balances local and global information in the embeddings, and obtains the embedding representation results. The association matching score is obtained by aggregating the embedding representation results of each graph attention layer and representing it using a hierarchical association matrix. Then, a greedy algorithm is used to obtain the final association matrix, thereby obtaining the association matching candidate users for each entity identity node. The feature extraction and representation module includes: The tree-structured unit consists of three layers: the first layer is the expression layer, which includes identity attribute information nodes, identity relationship topology nodes, and entity behavior detail nodes; the second layer is the type layer, which includes category nodes composed of attribute classification and relationship topology classification; and the third layer is the feature layer, which includes the specific results of feature extraction. The feature layer nodes are unique and universal. Graph embedding structure building units are configured to represent graph embedding structures as ,in For a set of nodes, The set of edges; the set of nodes From entity nodes Identity attribute information node Identity Relationship Topology Nodes Detailed nodes of entity behavior Category Node and feature nodes It consists of six types of nodes, namely ; The identity attribute information feature extraction unit is configured to directly construct the graph embedding structure as feature nodes for simple-dimensional identity attribute information; for short text information, it directly extracts characters as features and records the position of the characters. The entity behavior detail feature extraction unit is configured to concatenate and combine the fused and parsed entity behavior details, classify the entity behavior details using clustering, calculate word frequency using the TF-IDF method, extract behavior detail keywords as entity behavior detail nodes; for features with a sequence, the sequence information is recorded. The identity relationship topology feature extraction unit is configured to use the first-order similarity in LINE to embed each identity relationship and construct relationship embedding edges. Obtain the vector representation of identity in the latent space. And add a mapping function to obtain the feature vector of the relation embedding edges in the latent space. ; The node vectorized representation unit is configured to use a Glorot normal distribution for the node set. Vectorization is performed, attention weights of the extracted representations are calculated, internal correlations between features are captured, and embedding vectors in the latent space are dynamically updated during subsequent training.

6. The entity identity association mapping system based on graph neural networks according to claim 5, characterized in that, The feature extraction and representation module includes a tree structure determination unit, a graph embedding structure construction unit, an identity attribute information feature extraction unit, an entity behavior detail feature extraction unit, an identity relationship topology feature extraction unit, and a node vectorization representation unit, which are connected in sequence.

7. The entity identity association mapping system based on graph neural networks according to claim 5, characterized in that, The identity association mapping module includes, in sequence, an attribute noise simulation unit, an adaptive noise loss function introduction unit, an attention coefficient normalization unit, a node embedding fusion unit, a multi-layer feature expression unit, and an associated identity recognition calculation unit.

8. The entity identity association mapping system based on graph neural networks according to claim 5, characterized in that, The identity association mapping module includes: The attribute noise simulation unit is configured to build an adjacency matrix for reinforcement learning and set a zero-mask matrix. The Hadamard product of the zero-mask matrix and the adjacency matrix is ​​calculated to make the source entity identity node set... Able to use probability To randomly delete or add edges This simulates the structural noise present in real information systems. The adaptive noise loss function introduction unit is configured to introduce an adaptive noise loss function to minimize the difference in multi-level features between nodes before and after reinforcement learning, and to control whether the reinforcement learning process goes out of bounds based on a confidence mechanism. The attention coefficient normalization unit is configured to be passed through the parameterization matrix. node attribute matrix Perform a shared linear transformation and use the softmax function to adjust the attention coefficients. Normalization is performed to make it comparable across different identity nodes; the attention coefficient is... Used to reflect the degree of importance between identity nodes a and b; Nodes are embedded in fusion units and configured to build Each graph attention layer learns different attribute features in different neighborhood structures; layer 0 embedding. For the initial attribute matrix Embedded in each subsequent layer The attribute features are obtained by learning the structural information of its neighborhood, and the entity node set is integrated using a multi-layer graph attention mechanism. Identity attribute information node Identity Relationship Topology Nodes Detailed nodes of entity behavior Fusion is performed at different network layers; Multi-layer feature representation units are configured to use the embedding representation results of each layer as the feature representation of entity identity nodes. ,in This is a hyperparameter, and also the number of layers in the graph attention layer; The associated identity recognition calculation unit is configured to construct a hierarchical association matrix. ,in m This represents the current attention layer number. For information systems s The feature representation of entity identity nodes, For information systems t The feature representation of the entity identity node; based on all structural and attribute information of the current graph attention layer, obtain the associated matching candidate users for each entity identity node.

Images