Graph neural network training method based on depolarization topological attribute consciousness

Abstract

The invention provides a graph neural network training method based on depolarization topological attribute consciousness. The method comprises the steps of obtaining a relation network graph used for training a graph neural network; identifying sensitive attribute features by using an integral gradient method according to the set baseline feature matrix and the attribute feature matrix of the relation network diagram; calculating unbiased attribute similarity among the business object nodes according to the identified sensitive attribute characteristics; reconstructing the topological structure of the relation network diagram according to the unbiased attribute similarity between the business object nodes; inputting the reconstructed relation network graph into a graph neural network model, and estimating mutual information of the sensitive attribute characteristics and a prediction result according to the prediction result of the graph neural network model; and constructing a final loss function of the graph neural network model by combining the task loss of the graph neural network model and the mutual information of the sensitive attribute characteristics and the prediction result to update the parameters of the graph neural network model. According to the method, topology disturbance and attribute poisoning attacks can be effectively resisted, the prediction accuracy and stability of the graph neural network in an adversarial environment are guaranteed, and reliable model defense capability is provided for a key service scene.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular relates to a graph neural network training method based on debiased topological attribute awareness. Background Art

[0002] Graph neural networks (GNNs) have gained widespread application due to their ability to extract valuable information from complex graph-structured data, such as in social networks, recommender systems, natural language processing, and computer vision. However, numerous studies have shown that GNNs are susceptible to adversarial attacks: attackers can severely weaken model performance by modifying the graph's topology or imposing subtle perturbations on node attributes. Topological attacks primarily disrupt information propagation by adding or removing edges, while attribute attacks mislead the model by subtly modifying node features. These malicious manipulations alter the similarity relationships between nodes, disrupting the information flow within the GNN and leading to incorrect predictions.

[0003] To address these threats, existing research has focused on two main defense approaches: fine-tuning the graph's topology, and optimizing node attributes. For example, some methods leverage graph topology information to dynamically adjust node weights to enhance robustness. Others inject hierarchical Gaussian noise into the node position generator and combine it with a spectral graph convolution discriminator to protect against edge perturbations. Others link unlabeled nodes with labeled nodes with similar features to improve the model's resistance to label noise or label flipping attacks. Adversarial training methods also perturb input features through dynamic regularization, thereby enhancing the model's robustness against potential future attacks.

[0004] While existing methods have enhanced the robustness of graph neural networks to a certain extent, they generally overlook the issue of bias in node attributes. Nodes typically possess multiple attributes that characterize them. Certain sensitive attributes, such as gender, naturally divide individuals into two communities—male and female—resulting in dense intra-community connectivity and sparse cross-community connectivity. This segmentation not only impacts the overall connectivity and robustness of the network structure but can also lead to situations where nodes with similar non-sensitive attributes (such as age, height, and occupation) that should be in the same community are isolated due to gender differences. Meanwhile, nodes with significantly different attributes may be connected solely due to the shared sensitive attribute. Gender-driven community partitioning weakens similarity-based connectivity across communities, reducing overall cohesion and resilience. Attackers often target the most sensitive or important nodes for attribute attacks, making such structures more vulnerable to adversarial environments. Therefore, eliminating bias in sensitive attributes and optimizing the corresponding graph structure are crucial. Summary of the Invention

[0005] In order to solve the problems existing in the background technology, one aspect of the present invention provides a graph neural network training method based on debiased topological attribute awareness, comprising:

[0006] S1: Obtain a relational network graph for training a graph neural network, wherein the relational network graph includes a node set corresponding to multiple business object nodes, an attribute feature matrix, an edge set formed by association relationships between the business object nodes, and business labels of the business object nodes;

[0007] S2: Based on the set baseline feature matrix and the attribute feature matrix of the relationship network graph, the integral gradient method is used to calculate the attribution value of each attribute feature to the output of the graph neural network model, and the attribute feature with the largest attribution value is selected as the sensitive attribute feature;

[0008] S3: Calculate the unbiased attribute similarity between business object nodes based on whether adjacent business object nodes have the same sensitive attribute features, the attribute feature similarity between business object nodes, and the topological structure of the relationship network graph;

[0009] S4: reconstructing the topological structure of the relationship network graph according to the unbiased attribute similarity between the business object nodes to obtain a reconstructed relationship network graph;

[0010] S5: Input the reconstructed relationship network graph into the graph neural network model for forward propagation, and estimate the mutual information between the sensitive attribute features and the prediction results based on the prediction results of the graph neural network model;

[0011] S6: Construct the final loss function of the graph neural network model by combining the task loss of the graph neural network model, the business label of the business object node, and the mutual information between the sensitive attribute features and the prediction results;

[0012] S7: Perform backpropagation based on the final loss function of the graph neural network model to update the parameters of the graph neural network model, and repeat steps S2-S7 until the final loss function converges.

[0013] Another aspect of the present invention provides an ecological environment dynamic monitoring system based on remote sensing and geographic information system, including: a memory and a processor; the memory is used to store applications; the processor is used to run the applications and execute the graph neural network training method based on debiased topological attribute awareness.

[0014] Another aspect of the present invention provides a computer storage medium having a program stored thereon, which, when executed by a processor, implements the graph neural network training method based on debiased topological attribute awareness.

[0015] The present invention has at least the following beneficial effects

[0016] This invention achieves robustness enhancement and attack defense for graph neural networks through multi-dimensional technological innovation. This solution accurately identifies sensitive attribute features through the integral gradient method, locates the attribute dimensions of the model that are vulnerable to attack from the root, fuses topological transfer and attribute transfer probabilities with the help of the debiasing coefficient, and calculates unbiased attribute similarity in combination with the personalized page ranking algorithm. It can effectively distinguish between natural connections and adversarial edges, reconstruct an attack-resistant graph topology structure, and significantly reduce the impact of malicious edge modification on the model; by constructing a loss function containing a mutual information regularization term, it forces the model to reduce its dependence on sensitive attributes, improves the robustness of representation learning, and effectively resists topological perturbations and attribute poisoning attacks, ensuring the prediction accuracy and stability of graph neural networks in adversarial environments, and providing reliable model defense capabilities for key business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0019] See also Figure 1 The present invention provides a graph neural network training method based on debiased topological attribute awareness, which is characterized by comprising:

[0020] S1: Obtain a relational network graph for training a graph neural network, wherein the relational network graph includes a node set corresponding to multiple business object nodes, an attribute feature matrix, an edge set formed by association relationships between the business object nodes, and business labels of the business object nodes;

[0021] Preferably, the multiple business object nodes are business object nodes corresponding to multiple business objects, and the multiple business objects include: users, commodities, merchants or events.

[0022] In this embodiment, taking the e-commerce platform as an example, step S1 obtains a relationship network diagram containing 1,000 product nodes. The node attributes include features such as price and brand. The edge set is formed by user purchase behavior, and the business label is product classification, etc.

[0023] S2: Based on the set baseline feature matrix and the attribute feature matrix of the relationship network graph, the integral gradient method is used to calculate the attribution value of each attribute feature to the output of the graph neural network model, and the attribute feature with the largest attribution value is selected as the sensitive attribute feature;

[0024] Preferably, step S2 includes:

[0025] S21: Set the baseline characteristic matrix to meaningless values;

[0026] S22: For each attribute feature dimension, calculate its attribution value:

[0027]

[0028] in, Indicates the The attributed value of each attribute feature; The first column of the attribute feature matrix representing the relationship network graph attribute feature values; The first attribute feature values; m represents the number of sampling steps set; represents the baseline characteristics matrix; The attribute feature matrix representing the relationship network graph; Represents the output of the GNN model f(·) for the The gradient of an attribute feature; represents the input of the GNN model f(·);

[0029] S23: Select the attribute feature with the largest attribution value as the sensitive attribute feature:

[0030]

[0031] in, Represents sensitive attribute features, ||2 represents the L2 norm, and argmax represents the maximum value.

[0032] In this embodiment, step S2 calculates attribute feature attribution values ​​and selects sensitive attributes using the integral gradient method. This can accurately locate the attribute features that have the greatest impact on the output of the graph neural network, provide a clear target for subsequent debiasing processing, effectively avoid the model's over-reliance on sensitive attributes, enhance defense capabilities against attribute poisoning attacks, and lay the foundation for topology reconstruction and mutual information estimation. Taking social network user data as an example, after setting the baseline feature matrix to a meaningless value, for example, setting the baseline feature matrix to a matrix of all zeros, the attribution value is calculated for each user attribute (such as age, gender, and interest tags) according to the formula. It is assumed that the "gender" attribute is selected as a sensitive attribute feature after the attribution value is calculated, thereby achieving accurate identification of key attributes.

[0033] S3: Calculate the unbiased attribute similarity between business object nodes based on whether adjacent business object nodes have the same sensitive attribute features, the attribute feature similarity between business object nodes, and the topological structure of the relationship network graph;

[0034] Preferably, step S3 includes:

[0035] S31: dividing the edge set of the relationship network graph into homogeneous edge subsets and heterogeneous edge subsets according to whether the business object nodes have the same sensitive attribute characteristics, and calculating the debiasing coefficient based on the homogeneous edge subsets and the heterogeneous edge subsets;

[0036] In this embodiment, step S31 divides the edge set based on whether business object nodes share the same sensitive attribute characteristics and calculates the debiasing coefficient. This quantifies the impact of sensitive attributes on edge distribution and balances the weights of homogeneous and heterogeneous edges through the debiasing coefficient, providing a bias correction basis for subsequent transition probability calculations. This effectively weakens the dominant role of sensitive attributes in node associations, allowing the model to focus more on the true similarity of structure and non-sensitive attributes. Taking the e-commerce platform product edge set as an example, if the sensitive attribute is "brand," the edges between products with the same brand are divided into a homogeneous edge subset E0, and the edges with different brands are divided into a heterogeneous edge subset E1. The debiasing coefficient is calculated. For example, if E0 has 200 edges and E1 has 150 edges, then γ0 = 150 / 200 = 0.75 and γ1 = 200 / 150 ≈ 1.33. This coefficient is used to adjust the transition probability calculation and reduce the excessive influence of brand-sensitive attributes on product associations.

[0037] S32: Calculating the attribute transfer probability between the business object nodes based on the attribute feature similarity between the business object nodes;

[0038] Preferably, the attribute transfer probability between the business object nodes includes:

[0039]

[0040] Among them, P a [v i ,v j ] represents the business object node v i To business object node v j Attribute transition probability; V represents the node set corresponding to multiple business object nodes in the relationship network graph; Represents the business object node v i Attribute feature vector of ; Represents the business object node v j The transpose of the attribute feature vector of ; Represents the business object node v l The transpose of the attribute eigenvector; Indicates transpose.

[0041] In this embodiment, step S32 calculates the attribute transfer probability between business object nodes through attribute feature similarity, which can quantify the association strength of nodes based on attribute features, provide attribute similarity transfer basis for non-topological connections between nodes in the graph, and enable the model to capture node associations that are not directly reflected in the topological structure but have highly similar attributes, laying the foundation for subsequent unbiased attribute similarity calculation.

[0042] S33: Calculating the topological transition probability between adjacent business object nodes according to the topological structure of the relationship network graph;

[0043] Preferably, the topology transition probability between adjacent business object nodes includes:

[0044] P t [v i ,v k ]=1 / d(v i )

[0045] Among them, P t [v i ,v k ] represents the business object node v i To the adjacent business object node v k The topological transition probability of business object node v k Belongs to business object node v i Adjacent nodes; d(v i ) represents the business object node v i degree.

[0046] In this embodiment, step S33 calculates the topological transition probability of adjacent nodes based on the topological structure of the relationship network graph, quantifies the possibility of node transfer based on topological connections, and provides a probability measure for the information propagation of directly connected nodes in the graph, ensuring that the model can capture the local connection characteristics of the topological structure. When combined with the attribute transition probability, it provides topological dimension support for global unbiased similarity calculation.

[0047] S34: Calculating the transition probability between business nodes based on the attribute transition probability between business object nodes and the topological transition probability between adjacent business object nodes in combination with the debiasing coefficient;

[0048] Preferably, the transition probability between the service nodes includes:

[0049]

[0050] Among them, P T [v i ,v j ] represents the business object node v i To the adjacent business object node v jThe transition probability of γ0 and γ1 are debiasing coefficients; |ε0| is the number of homogeneous edges in the homogeneous edge subset; |ε1| is the number of heterogeneous edges in the heterogeneous edge subset; β is a free setting parameter with a value range of 0 to 1; P t [v i ,v j ] represents the business object node v i To the adjacent business object node v j The topological transition probability P a [v i ,v j ] represents the business object node v i To business object node v j Attribute transition probability; when the business object node v i and business object node v j When not adjacent, P t [v i ,v j ]=0.

[0051] In this embodiment, step S34 combines the attribute transfer probability, topology transfer probability and debiasing coefficient to calculate the business node transfer probability, which can balance the influence of sensitive attributes and topological structure on node association, weaken the sensitive attribute deviation through the debiasing coefficient, and make the transfer probability reflect the true association between topological connection and attribute similarity at the same time, providing a transfer model that integrates multi-dimensional information for unbiased attribute similarity calculation.

[0052] S35: using a personalized page ranking algorithm based on the transition probability between the business nodes, and calculating the unbiased attribute proximity between the business nodes by truncating the infinite series as the unbiased attribute similarity between the business object nodes;

[0053] Preferably, the unbiased attribute similarity between adjacent business object nodes includes:

[0054]

[0055] Among them, S[v i ,v j ] represents the business object node v i and business object node v j Unbiased attribute similarity between i ,v j ] represents the business object node v i and business object node v j The unbiased attribute closeness between them; L represents the number of truncation steps; α represents the stay probability; P T [v i ,v j ] represents the business object node v iTo the adjacent business object node v j The transition probability of Indicates P T [v i ,v j ] to the power of 1.

[0056] In this embodiment, step S35 uses a personalized page ranking algorithm combined with transition probability to calculate unbiased attribute similarity, which can capture the long-term association of multi-order transitions between nodes by truncating infinite series, eliminate local connection deviations caused by sensitive attributes, and enable the similarity measurement to integrate the global association information of topological structure and attribute features, providing accurate node similarity basis for topology reconstruction.

[0057] S4: reconstructing the topological structure of the relationship network graph according to the unbiased attribute similarity between the business object nodes to obtain a reconstructed relationship network graph;

[0058] Preferably, step S4 includes:

[0059] S41: constructing a similarity matrix based on the unbiased attribute similarity between business object nodes, selecting K pairs of business object nodes with the highest similarity and connecting them to construct a similarity edge set;

[0060] S42: Delete the edges that exist in the edge set of the relationship network graph but do not exist in the similarity edge set as abnormal edges, and add the edges that exist in the similarity edge set but do not exist in the edge set of the relationship network graph as legal edges; and obtain a reconstructed relationship network graph.

[0061] In this embodiment, step S4 reconstructs the relationship network graph based on the unbiased attribute similarity between business object nodes. This can identify and delete abnormal edges and add legitimate edges through the similarity matrix, effectively filtering out false connections caused by adversarial attacks, enhancing the graph structure's anti-attack resistance, making the reconstructed graph more consistent with the true association between node attributes and topology, and improving the model's representation learning ability in adversarial environments. Taking the e-commerce platform as an example, after constructing a matrix based on the unbiased attribute similarity of product nodes, the 50 pairs of product nodes with the highest similarity are selected to construct a similarity edge set. Edges that exist in the original edge set but not in the similarity edge set are deleted as abnormal edges, and edges that exist in the similarity edge set but not in the original edge set are added as legitimate edges. This results in a reconstructed relationship network graph that is resistant to attack and optimizes the product association structure.

[0062] S5: Input the reconstructed relationship network graph into the graph neural network model for forward propagation, and estimate the mutual information between the sensitive attribute features and the prediction results based on the prediction results of the graph neural network model;

[0063] S6: Construct the final loss function of the graph neural network model by combining the task loss of the graph neural network model, the business label of the business object node, and the mutual information between the sensitive attribute features and the prediction results;

[0064] Preferably, the final loss function of the graph neural network model includes:

[0065]

[0066] in, represents the final loss function of the graph neural network model, λ represents the weight parameter, represents the task loss constructed by the prediction result of the business object node and the business label; I represents the mutual information between the sensitive attribute feature and the prediction result; Represents sensitive attribute characteristics, Indicates the prediction result; Represents sensitive attribute characteristics and model prediction results The joint probability distribution of represents the discriminant function used to estimate the lower bound of mutual information; Indicates sensitive attributes and the predicted results The product of the marginal probability distributions of ; e represents the natural base; log represents the natural logarithm.

[0067] In this embodiment, steps S5-S6 input the reconstructed relationship network graph into the model forward propagation to estimate the mutual information between the sensitive attribute and the prediction result, and combine the task loss to construct the final loss function, which can quantify the influence of the sensitive attribute on the prediction result, and suppress the excessive dependence of the model on the sensitive attribute through the mutual information regularization term, so as to achieve robustness optimization under adversarial attacks. In this embodiment, D(z) represents the discriminant function used to estimate the lower bound of the mutual information, which is usually parameterized by a neural network (such as a multi-layer perceptron), with the input being the spliced ​​features of the sensitive attribute and the prediction result, and the output being a scalar value; by optimizing the target max D ∑ z P(z)D(z)-log(∑ z Q(z)e D(z) ) Adjust parameters and quantify sensitive attributes and the predicted results degree of association; Represents the task loss constructed by the prediction result of the business object node and the business label, usually using the cross entropy loss function; Represents sensitive attribute characteristics and model prediction results The joint probability distribution of And the prediction result is The number of nodes is divided by the number of summary points in the relationship network diagram; Indicates sensitive attributes and the predicted results The marginal probability distribution product of Indicates sensitive attributes The marginal probability distribution probability of , which is composed of sensitive attributes The number of nodes is divided by the number of summary points in the relationship network diagram; Represents the prediction results The marginal probability distribution probability of is given by the prediction result: The number of nodes is divided by the number of summary points in the relationship network diagram.

[0068] S7: Perform backpropagation based on the final loss function of the graph neural network model to update the parameters of the graph neural network model, and repeat steps S2-S7 until the final loss function converges.

[0069] Another aspect of the present invention provides an ecological environment dynamic monitoring system based on remote sensing and geographic information system, including: a memory and a processor; the memory is used to store applications; the processor is used to run the applications and execute the graph neural network training method based on debiased topological attribute awareness.

[0070] Another aspect of the present invention provides a computer storage medium having a program stored thereon, which, when executed by a processor, implements the graph neural network training method based on debiased topological attribute awareness.

[0071] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0072] In summary, the present invention achieves robustness enhancement and attack defense for graph neural networks through multi-dimensional technological innovation. This solution accurately identifies sensitive attribute features through the integral gradient method, locates the attribute dimensions of the model that are vulnerable to attack from the root, fuses topological transfer and attribute transfer probabilities with the help of the debiasing coefficient, and calculates unbiased attribute similarity in combination with the personalized page ranking algorithm. It can effectively distinguish between natural connections and adversarial edges, reconstruct an attack-resistant graph topology structure, and significantly reduce the impact of malicious edge modification on the model; by constructing a loss function containing a mutual information regularization term, the model is forced to reduce its dependence on sensitive attributes, improving the robustness of representation learning, effectively resisting topological perturbations and attribute poisoning attacks, ensuring the prediction accuracy and stability of graph neural networks in adversarial environments, and providing reliable model defense capabilities for key business scenarios.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A graph neural network training method based on debiased topological attribute awareness, characterized in that: include: S1: Obtain a relational network graph for training a graph neural network, wherein the relational network graph includes a node set corresponding to multiple business object nodes, an attribute feature matrix, an edge set formed by association relationships between the business object nodes, and business labels of the business object nodes; S2: Based on the set baseline feature matrix and the attribute feature matrix of the relationship network graph, the integral gradient method is used to calculate the attribution value of each attribute feature to the output of the graph neural network model, and the attribute feature with the largest attribution value is selected as the sensitive attribute feature; S3: Calculate the unbiased attribute similarity between business object nodes based on whether adjacent business object nodes have the same sensitive attribute features, the attribute feature similarity between business object nodes, and the topological structure of the relationship network graph; S4: reconstructing the topological structure of the relationship network graph according to the unbiased attribute similarity between the business object nodes to obtain a reconstructed relationship network graph; S5: Input the reconstructed relationship network graph into the graph neural network model for forward propagation, and estimate the mutual information between the sensitive attribute features and the prediction results based on the prediction results of the graph neural network model; S6: Construct the final loss function of the graph neural network model by combining the task loss of the graph neural network model, the business label of the business object node, and the mutual information between the sensitive attribute features and the prediction results; S7: Perform backpropagation based on the final loss function of the graph neural network model to update the parameters of the graph neural network model, and repeat steps S2-S7 until the final loss function converges.

2. A graph neural network training method based on debiased topological attribute awareness according to claim 1, characterized in that: The multiple business object nodes are business object nodes corresponding to multiple business objects, and the multiple business objects include: users, commodities, merchants or events.

3. A graph neural network training method based on debiased topological attribute awareness according to claim 1, characterized in that: The step S2 comprises: S21: Set the baseline characteristic matrix to meaningless values; S22: For each attribute feature dimension, calculate its attribution value: in, Indicates the The attributed value of each attribute feature; The first column of the attribute feature matrix representing the relationship network graph attribute feature values; The first attribute feature values; m represents the number of sampling steps set; represents the baseline characteristics matrix; Attribute feature matrix representing the relationship network graph; Represents the output of the GNN model f(·) for the The gradient of an attribute feature; represents the input of the GNN model f(·); S23: Select the attribute feature with the largest attribution value as the sensitive attribute feature: in, Represents sensitive attribute features, ||2 represents the L2 norm, and argmax represents the maximum value.

4. A graph neural network training method based on debiased topological attribute awareness according to claim 1, characterized in that: The step S3 comprises: S31: dividing the edge set of the relationship network graph into homogeneous edge subsets and heterogeneous edge subsets according to whether the business object nodes have the same sensitive attribute characteristics, and calculating the debiasing coefficient based on the homogeneous edge subsets and the heterogeneous edge subsets; S32: Calculating the attribute transfer probability between the business object nodes based on the attribute feature similarity between the business object nodes; S33: Calculating the topological transition probability between adjacent business object nodes according to the topological structure of the relationship network graph; S34: Calculating the transition probability between business nodes based on the attribute transition probability between business object nodes and the topological transition probability between adjacent business object nodes in combination with the debiasing coefficient; S35: Using a personalized page ranking algorithm based on the transition probability between the business nodes, the unbiased attribute proximity between the business nodes is calculated by truncating the infinite series as the unbiased attribute similarity between the business object nodes.

5. A graph neural network training method based on debiased topological attribute awareness according to claim 4, characterized in that: The attribute transfer probability between the business object nodes includes: Among them, P a [v i ,v j ] represents the business object node v i To business object node v j Attribute transition probability; V represents the node set corresponding to multiple business object nodes in the relationship network graph; Represents the business object node v i Attribute feature vector of ; Represents the business object node v j The transpose of the attribute feature vector of ; Represents the business object node v l The transpose of the attribute eigenvector; Indicates transpose.

6. A graph neural network training method based on debiased topological attribute awareness according to claim 4, characterized in that: The topology transition probability between adjacent business object nodes includes: P t [v i ,v k ]=1 / d(v i ) Among them, P t [v i ,v k ] represents the business object node v i To the adjacent business object node v k The topological transition probability of business object node v k Belongs to business object node v i Adjacent nodes; d(v i ) represents the business object node v i degree.

7. A graph neural network training method based on debiased topological attribute awareness according to claim 4, characterized in that: The transition probability between the service nodes includes: Among them, P T [v i ,v j ] represents the business object node v i To the adjacent business object node v j The transition probability of γ0 and γ1 are debiasing coefficients; |ε0| is the number of homogeneous edges in the homogeneous edge subset; |ε1| is the number of heterogeneous edges in the heterogeneous edge subset; β is a free setting parameter with a value range of 0 to 1; P t [v i ,v j ] represents the business object node v i To the adjacent business object node v j The topological transition probability P a [v i ,v j ] represents the business object node v i To business object node v j Attribute transition probability; when the business object node v i and business object node v j When not adjacent, P t [v i ,v j ]=0.

8. A graph neural network training method based on debiased topological attribute awareness according to claim 4, characterized in that: The unbiased attribute similarity between adjacent business object nodes includes: Among them, S[v i ,v j ] represents the business object node v i and business object node v j Unbiased attribute similarity between i ,v j ] represents the business object node v i and business object node v j The unbiased attribute closeness between them; L represents the number of truncation steps; α represents the stay probability; P T [v i ,v j ] represents the business object node v i To the adjacent business object node v j The transition probability of Indicates P T [v i ,v j ] to the power of 1.

9. A graph neural network training method based on debiased topological attribute awareness according to claim 4, characterized in that: The step S4 comprises: S41: constructing a similarity matrix based on the unbiased attribute similarity between business object nodes, selecting K pairs of business object nodes with the highest similarity and connecting them to construct a similarity edge set; S42: Delete the edges that exist in the edge set of the relationship network graph but do not exist in the similarity edge set as abnormal edges, and add the edges that exist in the similarity edge set but do not exist in the edge set of the relationship network graph as legal edges; and obtain a reconstructed relationship network graph.

10. A graph neural network training method based on debiased topological attribute awareness according to claim 1, characterized in that: The final loss function of the graph neural network model includes: in, represents the final loss function of the graph neural network model, λ represents the weight parameter, represents the task loss constructed by the prediction result of the business object node and the business label; I represents the mutual information between the sensitive attribute feature and the prediction result; Represents sensitive attribute characteristics, Indicates the prediction result; Represents sensitive attribute characteristics and model prediction results The joint probability distribution of represents the discriminant function used to estimate the lower bound of mutual information; Indicates sensitive attributes and the predicted results The product of the marginal probability distributions of ; e represents the natural base; log represents the natural logarithm.

Images