Improved graph algorithm knowledge mining and dynamic visualization analysis method for credit
By constructing a heterogeneous network for credit business and employing graph attention network algorithms and dynamic visualization methods, the problems of existing credit graph algorithms being unable to distinguish the types of relationships and lacking visualization are solved, thereby improving the accuracy of risk identification and the efficiency of risk control decisions.
Patent Information
- Application Number
- CN202611085317.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing credit graph algorithm models cannot effectively distinguish the differences between different types of relationships, lack dynamic adjustment capabilities, have high computational costs, and lack visualization, resulting in inaccurate risk assessments and difficulty in being understood by credit risk control personnel.
A heterogeneous network for credit business is constructed, semantic labels for risk transmission are defined, a graph attention network algorithm is used for differentiated aggregation, dynamic weights are calculated based on preset meta-paths, local updates are performed in conjunction with an incremental community discovery method, and a multi-view dynamic visualization is generated.
It achieves accurate expression of risk transmission logic for different relationship types, improves risk identification accuracy and interpretability, reduces computational overhead, and improves risk control decision-making efficiency.
Smart Images

Figure CN123115259A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of financial technology, knowledge graphs and graph data mining, and in particular to an improved graph algorithm for knowledge mining and dynamic visualization analysis for credit. Background Technology
[0002] Credit business involves complex networks of relationships, including guarantee chains between borrowers and guarantors, relationships between borrowers and existing clients, equity control chains between companies, and employment-related chains among directors, supervisors, and senior management. Traditional expert scoring models and scoring cards can only address single behavioral characteristics of clients and cannot handle the massive and heterogeneous data relationships described above, making it difficult to accurately assess the transmission effect of credit risk within these complex networks.
[0003] In recent years, graph algorithms have been increasingly applied to the field of credit business. The first approach involves using ordinary graph databases and path search algorithms to explicitly discover related chains. The second approach uses community detection algorithms to uncover subgraph structures and community customer attributes within relationships. The third approach uses graph neural networks to uncover potential implicit risk relationships. These existing technologies explore the application possibilities of graph algorithms in credit scenarios from different perspectives.
[0004] However, the aforementioned existing technologies still have the following shortcomings. First, most existing graph algorithm models construct credit network models as isomorphic graphs, meaning that different customer relationships ultimately form the same network. This simplification loses the key differences in business semantics. For example, guarantee relationships and mortgage relationships differ fundamentally in their transmission mechanisms. Guarantee relationships involve joint and several liability, resulting in high risk transmission intensity and symmetrical direction, while mortgage relationships only involve asset coverage for specific debts, leading to low risk transmission intensity and unidirectionality. Existing technologies cannot distinguish the differentiated transmission logic of different relationship types within the same model. Second, some advanced solutions use meta-paths to explicitly construct models to handle complex network relationships, but existing solutions use fixed meta-path weights for all business scenarios, lacking dynamic adjustment capabilities. In actual credit business, the risk sensitivity of the same meta-path varies greatly depending on the period, customer group, and macroeconomic environment. For example, during a real estate downturn, the risk weight of the enterprise-collateral-land meta-path should increase significantly, but existing technologies cannot dynamically adjust the focus of mining based on real-time business indicators. Third, credit relationships are highly dynamic, with events such as changes in ownership, new loans, and corporate defaults constantly occurring. Existing technologies that employ periodic full recalculation suffer from high computational overhead and response latency, while simple incremental updates lose historical evolution patterns. More critically, most solutions only output the final risk score or node classification results, lacking a dynamic visualization of the mining process. Credit risk control personnel cannot intuitively understand how risk is transmitted through the network, which intermediary nodes play a key role in propagation, and how the risk community expands or contracts over time, making it difficult for business stakeholders to trust and use the model results. Summary of the Invention
[0005] This invention aims to overcome the shortcomings of the prior art, and specifically provides the following technical solution: 1) In a first aspect, the present invention provides an improved graph algorithm knowledge mining and dynamic visualization analysis method for credit, the specific technical solution of which is as follows: S1. Construct a heterogeneous network for credit business, wherein the heterogeneous network for credit business includes multiple entity nodes and multiple relation edges, and define a risk transmission semantic label for each type of relation edge; S2. Determine the initial transmission strength weight value for each of the relation edges based on the type of each relation edge and the risk transmission semantic label; S3. Execute the graph attention network algorithm on the heterogeneous network of the credit business. In the attention aggregation process of the graph attention network algorithm, according to the type of each relation edge, the initial transmission strength weight value of each relation edge and the current aggregation weight of each relation edge, the feature vector of each entity node is updated by the corresponding differentiated aggregation formula for each type of relation edge, so as to obtain the updated feature vector of each entity node. S4. In the heterogeneous network of credit business, multiple meta-paths are preset, the dynamic weight of each meta-path at the current time is calculated, the aggregation weight of the corresponding relation edge is adjusted according to the dynamic weight of each meta-path, the process returns to step S3 and re-executes step S3 according to the adjusted aggregation weight until the preset conditions are met, and then the final updated feature vector is output. S5. Perform community partitioning on the current credit network snapshot based on the final updated feature vector to obtain the current community set. When the entity nodes or relation edges in the heterogeneous credit business network change, determine the set of affected entity nodes based on the change content, and perform local update on the community set based on the set of affected entity nodes to obtain the updated community set and record the community change data. S6. Generate multi-view dynamic visualization display data based on the final updated feature vector, the community change data, and the updated community set.
[0006] The beneficial effects of the improved graph algorithm knowledge mining and dynamic visualization analysis method for credit-oriented applications provided by this invention are as follows: By constructing a heterogeneous credit business network containing various entity nodes and relational edges, and defining a risk transmission semantic label for each relational edge, the graph mining algorithm can distinguish the differences in risk transmission mechanisms among different relation types. By setting an initial transmission strength weight function for each relational edge, the initial transmission strength of the relational edge is quantitatively calculated, providing a data foundation for subsequent differentiated aggregation. During the attention aggregation process of the graph attention network algorithm, the feature vectors of entity nodes are updated using guarantee aggregation formulas, mortgage aggregation formulas, and employment aggregation formulas for guarantee relational edges, mortgage aggregation formulas, and employment aggregation formulas, respectively. This ensures that the risk transmission logic of different relational types can be accurately expressed in feature updates, thereby improving the accuracy of risk identification. By pre-setting multiple meta-paths and calculating the dynamic weight of each meta-path at the current time, the aggregation weight of relational edges is adjusted according to the dynamic weights, enabling the graph mining model to adaptively adjust the mining focus based on real-time business indicators, overcoming the limitations of static weights. Through an incremental community detection method, community updates are performed only on local areas affected by changes, preserving historical evolution patterns while ensuring computational efficiency, and achieving dynamic tracking of risk evolution paths. By generating topological view data, risk heat propagation view data, and time-series propagation trend view data, the mining process is organically integrated with multi-view dynamic visualization, enabling credit risk control personnel to intuitively understand the risk transmission path and community evolution trend, thereby improving the interpretability of risk control decisions and the efficiency of human-machine collaboration.
[0007] 2) Secondly, the present invention also provides an improved graph algorithm knowledge mining and dynamic visualization analysis system for credit, the specific technical solution of which is as follows: This includes building a definition module, determining a module, executing an aggregation module, updating a module, dividing an update module, and generating a module; The definition module is used to: construct a heterogeneous network for credit business, which contains multiple entity nodes and multiple relation edges, and define risk transmission semantic labels for each type of relation edge; The determination module is used to: determine the initial transmission strength weight value of each relation edge based on the type of each relation edge and the risk transmission semantic label; The aggregation module is used to: execute the graph attention network algorithm on the heterogeneous network of credit business. In the attention aggregation process of the graph attention network algorithm, according to the type of each relation edge, the initial transmission strength weight value of each relation edge, and the current aggregation weight of each relation edge, the feature vector of each entity node is updated by the corresponding differentiated aggregation formula for each type of relation edge, so as to obtain the updated feature vector of each entity node. The update module is used to: preset multiple meta-paths in the heterogeneous network of credit business, calculate the dynamic weight of each meta-path at the current time, adjust the aggregation weight of the corresponding relation edges according to the dynamic weight of each meta-path, call and execute the aggregation module again and process according to the adjusted aggregation weight, until the preset conditions are met and the final updated feature vector is output. The partitioning and updating module is used to: perform community partitioning on the current credit network snapshot based on the final updated feature vector to obtain the current community set; when the entity nodes or relation edges in the heterogeneous credit business network change, determine the set of affected entity nodes based on the change content, perform local updates on the community set based on the set of affected entity nodes, obtain the updated community set, and record the community change data. The generation module is used to generate multi-view dynamic visualization data based on the final updated feature vector, community change data, and updated community set.
[0008] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device implements any of the above-mentioned improved graph algorithm knowledge mining and dynamic visualization analysis methods for credit.
[0009] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned improved graph algorithm knowledge mining and dynamic visualization analysis methods for credit.
[0010] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description
[0011] Figure 1 This is a flowchart illustrating an improved graph algorithm-based knowledge mining and dynamic visualization analysis method for credit in accordance with an embodiment of the present invention. Figure 2 A schematic diagram of entity node risk scoring and relation edge weights; Figure 3 A heterogeneous network topology view for credit operations; Figure 4 A schematic diagram illustrating the heat dissipation of heterogeneous network risks in credit operations; Figure 5 A schematic diagram illustrating the temporal propagation trend of heterogeneous networks in credit transactions; Figure 6 This is a schematic diagram of the structure of an improved graph algorithm knowledge mining and dynamic visualization analysis system for credit in accordance with an embodiment of the present invention. Detailed Implementation
[0012] like Figure 1 As shown in the figure, an improved graph algorithm knowledge mining and dynamic visualization analysis method for credit-related applications according to an embodiment of the present invention includes the following steps: S1. Construct a heterogeneous network for credit business, wherein the heterogeneous network for credit business includes multiple entity nodes and multiple relation edges, and define a risk transmission semantic label for each type of relation edge; The types of entity nodes include borrower nodes, enterprise nodes, collateral nodes, personnel nodes, and credit product nodes; the types of relationship edges include mortgage relationship edges, guarantee relationship edges, and employment relationship edges; a risk transmission semantic label is defined for each type of relationship edge, including: joint liability transmission for guarantee relationship edges, which means that risk is transmitted from the guarantor to the borrower along the direction of the guarantee relationship, and the guarantor's default risk will directly affect the debt repayment ability of the guaranteed party; asset coverage transmission for mortgage relationship edges, which means that changes in the value of collateral affect the borrower's credit limit, and the direction of risk transmission is negatively correlated with the loan-to-value ratio, with a high loan-to-value ratio corresponding to low risk; and key personnel transmission for employment relationship edges, which means that risk is transmitted bidirectionally between personnel nodes and enterprise nodes, and that if a key personnel is involved in a risk event, all the enterprises they serve will be affected, with risk being transmitted bidirectionally between personnel nodes and enterprise nodes.
[0013] The heterogeneous network for credit business is a graph structure network constructed by using different types of entities in the credit business as nodes and different business relationships between entities as edges. The types of nodes and edges in the network are not uniform, thus preserving the semantic differences between different business relationships. Entity nodes are the vertices in the heterogeneous network for credit business, representing the actual business objects in the credit business, including different types such as borrowers, enterprises, collateral, personnel, and credit products. Relationship edges are the edges connecting two entity nodes in the heterogeneous network for credit business, representing the specific business relationship between the two entity nodes, including different types such as mortgage relationships, guarantee relationships, and employment relationships. Risk transmission semantic labels are business semantic identifiers assigned to each relationship edge, determining the logic and direction of risk propagation in the graph algorithm, and are used to distinguish the differences in risk transmission mechanisms between different relationship types. Borrower nodes are node types in the heterogeneous network for credit business, representing natural or legal entities that apply for or have received loans. Their English abbreviation is BORROWER, and their attributes include information such as income, debt ratio, historical delinquency count, and number of loan accounts. Enterprise nodes, in the heterogeneous network of credit business, represent the enterprise to which the borrower belongs. Their attributes include industry, tax class, social security and housing provident fund contribution ratios, and years of operation. Collateral nodes, in the heterogeneous network of credit business, represent the physical assets used as collateral or pledge by the borrower. Their attributes include appraised value and loan-to-value ratio. Personnel nodes, in the heterogeneous network of credit business, represent the directors, supervisors, and senior management personnel of the enterprise. Their attributes include job title and a list of companies they work for. Credit product nodes, in the heterogeneous network of credit business, represent loan products offered by banks or financial institutions. Their attributes include product type, term, and interest rate. Mortgage edges, in the heterogeneous network of credit business, connect borrower nodes and collateral nodes, with the direction from borrower to collateral. Their attributes include mortgage amount and loan-to-value ratio. A guarantee relationship edge connects the guarantor node and the borrower node, with the direction from the guarantor to the borrower. Its attributes include the guarantee amount, guarantee method, and guarantee start date. An employment relationship edge connects the personnel node and the enterprise node, with the direction from the personnel to the enterprise. Its attributes include job title and employment status. Joint liability transmission is the risk transmission semantic tag corresponding to the guarantee relationship edge, represented by JOINT_LIABILITY. This indicates that the guarantor bears joint liability for the borrower's debt. When the borrower defaults, the guarantor must repay the debt on their behalf. The risk is transmitted from the guarantor to the borrower along the guarantee relationship direction, with high transmission strength and symmetrical direction.Asset Coverage Transmission is the semantic tag for risk transmission corresponding to the mortgage relationship edge, represented by ASSET_COVER. It indicates that changes in the value of the collateral will affect the borrower's credit limit and debt repayment ability. The direction of risk transmission is negatively correlated with the loan-to-value ratio, meaning that the higher the loan-to-value ratio, the lower the risk. The transmission strength is low and unidirectional. Key Personnel Transmission is the semantic tag for risk transmission corresponding to the employment relationship edge, represented by PERSON_LINK. It indicates that the companies employed by a person are all affected by the person's risk status, and the company's risk will also affect the person in turn. Risk is transmitted bidirectionally between the person and company nodes.
[0014] S2. Based on the type of each relation edge and the risk transmission semantic label, determine the initial transmission strength weight value for each relation edge, including: An initial transmission strength weight function is defined for each type of relation edge. The attribute information of each relation edge is used as the input parameter of the initial transmission strength weight function to calculate the initial transmission strength weight value of each relation edge. The specific implementation process is as follows: Define the initial transmission strength weight function for each type of relation edge. For each relation edge of type "guarantee relation edge", calculate the initial transmission strength weight value of the relation edge using the initial transmission strength weight function for guarantee relations. The calculation formula is as follows: ,in, This represents a specific guarantee relationship. This indicates the starting node of this guarantee relationship edge, which is the guarantor node. This indicates that the target node of this guarantee relationship edge is the borrower node. This indicates the type of the edge relationship, i.e., the guarantee relationship. This represents the set of attribute information for this guarantee relationship edge. This represents the initial transmission strength weight value corresponding to this guarantee relationship edge. The guarantee amount is obtained from the attribute information of this guarantee relationship edge. The guarantor's net assets are obtained from the attribute information of the guarantor node. The borrower's liabilities are obtained from the attribute information of the borrower node. This indicates that the minimum value between the guarantor's net assets and the borrower's liabilities is taken. The actual risk transmission is limited by the weaker link between the amount the guarantor can pay and the amount the borrower owes, hence the minimum value is taken. The joint liability factor ranges from 0.5 to 1.
[0015] For each relation edge of type "mortgage relation edge", the initial transmission strength weight value of the relation edge is calculated using the initial transmission strength weight function of the mortgage relation. The calculation formula is as follows: ,in, This represents a specific mortgage relationship edge. This indicates the starting node of this mortgage relationship edge, which is the borrower node. This indicates that the target node of this mortgage relationship edge is the mortgaged node. This indicates the type of this relation edge, namely, a mortgage relation. This represents the set of attribute information for this mortgage relationship edge. This represents the initial transmission strength weight value corresponding to this mortgage relationship edge. The mortgage guarantee amount is obtained from the attribute information of this mortgage relationship edge. The appraised value of the collateral is obtained from the attribute information of the collateral node.
[0016] For each relation edge of type "employment relationship edge", the initial transmission strength weight value of the relation edge is calculated using the initial transmission strength weight function of the employment relationship. The calculation formula is as follows: ,in, This indicates a specific employment relationship. This indicates the starting node of this employment relationship edge, which is the personnel node. This indicates that the target node of this employment relationship edge is the enterprise node. This indicates the type of this relationship edge, namely, the employment relationship. This represents the set of attribute information for this employment relationship edge. This represents the initial transmission strength weight value corresponding to this employment relationship edge. The personnel position level weight is mapped from the position information in the attribute information of this employment relationship edge, and the value ranges from 0 to 1. The positive and negative impact factors of personnel are obtained from the attribute information of personnel nodes and are used to reflect whether the direction of the personnel's influence on the employing company is positive or negative.
[0017] After setting the above three initial conduction strength weight functions, the attribute information of each relation edge is used as the input parameter of the corresponding type of initial conduction strength weight function to calculate the initial conduction strength weight value of each relation edge.
[0018] Attribute information: Attribute information is the business data field carried by the relation edge, including information such as guarantee amount, guarantee method, guarantee start date, mortgage amount, mortgage rate, position and employment status. Different types of relation edges have different sets of attribute information.
[0019] S3. Execute the graph attention network algorithm on the heterogeneous network of the credit business. In the attention aggregation process of the graph attention network algorithm, according to the type of each relation edge, the initial transmission strength weight value of each relation edge and the current aggregation weight of each relation edge, the feature vector of each entity node is updated by the corresponding differentiated aggregation formula for each type of relation edge, so as to obtain the updated feature vector of each entity node. Traditional graph attention network algorithms use a uniform linear transformation for different types of edges during the aggregation process, and their aggregation update formula is: ,in, Represents a node Updated feature vector, Representing neighboring nodes eigenvectors, Represents a node The set of first-order neighbors, Indicates attention weights. This represents a trainable weight matrix. This represents the activation function. Attention weights. Use the softmax function to evaluate the nodes. All neighbors are normalized, and the calculation formula is as follows: ,in, Represents a node With nodes The attention coefficients of corresponding edges are formed by the dot product of the weight matrices of the nodes. This represents the natural exponential function. The original attribute information of each entity node in the heterogeneous network of credit business is encoded into an initial feature vector, denoted as... ,in Represents any entity node. Represents entity nodes The initial feature vector is obtained by using the initial propagation strength weight of each relation edge as the initial aggregation weight of the corresponding relation edge in the graph attention network algorithm. The graph attention network algorithm updates the feature vector of entity nodes by performing the attention aggregation process through multiple rounds of iteration. In each iteration, for each entity node, the feature vectors of all its first-order neighbor entity nodes are collected, and according to the type of relation edge, the initial propagation strength weight, and the current aggregation weight, the corresponding differential aggregation formula is used to aggregate the feature vectors of the neighbor nodes to the current node, resulting in the updated feature vector. The iteration is repeated until the preset number of rounds is reached.
[0020] Specifically, the feature vector of each entity node is updated using a corresponding differentiated aggregation formula for different types of relation edges, including: 1) For relationship edges of type "guarantee relationship edge", the feature vector of the entity node connected to the guarantee relationship edge is updated using the guarantee aggregation formula. The guarantee aggregation formula includes a guarantee strength coefficient, which is determined by the initial propagation strength weight value of the guarantee relationship edge. The guarantee strength coefficient is used to control the degree of influence of the guarantee relationship edge in the aggregation process, as follows: For the target entity node Get all its first-order neighbor entity nodes connected by guarantee relationship edges. The set of, denoted as For each neighboring entity node Based on the initial transmission strength weight value of the guarantee relationship edge. Calculate the guarantee strength coefficient The calculation formula is: ,in, Represents neighboring entity nodes When acting as a guarantor for the target entity node The guarantee strength coefficient, Indicates the connection of entity nodes With the target entity node The initial transmission strength weight value of the guarantee relationship edge. This indicates that the value is the minimum between 1 and the initial conduction strength weight value, which is used to limit the range of the guarantee strength coefficient to between 0 and 1.
[0021] Then, the target entity node is calculated according to the following guarantee aggregation formula. Updated feature vector : ,in, This represents the entity node after the aggregation of guarantee relationships. The updated feature vector. This represents the activation function, used to perform nonlinear transformations on the aggregation results. This represents a trainable weight matrix used for entity nodes. Perform a linear transformation on its own eigenvectors. Represents entity nodes The feature vectors before aggregation in the current round. Represents entity nodes The set of all first-order neighbor entity nodes connected by guarantee relationship edges. Represents neighboring entity nodes For entity nodes The guarantee strength coefficient. Represents entity nodes Feature vectors and entity nodes The difference projection matrix between the feature vectors is used to map the feature differences between the two to a suitable feature space. Represents neighboring entity nodes The feature vectors before aggregation in the current round. Represents neighboring entity nodes The gating selection feature vector is used to control neighboring entity nodes. The degree to which feature information is retained during the aggregation process.
[0022] 2) For relationship edges of type "mortgage relationship edge", the feature vector of the entity node connected to the mortgage relationship edge is updated using the mortgage aggregation formula. The mortgage aggregation formula includes the collateral value ratio and the risk gap coefficient. The collateral value ratio is determined by the proportion of the appraised value of the collateral to the total value of all collateral of the corresponding borrower. The risk gap coefficient is determined by the proportion of the appraised value of the collateral to the loan amount. The specific implementation process is as follows: For the target entity node Get all its first-order neighbor entity nodes connected by the mortgage relationship edge. The set of, denoted as ,in This represents the collateral node. For each neighboring entity node... Obtain the appraised value of the collateral for this collateral node. Obtain the target entity node Total value of all collateral The value of the collateral is calculated as a percentage of the total value. Obtain the target entity node. Calculate the risk gap coefficient based on the loan amount. The calculation formula is: ,in, Represents the collateral node The risk gap coefficient is obtained from the attribute information of the collateral node, and the loan amount is obtained from the attribute information of the target entity node, i.e., the borrower node.
[0023] Then, the target entity node is calculated according to the following collateral aggregation formula. Updated feature vector : ,in, The entity node represents the aggregated mortgage relationship. The updated feature vector. Represents entity nodes The feature vectors before aggregation in the current round. Represents entity nodes The set of all first-order neighbor entity nodes connected by collateral relationship edges, where each neighbor entity node All of these are nodes representing collateral. Represents the collateral node The appraised value of the collateral. Represents entity nodes The total value of all collateral, i.e., the entity nodes The sum of the appraised value of the collateral of all associated collateral nodes. Represents the collateral node The risk gap coefficient. This represents the hyperbolic tangent activation function, used for nonlinear mapping of feature differences. Indicates the connection of entity nodes With collateral node The trainable transformation matrix corresponding to the mortgage relationship edge is used to transform the entity node. and The feature differences are projected onto a suitable feature space. Represents the collateral node The feature vectors before aggregation in the current round.
[0024] 3) For relationship edges of type "employment relationship edge", the feature vectors of the personnel nodes and enterprise nodes connected to the employment relationship edge are updated using the employment aggregation formula. The employment aggregation formula includes a first calculation step and a second calculation step. In the first calculation step, the feature vectors of each enterprise node to which the personnel node is employed are merged into the feature vector of the personnel node to obtain a temporary feature vector of the personnel node. In the second calculation step, the risk information in the temporary feature vector of the personnel node is fed back into the feature vectors of each enterprise node. The specific implementation process is as follows: The job aggregation formula includes a first calculation step and a second calculation step. The first calculation step is a fusion step, which merges the feature vectors of all enterprise nodes to which a person node works into the feature vector of that person node, obtaining a temporary feature vector for that person node. The fusion step calculates the feature vector of the person node according to the following formula. Temporary feature vectors : ,in, Represents personnel nodes Temporary eigenvectors. Represents personnel nodes The feature vectors before aggregation in the current round. Represents personnel nodes The set of all enterprise nodes for which the employee serves. Represents enterprise nodes For personnel nodes The influence weight reflects the enterprise node Risks to personnel nodes The importance of the enterprise node is determined by the value. With personnel nodes The information is obtained from the attribute information of the employment relationship between the parties. This represents a trainable weight matrix used for enterprise nodes. Perform a linear transformation on the eigenvectors. Represents enterprise nodes The feature vectors before aggregation in the current round.
[0025] The second calculation step is a divergence step, which involves personnel nodes. The risk information in the temporary feature vector is fed back to the feature vector of each enterprise node. The divergence step calculates the enterprise node's risk information according to the following formula. Updated feature vector : ,in, This indicates the enterprise node after the aggregation of employment relationships. The updated feature vector. Represents enterprise nodes The feature vectors before aggregation in the current round. This indicates all employees working at enterprise nodes. A set of personnel nodes. Represents personnel nodes For enterprise nodes The risk impact weight, this value is derived from the personnel node. With enterprise nodes The information is obtained from the attribute information of the employment relationship between the parties. This indicates that the personnel node The temporary feature vector difference is projected onto the enterprise node. The matrix of the linear characteristic space. Represents personnel nodes Temporary eigenvectors.
[0026] Graph attention network (GAN) is a neural network algorithm that operates on graph-structured data. It updates node features by assigning different attention weights to different neighbors when aggregating their features, thus capturing the differences in importance between nodes. The attention aggregation process is the core computational step in updating node feature vectors in GAN. In this process, each node collects the feature vectors of its neighbors and aggregates them into the current node's feature vector based on the attention weights, resulting in the updated node feature vector.
[0027] S4. In the heterogeneous network of credit business, multiple meta-paths are preset, the dynamic weight of each meta-path at the current time is calculated, the aggregation weight of the corresponding relation edge is adjusted according to the dynamic weight of each meta-path, the process returns to step S3 and re-executes step S3 according to the adjusted aggregation weight until the preset conditions are met, and then the final updated feature vector is output. Wherein, the initial transmission strength weight value is the static baseline value calculated in step S2, the aggregate weight is the weight parameter actually used by the graph attention network algorithm in step S3, and the dynamic weight is the time-varying weight value calculated in step S4 based on real-time business indicators and used to adjust the aggregate weight.
[0028] In this process, multiple meta-paths are pre-defined in the heterogeneous network of credit business, and the specific implementation process is as follows: A meta-path is defined as a sequence of alternating node types and relation types, used to describe a composite association path between two entity nodes formed by specific types of relation edges and intermediate node types. Based on node type and relation type combinations that have significance for risk transmission or association mining in credit business scenarios, multiple meta-paths are predefined in the heterogeneous network of credit business. The preset method for each meta-path is as follows: determine the starting node type, then alternately specify the relation edge type and intermediate node type until the ending node type is reached, forming a complete sequence of alternating node and relation types. For example, in a personal credit scenario, four meta-paths are pre-defined: The first meta-path sequence is borrower node, guarantee relationship edge, guarantor node, guarantee relationship edge, borrower node; this meta-path is used to discover the association between two borrowers through a joint guarantor. The second meta-path sequence is borrower node, employment relationship edge, company node, employment relationship edge, borrower node; this meta-path is used to discover the association between two borrowers through belonging to the same company. The third meta-path sequence is borrower node, guarantee relationship edge, guarantor node, employment relationship edge, company node, employment relationship edge, guarantor node, guarantee relationship edge, borrower node; this meta-path is used to discover the indirect association between two borrowers through the company where the guarantor works. The fourth meta-path sequence is collateral node, mortgage relationship edge, borrower node, mortgage relationship edge, collateral node; this meta-path is used to discover the association between multiple collaterals under the same borrower's name. After pre-setting, each meta-path is stored in the form of its node type sequence and relationship type sequence, serving as input data for subsequent dynamic weight calculation steps.
[0029] Metapaths are sequences of alternating node and relation types defined in heterogeneous networks of credit business. They describe a composite path between two entity nodes that passes through specific types of nodes and relation edges. Different metapaths correspond to different business semantic association patterns and are the foundational objects for subsequent dynamic weight calculation steps.
[0030] The calculation of the dynamic weight of each meta-path at the current moment includes: obtaining the risk semantic score corresponding to each meta-path and calculating the basic weight factor of each meta-path based on the risk semantic score; obtaining the real-time business indicator value at the current moment and calculating the time dynamic adjustment factor of each meta-path based on the real-time business indicator value; obtaining the dynamic weight of each meta-path at historical moments and calculating the historical evolution trend factor of each meta-path based on the historical dynamic weight; and determining the dynamic weight of each meta-path at the current moment based on the basic weight factor, the time dynamic adjustment factor, and the historical evolution trend factor. In traditional meta-path weighting methods, the weights are statically preset and cannot respond to the dynamic changes in risk patterns in the credit business environment. This method uses real-time business indicators as risk sensors and dynamically adjusts the importance of each meta-path in knowledge mining through an adaptive function. The specific implementation process is as follows: 1) Obtain the risk semantic score corresponding to each meta-path, using symbols. It means that among them This represents a specific meta-path. The risk semantic score is a pre-set score based on the risk transmission semantics involved in the credit business within that meta-path, ranging from 0 to 10. This score is determined by the risk transmission semantic label defined in step 2 of Technical Problem 1. The stronger the risk transmission semantics contained in the meta-path, the higher the score. Based on the risk semantic score... Calculate the basic weight factor of this metapath The calculation formula is: ,in, Metapath The basic weighting factor, Metapath Risk semantic score, Represents the natural constant. Represents the negative of the natural constant The formula maps the risk semantic score to a value between 0 and 1 using the sigmoid function, serving as the static weight basis for this meta-path without considering real-time business metrics and historical evolution.
[0031] 2) Obtain the real-time business metric values at the current moment, using symbols. It means that among them Indicates the current moment. This indicates the type of business relationship indicator, which includes indicators for guarantee relationships, mortgage relationships, and employment relationships. Real-time business indicator values are obtained from real-time data in the credit business system. A time-dynamic adjustment factor is calculated for each meta-path based on the real-time business indicator values. The calculation formula is: ,in, Metapath At any moment The time-dynamic adjustment factor. This represents a set of business relationship indicators, including three types: guarantee relationship, mortgage relationship, and employment relationship. Indicates the type of business relationship indicator The global relationship indicator weighting coefficients satisfy all business relationship indicator types. The sum of them equals 1. Indicates the type of business relationship indicator At any moment The value of the metapath Sensitive mapping functions. The calculation formula is: ,in, Indicates the type of business relationship indicator At any moment The value of the metapath The sensitive mapping value. Indicates the type of business relationship indicator Metapath The sensitivity to the impact of changes in this indicator is a value that is pre-set based on business experience, reflecting the impact of changes in this indicator on the metapath. The importance and intensity of its influence. Indicates the type of business relationship indicator At any moment Real-time business metric values. Indicates the type of business relationship indicator The average value over a historical period. Indicates the type of business relationship indicator Standard deviation over a historical period. This is achieved by subtracting the historical mean from the real-time business metric value and then dividing by the historical standard deviation, thus standardizing the real-time business metric value.
[0032] 3) Obtain the dynamic weight of each metapath at historical moments, using symbols. This indicates the metapath. In the previous moment The dynamic weights. Obtain the hidden state of each metapath at the previous time step, using symbols... This indicates the metapath. At any moment The hidden state of the gated loop unit. Obtain the rate of change of the business relationship indicator at the current moment, using symbols... This represents the vector of changes in each business relationship indicator at the current moment relative to the previous moment. Based on the above information, the historical evolution trend factor for each meta-path is calculated. The calculation formula is: ,in, Metapath At any moment Historical evolution trend factors. This represents a gated recurrent unit network that takes sequence data as input and outputs the predicted value at the current time step. Metapath At any moment The gated loop unit is in a hidden state. Metapath In the previous moment Dynamic weights. This represents the rate of change of business relationship indicators at the current moment. The gated recurrent unit network, through its internal update and reset gate structures, fuses and calculates the above three input information to output the meta-path. At any moment Historical evolution trend factors.
[0033] 4) Adjust the basic weighting factor Time-based dynamic adjustment factor and historical evolution trend factors The dynamic weight of each meta-path at the current time is calculated by weighting and combining the results according to the following formula. : ,in, Metapath At any moment Dynamic weights. This represents the sigmoid activation function, used to map the weighted summation result to a value between 0 and 1, serving as the dynamic weights of the metapath. This represents the weight coefficient of the basic weight factor. This represents the weighting coefficient of the time-dynamic adjustment factor. The weighting coefficients represent the factors that indicate the historical evolution trend. , and The sum of the three weighting coefficients equals 1. Metapath The basic weighting factor. Metapath At any moment The time-dynamic adjustment factor. Metapath At any moment The historical evolution trend factor is used. Through the above calculations, the dynamic weight of each meta-path at the current moment is obtained. Through the above formula, a computable functional mapping is established between the changes in dynamic business indicators and the calculation of meta-path weights, enabling the graph algorithm to perform in-depth mining based on changes in the credit business environment.
[0034] Among them, real-time business indicator values are quantitative data obtained from the credit business system at the current moment, including indicators of guarantee relationships, mortgage relationships, and employment relationships, which are used to reflect the latest status of various relationships in the current credit business environment.
[0035] The process of adjusting the aggregate weights of corresponding relation edges based on the dynamic weights of each meta-path includes: determining each relation edge corresponding to each meta-path based on the sequence of relation edge types contained in each meta-path; assigning the dynamic weight of the meta-path at the current moment to each relation edge corresponding to the meta-path as the updated aggregate weight of each relation edge; and when a relation edge corresponds to multiple meta-paths, using the weighted sum of the dynamic weights of the multiple meta-paths corresponding to the relation edge as the updated aggregate weight of the relation edge. The specific implementation process is as follows: 1) Each meta-path consists of alternating node types and relation edge types, with at least one relation edge type between the starting and ending node types. For each predefined meta-path, all relation edge types are extracted and arranged into a relation edge type sequence according to their order of appearance in the meta-path. Each existing relation edge in the heterogeneous credit business network is traversed to obtain its type. The type of an actual relation edge is compared one by one with each relation edge type in the relation edge type sequence of a meta-path. When the type of the actual relation edge matches any relation edge type in the meta-path, the actual relation edge is identified as the relation edge corresponding to that meta-path. This traversal and comparison process is performed on each meta-path to obtain the set of relation edges corresponding to each meta-path.
[0036] 2) For each meta-path, obtain its dynamic weight at the current moment. This dynamic weight is determined from the dynamic weights of each meta-path calculated in the previous step. Iterate through each relation edge corresponding to the meta-path and directly assign the current dynamic weight value to the current aggregate weight of these relation edges. After this assignment operation, each relation edge corresponding to each meta-path uses the dynamic weight of the meta-path at the current moment as its updated aggregate weight.
[0037] 3) Traverse each existing relation edge in the heterogeneous network of credit business and count the number of meta-paths corresponding to that relation edge. When a relation edge corresponds to only one meta-path, use the dynamically assigned weight as the updated aggregate weight of that relation edge. When a relation edge corresponds to multiple meta-paths, obtain the dynamic weights of all meta-paths corresponding to that relation edge at the current time, add these dynamic weights together to obtain a weighted sum, and use this weighted sum as the updated aggregate weight of that relation edge. After all relation edges have been updated, use the updated aggregate weights as the aggregate weights of the corresponding relation edges in the next iteration of the graph attention network algorithm, return to the step of executing the graph attention network algorithm on the heterogeneous network of credit business, and re-execute the attention aggregation process of the graph attention network algorithm according to the adjusted aggregate weights.
[0038] In this context, the relation edge corresponding to a meta-path refers to the actual relation edge in the heterogeneous network of credit business whose relation edge type matches any relation edge type in the relation edge type sequence of a certain meta-path. A meta-path can correspond to multiple actual relation edges, and an actual relation edge can also correspond to multiple meta-paths at the same time.
[0039] The process involves returning to step S3 and re-executing step S3 according to the adjusted aggregation weights until the preset conditions are met, at which point the final updated feature vector is output. The specific implementation process is as follows: 1) The updated aggregate weight of each relation edge is used as the aggregate weight of that relation edge in the next iteration of the graph attention network algorithm, and substituted into the differentiated aggregation formula in the graph attention network algorithm executed on the heterogeneous network of credit business. For each relation edge, its updated aggregate weight replaces the original aggregate weight and participates in the feature vector update calculation of the entity node connected to that relation edge. Specifically, for relation edges of the type of guarantee relation edge, its updated aggregate weight is used as an adjustment parameter for the guarantee strength coefficient and participates in the calculation of the updated feature vector of the target entity node in the guarantee aggregation formula; for relation edges of the type of mortgage relation edge, its updated aggregate weight is used as an adjustment parameter for the proportion of collateral value and the risk gap coefficient and participates in the calculation of the updated feature vector of the target entity node in the mortgage aggregation formula; for relation edges of the type of employment relation edge, its updated aggregate weight is used as an adjustment parameter for the influence weight and the risk influence weight and participates in the calculation of the temporary feature vector of the personnel node and the updated feature vector of the enterprise node in the first and second calculation steps of the employment aggregation formula, respectively. After completing a full attention aggregation process, the feature vector of each entity node in the heterogeneous network of credit business is updated.
[0040] 2) The preset condition is reaching a preset maximum number of iterations. Each complete attention aggregation process is counted as one iteration. After each iteration, the number of iterations executed is incremented by one and compared with the preset maximum number of iterations. If the number of iterations executed is less than the preset maximum number of iterations, the preset condition is not met, and the next attention aggregation process continues using the latest aggregation weight. If the number of iterations executed is equal to the preset maximum number of iterations, the preset condition is met, and iteration stops.
[0041] 3) After determining that the preset conditions are met and stopping the iteration, the feature vector obtained by each entity node in the heterogeneous network of credit business after the last iteration is used as the final updated feature vector output for subsequent steps to perform community partitioning.
[0042] S5. Perform community partitioning on the current credit network snapshot based on the final updated feature vector to obtain the current community set. When the entity nodes or relation edges in the heterogeneous credit business network change, determine the set of affected entity nodes based on the change content, and perform local update on the community set based on the set of affected entity nodes to obtain the updated community set and record the community change data. Specifically, community partitioning is performed on the current time-to-time credit network snapshot based on the final updated feature vector, resulting in the current time-to-time community set, which includes: Based on the final updated feature vector, the node features of each entity node in the credit network snapshot are determined. According to the connection relationships between entity nodes and the node features of each entity node, all entity nodes in the credit network snapshot are divided into multiple communities to obtain the community set at the current time. The specific implementation process is as follows: A credit time series model is constructed based on the time series data of credit business, and the heterogeneous network of credit business is defined as a time series set. , of which each Indicates time The status of the credit network Represents a set of entity nodes. Represents the set of relation edges. Indicates the time interval between entity nodes Relationship weight matrix, The attribute feature matrix representing the entity node.
[0043] Extract the feature vector corresponding to each entity node from the final updated feature vector output by the steps, and use it as the node feature of that entity node in the credit network snapshot at the current time.
[0044] Each entity node in the credit network snapshot is treated as an independent initial community. For each entity node, the modularity increment resulting from moving it from its current community to the communities of each of its neighboring entity nodes is calculated. The neighboring community with the largest modularity increment is selected; if the largest modularity increment is positive, the entity node is moved to that neighboring community; otherwise, the entity node remains in its original community. This process is repeated for all entity nodes until no further community moves occur. Based on this, each of the resulting communities is shrunk into a new node. The sum of edge weights between all entity nodes within a community is used as the weight of the new node's self-loop edges, and the sum of edge weights between communities is used as the weight of the edges between the new nodes, forming a new network. The community movement and shrinking process is repeated on the new network, iterating until the modularity of the entire network no longer increases. Finally, each entity node in the original credit network snapshot is assigned a unique community identifier, and entity nodes with the same community identifier belong to the same community.
[0045] All entity nodes with the same community identifier are grouped into the same community, and all communities are organized into a set as the community set at the current moment. Incremental community updates address the high-frequency and dynamic nature of credit business, including daily new loans, guarantees, repayments, defaults, changes in corporate relationships, and the addition or removal of collateral. This makes it impractical to re-perform a complete community partitioning after each change on a large dataset with a large number of nodes. Therefore, this method only updates the local regions affected by the changes, maintaining the consistency of the overall community structure. Considering the frequent updates to credit business data, re-performing community partitioning on the global network after each change would be computationally expensive and difficult to meet real-time requirements. Therefore, an incremental community discovery method is designed based on the aforementioned credit time-series model.
[0046] The credit network snapshot is a complete graph structure state in the heterogeneous network of credit business at the current moment, consisting of all entity nodes, all relation edges, node characteristics of each entity node, and aggregate weights of each relation edge. A community is a subset of entity nodes in the credit network snapshot, where the connection strength between entity nodes within a subset is greater than the connection strength between the subset and external entity nodes, and entity nodes within a community have similar node characteristics. The community set is the overall set of all communities obtained after dividing all entity nodes in the credit network snapshot. Each community in the set covers a portion of the entity nodes, and all communities collectively cover all entity nodes.
[0047] Specifically, when an entity node or relation edge in the heterogeneous network of credit business changes, the set of affected entity nodes is determined based on the content of the change, including: When a relationship edge changes in the heterogeneous network of the credit business, the changed relationship edge and the first and second entity nodes connected by the changed relationship edge are obtained. The first and second entity nodes are added to the initial set of affected entity nodes. For each entity node in the initial set of affected entity nodes, the difference in community core degree before and after the change of the relationship edge is calculated. The community core degree is used to represent the average connection strength between the entity node and other entity nodes in its community. When the difference in community core degree is greater than or equal to a preset perturbation threshold, all first-order neighbor entity nodes of the entity node are added to the initial set of affected entity nodes. The community core degree difference is calculated repeatedly, and first-order neighbor entity nodes that meet the condition are added to the initial set of affected entity nodes until the initial set of affected entity nodes no longer changes. The final initial set of affected entity nodes is then determined as the set of affected entity nodes. The specific implementation process is as follows: 1) Changes in relation edges in the heterogeneous network of credit business include adding a relation edge, deleting an existing relation edge, or modifying the attribute information of a relation edge. When a change is detected, the relation edge that has changed is recorded as the changed relation edge, the starting node of the changed relation edge is recorded as the first entity node, and the target node of the changed relation edge is recorded as the second entity node. An empty set of nodes is created as the initial set of affected entity nodes, and the first entity node and the second entity node are added to this set respectively.
[0048] 2) Community coreness represents the average connection strength between an entity node and other entity nodes within its community. For any entity node v in the initial set of affected entity nodes, before any change occurs in the relationship edge, the current community to which entity node v belongs is obtained. To calculate the community core degree before the change, sum the weights of all relationship edges between entity nodes and entity node v within the community, divide by the number of nodes in the community minus one. The community core degree before the change is calculated using the following formula: ,in, This indicates the community coreness of entity node v before the change. This represents the set of all entity nodes within the community that entity node v belonged to before the change. This represents the weight of the edge representing the relationship between entity node u and entity node v. This weight is consistent with the aggregate weight of the corresponding edge in the step of executing the graph attention network algorithm on the heterogeneous network of credit business. This represents the total number of entity nodes within the community. It also defines the community membership degree. , representing entity nodes At any moment For the community The degree of membership is calculated using the following formula: ,in, Represents entity nodes At any moment For the community membership degree Represents entity nodes At any moment The set of all entity nodes within the community. Represents entity nodes With entity nodes The weight of the edges relating to each other. Indicates time The set of all entity nodes in the heterogeneous network of the entire credit business.
[0049] After the relationship edges are changed, the topology of the heterogeneous network for credit business changes, requiring a recalculation of the community to which entity node v belongs and its community core degree. The same formula as before the change is used to calculate the revised community core degree, but the community set and edge weights are updated to reflect the latest state after the change. The difference in community coreness is obtained by subtracting the changed community coreness from the original value and taking the absolute value.
[0050] 3) Preset a perturbation threshold For each entity node in the initial set of entity nodes, its community core degree difference is compared with... Compare the values. When the difference is greater than or equal to... When the difference is less than 1, obtain all first-order neighbor entity nodes of the entity node in the heterogeneous network of the credit business after the change, and add all these first-order neighbor entity nodes to the initial set of affected entity nodes. At that time, no first-order neighbor entity node is added.
[0051] In each round of calculation, the above process is sequentially executed for all entity nodes in the currently initialized set of affected entity nodes, calculating the community core degree difference for each entity node, and adding first-order neighbor entity nodes that meet the conditions to the set. After one round of calculation is completed, it is determined whether there are any newly added entity nodes in the initialized set of affected entity nodes. If there are, the next round of calculation begins, and the community core degree difference calculation and neighbor addition operation are performed on the newly added entity nodes. If there are no new additions, that is, the initialized set of affected entity nodes no longer changes, the iteration stops.
[0052] When the iteration stops, the currently initialized set of affected entity nodes is output as the final set of affected entity nodes for subsequent local update steps.
[0053] In this context, a changed relationship edge refers to the specific relationship edge in the heterogeneous network of credit business where an addition, deletion, or attribute modification operation occurs. Changes to this relationship edge alter the network's topology or edge weights. The initial set of affected entity nodes is a dynamically changing set of nodes. Initially, it only includes the first and second entity nodes connected by the changed relationship edge. Subsequently, other entity nodes affected by the change are continuously added during the iteration process to track the propagation range of the change in the heterogeneous network of credit business. First-order neighbor entity nodes are other entity nodes directly connected to the current entity node through a relationship edge; that is, all entity nodes with a direct connection to the current entity node. The set of affected entity nodes is the final set of all entity nodes affected by the changed relationship edge, determined through iterative propagation. Entity nodes in this set will be the objects of subsequent local community updates.
[0054] Specifically, the community set is partially updated based on the set of affected entity nodes to obtain the updated community set, and the community change data is recorded, including: Within the subgraph covered by the set of affected entity nodes, calculate the local modularity change value of each entity node in the set when it moves from its current community to another community. Assign each entity node to the community with the largest local modularity change value to obtain the updated community set. During the local update process, record the changes in community member composition of the updated community set relative to the current community set as community change data. The specific implementation process is as follows: 1) Extract all entity nodes in the affected entity node set and all relational edges between these entity nodes to form a subgraph. This subgraph only contains entity nodes within the affected entity node set and relational edges whose ends are both within the affected entity node set.
[0055] 2) For any entity node in the set of entity nodes that are affected Get the community to which the entity node currently belongs, and denot it as the community. Retrieve other communities that the entity node might move to, denoted as communities. ,in, For the current moment, the community collection excluding Any community other than [the one mentioned above]. Calculate the entity node according to the following formula for local modularity change. From the community Move to community Local modularity change value: ,in, Represents entity nodes From the current community Move to Community The change in local modularity over time. Represents entity nodes With the community The sum of the weights of the relation edges between all entity nodes. Represents entity nodes With the community The sum of the weights of the relation edges between all other entity nodes. This represents the sum of the weights of all relation edges in the subgraph covered by the set of entity nodes. Represents entity nodes The weighted degree in the subgraph, i.e., the entity node. The sum of the weights of the relation edges between the node and all other entity nodes in the subgraph. Indicates community The sum of the weighted degrees of all entity nodes within the node. Indicates community The sum of the weighted degrees of all entity nodes within the boundary. Define the boundary strength. Used to quantify communities With the community The degree of coupling between them is calculated using the following formula: ,in, Indicates community With the community Boundary strength between Represents entity nodes With entity nodes The weights of the edges relating to the relationships are consistent with the aggregated weights of the corresponding edges in the graph attention network algorithm performed on the heterogeneous network of credit business. Indicates community The number of internal entity nodes, Indicates community The number of internal entity nodes, This indicates taking the smaller of the two community node counts. By influencing the set of entity nodes, the local modularity change value of each entity node is recalculated, and its community is reassigned, returning the overall community set for the next time step.
[0056] For each entity node Iterate through the community set at the current moment, excluding For all communities outside the local area, calculate the corresponding local modularity change value.
[0057] 3) For each entity node in the set of entity nodes that are affected The system compares all possible changes in local modularity and selects the community with the largest value as the target community. If the maximum value is positive, entity node i is moved from its current community to the target community; otherwise, entity node i remains in its current community. After all entity nodes in the affected entity node set have completed the above allocation operation, the updated community set is obtained.
[0058] 4) For each community in the updated community set, compare it with each community in the current community set, and record the following changes: newly added communities (communities that appear after the update but do not exist at the current time); disappeared communities (communities that exist at the current time but no longer exist after the update); and newly added and removed entity nodes in each community. Summarize this change information into community change data, which will be used in subsequent steps to generate multi-view dynamic visualization data. Based on the incrementally updated community sequence, the spatial trajectory of risky entity nodes propagating over time and the changes in community risk levels can be tracked. The formula for calculating the overall risk level of a single community is: ,in, Indicates community At any moment The overall risk level. Indicates community The number of internal entity nodes. Represents entity nodes At any moment The individual risk level is determined by the attribute characteristics of the entity node. This represents the cross-community risk attenuation coefficient, with a value ranging from 0.3 to 0.5. This indicates that they do not belong to the community. Other communities. Indicates time Community With the community The boundary strength between them is calculated using the boundary strength calculation formula mentioned above. Indicates community The number of internal entity nodes. Indicates community At any moment The overall risk level. Based on the risk evolution path tracking formula mentioned above, different types of risk propagation methods in the heterogeneous network of credit business can be viewed and early warnings can be issued.
[0059] The local modularity change value is a quantitative indicator that measures the impact of a single entity node moving from one community to another on the quality of community partitioning within the subgraph. A larger value indicates a greater improvement in community partitioning quality after the move, guiding the community reallocation decision for entity nodes. The subgraph is a graph structure composed of all entity nodes in the affected entity node set and all relational edges between these entity nodes. It is a local subset of the heterogeneous network of credit business, used to define the scope of local updates. The weighted degree is the sum of the weights of the relational edges between an entity node and all other entity nodes in the subgraph, reflecting the overall connectivity strength of that entity node in the subgraph. Community change data is a dataset recording the differences in community member composition between the updated community set and the current community set, including newly added communities, disappeared communities, and the addition and removal of entity nodes within each community. This data is used for subsequent visualization of the dynamic evolution of the community structure.
[0060] S6. Based on the final updated feature vector, the community change data, and the updated community set, generate multi-view dynamic visualization display data, including: 1) Determine the risk score and risk influence of each entity node based on the final updated feature vector; determine the risk level of each entity node based on its risk score; determine the display size of each entity node based on its risk influence; determine the display color of each entity node based on its risk level; determine the display thickness of each relation edge based on its initial transmission strength weight value; determine the display color of each relation edge based on its type; generate topology view data based on the display size and color of each entity node and the display thickness and color of each relation edge. The specific implementation process is as follows: ① The final updated feature vector of each entity node is subjected to a dot product operation through a preset linear projection vector to obtain a scalar value, which is used as the risk score of the entity node. The risk score ranges from 0 to 1. For each entity node, the sum of the aggregate weights of all connection edges of the node in the heterogeneous network of credit business is calculated to obtain the weighted degree value of the node, which is used as the risk influence of the entity node.
[0061] Figure 2 The document displays specific risk scores for some entity nodes, including different types such as borrower nodes, corporate nodes, guarantor nodes, and collateral nodes. Each entity node is labeled with its risk score; a higher risk score indicates a higher individual risk level for that entity node. Additionally, Figure 2It displays the relationship edges connecting entity nodes, along with their corresponding relationship types and weight values. Relationship types include equity relationship edges, guarantee relationship edges, employment relationship edges, and mortgage relationship edges. Each relationship edge is labeled with its weight, and a higher weight value indicates a stronger risk transmission intensity for that relationship edge.
[0062] ② Three risk level ranges are preset: the low-risk range is a risk score greater than or equal to 0 and less than 0.4, the medium-risk range is a risk score greater than or equal to 0.4 and less than 0.7, and the high-risk range is a risk score greater than or equal to 0.7 and less than or equal to 1. For each entity node, determine which range its risk score falls into, and assign the corresponding risk level to the entity node.
[0063] ③ Obtain the maximum and minimum risk influence values of all entity nodes in the heterogeneous network of credit business. For each entity node, map its risk influence value to a preset minimum and maximum display size using a linear mapping formula to obtain the display size value of that entity node. Nodes with greater risk influence have larger display sizes.
[0064] ④ Preset the correspondence between risk levels and display colors: low risk level corresponds to green, medium risk level corresponds to yellow, and high risk level corresponds to red. For each entity node, find the corresponding display color based on its risk level and use that color as the display color for that entity node.
[0065] ⑤ Obtain the maximum and minimum values of the initial transmission strength weights for all relation edges in the heterogeneous network of credit business. For each relation edge, map its initial transmission strength weight to a preset minimum and maximum display thickness using a linear mapping formula to obtain the display thickness value of that relation edge. Relation edges with larger initial transmission strength weights are displayed thicker.
[0066] ⑥ Preset the correspondence between relationship edge types and display colors: guarantee relationship edges correspond to red, mortgage relationship edges correspond to blue, and employment relationship edges correspond to green. For each relationship edge, find the corresponding display color based on its type and use that color as the display color for that relationship edge.
[0067] ⑦ Draw each entity node on a 2D plane with its display size as a circle and its display color as its fill color. Draw each relationship edge between the corresponding two entity nodes with its display thickness as the line segment width and its display color as the line segment color. Use a force-directed layout algorithm to calculate the 2D coordinates of all entity nodes, so that connected nodes are as close as possible and unconnected nodes are as far apart as possible. Encode all the above drawing information into topology view data for subsequent visualization rendering.
[0068] like Figure 3As shown, entity nodes are represented by circles. The size of a node maps to its risk influence; nodes with greater risk influence appear larger in the view. The color of a node maps to its risk level: red for high risk, yellow for medium risk, and green for low risk. Relationship edges are represented by line segments. The thickness of the line segment maps to the initial transmission strength weight value of the relationship edge; relationships with higher initial transmission strength weight values appear thicker. The color of the line segment maps to the type of relationship edge: red for guarantee relationships, blue for mortgage relationships, and green for employment relationships. Figure 3 This demonstrates the global topology of the heterogeneous network of credit business, where entity nodes are interconnected through different types of relational edges. The topology view data is a complete set of information including the coordinates, size, and color of all entity nodes on a two-dimensional plane, as well as the connection endpoints, thickness, and color of all relational edges. This data drives the visualization interface to draw the global topology of the heterogeneous network of credit business.
[0069] 2) Based on the risk score of each entity node and the connection relationships between entity nodes in the updated community set, calculate the risk heat value of each entity node, determine the risk heat display color of each entity node based on the risk heat value, and generate risk heat propagation view data based on the risk heat display color of each entity node, the connection relationships between entity nodes in the updated community set, and the propagation direction. The specific implementation process is as follows: ① For any entity node in the updated community set Obtain the risk score of the entity node itself, and record it as . ,in Represents the current time. Retrieves all entity nodes. The set of neighboring entity nodes directly connected by relation edges is denoted as For neighbor sets Each neighbor entity node in Get the risk heat value of the neighboring entity node at the previous time. Get the out-degree of the neighboring entity node. Out-degree refers to the number of relation edges emitted by the neighboring entity node in the heterogeneous network of credit business. (Obtain entity node) with neighboring entity nodes Risk semantic transmission weight between the edges of the relationship This weight is consistent with the aggregated weight of the corresponding relational edges in the graph attention network algorithm performed on the heterogeneous network of credit business. The entity node is calculated according to the following risk heat value calculation formula. Risk heat value at the current moment : ,in, Represents entity nodes At any moment The risk heat value. This represents the risk weight coefficient, which is used to control the contribution of the entity node's own risk score to the risk heat value. Represents entity nodes At any moment Its own risk score. This represents the risk weighting coefficient, used to control the contribution of risks transmitted from neighboring entity nodes to the risk heat value. and The sum of them equals 1. Represents entity nodes The set of all directly connected neighboring entity nodes. Represents neighboring entity nodes In the previous moment The risk heat value. Represents neighboring entity nodes The degree of departure. Represents entity nodes with neighboring entity nodes The risk semantic propagation weight of the relationship edges between nodes is calculated for each neighboring entity node. Divide by Multiply by The result is the sum of these results multiplied by In addition Multiply Obtain entity nodes Risk heat value at the current moment .
[0070] ② A preset mapping range between risk heat values and display colors is established: a risk heat value greater than or equal to 0 and less than 0.3 corresponds to blue; a risk heat value greater than or equal to 0.3 and less than 0.5 corresponds to green; a risk heat value greater than or equal to 0.5 and less than 0.7 corresponds to yellow; a risk heat value greater than or equal to 0.7 and less than 0.85 corresponds to orange; and a risk heat value greater than or equal to 0.85 and less than or equal to 1 corresponds to red. For each entity node, the corresponding display color is found based on its risk heat value, and this color is used as the risk heat display color for that entity node.
[0071] ③ Draw each entity node in the updated community set on a 2D plane using its risk heat display color as the fill color. For each relationship edge in the updated community set, determine the propagation direction based on the starting and target nodes of the relationship edge, and draw it as an arrow on the relationship edge, with the arrow pointing from the starting node to the target node. Use a force-directed layout algorithm to calculate the 2D coordinates of all entity nodes. Encode all the above drawing information into risk heat propagation view data for subsequent visualization rendering.
[0072] like Figure 4 As shown, the color fill of entity nodes represents their risk intensity value, with different colors corresponding to different risk intensity ranges. A risk intensity value greater than or equal to 0 and less than 0.3 corresponds to blue; greater than or equal to 0.3 and less than 0.5 corresponds to green; greater than or equal to 0.5 and less than 0.7 corresponds to yellow; greater than or equal to 0.7 and less than 0.85 corresponds to orange; and greater than or equal to 0.85 and less than or equal to 1 corresponds to red. Arrows on relationship edges indicate the direction of risk propagation, pointing from the risk source node to the target node affected by the risk, with the arrow direction consistent with the direction of the relationship edge. The thickness of the edge represents the magnitude of the risk semantic transmission weight; the greater the transmission weight, the thicker the edge appears in the view. Risk source nodes are highlighted with special markers in the graph. Figure 4 It visually demonstrates the heat distribution and transmission path of risk in the heterogeneous network of credit business.
[0073] The risk heat value is a comprehensive risk metric for each entity node at the current moment. It is calculated by weighting the node's own risk score and the risk transmitted from its neighboring nodes, ranging from 0 to 1, and reflects the node's risk intensity in the current network. The self-risk weight coefficient is a preset coefficient in the risk heat value calculation formula, used to control the weight proportion of the entity node's own risk score in the final risk heat value, and its value is between 0 and 1. The transmitted risk weight coefficient is also a preset coefficient in the risk heat value calculation formula, used to control the weight proportion of the risk transmitted from neighboring entity nodes in the final risk heat value, and its value is between 0 and 1, and its sum with the self-risk weight coefficient equals 1. The risk heat display color is the display color obtained by mapping the entity node's risk heat value to a preset color range, used to intuitively reflect the relative level of node risk heat through hue; a gradient from blue to red indicates a risk heat from low to high. The propagation direction is indicated by the arrows displayed on the risk heat map in the relationship edge diagram, pointing from the starting node of the relationship edge to the target node. This visually reflects the transmission path and direction of risk within the heterogeneous network of credit business. The risk heat map data is a complete set of information including the two-dimensional coordinates of all entity nodes in the updated community set, the risk heat display color, and the connection endpoints and propagation directions of all relationship edges. This information drives the visualization interface to draw the risk heat distribution and propagation path within the heterogeneous network of credit business.
[0074] 3) Calculate the community evolution rate based on community change data, calculate the network risk entropy based on the risk heat value of each entity node, calculate the risk proportion of key nodes based on the risk heat value of each entity node, and generate a time-series propagation trend view based on the changing trends of the community evolution rate, network risk entropy, and risk proportion of key nodes over time. The specific implementation process is as follows: ① Regarding the current moment Obtain the current community set from community change data. and the previous moment community collection Obtain the set of newly added communities at the current moment from the community change data. Newly added communities refer to those existing in In but not existing The community data. Extract the set of communities that have disappeared at the current moment from the community change data. Disappearing communities refer to communities that exist in the past. In but not existing The communities in the set. Calculate the number of elements in the union of the newly added community set and the disappeared community set, denoted as . Calculate the number of elements in the intersection of the current community set and the previous community set, denoted as . This value represents the number of communities shared between two time points. The community evolution rate at the current time is calculated using the following formula. : ,in, Indicates time Community evolution rate. Indicates time New community collections. Indicates time A collection of vanished communities. This refers to the community collection at the previous moment. Represents the community set at the current moment. When the denominator... When it equals 0, The value is assigned to 0.
[0075] ② Regarding the current moment Obtain the set of all entity nodes in the heterogeneous network of credit business, denoted as . Get each entity node. At any moment Risk heat value Calculate the sum of the risk heat values of all entity nodes, denoted as . For each entity node Calculate the proportion of the risk heat value of this node to the total. The calculation formula is: ,in, Represents entity nodes At any moment The proportion of the risk heat value to the total risk heat value of all nodes. Represents entity nodes At any moment The risk heat value. Indicates time The sum of the risk heat values of all entity nodes.
[0076] Based on each entity node Calculate the time according to the following network risk entropy formula. Network risk entropy : ,in, Indicates time Network risk entropy. Indicates time The set of all entity nodes in a heterogeneous network for credit operations. Represents entity nodes At any moment The proportion of risk heat value. Indicates base 2 The logarithm of . When When equal to 0, The value is 0.
[0077] ③ Regarding the current moment Obtain the risk heat value of all entity nodes in the heterogeneous network of credit business. Sort all entity nodes according to their risk intensity value from highest to lowest, and select the top-ranked nodes. A set of entity nodes ,in The value is a preset positive integer. Calculate. The sum of the risk heat values of all entity nodes in the data is denoted as: Calculation time The sum of the risk heat values of all entity nodes Calculate the time frame according to the following formula for the risk percentage of key nodes. Key node risk percentage : ,in, Indicates time The proportion of risk at key nodes. Indicates time Top Risk Popularity Ranking A collection of entity nodes. Represents entity nodes At any moment The risk heat value. Indicates time The set of all entity nodes in a heterogeneous network for credit operations.
[0078] ④ Collect multiple consecutive moments Community evolution rate Network risk entropy and the proportion of risks at key nodes The numerical sequence. For each time step, , , These are the values of three data sequences at corresponding times. With time as the horizontal axis and the three indicator values as the vertical axis, each indicator is plotted as a line graph changing over time, forming three line graphs. The horizontal axis time, vertical axis indicator value, and indicator name of each data point in the line graphs are encoded into time-series propagation trend view data for subsequent visualization rendering.
[0079] Figure 5Using time as the horizontal axis and the values of the three indicators as the vertical axis, three line graphs are plotted: community evolution rate, network risk entropy, and risk share of key nodes. The community evolution rate line reflects the drastic changes in community structure over time; an increase in the community evolution rate indicates drastic changes in community structure, while a decrease indicates a more stable network structure. The network risk entropy line reflects the degree of risk distribution in the heterogeneous network of credit business; an increase in network risk entropy indicates that risk has spread throughout the network, while a decrease indicates that risk is concentrated in a few entity nodes. The risk share of key nodes reflects the degree to which risk is concentrated in high-risk nodes; an increase in the risk share of key nodes indicates that risk is beginning to concentrate at the top, while a decrease indicates that risk is more evenly distributed. Figure 5 The trends of the three indicators mentioned above over time are shown, which can be used to analyze the propagation and evolution of risks in the time series dimension.
[0080] Among them, the community evolution rate is an indicator that measures the drastic change in community structure between two adjacent time points. It is equal to the union of newly added and disappeared communities at the current time point divided by the number of communities shared between the previous and next time points. A larger value indicates a more drastic change in community structure. Network risk entropy is an indicator that measures the degree of risk distribution dispersion in the heterogeneous network of the entire credit business. It is calculated based on the proportion of each entity node's risk intensity value to the total. A larger value indicates a more dispersed risk distribution across the entire network, while a smaller value indicates that the risk is concentrated in a few nodes. The risk proportion of key nodes is an indicator that measures whether the risk is concentrated in the top high-risk nodes. It is equal to the risk intensity value of the top... The sum of the node risk heat values is the proportion of the total node risk heat values. A larger value indicates that the risk is more concentrated on a few key nodes. New communities are those that appear in the community set at the current moment but did not appear in the community set at the previous moment, representing communities created by the growth or split of the community structure. Disappearing communities are those that existed in the community set at the previous moment but disappeared from the community set at the current moment, representing community removal caused by the shrinkage or merging of the community structure.
[0081] This method integrates graph algorithm mining with visualization, offering key advantages including: ① Process traceability: in addition to displaying the final mining results, it also records the intermediate states generated by each algorithm iteration; ② View collaboration: the same heterogeneous network data of credit business is presented in multiple view formats, such as topology view, risk heat propagation view, and time-series propagation trend view; ③ Incremental updates: when the relationship edges in the heterogeneous network of credit business change, the visualization interface only updates the affected parts rather than updating the entire dataset. The data sources relied upon for visualization mainly include three categories: the first is the final updated feature vector of entity nodes output by the graph attention network algorithm, used for coloring entity nodes by risk level and mapping the magnitude of risk influence; the second is community change data recorded during incremental community updates, used for animated demonstrations of community boundary changes; and the third is data on changes in the overall risk level of the community and data on the propagation path of individual risk changes, output according to the risk evolution path tracking formula.
[0082] In personal lending, taking a bank's personal loan risk control model as an example, the network is constructed using the aforementioned method for building a heterogeneous network for lending, specifically including the following steps: 1) Based on customer basic information data, credit data, and relationship data, borrower nodes, credit product nodes, guarantor nodes, collateral nodes, and enterprise nodes are extracted as entity nodes, and application relationship edges, holding relationship edges, guarantee relationship edges, contact relationship edges, colleague relationship edges, and mortgage relationship edges are constructed as relationship edges. Customer basic information data includes customer identifier, employer, residential address, and credit information, with the corresponding entity node being the borrower node. Credit data includes loan amount, loan term, and overdue status, with the corresponding entity nodes being the borrower node and credit product node, and the corresponding relationship edges being the application relationship edge and holding relationship edge. Relationship data includes guarantor relationships, contact person relationships, same-enterprise relationships, and collateral / pledged asset relationships, with the corresponding entity nodes being the borrower node, guarantor node, collateral node, and enterprise node, and the corresponding relationship edges being the guarantee relationship edge, contact relationship edge, colleague relationship edge, and mortgage relationship edge. Based on the above data, a heterogeneous network for credit business was constructed, comprising approximately 250,000 borrower nodes, approximately 40,000 guarantor nodes, 88,000 enterprise nodes, and 30,000 collateral nodes, totaling 420,000 entity nodes and 1.37 million relationship edges.
[0083] 2) Four meta-paths related to personal credit scenarios are predefined. The first meta-path sequence is: Borrower node, Guarantee relationship edge, Guarantor node, Guarantee relationship edge, Borrower node, with the business meaning of joint guarantor association, and an initial dynamic weight of 0.4. The second meta-path sequence is: Borrower node, Employment relationship edge or Colleague relationship edge, Company node, Employment relationship edge or Colleague relationship edge, Borrower node, with the business meaning of association within the same company, and an initial dynamic weight of 0.2. The third meta-path sequence is: Borrower node, Guarantee relationship edge, Guarantor node, Employment relationship edge, Company node, Employment relationship edge, Guarantor node, Guarantee relationship edge, Borrower node, with the business meaning of other borrowers within the guarantor's company, and an initial dynamic weight of 0.1. The fourth meta-path sequence is: Collateral node, Collateral relationship edge, Borrower node, Collateral relationship edge, Collateral node, with the business meaning of multiple collaterals for the same borrower, and an initial dynamic weight of 0.3. The dynamic weighting strategy is triggered as follows: if a company has three overdue customers, and all three are guaranteed by the same guarantee company, the company's associated overdue rate increases from 1.2% to 7%, triggering the dynamic weighting process. Following the aforementioned meta-path dynamic weight calculation method, the dynamic weights of each meta-path are recalculated based on real-time business indicators. Specifically, the dynamic weight of the first meta-path increases from 0.4 to 0.5, the dynamic weight of the second meta-path decreases from 0.2 to 0.1, and the dynamic weights of other meta-paths remain unchanged. The recalculated meta-path dynamic weights are used to adjust the aggregation weights of corresponding relational edges, and the attention aggregation process of the graph attention network algorithm is re-executed to update the feature vectors of borrower nodes, embedding risk information into the node representation.
[0084] 3) After adjusting the meta-path weights, community segmentation was performed on the heterogeneous network of the credit business. Using the aforementioned incremental community discovery method, three risk communities were identified. The first community contained 359 entity nodes, the core characteristic of which were all related clients of a certain guarantee company. Its community risk heat value was 0.82, belonging to the high-risk level. The second community contained 188 entity nodes, the core characteristic of which were all belonging to the same enterprise. Its community risk heat value was 0.42, belonging to the medium-risk level. The third community contained 72 entity nodes, with no obvious clustering characteristics. Its community risk heat value was 0.18, belonging to the low-risk level. The community risk heat value was calculated according to the aforementioned risk heat value formula. H v (t)The risk intensity values of entity nodes within each community were aggregated, specifically calculated as the arithmetic mean of the risk intensity values of all entity nodes within the community. Combined with dynamic visualization analysis, the first and second communities identified in the incremental community discovery were given focused monitoring. Abnormal entity nodes within these risky communities were tracked, and key influencing nodes were identified, thus shifting the focus from risk community risk control to risk node risk control. Ultimately, the bank's delinquency rate for personal loans decreased from 1.2% to 0.87%.
[0085] This invention constructs a heterogeneous network for credit business, assigning differentiated risk transmission semantic labels to different types of entity nodes and relationship edges. This solves the problem that traditional isomorphic graphs cannot distinguish between different business semantics such as guarantee transmission and mortgage transmission, enabling graph attention network algorithms to accurately express the substantial differences in risk propagation mechanisms among various relationships. Building on this, the invention introduces a dynamic weighting strategy for meta-paths, utilizing a weighted combination of three dimensions: basic weight factors, time-dynamic adjustment factors, and historical evolution trend factors. This allows the weights of each meta-path to dynamically change with real-time business indicators, enabling the graph mining model to adapt to changes in risk-sensitive characteristics under different credit scenarios. Addressing the temporal dynamic characteristics of credit data, this invention designs an incremental community detection method. This method determines the set of influencing entity nodes through a perturbation propagation mechanism, performing community updates only on local subgraphs. This ensures computational efficiency while retaining the ability to dynamically track risk evolution paths. Furthermore, this invention organically integrates the graph mining process with a multi-view dynamic visualization method. It displays the global relational structure through a topological view, shows the risk heat propagation view as the distribution and direction of risk intensity, and displays the changing trends of community evolution rate, network risk entropy, and the risk proportion of key nodes over time through a time-series propagation trend view, thus achieving a visualized presentation of the mining process. In a practical application of this invention in a bank's personal loan business, the delinquency rate decreased from 1.2% to 0.87% after adopting the method of this invention, verifying the practical effect of this invention in improving the accuracy of credit risk identification and the interpretability of risk control decisions.
[0086] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.
[0087] like Figure 6 As shown in the figure, an improved graph algorithm knowledge mining and dynamic visualization analysis system for credit in an embodiment of the present invention includes a construction definition module, a determination module, an execution aggregation module, an update module, a partitioning update module, and a generation module; The definition module is used to: construct a heterogeneous network for credit business, which contains multiple entity nodes and multiple relation edges, and define risk transmission semantic labels for each type of relation edge; The determination module is used to: determine the initial transmission strength weight value of each relation edge based on the type of each relation edge and the risk transmission semantic label; The aggregation module is used to: execute the graph attention network algorithm on the heterogeneous network of credit business. In the attention aggregation process of the graph attention network algorithm, according to the type of each relation edge, the initial transmission strength weight value of each relation edge, and the current aggregation weight of each relation edge, the feature vector of each entity node is updated by the corresponding differentiated aggregation formula for each type of relation edge, so as to obtain the updated feature vector of each entity node. The update module is used to: preset multiple meta-paths in the heterogeneous network of credit business, calculate the dynamic weight of each meta-path at the current time, adjust the aggregation weight of the corresponding relation edges according to the dynamic weight of each meta-path, call and execute the aggregation module again and process according to the adjusted aggregation weight, until the preset conditions are met and the final updated feature vector is output. The partitioning and updating module is used to: perform community partitioning on the current credit network snapshot based on the final updated feature vector to obtain the current community set; when the entity nodes or relation edges in the heterogeneous credit business network change, determine the set of affected entity nodes based on the change content, perform local updates on the community set based on the set of affected entity nodes, obtain the updated community set, and record the community change data. The generation module is used to generate multi-view dynamic visualization data based on the final updated feature vector, community change data, and updated community set.
[0088] Optionally, in the above technical solution, the types of entity nodes include borrower nodes, enterprise nodes, collateral nodes, personnel nodes, and credit product nodes; the types of relationship edges include mortgage relationship edges, guarantee relationship edges, and employment relationship edges; and a risk transmission semantic label is defined for each type of relationship edge, including: joint liability transmission for guarantee relationship edges, which means that risk is transmitted from the guarantor to the borrower along the direction of the guarantee relationship; asset coverage transmission for mortgage relationship edges, which means that changes in the value of collateral affect the borrower's credit limit, and the direction of risk transmission is negatively correlated with the loan-to-value ratio; and key personnel transmission for employment relationship edges, which means that risk is transmitted bidirectionally between personnel nodes and enterprise nodes.
[0089] Optionally, in the above technical solution, the determining module is specifically used to: set the initial transmission strength weight function corresponding to each type of relation edge, take the attribute information of each relation edge as the input parameter of the initial transmission strength weight function, and calculate the initial transmission strength weight value of each relation edge.
[0090] Optionally, in the above technical solution, the aggregation module is specifically used for: For relationship edges of type "guarantee relationship edge", the feature vector of the entity node connected to the guarantee relationship edge is updated using the guarantee aggregation formula. The guarantee aggregation formula includes a guarantee strength coefficient, which is determined by the initial transmission strength weight value of the guarantee relationship edge. The guarantee strength coefficient is used to control the degree of influence of the guarantee relationship edge in the aggregation process. For a relationship edge of type mortgage relationship edge, the feature vector of the entity node connected to the mortgage relationship edge is updated using the mortgage aggregation formula. The mortgage aggregation formula includes the proportion of collateral value and the risk gap coefficient. The proportion of collateral value is determined by the proportion of the appraised value of the collateral to the total value of all collateral of the corresponding borrower. The risk gap coefficient is determined by the proportion of the appraised value of the collateral to the loan amount. For relationship edges of type "employment relationship edge", the feature vectors of personnel nodes and enterprise nodes connected to the employment relationship edge are updated using the employment aggregation formula. The employment aggregation formula includes a first calculation step and a second calculation step. In the first calculation step, the feature vectors of each enterprise node to which the personnel node is employed are merged into the feature vector of the personnel node to obtain a temporary feature vector of the personnel node. In the second calculation step, the risk information in the temporary feature vector of the personnel node is fed back to the feature vectors of each enterprise node.
[0091] Optionally, in the above technical solution, the update module is specifically used to: obtain the risk semantic score corresponding to each meta-path, and calculate the basic weight factor of each meta-path based on the risk semantic score; obtain the real-time business indicator value at the current moment, and calculate the time dynamic adjustment factor of each meta-path based on the real-time business indicator value; obtain the dynamic weight of each meta-path at historical moments, and calculate the historical evolution trend factor of each meta-path based on the dynamic weight at historical moments; and determine the dynamic weight of each meta-path at the current moment based on the basic weight factor, the time dynamic adjustment factor, and the historical evolution trend factor.
[0092] Optionally, in the above technical solution, the updating module is further specifically used to: determine each relation edge corresponding to each meta-path based on the relation edge type sequence contained in each meta-path; for each meta-path, assign the dynamic weight of the meta-path at the current moment to each relation edge corresponding to the meta-path as the updated aggregate weight of each relation edge; when a relation edge corresponds to multiple meta-paths at the same time, use the weighted sum of the dynamic weights of the multiple meta-paths corresponding to the relation edge as the updated aggregate weight of the relation edge.
[0093] Optionally, in the above technical solution, the partitioning update module is specifically used to: determine the node features of each entity node in the credit network snapshot based on the final updated feature vector, and divide all entity nodes in the credit network snapshot into multiple communities based on the connection relationship between each entity node and the node features of each entity node, thereby obtaining the community set at the current moment. When a relationship edge changes in the heterogeneous network of the credit business, the changed relationship edge and the first and second entity nodes connected by the changed relationship edge are obtained. The first and second entity nodes are added to the initial set of affected entity nodes. For each entity node in the initial set of affected entity nodes, the difference in community core degree before and after the change of the relationship edge is calculated. The community core degree is used to represent the average connection strength between the entity node and other entity nodes in its community. When the difference in community core degree is greater than or equal to a preset perturbation threshold, all first-order neighbor entity nodes of the entity node are added to the initial set of affected entity nodes. The difference in community core degree is calculated repeatedly and first-order neighbor entity nodes that meet the condition are added to the initial set of affected entity nodes until the initial set of affected entity nodes no longer changes. The final initial set of affected entity nodes is then determined as the set of affected entity nodes. Within the subgraph covered by the set of affected entity nodes, calculate the local modularity change value of each entity node in the set when it moves from its current community to another community. Assign each entity node to the community with the largest local modularity change value to obtain the updated community set. During the local update process, record the changes in community member composition of the updated community set relative to the current community set as community change data.
[0094] Optionally, in the above technical solution, the generation module is specifically used for: The risk score and risk influence of each entity node are determined based on the final updated feature vector. The risk level of each entity node is determined based on its risk score. The display size of each entity node is determined based on its risk influence. The display color of each entity node is determined based on its risk level. The display thickness of each relation edge is determined based on its initial transmission strength weight value. The display color of each relation edge is determined based on its type. Topology view data is generated based on the display size and color of each entity node and the display thickness and color of each relation edge. Based on the risk score of each entity node and the connection relationship between each entity node in the updated community set, calculate the risk heat value of each entity node, determine the risk heat display color of each entity node based on the risk heat value, and generate risk heat propagation view data based on the risk heat display color of each entity node, the connection relationship between each entity node in the updated community set, and the propagation direction. The community evolution rate is calculated based on community change data, the network risk entropy is calculated based on the risk heat value of each entity node, the risk proportion of key nodes is calculated based on the risk heat value of each entity node, and a time-series propagation trend view is generated based on the changing trends of community evolution rate, network risk entropy, and risk proportion of key nodes over time.
[0095] It should be noted that the beneficial effects of the improved graph algorithm knowledge mining and dynamic visualization analysis system for credit-oriented systems provided in the above embodiments are the same as those of the improved graph algorithm knowledge mining and dynamic visualization analysis method for credit-oriented systems, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.
[0096] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned improved graph algorithm knowledge mining and dynamic visualization analysis methods for credit.
[0097] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned improved graph algorithm knowledge mining and dynamic visualization analysis methods for credit.
[0098] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.
[0099] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. An improved graph algorithm-based knowledge mining and dynamic visualization analysis method for credit, characterized in that, include: S1. Construct a heterogeneous network for credit business, wherein the heterogeneous network for credit business includes multiple entity nodes and multiple relation edges, and define a risk transmission semantic label for each type of relation edge; S2. Determine the initial transmission strength weight value for each of the relation edges based on the type of each relation edge and the risk transmission semantic label; S3. Execute the graph attention network algorithm on the heterogeneous network of the credit business. In the attention aggregation process of the graph attention network algorithm, according to the type of each relation edge, the initial transmission strength weight value of each relation edge and the current aggregation weight of each relation edge, the feature vector of each entity node is updated by the corresponding differentiated aggregation formula for each type of relation edge, so as to obtain the updated feature vector of each entity node. S4. In the heterogeneous network of credit business, multiple meta-paths are preset, the dynamic weight of each meta-path at the current time is calculated, the aggregation weight of the corresponding relation edge is adjusted according to the dynamic weight of each meta-path, the process returns to step S3 and re-executes step S3 according to the adjusted aggregation weight until the preset conditions are met, and then the final updated feature vector is output. S5. Perform community partitioning on the current credit network snapshot based on the final updated feature vector to obtain the current community set. When the entity nodes or relation edges in the heterogeneous credit business network change, determine the set of affected entity nodes based on the change content, and perform local update on the community set based on the set of affected entity nodes to obtain the updated community set and record the community change data. S6. Generate multi-view dynamic visualization display data based on the final updated feature vector, the community change data, and the updated community set.
2. The improved graph algorithm knowledge mining and dynamic visualization analysis method for credit-oriented applications according to claim 1, characterized in that, In step S1, the types of entity nodes include borrower nodes, enterprise nodes, collateral nodes, personnel nodes, and credit product nodes; the types of relationship edges include mortgage relationship edges, guarantee relationship edges, and employment relationship edges. Define risk transmission semantic labels for each type of relationship edge, including: defining joint liability transmission for the guarantee relationship edge, where joint liability transmission indicates that risk is transmitted from the guarantor to the borrower along the direction of the guarantee relationship; defining asset coverage transmission for the mortgage relationship edge, where asset coverage transmission indicates that changes in the value of the collateral affect the borrower's credit limit, and the direction of risk transmission is negatively correlated with the loan-to-value ratio; and defining key personnel transmission for the employment relationship edge, where key personnel transmission indicates that risk is transmitted bidirectionally between personnel nodes and enterprise nodes.
3. The improved graph algorithm knowledge mining and dynamic visualization analysis method for credit-oriented systems according to claim 2, characterized in that, Based on the type of each relation edge and the risk transmission semantic label, determine the initial transmission strength weight value of each relation edge, including: setting an initial transmission strength weight function corresponding to each type of relation edge, using the attribute information of each relation edge as the input parameter of the initial transmission strength weight function, and calculating the initial transmission strength weight value of each relation edge.
4. The improved graph algorithm knowledge mining and dynamic visualization analysis method for credit-oriented applications according to claim 2, characterized in that, In step S3, the feature vector of each entity node is updated using the corresponding differentiated aggregation formula for different types of relation edges, including: For a relationship edge of type "guarantee relationship edge", the feature vector of the entity node connected to the guarantee relationship edge is updated using a guarantee aggregation formula. The guarantee aggregation formula includes a guarantee strength coefficient, which is determined by the initial transmission strength weight value of the guarantee relationship edge. The guarantee strength coefficient is used to control the degree of influence of the guarantee relationship edge in the aggregation process. For a relationship edge of type mortgage relationship edge, the feature vector of the entity node connected to the mortgage relationship edge is updated using a mortgage aggregation formula. The mortgage aggregation formula includes the proportion of collateral value and the risk gap coefficient. The proportion of collateral value is determined by the proportion of the appraised value of the collateral to the total value of all collateral of the corresponding borrower. The risk gap coefficient is determined by the proportion of the appraised value of the collateral to the loan amount. For relationship edges of type "employment relationship edge", the feature vectors of personnel nodes and enterprise nodes connected to the employment relationship edge are updated using an employment aggregation formula. The employment aggregation formula includes a first calculation step and a second calculation step. In the first calculation step, the feature vectors of each enterprise node to which the personnel node is employed are merged into the feature vector of the personnel node to obtain a temporary feature vector of the personnel node. In the second calculation step, the risk information in the temporary feature vector of the personnel node is fed back to the feature vectors of each enterprise node.
5. An improved graph algorithm-based knowledge mining and dynamic visualization analysis system for credit, characterized in that, This includes building a definition module, determining a module, executing an aggregation module, updating a module, dividing an update module, and generating a module; The construction definition module is used to: construct a heterogeneous network for credit business, the heterogeneous network for credit business includes multiple entity nodes and multiple relation edges, and define a risk transmission semantic label for each type of relation edge; The determining module is used to: determine the initial transmission strength weight value of each relation edge according to the type of each relation edge and the risk transmission semantic label; The execution aggregation module is used to: execute a graph attention network algorithm on the heterogeneous network of credit business; during the attention aggregation process of the graph attention network algorithm, according to the type of each relation edge, the initial transmission strength weight value of each relation edge and the current aggregation weight of each relation edge, update the feature vector of each entity node by using the corresponding differentiated aggregation formula for each type of relation edge, so as to obtain the updated feature vector of each entity node. The update module is used to: preset multiple meta-paths in the heterogeneous network of the credit business, calculate the dynamic weight of each meta-path at the current time, adjust the aggregation weight of the corresponding relation edge according to the dynamic weight of each meta-path, call the execution aggregation module again and process it according to the adjusted aggregation weight, until the preset conditions are met and the final updated feature vector is output. The partitioning and updating module is used to: perform community partitioning on the current credit network snapshot based on the final updated feature vector to obtain the current community set; when the entity nodes or relation edges in the heterogeneous credit business network change, determine the set of affected entity nodes based on the change content, perform local updates on the community set based on the set of affected entity nodes, obtain the updated community set, and record the community change data. The generation module is used to generate multi-view dynamic visualization display data based on the final updated feature vector, the community change data, and the updated community set.
6. The improved graph algorithm knowledge mining and dynamic visualization analysis system for credit as described in claim 5, characterized in that, The types of entity nodes include borrower nodes, enterprise nodes, collateral nodes, personnel nodes, and credit product nodes; the types of relationship edges include mortgage relationship edges, guarantee relationship edges, and employment relationship edges. Define risk transmission semantic labels for each type of relationship edge, including: defining joint liability transmission for the guarantee relationship edge, where joint liability transmission indicates that risk is transmitted from the guarantor to the borrower along the direction of the guarantee relationship; defining asset coverage transmission for the mortgage relationship edge, where asset coverage transmission indicates that changes in the value of the collateral affect the borrower's credit limit, and the direction of risk transmission is negatively correlated with the loan-to-value ratio; and defining key personnel transmission for the employment relationship edge, where key personnel transmission indicates that risk is transmitted bidirectionally between personnel nodes and enterprise nodes.
7. The improved graph algorithm knowledge mining and dynamic visualization analysis system for credit as described in claim 6, characterized in that, The determining module is specifically used to: set an initial conduction strength weight function corresponding to each type of relation edge, use the attribute information of each relation edge as the input parameter of the initial conduction strength weight function, and calculate the initial conduction strength weight value of each relation edge.
8. The improved graph algorithm knowledge mining and dynamic visualization analysis system for credit as described in claim 6, characterized in that... The execution aggregation module is specifically used for: For a relationship edge of type "guarantee relationship edge", the feature vector of the entity node connected to the guarantee relationship edge is updated using a guarantee aggregation formula. The guarantee aggregation formula includes a guarantee strength coefficient, which is determined by the initial transmission strength weight value of the guarantee relationship edge. The guarantee strength coefficient is used to control the degree of influence of the guarantee relationship edge in the aggregation process. For a relationship edge of type mortgage relationship edge, the feature vector of the entity node connected to the mortgage relationship edge is updated using a mortgage aggregation formula. The mortgage aggregation formula includes the proportion of collateral value and the risk gap coefficient. The proportion of collateral value is determined by the proportion of the appraised value of the collateral to the total value of all collateral of the corresponding borrower. The risk gap coefficient is determined by the proportion of the appraised value of the collateral to the loan amount. For relationship edges of type "employment relationship edge", the feature vectors of personnel nodes and enterprise nodes connected to the employment relationship edge are updated using an employment aggregation formula. The employment aggregation formula includes a first calculation step and a second calculation step. In the first calculation step, the feature vectors of each enterprise node to which the personnel node is employed are merged into the feature vector of the personnel node to obtain a temporary feature vector of the personnel node. In the second calculation step, the risk information in the temporary feature vector of the personnel node is fed back to the feature vectors of each enterprise node.
9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the improved graph algorithm knowledge mining and dynamic visualization analysis method for credit as described in any one of claims 1 to 4.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the improved graph algorithm knowledge mining and dynamic visualization analysis method for credit as described in any one of claims 1 to 4.
