A business feature multi-dimensional graph correlation method for business risk identification

Abstract

The application relates to the technical field of data processing, and discloses a business feature multi-dimensional graph correlation method for business risk identification, which comprises the following steps: acquiring feature node business data in the dimensions of funds, logistics and control right, and constructing an initial heterogeneous graph containing initial correlation strength; calculating the actual variation time difference between adjacent nodes in the cross-dimension, combining a predetermined business lag coefficient to generate a dynamic correlation edge weight with inverse proportion attenuation; updating the initial heterogeneous graph according to the dynamic correlation edge weight to generate a dynamic evolution graph; extracting the topological feature variation amount of a target node and comparing the reference risk trigger threshold value to output a business risk identification result. According to the application, the dimensional transmission delay is converted into a topological edge parameter constraint, false correlation is eliminated, a dynamic evolution graph with time sequence causality is reconstructed, cross-dimension risks are accurately identified, and the graph topological evolution stability under the condition of data sparseness is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method for associating multidimensional maps of business characteristics for business risk identification. Background Technology

[0002] Currently, when dealing with multidimensional business relationships, heterogeneous knowledge graphs are typically used to integrate heterogeneous data from multiple sources, such as business registration, equity control, transaction settlement, and warehousing and logistics. Graph topology connectivity retrieval algorithms are then used to identify the association paths between target feature nodes. In the heterogeneous graph structure, entities of different dimensions, such as capital, logistics, and control, can be set as feature nodes, and association edges and initial association strength parameters can be established between feature nodes, thereby providing a structured representation of the business relationship network between different enterprise entities.

[0003] However, existing knowledge graph association technologies typically treat the relationships between different business dimensions as synchronous changes at the same time point, failing to fully consider the objectively existing response lags and time differences between different business dimensions. In actual supply chain operations, the transmission of warehousing or shipping disruptions in the logistics dimension to the freezing of accounts receivable in the financial dimension usually takes a certain amount of time; changes in equity ownership in the control dimension also have a corresponding delay in impacting corporate business decisions. Using static weights for graph topology traversal makes it difficult to reflect the asynchronous changes of different business dimensions over time, easily causing heterogeneous data to lose its original temporal causal relationships during aggregation calculations. This leads to early, weak risk warning signals being interfered with by false connectivity relationships. For example, the public number CN Chinese invention patent application 120298117A discloses a method and system for identifying transaction risks. It receives multi-source event streams through a message queue and updates the dynamic transaction graph in real time. It also extracts topological risk features and outputs a risk score by combining graph neural networks. However, when processing cross-dimensional business data, this prior art often treats the state changes of heterogeneous business dimensions as occurring synchronously, without fully considering the objective time lag effect between funds, logistics and control. It lacks a dynamic attenuation factor adjustment mechanism based on the actual time difference of changes. Under the condition of sparse multi-source data or local sampling interruption, it is easy to introduce false associations due to the lack of temporal causal constraints, and cannot effectively guarantee the stability of graph topology evolution and the accuracy of temporal causal chains.

[0004] Therefore, the technical problem to be solved by this invention is how to reconstruct the dynamic associated edge weights based on the transmission lag between different business dimensions while maintaining low computational complexity, and to maintain the temporal causal chain in the cross-dimensional risk transmission process. Summary of the Invention

[0005] This invention provides a method for associating multidimensional maps of business characteristics for business risk identification, comprising the following steps: Step S1: Obtain the underlying business data of the target related enterprise. The underlying business data includes enterprise resource planning system data, supply chain management system data, and business and legal data. Extract the first feature node in the funding dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension. Step S2: Based on the objective transaction records and entity ownership relationships in the underlying business data, construct the fund association edges, logistics association edges, and control rights association edges connecting each feature node to form an initial heterogeneous graph. The initial association edges include the initial association strength parameter. Step S3: Extract the state change timestamps of each feature node in the initial heterogeneous graph. For cross-dimensional adjacent feature nodes that are directly connected by related edges and belong to different dimensions such as the capital dimension, logistics dimension and control dimension, calculate their actual change time difference. Step S4: Calculate the dynamic attenuation factor based on the actual time difference of change and the predetermined business hysteresis coefficient. The dynamic attenuation factor decreases inversely as the product of the actual time difference of change and the business hysteresis coefficient increases. Multiply the initial association strength parameter with the dynamic attenuation factor to generate the dynamic association edge weight of the corresponding association edge. Step S5: Update the edge weights of each associated edge in the initial heterogeneous graph according to the dynamic associated edge weights to generate a dynamic evolution graph. Step S6: Extract the topological feature changes of the target feature nodes in the dynamic evolution graph that serve as risk detection targets, compare the topological feature changes with the benchmark risk trigger threshold, and output the business risk identification results.

[0006] Preferably, step S5 includes the following node compensation steps: Step S51, when the data sampling interruption time of the second feature node of the target associated enterprise exceeds the preset period, determine the control radiation range of the third feature node corresponding to the target associated enterprise; Step S52, extract the state change timestamps of the second feature nodes of other enterprises within the control radiation range that have the same third feature node association with the target associated enterprise, and calculate the weighted average timestamp based on the initial association strength parameter; Step S53, use the weighted average timestamp as the compensation timestamp of the second feature node with sampling interruption, so as to provide the time boundary of the graph evolution and maintain the topological connectivity of the second feature node in the dynamic evolution graph; wherein, the topological compensation timestamp does not participate in the determination of the cross-dimensional actual change time difference and dynamic decay factor.

[0007] Preferably, step S4 includes the following edge truncation steps: Step S41, when the actual change time difference exceeds the set critical hysteresis window, adjust the weight of the corresponding dynamic associated edge to 0; Step S42, stop the weight transfer of the associated edge with a dynamic associated edge weight of 0 in the initial heterogeneous graph, and update the topological convergence boundary.

[0008] Preferably, in step S2, the initial correlation strength parameter is determined in the following way: step S21, extract the basic transaction frequency data in the funding dimension and the control ratio data in the control dimension; step S22, normalize the basic transaction frequency and control ratio data to determine the initial correlation strength parameter.

[0009] Preferably, in step S6, the baseline risk trigger threshold is set to an initial scalar base on the same numerical scale as the topological feature change, and is determined according to the following adaptive adjustment steps: Step S61, obtain the macroeconomic prosperity index of the industry in which the target related enterprise is located, and calculate the gradient rate of change of the macroeconomic prosperity index within the continuous sampling period; Step S62, when the gradient rate of change is negative and the absolute value exceeds the preset steady-state fluctuation boundary, calculate the bias compensation factor on the same numerical scale based on the product of the gradient rate of change and the initial scalar base; Step S63, add the baseline risk trigger threshold and the bias compensation factor to update the baseline risk trigger threshold.

[0010] Preferably, in step S6, the topological feature change is the change in the feature vector centrality of the target feature node in the dynamic evolution graph; the output business risk identification result includes the following judgment steps: step S64, when the change in feature vector centrality is lower than the baseline risk trigger threshold, it is determined that there is risk transmission; step S65, generate and output a risk warning signal containing the target feature node identifier and the risk transmission path.

[0011] Preferably, in step S1, the first feature node is a funds account node, the second feature node is a warehousing and logistics node, and the third feature node is a legal representative and equity control node. The process of obtaining underlying business data includes: step S11, reading original business messages from the enterprise resource planning system, supply chain management system, and business and legal database respectively through a combination of streaming parsing and timed batch processing via a preset data acquisition interface and network transmission protocol; step S12, performing structured parsing on the original business messages, extracting target field data including transaction amount, warehousing and logistics records, equity ratio, and legal representative, and performing entity attribution alignment on the multi-source heterogeneous target field data using the unified social credit code of the target associated enterprise as the unique entity identifier; step S13, constructing the first feature node in the fund dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension based on the target field data after entity alignment and the preset graph attribute mapping relationship.

[0012] Preferably, step S65 includes the following tiered response steps: Step S651, setting the first risk level threshold to be the same as the benchmark risk trigger threshold, and setting a second risk level threshold with a value lower than the first risk level threshold, forming a progressive risk judgment interval; Step S652, when the change in feature vector centrality is between the first risk level threshold and the second risk level threshold, outputting a warning data packet containing the target feature node identifier, risk transmission path, and medium risk identifier to the business risk control terminal, triggering the business risk control terminal to generate a manual review task; Step S653, when the change in feature vector centrality is lower than the second risk level threshold, outputting an interception instruction data packet containing the target feature node identifier and high-risk interception identifier to the transaction settlement system, triggering the transaction settlement system to execute the transaction settlement blocking and fund account freezing actions of the corresponding target related enterprise.

[0013] Preferably, in step S3, the process of extracting the state change timestamp and calculating the actual change time difference is based on the set sliding time window period update, and includes the following steps: Step S31, when new business data input is detected, update the node state change timestamp within the set sliding time window; Step S32, recalculate the actual change time difference between cross-dimensional adjacent feature nodes within the set sliding time window, and trigger dynamic association edge weight update.

[0014] The present invention has at least the following beneficial effects: 1. In the multi-dimensional graph association of business characteristics for business risk identification, the weights of dynamic association edges are adjusted by the actual change time difference and business lag coefficient. This allows the association strength between different business dimensions such as capital, logistics, and control to change with the time difference, rather than simply treating business changes in each dimension as occurring simultaneously. As a result, false connectivity caused by static associations can be reduced, and the dynamic evolution graph can more accurately reflect the sequential transmission relationship of risks between different business dimensions. This avoids early and weak risk signals being masked by irrelevant associations, and improves the accuracy and timeliness of cross-dimensional risk identification.

[0015] 2. When the data sampling of the second feature node is interrupted, a compensatory timestamp is generated by using the timestamps of the node status changes of other related enterprises within the control range. This can supplement the time sequence information of the missing node and prevent the data interruption of a single node from causing the entire map update to stagnate. This method can maintain the continuous operation of the dynamic evolution map when the data is incomplete or relatively sparse, reduce the impact of local data loss on risk analysis results, and improve the stability of the map structure and risk transmission path.

[0016] 3. When the actual time difference of change exceeds the critical hysteresis window, the weight of the corresponding dynamic associated edge is adjusted to zero. This can promptly eliminate associated edges with excessively long time intervals and weak temporal correlation, preventing low-correlation noise from participating in subsequent graph calculations. As a result, the topological feature changes of the target feature node can better reflect the risk transmission paths with a high degree of correlation, while reducing redundant calculations caused by invalid associated edges, reducing the computational overhead of graph iteration updates, and improving the convergence efficiency of the dynamic evolution graph and the stability of the risk judgment boundary. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the invention are illustrated by way of example and not limitation, wherein: Figure 1 This is a flowchart illustrating the dynamic association and risk identification process of the multidimensional map of business characteristics in this invention. Figure 2 This is a logical structure diagram of the multidimensional graph of the operating characteristics of this invention. Detailed Implementation

[0018] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments in conjunction with the accompanying drawings. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0019] A method for associating operational characteristics in a multidimensional graph for business risk identification includes the following steps: Step S1: Obtain the underlying business data of the target related enterprise. The underlying business data includes enterprise resource planning system data, supply chain management system data, and business and legal data. Extract the first feature node in the funding dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension. Step S2: Based on the objective transaction records and entity ownership relationships in the underlying business data, construct the fund association edges, logistics association edges, and control rights association edges connecting each feature node to form an initial heterogeneous graph. The initial association edges include the initial association strength parameter. Step S3: Extract the state change timestamps of each feature node in the initial heterogeneous graph. For cross-dimensional adjacent feature nodes that are directly connected by related edges and belong to different dimensions such as the capital dimension, logistics dimension and control dimension, calculate their actual change time difference. Step S4: Calculate the dynamic attenuation factor based on the actual time difference of change and the predetermined business hysteresis coefficient. The dynamic attenuation factor decreases inversely as the product of the actual time difference of change and the business hysteresis coefficient increases. Multiply the initial association strength parameter with the dynamic attenuation factor to generate the dynamic association edge weight of the corresponding association edge. Step S5: Update the edge weights of each associated edge in the initial heterogeneous graph according to the dynamic associated edge weights to generate a dynamic evolution graph. Step S6: Extract the topological feature changes of the target feature nodes in the dynamic evolution graph that serve as risk detection targets, compare the topological feature changes with the benchmark risk trigger threshold, and output the business risk identification results.

[0020] Preferably, step S5 includes the following node compensation steps: Step S51, when the data sampling interruption time of the second feature node of the target associated enterprise exceeds the preset period, determine the control radiation range of the third feature node corresponding to the target associated enterprise; Step S52, extract the state change timestamps of the second feature nodes of other enterprises within the control radiation range that have the same third feature node association with the target associated enterprise, and calculate the weighted average timestamp based on the initial association strength parameter; Step S53, use the weighted average timestamp as the compensation timestamp of the second feature node with sampling interruption, so as to provide the time boundary of the graph evolution and maintain the topological connectivity of the second feature node in the dynamic evolution graph; wherein, the topological compensation timestamp does not participate in the determination of the cross-dimensional actual change time difference and dynamic decay factor.

[0021] Preferably, step S4 includes the following edge truncation steps: Step S41, when the actual change time difference exceeds the set critical hysteresis window, adjust the weight of the corresponding dynamic associated edge to 0; Step S42, stop the weight transfer of the associated edge with a dynamic associated edge weight of 0 in the initial heterogeneous graph, and update the topological convergence boundary.

[0022] Preferably, in step S2, the initial correlation strength parameter is determined in the following way: step S21, extract the basic transaction frequency data in the funding dimension and the control ratio data in the control dimension; step S22, normalize the basic transaction frequency and control ratio data to determine the initial correlation strength parameter.

[0023] Preferably, in step S6, the baseline risk trigger threshold is set to an initial scalar base on the same numerical scale as the topological feature change, and is determined according to the following adaptive adjustment steps: Step S61, obtain the macroeconomic prosperity index of the industry in which the target related enterprise is located, and calculate the gradient rate of change of the macroeconomic prosperity index within the continuous sampling period; Step S62, when the gradient rate of change is negative and the absolute value exceeds the preset steady-state fluctuation boundary, calculate the bias compensation factor on the same numerical scale based on the product of the gradient rate of change and the initial scalar base; Step S63, add the baseline risk trigger threshold and the bias compensation factor to update the baseline risk trigger threshold.

[0024] Preferably, in step S6, the topological feature change is the change in the feature vector centrality of the target feature node in the dynamic evolution graph; the output business risk identification result includes the following judgment steps: step S64, when the change in feature vector centrality is lower than the baseline risk trigger threshold, it is determined that there is risk transmission; step S65, generate and output a risk warning signal containing the target feature node identifier and the risk transmission path.

[0025] Preferably, in step S1, the first feature node is a funds account node, the second feature node is a warehousing and logistics node, and the third feature node is a legal representative and equity control node. The process of obtaining underlying business data includes: step S11, reading original business messages from the enterprise resource planning system, supply chain management system, and business and legal database respectively through a combination of streaming parsing and timed batch processing via a preset data acquisition interface and network transmission protocol; step S12, performing structured parsing on the original business messages, extracting target field data including transaction amount, warehousing and logistics records, equity ratio, and legal representative, and performing entity attribution alignment on the multi-source heterogeneous target field data using the unified social credit code of the target associated enterprise as the unique entity identifier; step S13, constructing the first feature node in the fund dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension based on the target field data after entity alignment and the preset graph attribute mapping relationship.

[0026] Preferably, step S65 includes the following tiered response steps: Step S651, setting the first risk level threshold to be the same as the benchmark risk trigger threshold, and setting a second risk level threshold with a value lower than the first risk level threshold, forming a progressive risk judgment interval; Step S652, when the change in feature vector centrality is between the first risk level threshold and the second risk level threshold, outputting a warning data packet containing the target feature node identifier, risk transmission path, and medium risk identifier to the business risk control terminal, triggering the business risk control terminal to generate a manual review task; Step S653, when the change in feature vector centrality is lower than the second risk level threshold, outputting an interception instruction data packet containing the target feature node identifier and high-risk interception identifier to the transaction settlement system, triggering the transaction settlement system to execute the transaction settlement blocking and fund account freezing actions of the corresponding target related enterprise.

[0027] Preferably, in step S3, the process of extracting the state change timestamp and calculating the actual change time difference is based on the set sliding time window period update, and includes the following steps: Step S31, when new business data input is detected, update the node state change timestamp within the set sliding time window; Step S32, recalculate the actual change time difference between cross-dimensional adjacent feature nodes within the set sliding time window, and trigger dynamic association edge weight update.

[0028] Example 1: The current distributed computing node processes a feature dataset containing 10,000 target related enterprises to identify cross-legal entity business risks. The computing node calls a preset user authorization interface to obtain historical transaction reconciliation records and fund transfer data from the enterprise resource planning system, warehousing and logistics records from the supply chain management system, and related guarantee relationships, equity and legal representative data, and judicial litigation announcements from the business and legal database. The built-in data desensitization unit de-identifies the legal representative's ID card number, bank account number, and sensitive communication details to generate underlying business data that meets security and privacy requirements.

[0029] The multidimensional graph association system for operational characteristics receives underlying business data, extracts fund account nodes from fund transaction data, warehousing and logistics nodes from warehousing and logistics records, and legal representative and equity control nodes from equity and legal representative data. It records the entity affiliation of each characteristic node with the identifier of the target related enterprise. The system constructs fund association edges based on objective transaction records, logistics association edges based on warehousing, transportation and delivery records, and control association edges based on equity relationships and legal representative affiliation relationships, forming an initial heterogeneous graph. This initial heterogeneous graph retains the transaction edge topology between the target related enterprises and constitutes the initial state of the multidimensional graph relationship network for operational characteristics with each characteristic node and its association edges.

[0030] When determining the initial association strength parameters, the system extracts basic transaction frequency data from the capital dimension and control ratio data from the control dimension, normalizes the two types of data respectively, and determines the initial association strength parameters of each association edge based on the normalization results. The graph association server then runs the implicit negative event record injection algorithm, using the adjacency matrix A between each feature node in the initial heterogeneous graph and the negative event occurrence probability distribution P on the historical time series as input. The adjacency matrix A represents the connection topology between each feature node, and the negative event occurrence probability distribution P represents the probability of negative events occurring in each target associated enterprise within the historical time period.

[0031] The implicit negative event recording injection algorithm sets a 90-day fund flow statistics window to extract the frequency of fund transactions and abnormal fund flow characteristics of the target related enterprises. When the abnormality degree D of fund flow between the target related enterprise and the marked risk related entities exceeds the benchmark threshold for fund flow abnormality degree. At that time, the server generates virtual event records representing implicit negative effects. The anomaly degree D of fund flow is used to quantify the degree of anomaly in fund flow, and the benchmark threshold for fund flow anomaly degree is... When set to 0.75, the server writes the virtual event record as an implicit negative event record to the feature node and associated edge of the corresponding target associated enterprise, updates the state change timestamp of the relevant node, and changes the local weight distribution in the initial heterogeneous graph.

[0032] The probability distribution P of negative events is obtained by retrieving the administrative penalty records, frequency of judicial litigation, and number of overdue repayment days of the target related enterprise in the past 365 days, and then normalizing them. The abnormality degree of fund flow D is calculated by weighting and summing the proportion of the transaction amount of the target related enterprise and the risk related entity during the nighttime non-working hours in the fund flow statistics window to the average daily transaction amount, as well as the ratio of a single large amount of funds transferred out quickly. The result is between 0 and 1.

[0033] During the event state update process of the initial heterogeneous graph, the implicit negative event injection coefficient To adjust the injection strength of virtual events, the system performs gradient comparison tests on the computing nodes, adjusting the implicit negative event injection coefficients accordingly. Setting the values ​​to 0.1, 0.5, and 1.2, the test results show that when the implicit negative event injection coefficient is... When the coefficient is below 0.2, the indirect risk diffusion characteristics between nodes cannot be extracted; when the implicit negative event injection coefficient is below 0.2, the indirect risk diffusion characteristics between nodes cannot be extracted. When the value is greater than 1.0, normal related entities will be converted to an abnormal state. After considering the combined risk of false negative rate and false positive rate, the implicit negative event injection coefficient will be used. Limiting it to between 0.4 and 0.6 keeps the graph topology stable and blocks potentially risky paths.

[0034] Baseline threshold for abnormality of fund flow The threshold is set at 0.75. This value was determined through offline calculations of 5,000 historical known risk cases. When the threshold is below 0.65, normal liquidity turnover is easily judged as abnormal fund flow, and the false alarm rate rises to more than 8%. When the threshold is above 0.85, weak abnormal flow is difficult to trigger virtual event injection, and the false alarm rate exceeds 12%. When the threshold is set at 0.75, the false alarm rate and the false alarm rate are in a relatively balanced range.

[0035] After completing the implicit negative event record injection, the server extracts the state change timestamps of each feature node in the initial heterogeneous graph, calculates the actual change time difference between adjacent feature nodes across dimensions according to the time sequence, configures the business hysteresis coefficient according to the business response cycle between the dimensions of funds, logistics and control, and calculates the dynamic decay factor based on the actual change time difference and the business hysteresis coefficient. The dynamic decay factor is between 0 and 1, and decreases as the product of the actual change time difference and the business hysteresis coefficient increases. The server multiplies the initial association strength parameter with the corresponding dynamic decay factor to generate the dynamic association edge weight of each association edge, and then updates the edge weight in the initial heterogeneous graph with the dynamic association edge weight to generate the dynamic evolution graph.

[0036] For a target feature node that serves as a risk detection target, the server calculates its feature vector centrality in the graph before and after the update, and uses the difference between the two calculation results as the change in feature vector centrality. When the change in feature vector centrality is lower than the baseline risk trigger threshold pre-configured by the system, it is determined that there is risk transmission. Starting from the target feature node, the server backtracks along the associated edges that participated in the feature vector centrality calculation and whose dynamic associated edge weights are not 0. It records the identifiers of adjacent feature nodes according to the order of state change timestamps, forming a risk transmission path, and outputs a risk warning signal containing the target feature node identifier and the risk transmission path.

[0037] While outputting business risk identification results, the multidimensional graph association system uses a risk propagation model based on topological random walks to calculate the risk probability of target related enterprises and outputs a risk characteristic index between 0 and 1. This provides real-time early warnings and risk characteristic indices for the backend management system. Used to indicate the risk level of target related enterprises, it does not replace the risk transmission judgment based on the change in feature vector centrality. The multidimensional graph association system uses the existing distributed database physical storage architecture to dynamically construct a virtual topology in memory and output the identification status of hidden risks between enterprises.

[0038] A business network containing 2,000 feature nodes and 5,000 association edges was selected from the feature dataset of 10,000 target related enterprises for testing. The test results show that after adopting the business feature multi-dimensional graph association method of this embodiment, the false alarm rate of risk identification is controlled within 3%, and the false alarm rate is not higher than 5%.

[0039] Example 2: This example uses a feature association server equipped with a dual-processor and mounted with an operational feature graph storage medium. The server receives raw operational feature data generated by 512 distributed data acquisition nodes over a continuous period of 168 hours. The data includes basic attribute features of the business entity, behavioral trajectory features, and related transaction features. This data is used to verify the association accuracy and convergence stability of the multi-dimensional operational feature graph association method under the condition of multi-source heterogeneous operational data input.

[0040] The feature association server calls the existing data acquisition interface to receive the original business behavior messages reported by each distributed data acquisition node. The streaming data parsing unit decodes the messages, extracts underlying physical quantities such as subject identifier, timestamp, transaction amount and network protocol address, and generates basic attribute tensors and behavior trajectory matrices based on these physical quantities, so that the original business data can be converted into mathematical vectors that the graph computing unit can process.

[0041] The server categorizes operational characteristics based on data sources and business attributes. It extracts fund transaction data from the Enterprise Resource Planning (ERP) system and constructs fund account nodes, extracts warehousing and logistics records from the Supply Chain Management (SLM) system and constructs warehousing and logistics nodes, and extracts equity and legal representative data from the business registration and legal affairs data and constructs legal representative and equity control nodes. Subsequently, based on objective transaction records and entity ownership relationships, it constructs fund association edges, logistics association edges, and control association edges between each characteristic node to form an initial heterogeneous graph.

[0042] In the initial heterogeneous graph, each feature node is marked with the unified social credit code of the target associated enterprise. The server extracts the basic transaction frequency data in the funding dimension and the control ratio data in the control dimension, normalizes the two types of data, and determines the initial association strength parameters of each associated edge by combining the spatial Euclidean distance and topological overlap coefficient between feature nodes. This stage only determines the initial state of each feature node and associated edge, without introducing subsequent temporal changes.

[0043] The graph association algorithm module includes a feature projection layer, a multidimensional association matrix product layer, and a threshold determination output layer. The feature projection layer linearly maps the basic attribute tensor and the behavior trajectory matrix to a unified low-dimensional association space. The multidimensional association matrix product layer calculates the spatial Euclidean distance and topological overlap coefficient between different feature nodes based on the mapped feature vectors. The threshold determination output layer judges the graph update status based on the preset critical convergence condition.

[0044] During the dynamic evolution process, the server extracts the timestamps of state changes of each feature node and calculates the actual time difference of changes between adjacent feature nodes across dimensions. The business lag coefficient is determined based on the industry average response cycle of different business dimensions. The benchmark lag cycle between the capital dimension and the logistics dimension is set to 72 hours, and the benchmark lag cycle between the control dimension and the business decision dimension is set to 180 days. The server standardizes the actual time difference of changes based on the corresponding benchmark lag cycle, so that the product of the actual time difference of changes and the business lag coefficient forms a standardized lag ratio.

[0045] The server calculates the dynamic attenuation factor between 0 and 1 using the following formula. ,in, This is the dynamic attenuation factor, with a value range of (0,1]. The actual time difference of change between adjacent feature nodes across dimensions, with values ​​satisfying... ; For a given business lag coefficient, in this embodiment, we take... ,ratio This is the dimensionless standardized lag ratio calculated by subtracting the actual time difference from the business lag coefficient. is the preset attenuation adjustment coefficient, and is a dimensionless scaling factor used to control the smoothness of the attenuation curve and the response sensitivity. Its value ranges from [0.5, 2.0]. In this embodiment, is preferably selected. ; with the actual time difference With business lag coefficient The product of these, i.e., the standardized hysteresis ratio. The increase of dynamic decay factor The decay rate exhibits an inverse relationship, gradually and smoothly decreasing from 1 to near 0. Subsequently, the initial association strength parameter of each associated edge is compared with the corresponding dynamic decay factor. Multiplication generates dynamic association edge weights, which are then used to update the edge weights in the initial heterogeneous graph, forming a dynamic evolution graph. Static feature fusion is used to determine the initial association strength parameters, and time decay processing is used to update the dynamic association edge weights. The two are applied sequentially to the graph construction and graph evolution stages.

[0046] The server updates the state change timestamps based on a set sliding time window period. Whenever new business data input is detected, it first updates the state change timestamps of the corresponding feature nodes within the set sliding time window, then recalculates the actual change time difference between adjacent feature nodes across dimensions within the window, and triggers dynamic association edge weight updates. To reduce input fluctuations caused by network jitter and clock asynchrony of distributed data acquisition nodes, the server uses a sliding window mean filtering algorithm to smooth the distributed transaction data before calculating the output volatility of the multidimensional association matrix product layer.

[0047] The output volatility of the multidimensional correlation matrix product layer is below the convergence threshold for three consecutive time steps. When the server determines that the dynamic evolution graph has reached a stable state, it sends a convergence completion result to the output terminal, and the convergence threshold is set. Configure the critical convergence conditions as set before the test, and keep them the same in each test sample group in this embodiment.

[0048] During the risk identification phase, the server calculates the change in eigenvector centrality of the target feature node in the dynamic evolution graph and compares it with the baseline risk trigger threshold. Eigenvector centrality is used to characterize the core connectivity and resource control degree of the target feature node in the associated network. When the change is negative and falls below the baseline risk trigger threshold, the core supply chain, funding channels, or control connections of the target associated enterprise are weakened. Based on this, the server determines that there is risk transmission and outputs the target feature node identifier and the corresponding risk transmission path.

[0049] For the adjustment of risk assessment boundaries caused by changes in the industry's macroeconomic climate, the server calculates the gradient rate of change of the macroeconomic climate index within a continuous sampling period, and converts the gradient rate of change into a dimensionless relative change ratio. The relative change ratio is multiplied by the initial scalar base of the benchmark risk trigger threshold to obtain a bias compensation factor that is on the same numerical scale as the change in the centrality of the feature vector. Then, the bias compensation factor is added to the benchmark risk trigger threshold to complete the update of the benchmark risk trigger threshold.

[0050] This embodiment further verifies the risk identification effect and algorithm convergence performance of the proposed method through experiments. All tests are conducted on a labeled business risk dataset containing 10,000 enterprise feature nodes and 50,000 dynamic related edges. The risk transmission path in the dataset is used as the benchmark for judgment. The test server is configured with dual processors and 128GB of memory, and all test groups use the same graph traversal depth and threshold judgment criteria.

[0051] First, under a fixed 1.5% base perturbation noise environment, the feature dimension scalability was tested. Test group one set the feature dimension to 32 dimensions. After dynamic update of the associated edge weights, the algorithm converged in 125.6 ms, with a risk identification accuracy of 92.4%. Test group two expanded the feature dimension to 64 dimensions, with a convergence time of 142.1 ms and an identification accuracy of 95.1%. Test group three further increased the feature dimension to 128 dimensions, with a convergence time of 165.3 ms and an identification accuracy of 97.2%. The test results show that with the integration of multi-dimensional feature information such as funds, logistics, and control, the graph is supplemented with richer cross-dimensional topological associations, which significantly improves the risk identification accuracy, while the increase in convergence time remains within a controllable range of milliseconds.

[0052] Secondly, to verify the ability of the dynamic attenuation and truncation mechanism of this invention to suppress pseudo-association noise, while keeping the 128-dimensional feature configuration unchanged, the external disturbance noise intensity in test sample group four was increased to 8.0%. The test results showed that the algorithm convergence time was 178.2ms, and the risk identification accuracy remained at a high level of 96.5%. When the noise increases and introduces false transactions or abnormal associations, this method extracts the actual change time difference through steps S3 to S4 and calculates the dynamic attenuation factor using the business hysteresis coefficient. For pseudo-association edges whose time difference exceeds the set critical hysteresis window, step S41 directly truncates their weight to 0 and stops the transmission. This mechanism automatically cuts off the diffusion path of noise pseudo-associations in the graph topology, so that even when the noise intensity is greatly increased, the identification accuracy of the system only fluctuates slightly, demonstrating extremely strong noise resistance.

[0053] In contrast, under the same test environment with identical node size, 128-dimensional feature configuration, and 8.0% noise intensity, a traditional static graph topology traversal method without a temporal hysteresis decay mechanism was used for comparative testing. The experimental results show that the traditional static graph method cannot identify cross-dimensional time delays and false connectivity relationships, and a large amount of pseudo-association noise participates in the graph topology calculation, resulting in a risk identification accuracy rate of 78.6%, and the algorithm's iterative convergence is unstable.

[0054] The above test results fully demonstrate that by transforming dimensional propagation delay into topological edge parameter constraints, this invention can effectively eliminate false association noise while maintaining low computational complexity, thus ensuring the stability and risk identification accuracy of the dynamic evolution graph under complex conditions.

[0055] Example 3: The current processor obtains the underlying business data of the target related enterprises from the enterprise resource planning system, supply chain management system, and business and legal database. The underlying business data includes business registration information, tax declaration data, bank transaction records, warehousing and logistics records, equity and legal representative data, and upstream and downstream transaction vouchers. The processor performs structured parsing on the underlying business data, extracts fund account nodes from fund transaction data, extracts warehousing and logistics nodes from warehousing and logistics records, and extracts legal representative and equity control nodes from equity and legal representative data, and records the target related enterprises corresponding to each feature node.

[0056] The processor establishes capital and logistics links based on objective transaction records, and control links based on equity relationships and legal representative attribution. The initial heterogeneous graph is composed of each feature node and link. For raw data lacking specific key fields, the processor uses a general overarching concept corresponding to the data category to which the field belongs to complete the structured replacement, and retains the data missing marker, so that the relevant feature nodes can continue to participate in the construction of the initial heterogeneous graph and subsequent updates.

[0057] When determining the scope of control, the processor includes entities with a direct equity penetration shareholding ratio exceeding 50% of the target related enterprise, as well as corporate entities with the same natural person as the legal representative, into the same scope of control. When the data sampling of the warehousing and logistics node of the target related enterprise is interrupted, the processor extracts the effective warehousing and logistics node status change timestamps of other related enterprises within the scope of control. Using the initial association strength parameter of the corresponding association edge as the weight coefficient, the processor accumulates the product of each status change timestamp and the corresponding weight coefficient, and then divides it by the sum of all weight coefficients to obtain the weighted average timestamp. The warehousing and logistics node with the sampling interruption is used as the compensation timestamp.

[0058] The processor extracts physical quantities that characterize the operational status of the target related enterprises and converts them into a feature matrix that can be computed. The model used to process this feature matrix includes a feature extraction layer, a multidimensional matrix multiplication layer, and a risk mapping layer. The feature extraction layer receives structured parsed data on funds, logistics, and control rights, and outputs normalized values ​​for each operational feature. The multidimensional matrix multiplication layer determines the initial correlation strength parameters based on the normalized values. The risk mapping layer receives the topological feature changes in the dynamic evolution map and outputs the business risk identification results according to preset risk judgment conditions.

[0059] In the static feature fusion stage, the processor extracts the normalized values ​​of the basic transaction frequency and the normalized values ​​of the control rights ratio data from the feature matrix, and calculates the initial correlation strength parameter according to the following formula: ,in, The initial correlation strength parameter is represented by X, which represents the normalized value of the basic transaction frequency, Y, which represents the normalized value of the control ratio data, a, which represents the first weight coefficient, and b, which represents the second weight coefficient. The first weight coefficient a and the second weight coefficient b satisfy: a + b = 1.0. The first weight coefficient a and the second weight coefficient b are adjusted in conjunction with the system's preset risk sensitivity. The processor writes the calculated initial correlation strength parameter into the corresponding capital correlation edge, logistics correlation edge, and control correlation edge to form the initial edge weights of the initial heterogeneous graph.

[0060] During the dynamic evolution phase of the graph, the processor extracts the timestamps of state changes of each feature node, calculates the actual time difference of change between adjacent feature nodes across dimensions, and calculates the dynamic decay factor based on the actual time difference of change and the business hysteresis coefficient. The dynamic decay factor decreases inversely as the product of the actual time difference of change and the business hysteresis coefficient increases. The processor multiplies the dynamic decay factor with the initial association strength parameter of the corresponding associated edge to generate the dynamic association edge weight, and then updates the initial heterogeneous graph with the dynamic association edge weight to obtain the dynamically evolved graph.

[0061] To filter target feature nodes that need to enter topology risk detection, the processor will initialize the association strength parameters. Related filtering threshold Compare, when satisfied At that time, the corresponding feature nodes are included in the candidate detection range, and the associated screening threshold is set. It is only used to narrow down the detection range of target feature nodes and is not directly used as the basis for determining the results of business risk identification. When the real-time load of the system increases and the historical incidence rate of potential risks rises, the processor will associate the screening threshold according to the existing configuration. The value was adjusted from 0.85 to 0.72 to increase the number of feature nodes that enter the candidate detection range.

[0062] For target feature nodes within the candidate detection range, the processor calculates their feature vector centrality in adjacent sampling periods and uses the difference between the two calculation results as the feature vector centrality change. When the feature vector centrality change is lower than the baseline risk trigger threshold, the risk mapping layer determines that there is risk propagation and generates a risk propagation path based on the associated edges involved in the centrality calculation and their state change timestamps.

[0063] The benchmark risk trigger threshold is updated based on the macroeconomic prosperity index of the industry in which the target related enterprise is located. The processor calculates the gradient rate of change of the macroeconomic prosperity index within the continuous sampling period. When the gradient rate of change is negative and the absolute value is between 5% and 15%, the bias compensation factor is 0.05; when the gradient rate of change is negative and the absolute value exceeds 15%, the bias compensation factor is 0.12. The processor adds the bias compensation factor to the benchmark risk trigger threshold to obtain the updated benchmark risk trigger threshold.

[0064] The risk mapping layer also sets a first risk level threshold and a second risk level threshold, where the first risk level threshold is higher than the second risk level threshold. When the change in feature vector centrality is between the first risk level threshold and the second risk level threshold, the risk mapping layer outputs a medium risk warning instruction; when the change in feature vector centrality is lower than the second risk level threshold, it outputs a high risk interception instruction, and writes high risk / manual intervention required, target feature node identifier, and risk transmission path into the output information. Operation and maintenance personnel make corresponding business handling judgments based on this output information.

[0065] Example 4: This example combines Figures 1 to 2 This paper describes a method for associating operational characteristics using a multidimensional graph for business risk identification, such as... Figure 1 As shown, step S1 obtains the underlying business data of the target related enterprise, including enterprise resource planning system, supply chain management system, and business and legal data, and extracts the first feature node in the capital dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension; step S2 constructs capital association edges, logistics association edges, and control association edges connecting each feature node based on the objective transaction records and entity ownership relationships in the underlying business data, forming an initial heterogeneous graph containing initial association strength parameters; step S3 extracts the state change timestamps of each feature node in the initial heterogeneous graph and calculates the actual change time difference between adjacent feature nodes across dimensions; step S4 calculates a dynamic decay factor that decays inversely based on the actual change time difference and a predetermined business lag coefficient, multiplies the initial association strength parameters by the dynamic decay factor, and generates the dynamic association edge weights of the corresponding association edges; step S5 updates the edge weights of each association edge in the initial heterogeneous graph based on the dynamic association edge weights, generating a dynamic evolution graph; step S6 extracts the topological feature changes of the target feature nodes in the dynamic evolution graph, compares them with the benchmark risk trigger threshold, and outputs the business risk identification results.

[0066] like Figure 2 As shown, the underlying business data includes enterprise resource planning system data, supply chain management system data, and business and legal data; feature node extraction includes the first feature node in the funding dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension; the initial heterogeneous graph construction includes constructing funding association edges, logistics association edges, and control association edges; the dynamic graph evolution includes calculating the actual change time difference across dimensions, generating dynamic decay factors and association edge weights, and updating the graph based on dynamic weights; the threshold comparison judgment includes extracting the topological feature change amount of the target feature node and implementing the comparison benchmark risk trigger threshold; all parts interact to ultimately form the business risk identification result output.

[0067] Example 5: Before the multi-dimensional graph association method of business characteristics is put into operation, the controller reads the basic time series of graph feature nodes from the designated address of the global memory, and calls the hardware timer to perform synchronous sampling rate calibration of the data bus of the connected enterprise resource planning system, supply chain management system and business and legal database with a period of 10ms.

[0068] The controller converts the characteristic transition edges of different business data streams within the same calibration window into pulse sequences. By calculating the cross-correlation coefficient of the two sets of characteristic pulse sequences, the time alignment deviation between each business data stream is determined. When the business characteristic flow record of the financial and tax interface and the behavioral characteristic record of the supply chain master file are compared within a calibration window with a length of 120 sampling points, the cross-correlation coefficient reaches an extreme value at a position with a shift of 3 sampling points. Based on this, the controller determines the time alignment deviation to be 3 time units and moves the read pointer of the financial and tax interface forward by 3 storage units, so that the two business characteristic data streams enter subsequent processing according to the same time base.

[0069] After completing time alignment, the controller extracts fund account nodes from enterprise resource planning system data, warehousing and logistics nodes from supply chain management system data, and legal representative and equity control nodes from business and legal data. Based on objective transaction records and entity ownership relationships, it constructs fund association edges, logistics association edges, and control rights association edges to form an initial heterogeneous graph. The processor extracts the state change timestamps of each feature node, calculates the actual change time difference between adjacent feature nodes across dimensions, and then generates dynamic association edge weights based on the actual change time difference and business lag coefficient to update the initial heterogeneous graph and form a dynamic evolution graph.

[0070] In this embodiment, the critical hysteresis window is set to 180 days. The longest cycle for corporate financial and tax audits and supply chain payment settlements is usually no more than half a year. When the actual change time difference between adjacent feature nodes across dimensions exceeds 180 days, the two types of business behaviors are no longer treated as continuous events in the same temporal causal chain. The controller adjusts the dynamic associated edge weight of the corresponding associated edge to 0, stops the weight transfer of the associated edge in the initial heterogeneous graph, and updates the topological convergence boundary based on the retained associated edge.

[0071] The steady-state fluctuation boundary is set at 5%. The controller calculates the gradient rate of change of the macroeconomic prosperity index of the industry in which the target related enterprise is located within the continuous sampling period. When the absolute value of the gradient rate of change is less than 5%, the external environment is within the set steady-state fluctuation range, and the current benchmark risk trigger threshold is maintained. When the gradient rate of change is negative and the absolute value exceeds 5%, the bias compensation factor is calculated according to the gradient rate of change, and the bias compensation factor is added to the benchmark risk trigger threshold to obtain the updated benchmark risk trigger threshold.

[0072] When data sampling at a warehousing and logistics node is interrupted, the controller first determines the control scope covered by the legal representative and equity control node of the target related enterprise, and then extracts the state change timestamps of other enterprises within that scope that have the same legal representative and equity control node as the target related enterprise. The processor uses the initial association strength parameter of each association edge as a weighting coefficient to calculate the weighted average timestamp of these state change timestamps, and uses the warehousing and logistics node with the sampling interruption as the compensation timestamp.

[0073] The compensation timestamp is used to define the temporal causal position of warehousing and logistics nodes in the dynamic evolution graph. It is not written into the underlying business data as the occurrence time of actual warehousing, transportation or delivery events. Related enterprises within the same control scope are constrained by common equity and management relationships. When the warehousing and logistics node of the target related enterprise temporarily lacks sampling data, the weighted average timestamp provides the time boundary required for the graph evolution of that node, enabling the dynamic evolution graph to continue to be updated and avoiding the stagnation of topology operations caused by the data interruption of a single warehousing and logistics node.

[0074] Example 6: When the multidimensional business characteristic mapping method is first deployed in a new business environment, or when it is necessary to establish a baseline reference for a target related enterprise, the processor obtains the historical business characteristic data of the target related enterprise, extracts principal components using principal component analysis, and constructs a business characteristic baseline reference model. The historical business characteristic dataset contains N sample data points, and each sample data point contains M-dimensional feature parameters; where N represents the number of sample data points, and M represents the number of feature parameters contained in each sample data point.

[0075] The processor centers and standardizes the historical business characteristic dataset, calculates the covariance matrix C, and solves for the eigenvalues ​​and eigenvectors of the covariance matrix C. The covariance matrix C is calculated according to the following formula: ,in, This represents the feature vector of the i-th sample. Let represent the sample mean vector, i represent the sample index, and the superscript T represents the matrix transpose.

[0076] The processor sorts the solved eigenvalues ​​according to their numerical values ​​based on an 85% cumulative variance contribution rate threshold, selects the eigenvectors corresponding to the first k eigenvalues, and constructs a dimension-reduced projection matrix W using these eigenvectors; where k represents the number of eigenvalues ​​to be retained when the cumulative variance contribution rate reaches 85%.

[0077] Subsequently, the historical operational feature dataset is projected onto a low-dimensional subspace defined by the dimension-reduced projection matrix W, and the low-dimensional baseline feature vector after projection of the i-th sample is obtained. Calculate according to the following formula: The processor arranges the low-dimensional baseline feature vectors according to the sample order. Forming a baseline feature matrix This baseline feature matrix is ​​used to characterize the health baseline status of the target related enterprises in subsequent daily business operations.

[0078] After generating low-dimensional baseline feature vectors, the processor writes them into the initial heterogeneous graph as the initial attribute embedding vectors of the corresponding feature nodes. For the capital dimension, logistics dimension, and control dimension, the initial attribute embedding vectors are written into the capital account node, warehousing and logistics node, and legal representative and equity control node, respectively. The cosine similarity between the initial attribute embedding vectors of adjacent feature nodes across dimensions is calculated to verify the degree of correlation between the initial states of adjacent feature nodes.

[0079] Based on objective transaction records and entity ownership relationships in the underlying business data, the processor constructs capital association edges, logistics association edges, and control rights association edges respectively. For each association edge, it extracts the basic transaction frequency data in the capital dimension and the control rights ratio data in the control rights dimension, normalizes the two types of data, and determines the initial association strength parameter based on the normalization result. Then, it writes the parameter into the corresponding association edge. Thus, the low-dimensional baseline feature vector obtained from principal component analysis forms the initial attributes of each feature node, and the normalized basic transaction frequency and control rights ratio data form the initial association strength parameter of each association edge, which together constitute the initial state of the initial heterogeneous graph.

[0080] The above description is only a few preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with the technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A method for associating multidimensional maps of operational characteristics for business risk identification, characterized in that, Includes the following steps: Step S1: Obtain the underlying business data of the target related enterprise. The underlying business data includes enterprise resource planning system data, supply chain management system data, and business and legal data. Extract the first feature node in the funding dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension. Step S2: Based on the objective transaction records and entity ownership relationships in the underlying business data, construct the fund association edges, logistics association edges, and control rights association edges connecting each feature node to form an initial heterogeneous graph. The initial association edges include the initial association strength parameter. Step S3: Extract the state change timestamps of each feature node in the initial heterogeneous graph. For cross-dimensional adjacent feature nodes that are directly connected by related edges and belong to different dimensions such as the capital dimension, logistics dimension and control dimension, calculate their actual change time difference. Step S4: Calculate the dynamic attenuation factor based on the actual time difference of change and the predetermined business hysteresis coefficient. The dynamic attenuation factor decreases inversely as the product of the actual time difference of change and the business hysteresis coefficient increases. Multiply the initial association strength parameter with the dynamic attenuation factor to generate the dynamic association edge weight of the corresponding association edge. Step S5: Update the edge weights of each associated edge in the initial heterogeneous graph according to the dynamic associated edge weights to generate a dynamic evolution graph. Step S6: Extract the topological feature changes of the target feature nodes in the dynamic evolution graph that serve as risk detection targets, compare the topological feature changes with the benchmark risk trigger threshold, and output the business risk identification results.

2. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 1, characterized in that, Step S5 includes the following node compensation steps: Step S51, when the data sampling interruption time of the second feature node of the target associated enterprise exceeds the preset period, determine the control radiation range of the third feature node corresponding to the target associated enterprise; Step S52, extract the state change timestamps of the second feature nodes of other enterprises within the control radiation range that have the same third feature node association with the target associated enterprise, and calculate the weighted average timestamp based on the initial association strength parameter; Step S53, use the weighted average timestamp as the compensation timestamp of the second feature node with sampling interruption, so as to provide the time boundary of the graph evolution and maintain the topological connectivity of the second feature node in the dynamic evolution graph; wherein, the topological compensation timestamp does not participate in the determination of the cross-dimensional actual change time difference and dynamic decay factor.

3. The method for associating multidimensional operational characteristics for business risk identification according to claim 1, characterized in that, Step S4 includes the following edge truncation steps: Step S41, when the actual change time difference exceeds the set critical hysteresis window, adjust the weight of the corresponding dynamic associated edge to 0; Step S42, stop the weight transfer of associated edges with dynamic associated edge weight of 0 in the initial heterogeneous graph, and update the topological convergence boundary.

4. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 1, characterized in that, In step S2, the initial correlation strength parameter is determined in the following way: Step S21, extract the basic transaction frequency data in the funding dimension and the control ratio data in the control dimension; Step S22, normalize the basic transaction frequency and control ratio data to determine the initial correlation strength parameter.

5. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 1, characterized in that, In step S6, the baseline risk trigger threshold is set to an initial scalar base on the same numerical scale as the topological feature change, determined according to the following adaptive adjustment steps: Step S61, obtain the macroeconomic prosperity index of the industry in which the target related enterprise is located, and calculate the gradient rate of change of the macroeconomic prosperity index within the continuous sampling period; Step S62, when the gradient rate of change is negative and the absolute value exceeds the preset steady-state fluctuation boundary, calculate the bias compensation factor on the same numerical scale based on the product of the gradient rate of change and the initial scalar base; Step S63, add the baseline risk trigger threshold and the bias compensation factor to update the baseline risk trigger threshold.

6. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 1, characterized in that, In step S6, the topological feature change is the change in the feature vector centrality of the target feature node in the dynamic evolution graph; the output business risk identification result includes the following judgment steps: step S64, when the change in feature vector centrality is lower than the baseline risk trigger threshold, it is determined that there is risk transmission; step S65, generate and output a risk warning signal containing the target feature node identifier and the risk transmission path.

7. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 1, characterized in that, In step S1, the first feature node is the funds account node, the second feature node is the warehousing and logistics node, and the third feature node is the legal representative and equity control node. The process of obtaining the underlying business data includes: Step S11, reading the original business messages from the enterprise resource planning system, supply chain management system, and business and legal database respectively through a combination of streaming parsing and timed batch processing via a preset data acquisition interface and network transmission protocol; Step S12, performing structured parsing on the original business messages, extracting target field data including transaction amount, warehousing and logistics records, equity ratio, and legal representative, and using the unified social credit code of the target associated enterprise as the unique entity identifier to perform entity attribution alignment on the multi-source heterogeneous target field data; Step S13, constructing the first feature node in the funds dimension, the second feature node in the logistics dimension, and the third feature node in the control dimension based on the target field data after entity alignment and the preset graph attribute mapping relationship.

8. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 6, characterized in that, Step S65 includes the following tiered response steps: Step S651, setting the first risk level threshold to be the same as the benchmark risk trigger threshold, and setting a second risk level threshold with a value lower than the first risk level threshold, forming a progressive risk judgment interval; Step S652, when the change in feature vector centrality is between the first risk level threshold and the second risk level threshold, outputting a warning data packet containing the target feature node identifier, risk transmission path, and medium risk identifier to the business risk control terminal, triggering the business risk control terminal to generate a manual review task; Step S653, when the change in feature vector centrality is lower than the second risk level threshold, outputting an interception instruction data packet containing the target feature node identifier and high-risk interception identifier to the transaction settlement system, triggering the transaction settlement system to execute the transaction settlement blocking and fund account freezing actions of the corresponding target related enterprise.

9. The method for associating multidimensional maps of business characteristics for business risk identification according to claim 1, characterized in that, In step S3, the process of extracting the state change timestamp and calculating the actual change time difference is based on the set sliding time window period update, and includes the following steps: Step S31, when new business data input is detected, update the node state change timestamp within the set sliding time window; Step S32, recalculate the actual change time difference between cross-dimensional adjacent feature nodes within the set sliding time window, and trigger dynamic association edge weight update.

Images