An import and export cross-border payment chain supervision method based on a dynamic graph

Abstract

The application belongs to the technical field of payment security, and provides an import and export cross-border payment chain supervision method based on a dynamic graph, comprising: payment knowledge graph construction, payment node portrait construction, initial graph risk data acquisition, influence diffusion processing, FinBERT model identification, decision generation and execution; the application upgrades the risk monitoring of cross-border payment from single node monitoring to risk tracking of the whole link and correlation by building a cross-border payment exclusive knowledge graph, adding an influence diffusion processing mechanism and combining a large language model; and the application realizes dynamic and scenario-based grading of cross-border payment risks by calculating the risk value and risk level interval boundary of the node and the payment chain, so that the grading can accurately adapt to the dynamic risk changes of cross-border payment and fit the scenario differences of different nodes and payment chains.

Description

Technical Field

[0001] This invention relates to the field of payment security technology, and in particular to a method for supervising cross-border payment chains for import and export based on dynamic graphs. Background Technology

[0002] In import and export cross-border payment scenarios, the transaction chain is long and the participating nodes are complex, involving domestic and foreign enterprises, payment institutions, clearing nodes, etc., and there are common problems such as illegal cross-border fund flows, payment risks under the background of fraudulent trade, and risk propagation and diffusion across nodes. At present, the conventional methods for controlling import and export cross-border payments mainly include manual review of transaction documents, setting fixed transaction amount thresholds, risk monitoring of individual nodes, and simple verification of the qualifications of both parties to the transaction.

[0003] Existing control methods have many problems in practical applications, including: 1) Risk assessment is based only on transaction data of a single node, without considering the correlation characteristics between cross-border payment nodes, making it impossible to identify the cross-node propagation patterns of risks, and difficult to predict the spread trend of impacts and accurately locate the spread path of impacts; 2) Risk level classification uses fixed thresholds or integer levels, which cannot adapt to the dynamic changes in cross-border payment risks and cannot match the differences in the impact of risks on surrounding nodes under different cross-border transaction scenarios; 3) Node profiles are constructed too simply, without combining cross-border payment-related transaction data, import and export qualification data, cross-border relationship data, etc., and cannot truly reflect the risk attributes of cross-border payment nodes; 4) Risk identification and control decisions are too simplistic, without designing targeted solutions based on the scenario characteristics of cross-border payments, resulting in insufficient targeting and effectiveness in preventing and controlling cross-border illegal payments and the spread of impacts. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method for supervising cross-border payment chains for import and export based on dynamic graphs, which solves the problems of existing control methods lacking correlation analysis, being unable to cope with dynamic changes, and having overly simplistic profile construction.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for supervising cross-border payment chains for imports and exports based on dynamic graphs, comprising: Construct a payment knowledge graph for cross-border payment scenarios; A payment node profile is constructed based on the payment knowledge graph and the pre-collected raw node multi-source data; Based on the payment knowledge graph and the payment node profile, risk values ​​are calculated and levels are classified for each node and payment chain to obtain initial graph risk data. The initial graph risk data is then dynamically adjusted in terms of risk level and supplemented with diffusion characteristics to obtain optimized graph risk data. The initial graph risk data includes node risk data and payment chain risk data. Based on the optimized graph risk data, the nodes and payment chains in the payment knowledge graph are subjected to negative or positive influence diffusion processing to obtain diffused data; The diffused data is input into a pre-trained FinBERT model for identification processing to obtain multi-dimensional risk identification and diffusion trend prediction results. Based on the risk map data, the post-diffusion data, and the multi-dimensional risk identification and diffusion trend prediction results, cross-border payment risk prevention and control decisions are generated and executed using pre-set hierarchical differentiated prevention and control decisions.

[0006] Preferably, a payment knowledge graph for cross-border payment scenarios is constructed, including: Invalid and duplicate data are removed from the pre-collected multi-source data of cross-border payment nodes to obtain standardized collected data; Use the Protégé tool to set the node and relationship types to obtain the ontology model; Based on the ontology model, transaction subgraphs and enterprise business subgraphs are constructed using transaction behavior data and business association data from the standardized collected data. The transaction subgraphs and enterprise business subgraphs are then fused using a single node identifier to obtain an initial dual graph. The initial dual graph is subjected to invalid edge removal and connectivity verification using the pruning algorithm to obtain the payment knowledge graph; The payment knowledge graph is periodically updated and its connectivity is verified using a preset dynamic graph update formula; the dynamic graph update formula is as follows: ;in, The updated map; The map before fusion; This includes newly added nodes and associated edge data; These are invalid associated edges.

[0007] Preferably, a payment node profile is constructed based on the payment knowledge graph and pre-collected multi-source data of raw nodes, including: The original node multi-source data is obtained by collecting basic node attributes, transaction behavior, correlation relationships, risk and compliance data, and supplementary data; the supplementary data includes: personal credit and enterprise operating status. The original node multi-source data is cleaned and normalized to obtain preprocessed node multi-source data; Static features, behavioral dynamic features, correlation features, and risk compliance features are extracted from the multi-source data of the preprocessing nodes to obtain multi-dimensional feature data. The multidimensional feature data is weighted and fused to obtain the payment node profile, and the payment node profile is periodically updated using a preset dynamic profile update formula; the dynamic profile update formula is: ;in, The updated node comprehensive profile feature values; To update the comprehensive profile feature values ​​of the previous node; This refers to the adjustment amount for the feature weights; This represents the change in eigenvalues.

[0008] Preferably, the formula for calculating the node risk data is: ;in, ; ; Real-time risk value for nodes; Risk contribution value of associated nodes; This serves as the baseline value for its own risk. These are conversion factors; Node comprehensive profile feature values; The real-time risk value of the target associated node; This represents the affinity coefficient between the current node and the target associated node. This affects the efficiency of diffusion and transfer.

[0009] Preferably, the formula for calculating the payment chain risk data is: ;in, Real-time risk value for the payment chain; The number of nodes on the payment chain; This represents the real-time risk value of the i-th node on the payment chain. Let be the weight of the i-th node.

[0010] Preferably, the expression corresponding to the dynamic adjustment of the risk level is: The expression corresponding to the diffusion characteristic is: ;in, Adjustment coefficient for interval boundary values; The weight of node correlation; Weighted by historical risk frequency; Weighting of the transaction amount fluctuation coefficient; Node correlation degree; Historical risk frequency; This refers to the fluctuation coefficient of the transaction amount. The intensity of risk diffusion; This represents the real-time risk value for the current node or payment chain. The correlation coefficient; The diffusion attenuation coefficient; This represents the correlation distance.

[0011] Preferably, the process of the negative influence spreading includes: When the risk value of the target object exceeds that of its associated objects within the domain, if the influence diffusion and transmission efficiency and risk duration of the target object and its associated objects exceed a set threshold, then the target object and its associated objects are designated as a negative influence pair. The target object includes either a node or a payment chain. The expression for the influence diffusion and transmission efficiency is: ;in, This affects the diffusion and transfer efficiency; This represents the real-time risk value for the current node or payment chain. This represents the real-time risk value of the associated node or payment chain. The correlation coefficient; The duration of the risk; Collect all the negative impact pairs within the payment knowledge graph; A diffusion efficiency gradient attenuation process is applied to each negative influence pair using an attenuation formula. If the risk value difference and diffusion intensity corresponding to the negative influence pair are less than a set threshold, the state of the negative influence pair is switched to the completed state. The expression of the attenuation formula is: ;in, Let be the propagation efficiency of the k-th stage diffusion; For initial transmission efficiency; is the diffusion attenuation coefficient.

[0012] Preferably, the process of positive influence diffusion includes: When the risk value of the target object is lower than that of the related objects in the domain, if the target object meets the preset low-risk standard, the target object and the related objects are set as a positive influence pair; the target object includes either a node or a payment chain; the low-risk standard includes: compliance score exceeding a preset threshold, no violation record in the previous six months, and risk value lower than a preset threshold for three consecutive months; Statistically analyze all the positive impact pairs within the payment knowledge graph; The positive impact pair is diffused using an attenuation formula. If the risk value difference between the positive impact pair and the positive impact pair is less than a set threshold, the state of the positive impact pair is switched to the completed state.

[0013] Preferably, the diffused data is input into a pre-trained FinBERT model for identification processing to obtain multi-dimensional risk identification and diffusion trend prediction results, including: The structured and unstructured data in the diffused data are transformed and concatenated to obtain a joint input vector; The joint input vector is adaptively fused using a multi-head self-attention mechanism to obtain optimized fused features. The optimized fusion features are mapped using a multi-task output layer to obtain the multi-dimensional risk identification and diffusion trend prediction results. The multi-dimensional risk identification and diffusion trend prediction results include: transaction violation identification results, diffusion trend prediction results, compliance experience effectiveness identification results, and profile supplementation results.

[0014] Preferably, the cross-border payment risk prevention and control management decision and the transaction violation identification result in the multi-dimensional risk identification and diffusion trend prediction result correspond one-to-one; the types of the cross-border payment risk prevention and control management decision include: low-risk compliance type, medium-low risk potential violation type, medium-risk minor violation type, medium-high risk obvious violation type, and high-risk major violation type.

[0015] The present invention discloses the following technical effects: This invention provides a method for supervising cross-border payment chains for import and export based on dynamic graphs. By constructing a knowledge graph specific to cross-border payments, adding an impact diffusion processing mechanism, and combining a large language model, it solves the problem that existing controls only assess risk at a single node and do not consider the correlation characteristics between nodes, thus enabling the analysis and processing of the impact of risk propagation across nodes. By calculating the risk value and risk level range boundary of nodes and payment chains, it solves the problem that existing methods cannot adapt to the dynamic changes of cross-border payment risks and are insufficient in meeting the differences in scenarios, thus achieving dynamic and scenario-based classification of cross-border payment risks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the import and export cross-border payment chain supervision process based on dynamic graphs provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the node profile construction process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the negative influence diffusion process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the positive influence diffusion process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the model recognition process provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a method for supervising the import and export cross-border payment chain based on dynamic graphs, which solves the problems of existing control methods lacking correlation analysis, being unable to cope with dynamic changes, and having overly simplistic profile construction.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Figure 1 This is a schematic diagram of the import and export cross-border payment chain supervision process based on dynamic graphs provided in an embodiment of the present invention, such as... Figure 1 As shown, this invention provides a method for supervising cross-border payment chains for imports and exports based on dynamic graphs, including: Step 100: Construct a payment knowledge graph for cross-border payment scenarios; Step 200: Construct a payment node profile based on the payment knowledge graph and the pre-collected original node multi-source data; Step 300: Calculate the risk value and classify the level of each node and payment chain according to the payment knowledge graph and the payment node profile to obtain initial graph risk data. Dynamically adjust the risk level and supplement the diffusion characteristics of the initial graph risk data to obtain optimized graph risk data. The initial graph risk data includes: node risk data and payment chain risk data. Step 400: Based on the optimized graph risk data, perform negative or positive influence diffusion processing on the nodes and payment chains in the payment knowledge graph to obtain the diffused data; Step 500: Input the diffused data into the pre-trained FinBERT model for identification processing to obtain multi-dimensional risk identification and diffusion trend prediction results; Step 600: Based on the map risk data, the post-diffusion data, and the multi-dimensional risk identification and diffusion trend prediction results, generate and execute cross-border payment risk prevention and control decisions using pre-set hierarchical differentiated prevention and control decisions.

[0022] Furthermore, a payment knowledge graph for cross-border payment scenarios will be constructed, including: Invalid and duplicate data are removed from the pre-collected multi-source data of cross-border payment nodes to obtain standardized collected data; Use the Protégé tool to set the node and relationship types to obtain the ontology model; Based on the ontology model, transaction subgraphs and enterprise business subgraphs are constructed using transaction behavior data and business association data from the standardized collected data. The transaction subgraphs and enterprise business subgraphs are then fused using a single node identifier to obtain an initial dual graph. The initial dual graph is subjected to invalid edge removal and connectivity verification using the pruning algorithm to obtain the payment knowledge graph; The payment knowledge graph is periodically updated and its connectivity is verified using a preset dynamic graph update formula; the dynamic graph update formula is as follows: ;in, The updated map; The map before fusion; This includes newly added nodes and associated edge data; These are invalid associated edges.

[0023] refer to Figure 2 Based on the payment knowledge graph and pre-collected multi-source data of raw nodes, a payment node profile is constructed, including: Step 201: Collect basic node attributes, transaction behavior, relationships, risk and compliance data, and supplementary data to obtain the original node multi-source data; the supplementary data includes: personal credit and enterprise operating status; Step 202: Perform data cleaning and normalization on the original node multi-source data to obtain preprocessed node multi-source data; Step 203: Extract static features, behavioral dynamic features, correlation features, and risk compliance features from the multi-source data of the preprocessed nodes to obtain multi-dimensional feature data; Step 204: Perform weighted fusion calculation on the multi-dimensional feature data to obtain the payment node profile, and periodically update the payment node profile using a preset dynamic profile update formula; the dynamic profile update formula is: ;in, The updated node comprehensive profile feature values; To update the comprehensive profile feature values ​​of the previous node; This refers to the adjustment amount for the feature weights; This represents the change in eigenvalues.

[0024] Furthermore, the formula for calculating the node risk data is as follows: ;in, ; ; Real-time risk value for nodes; Risk contribution value of associated nodes; This serves as the baseline value for its own risk. These are conversion factors; Node comprehensive profile feature values; The real-time risk value of the target associated node; This represents the affinity coefficient between the current node and the target associated node. This affects the efficiency of diffusion and transfer.

[0025] Specifically, the formula for calculating the payment chain risk data is as follows: ;in, Real-time risk value for the payment chain; The number of nodes on the payment chain; This represents the real-time risk value of the i-th node on the payment chain. Let be the weight of the i-th node.

[0026] Furthermore, the expression corresponding to the dynamic adjustment of the risk level is: The expression corresponding to the diffusion characteristic is: ;in, Adjustment coefficient for interval boundary values; The weight of node correlation; Weighted by historical risk frequency; Weighting of the transaction amount fluctuation coefficient; Node correlation degree; Historical risk frequency; This refers to the fluctuation coefficient of the transaction amount. The intensity of risk diffusion; This represents the real-time risk value for the current node or payment chain. The correlation coefficient; The diffusion attenuation coefficient; This represents the correlation distance.

[0027] refer to Figure 3 The process of the negative impact spreading includes: Step 401: When the risk value of the target object exceeds that of related objects within the domain, if the influence diffusion and transmission efficiency and risk duration of the target object and the related objects exceed a set threshold, then the target object and the related objects are set as a negative influence pair; the target object includes either a node or a payment chain; the expression for the influence diffusion and transmission efficiency is: ;in, This affects the diffusion and transfer efficiency; This represents the real-time risk value for the current node or payment chain. This represents the real-time risk value of the associated node or payment chain. The correlation coefficient; The duration of the risk; Step 402: Statistically analyze all the negative impact pairs within the payment knowledge graph; Step 403: Apply a diffusion efficiency gradient attenuation process to each negative influence pair using an attenuation formula. If the risk value difference and diffusion intensity corresponding to the negative influence pair are less than a set threshold, then switch the state of the negative influence pair to the completed state. The expression of the attenuation formula is: ;in, Let be the propagation efficiency of the k-th stage diffusion; For initial transmission efficiency; is the diffusion attenuation coefficient.

[0028] refer to Figure 4 The process of the positive influence diffusion includes: Step 404: When the risk value of the target object is lower than that of the related objects in the domain, if the target object meets the preset low-risk standard, the target object and the related objects are set as a positive influence pair; the target object includes either a node or a payment chain; the low-risk standard includes: compliance score exceeding a preset threshold, no violation record in the previous six months, and risk value lower than a preset threshold for three consecutive months; Step 405: Statistically analyze all the positive influence pairs within the payment knowledge graph; Step 406: Use the attenuation formula to diffuse each positive influence pair. If the risk value difference corresponding to the positive influence pair is less than a set threshold, then switch the state of the positive influence pair to the completed state.

[0029] refer to Figure 5 The diffused data is input into a pre-trained FinBERT model for identification processing to obtain multi-dimensional risk identification and diffusion trend prediction results, including: Step 501: Transform and concatenate the structured and unstructured data in the diffused data to obtain a joint input vector; Step 502: Use a multi-head self-attention mechanism to perform adaptive feature fusion processing on the joint input vector to obtain optimized fused features; Step 503: Use the multi-task output layer to map the optimized fusion features to obtain the multi-dimensional risk identification and diffusion trend prediction results; the multi-dimensional risk identification and diffusion trend prediction results include: transaction violation identification results, diffusion trend prediction results, compliance experience effectiveness identification results, and profile supplementation results.

[0030] Preferably, the cross-border payment risk prevention and control management decision and the transaction violation identification result in the multi-dimensional risk identification and diffusion trend prediction result correspond one-to-one; the types of the cross-border payment risk prevention and control management decision include: low-risk compliance type, medium-low risk potential violation type, medium-risk minor violation type, medium-high risk obvious violation type, and high-risk major violation type.

[0031] Specifically, in the knowledge graph construction process, this embodiment integrates transaction and business registration data to build a dedicated fusion graph, providing related data support for node profiling, risk level calculation, and diffusion path tracking, as follows: First, the basic attributes and transaction behaviors of cross-border payment nodes are standardized, and then... The formula removes invalid and duplicate data, where Standardized data that can be used after cleaning For unprocessed raw multi-source data, Invalid data (not limited to data missing core information or data with incorrect format). To avoid duplicate data, the core nodes and relationship types of the graph were defined, and an ontology model was constructed using the Protégé tool. Subsequently, a transaction subgraph was constructed based on transaction behavior data, and an enterprise business association subgraph was constructed based on business association data, forming a dual graph. The dual graphs were then merged using unique node identifiers, using the following formula: ,in To form a complete knowledge graph after integration. For the transaction subgraph, For the industrial and commercial sub-map, For invalid edges, the pruning algorithm is used to remove invalid edges and perform connectivity checks. This is done periodically. Update the graph and then check connectivity again. For the updated atlas, For newly added nodes and associated edge data, These are invalidated related edges, such as those resulting from transaction termination or dissolution of business relationships. Furthermore, the knowledge graph constructed in this embodiment has an automatic exchange rate settlement function. By constructing a dynamically updated knowledge graph in real time, it deeply integrates and dynamically analyzes various entities and relationships involved in the cross-border payment chain, including domestic and foreign transaction entities, payment institutions, channel banks, and settlement exchange rates. This addresses the problems of existing control methods lacking correlation analysis, being unable to cope with dynamic changes, and having overly simplistic profile construction.

[0032] Furthermore, this embodiment relies on knowledge graph-related data and A multi-dimensional dynamic node profile is constructed to provide accurate support for risk level calculation. Multi-source data is collected from nodes, cleaned, and normalized before use. Core data types include: basic attributes, transaction behavior, relationships, risk compliance, and supplementary data. Supplementary data includes personal credit, business operations, and institutional technical assessments. In the basic attribute data, individual user nodes include name and ID number, enterprise nodes include enterprise name and unified social credit code, and payment institution nodes include institutional qualifications and regulatory filings. Transaction behavior data includes transaction time, amount, and recipient, excluding reversals and invalid transactions. Relationship data includes inter-node transactions, business registration, and cooperation relationships; this data is consistent with the knowledge graph data. Risk and compliance data includes historical risk events and compliance scores.

[0033] Specifically, static features are extracted from the basic attribute data, expressed as follows: ,in The comprehensive value of basic attribute features, The number of basic attribute features, The weight of the k-th basic attribute feature. This represents the normalized value of the k-th basic attribute feature, covering the basic dimensions of individuals, enterprises, and payment institutions; dynamic features are extracted from transaction behavior data, utilizing... , , Calculate transaction activity, percentage of abnormal transactions, and transaction concentration, among which... To increase trading activity, This is the activity level weighting coefficient. To count the number of transactions within a period, To calculate the number of days in the statistical period, This represents the percentage of abnormal transactions. To count the number of abnormal transactions within the statistical period, For transaction concentration, For the number of trading partners, The transaction amount with the i-th object, To calculate the total transaction amount within the statistical period, then... The comprehensive value of behavioral characteristics is obtained, where This is a comprehensive value representing behavioral characteristics. , , The weights for transaction activity, percentage of abnormal transactions, and transaction concentration are respectively used; related features are extracted based on a pre-built knowledge graph, and then... Calculate the correlation coefficient, where Let be the correlation coefficient between node i and node j. Let be the transaction amount between node i and node j. Let be the industrial and commercial association weight between node i and node j. Let i be the total transaction amount of node i. The total industrial and commercial association weight of node i is then determined by... The comprehensive value of the associated features is obtained, where This is the comprehensive value of the associated features. , , These are the weights for the average correlation coefficient, the number of associated nodes, and the length of the associated link, respectively. This is the average affinity coefficient between the current node and all its associated nodes. The number of associated nodes. The average length of the associated links is given; then features are extracted from the risk and compliance data, and the calculation expression is given. ,in This is a comprehensive value representing risk and compliance characteristics. , , The weights for these factors are historical risk frequency, compliance score, and risk rectification status, respectively. As for the frequency of historical risks, For compliance scoring, the range is from 0 to 100. To assess the completion rate of risk rectification; Utilize Weighted fusion is performed to obtain the comprehensive feature values ​​of the node profile, where For the comprehensive portrait feature values ​​of nodes, , , , The weights are assigned to basic attributes, behaviors, relationships, and risk compliance characteristics, respectively.

[0034] Optionally, the node profile is dynamically updated, and the update formula is: ,in The updated node comprehensive profile feature values, To update the comprehensive profile feature values ​​of the previous node, This is the adjustment amount for the feature weights. The update cycle is set according to the node type, and can be set as follows: once a month for enterprises, once every half month for payment institutions, and once a week for individuals.

[0035] Specifically, based on node-based comprehensive profile feature values And the impact of associated node risks, the real-time risk value of the node is calculated using the following formula: .

[0036] .

[0037] in, Real-time risk value for nodes; The risk contribution value of associated nodes is the sum of the risk impact of all directly associated nodes on the current node; This serves as the baseline value for its own risk. This is the conversion factor.

[0038] .

[0039] in, This represents the risk contribution value of a single associated node to the current node. This represents the real-time risk value of the associated node. This represents the affinity coefficient between the current node and its associated node. To influence diffusion and transmission efficiency (initial value determined by risk level difference).

[0040] Risk levels are specifically categorized as: Low risk Low to medium risk Medium risk Medium and high risk High risk .

[0041] Furthermore, the entire cross-border payment chain is extracted based on the knowledge graph and marked as... The payment chain risk value is calculated by combining the risk values ​​and weights of on-chain nodes, with the classification criteria consistent with those of the nodes. Based on the real-time risk values ​​of all on-chain nodes and the tightness of the chain connections, the real-time risk value of the payment chain is calculated using the following formula: .

[0042] in, This represents the real-time risk value of the payment chain. The number of nodes on the payment chain. Let be the real-time risk value of the i-th node on the payment chain. Let be the weight of the i-th node.

[0043] The risk level classification of the payment chain is consistent with the risk level of the nodes, and its diffusion impact characteristics are synchronized with the risk level of the nodes.

[0044] Optionally, the risk level is not fixed and needs to be dynamically adjusted in conjunction with the node profile update cycle to ensure that the risk level matches the profile data in real time. The adjustment formula is as follows: .

[0045] in, Adjustment coefficient for interval boundary values; The weight of node correlation; Node correlation degree; Weighted by historical risk frequency; Historical risk frequency; Weighting of the transaction amount fluctuation coefficient; This is the fluctuation coefficient of the transaction amount.

[0046] .

[0047] in, These are the adjusted interval boundary values; As the baseline boundary value; Adjustment coefficient for interval boundary values.

[0048] The intensity of risk diffusion is positively correlated with the risk level and negatively correlated with the association distance, as shown in the formula: ,in As for the intensity of risk diffusion, This represents the real-time risk value of the current node / payment chain. This represents the affinity coefficient between the current node and its associated nodes. The diffusion attenuation coefficient is... This represents the correlation distance.

[0049] Specifically, the negative impact diffusion stage. The current risk value is higher than any directly related node / payment chain in the neighborhood, and , Triggered at time, where ; To affect diffusion and transfer efficiency, This represents the real-time risk value of the current node / payment chain. This represents the real-time risk value of the associated nodes / payment chain. The correlation coefficient between the two is... The duration of risk is defined; low-risk associated nodes calculate their risk contribution value according to the diffusion intensity formula and update their own risk value. If they become new high-risk centers, they continue to diffuse and record data throughout the process; the constraint phase is defined according to the decay formula. ( (Consistent with the diffusion intensity formula) to achieve gradient decay of diffusion efficiency, where Let be the propagation efficiency of the k-th stage diffusion. For initial transmission efficiency, The attenuation coefficient is... For diffusion levels; when , The diffusion was terminated at that time, among which The difference between the current risk value and the associated node. The diffusion intensity is denoted as .

[0050] Further, the positive impact diffusion phase. The current risk value is lower than any directly related node / payment chain within the neighborhood, and the compliance score... No violations in the past 6 months, for 3 consecutive months Triggered in time; high-risk associated nodes absorb compliance experience and reduce their own risk values. If they become new low-risk benchmarks, they continue to spread and record data throughout the process; in the impact phase, the diffusion model is used to disseminate compliance impact, summarize compliance experience to form an experience base, and supplement the model input data; when It then enters a stable phase.

[0051] Preferably, the large language model selection and risk identification involve inputting the standardized dataset after diffusion (including post-diffusion risk values, diffusion records, and profiling data) into a suitable large language model to achieve accurate risk identification and diffusion trend prediction. Specifically, the FinBERT model is selected, and the model structure is adjusted for post-diffusion data types: To adapt to the data types affected by diffusion processing in this embodiment (structured data: post-diffusion RRV, CCF; unstructured text: diffusion records, compliance experience, transaction notes), the FinBERT model structure needs to be adaptively adjusted. A structured data adaptation module is added to the input layer. , Structured data is converted into feature vectors and concatenated with unstructured text feature vectors to form a joint input vector. The intermediate layer sets up a cross-border payment feature fusion module based on a multi-head self-attention mechanism to assign targeted weights to the joint input vectors passed from the input layer and complete feature fusion optimization. The output layer receives the feature vectors fused by the intermediate layer and outputs the recognition results based on the multi-task output layer, which corresponds to four major scenarios: transaction violation recognition, diffusion trend prediction, compliance experience effectiveness recognition, and profile supplementation.

[0052] Furthermore, based on knowledge graph data, node profile data, and violation results identified by the model, decision-making processes are performed on target nodes and payment chains. Combining the violation results output by the model (including violation type, violation severity, risk diffusion tendency, and predicted diffusion intensity), target nodes / payment chains are categorized and processed according to type. Corresponding decisions include: 1) Low-risk compliant type: No violations identified by the model, stable associations shown by the knowledge graph, and low risk shown by the node profile (RRV∈(0,2], CS≥85). Linking related nodes to share compliance experience and providing compliance incentives simultaneously; 2) Medium-low risk potential violation type: For nodes / payment chains with potential violations identified by the model, close associations (CCF≥0.5), and medium-low risk shown by the node profile (RRV∈(2,3]), tracking related links based on the knowledge graph and increasing monitoring frequency; 3) Medium-risk minor violation type: Minor violations identified by the model, multiple node linkages shown by the knowledge graph, and node profiles... For example, if the risk is medium (RRV∈(3,4]), restrictions will be placed on the scale and scope of transactions, a rectification report will be required, and node profile data will be updated synchronously; 4) Medium-to-high risk with obvious violations: if the model identifies obvious violations, the knowledge graph shows cross-link diffusion, and the node profile shows medium-to-high risk (RRV∈(4,5]), risk blocking measures will be taken (suspending some payment functions or freezing illegal related funds), manual review will be initiated, and the status of related edges in the knowledge graph will be updated; 5) High risk with major violations: if the model identifies major violations, the knowledge graph shows large-scale diffusion, and the node profile shows high risk (RRV≥5), systemic prevention and control will be initiated, all payment functions will be suspended, all related funds will be frozen, regulatory authorities will be involved in investigation, the source of risk will be traced in conjunction with the node profile, and the knowledge graph and node profile data will be updated synchronously.

[0053] The beneficial effects of this invention are as follows: (1) This invention builds a knowledge graph dedicated to cross-border payments, designs a mechanism for handling the diffusion of positive and negative impacts, and combines it with an optimized model to predict the trend of risk diffusion, ultimately upgrading the risk monitoring of cross-border payments from monitoring a single node to tracking risks across the entire chain and in a related manner.

[0054] (2) This invention calculates the real-time risk value of nodes and payment chains, designs a dynamic adjustment formula for the risk level interval boundary, and sets the corresponding risk level update cycle according to different node types. This breaks the limitation of fixed threshold for risk level classification, so that the level classification can not only accurately adapt to the dynamic risk changes of cross-border payments, but also fit the scenario differences of different nodes and payment chains, making the risk classification more in line with the actual business situation.

[0055] (3) This invention integrates multi-source data related to cross-border payments to build a standardized multi-dimensional dynamic node profile system, sets the profile update cycle according to node type, and continuously supplements and improves the node profile by combining the data after diffusion, providing reliable core data support for subsequent core links such as risk level calculation and impact diffusion processing.

[0056] (4) This invention adapts and optimizes the FinBERT professional model, conducts multi-dimensional accurate risk identification based on the diffused data, designs a hierarchical and differentiated prevention and control decision mechanism, and builds a data linkage closed loop system for the whole process. This not only improves the accuracy of risk identification and the accuracy of diffusion trend prediction, but also enables prevention and control decisions to accurately adapt to the actual risk status, correlation characteristics and violations of nodes and payment chains, effectively preventing systemic risks in the field of cross-border payments.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0058] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for supervising cross-border payment chains for import and export based on dynamic graphs, characterized in that, include: Construct a payment knowledge graph for cross-border payment scenarios; A payment node profile is constructed based on the payment knowledge graph and the pre-collected raw node multi-source data; Based on the payment knowledge graph and the payment node profile, risk values ​​are calculated and levels are classified for each node and payment chain to obtain initial graph risk data. The initial graph risk data is then dynamically adjusted in terms of risk level and supplemented with diffusion characteristics to obtain optimized graph risk data. The initial graph risk data includes: node risk data and payment chain risk data; Based on the optimized graph risk data, the nodes and payment chains in the payment knowledge graph are subjected to negative or positive influence diffusion processing to obtain diffused data; The diffused data is input into a pre-trained FinBERT model for identification processing to obtain multi-dimensional risk identification and diffusion trend prediction results. Based on the risk map data, the post-diffusion data, and the multi-dimensional risk identification and diffusion trend prediction results, cross-border payment risk prevention and control decisions are generated and executed using pre-set hierarchical differentiated prevention and control decisions.

2. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, Constructing a payment knowledge graph for cross-border payment scenarios, including: Invalid and duplicate data are removed from the pre-collected multi-source data of cross-border payment nodes to obtain standardized collected data; Use the Protégé tool to set the node and relationship types to obtain the ontology model; Based on the ontology model, transaction subgraphs and enterprise business subgraphs are constructed using transaction behavior data and business association data from the standardized collected data. The transaction subgraphs and enterprise business subgraphs are then fused using a single node identifier to obtain an initial dual graph. The initial dual graph is subjected to invalid edge removal and connectivity verification using the pruning algorithm to obtain the payment knowledge graph; The payment knowledge graph is periodically updated and its connectivity is verified using a preset dynamic graph update formula; the dynamic graph update formula is as follows: ;in, The updated map; The map before fusion; This includes newly added nodes and associated edge data; These are invalid associated edges.

3. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, A payment node profile is constructed based on the payment knowledge graph and pre-collected multi-source data of raw nodes, including: The original node multi-source data is obtained by collecting basic node attributes, transaction behavior, correlation relationships, risk and compliance data, and supplementary data; the supplementary data includes: personal credit and enterprise operating status. The original node multi-source data is cleaned and normalized to obtain preprocessed node multi-source data; Static features, behavioral dynamic features, correlation features, and risk compliance features are extracted from the multi-source data of the preprocessing nodes to obtain multi-dimensional feature data. The multidimensional feature data is weighted and fused to obtain the payment node profile, and the payment node profile is periodically updated using a preset dynamic profile update formula; the dynamic profile update formula is: ;in, The updated node comprehensive profile feature values; To update the comprehensive profile feature values ​​of the previous node; This refers to the adjustment amount for the feature weights; This represents the change in eigenvalues.

4. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, The formula for calculating the node risk data is as follows: ;in, ; ; Real-time risk value for nodes; Risk contribution value of associated nodes; This serves as the baseline value for its own risk. These are conversion factors; Node comprehensive profile feature values; The real-time risk value of the target associated node; This represents the affinity coefficient between the current node and the target associated node. This affects the efficiency of diffusion and transfer.

5. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, The formula for calculating the payment chain risk data is as follows: ;in, Real-time risk value for the payment chain; The number of nodes on the payment chain; This represents the real-time risk value of the i-th node on the payment chain. Let be the weight of the i-th node.

6. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, The expression corresponding to the dynamic adjustment of the risk level is: The expression corresponding to the diffusion characteristic is: ;in, Adjustment coefficient for interval boundary values; The weight of node correlation; Weighted by historical risk frequency; Weighting of the transaction amount fluctuation coefficient; Node correlation degree; Historical risk frequency; This refers to the fluctuation coefficient of the transaction amount. The intensity of risk diffusion; This represents the real-time risk value for the current node or payment chain. The correlation coefficient; The diffusion attenuation coefficient; This represents the correlation distance.

7. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, The process of the negative impact spreading includes: When the risk value of the target object exceeds that of its associated objects within the domain, if the influence diffusion and transmission efficiency and risk duration of the target object and its associated objects exceed a set threshold, then the target object and its associated objects are designated as a negative influence pair. The target object includes either a node or a payment chain. The expression for the influence diffusion and transmission efficiency is: ;in, This affects the diffusion and transfer efficiency; This represents the real-time risk value for the current node or payment chain. This represents the real-time risk value of the associated node or payment chain. The correlation coefficient; The duration of the risk; Collect all the negative impact pairs within the payment knowledge graph; A diffusion efficiency gradient attenuation process is applied to each negative influence pair using an attenuation formula. If the risk value difference and diffusion intensity corresponding to the negative influence pair are less than a set threshold, the state of the negative influence pair is switched to the completed state. The expression of the attenuation formula is: ;in, Let be the propagation efficiency of the k-th stage diffusion; For initial transmission efficiency; is the diffusion attenuation coefficient.

8. The method for supervising import and export cross-border payment chains based on dynamic graphs according to claim 1, characterized in that, The process of positive influence diffusion includes: When the risk value of the target object is lower than that of the related objects in the domain, if the target object meets the preset low-risk standard, the target object and the related objects are set as a positive influence pair; the target object includes either a node or a payment chain; the low-risk standard includes: compliance score exceeding a preset threshold, no violation record in the previous six months, and risk value lower than a preset threshold for three consecutive months; Statistically analyze all the positive impact pairs within the payment knowledge graph; The positive impact pair is diffused using an attenuation formula. If the risk value difference between the positive impact pair and the positive impact pair is less than a set threshold, the state of the positive impact pair is switched to the completed state.

9. A method for supervising cross-border payment chains for imports and exports based on dynamic graphs according to claim 1, characterized in that, The diffused data is input into a pre-trained FinBERT model for identification processing to obtain multi-dimensional risk identification and diffusion trend prediction results, including: The structured and unstructured data in the diffused data are transformed and concatenated to obtain a joint input vector; The joint input vector is adaptively fused using a multi-head self-attention mechanism to obtain optimized fused features. The optimized fusion features are mapped using a multi-task output layer to obtain the multi-dimensional risk identification and diffusion trend prediction results. The multi-dimensional risk identification and diffusion trend prediction results include: transaction violation identification results, diffusion trend prediction results, compliance experience effectiveness identification results, and profile supplementation results.

10. A method for supervising cross-border payment chains for imports and exports based on dynamic graphs according to claim 1, characterized in that, The decision-making process for cross-border payment risk prevention and control corresponds one-to-one with the transaction violation identification results in the multi-dimensional risk identification and diffusion trend prediction results. The types of cross-border payment risk prevention and control decisions include: low-risk compliance, medium-low risk potential violations, medium-risk minor violations, medium-high risk obvious violations, and high-risk major violations.

Images