A trigger method for merchant credit evaluation based on multi-modal data fusion

Abstract

The application provides a kind of trigger method for merchant credit evaluation based on multi-modal data fusion, obtains the multi-modal data of target merchant and extracts multi-modal feature;Based on multi-modal feature, the network of merchant association is constructed, and the risk propagation index of target merchant is calculated;Based on the change analysis of risk propagation index in continuous time window, the risk propagation change sequence is generated, and the risk propagation evolution graph is constructed according to the risk propagation change sequence;Based on the distribution difference of multi-modal feature in different time stages of risk propagation evolution graph, the optimal transport distance or energy distance is used to calculate the distribution migration degree;When the distribution migration degree meets the preset structure mutation judgment condition, trigger the credit evaluation process of target merchant.The application can improve the identification ability of hidden risk and sudden risk, reduce invalid repeated evaluation, improve the resource utilization efficiency and response efficiency of credit evaluation system.

Description

Technical Field

[0001] This application relates to the field of credit assessment, specifically to a triggering method for merchant credit assessment based on multimodal data fusion. Background Technology

[0002] With the development of internet finance, e-commerce, and supply chain finance, an increasing number of merchant credit assessment businesses are relying on data-driven approaches for risk analysis. Existing merchant credit assessment systems typically use data such as transaction records, performance history, and historical default information to score merchants' creditworthiness and update the results periodically according to preset time cycles. For example, some systems perform credit score calculations on a daily, weekly, or monthly basis to determine whether a target merchant poses a credit risk.

[0003] However, most existing merchant credit assessment methods focus on static credit indicator analysis, typically making risk judgments based on data from a single point in time, which fails to reflect the dynamic changes in the merchant risk propagation structure over time. Furthermore, existing periodic assessment methods often perform repetitive credit assessments on all merchants, consuming significant computational resources even if the target merchant's risk status remains unchanged, resulting in low system resource utilization efficiency.

[0004] In addition, existing technologies typically analyze structured data such as transaction amounts and performance status, but they do not make sufficient use of multimodal data such as textual public opinion, complaint information, and changes in relationships with related merchants, making it difficult to identify hidden risks caused by changes in relationships, the spread of public opinion, or abnormal propagation in upstream and downstream. Summary of the Invention

[0005] In view of the above problems, this application provides a triggering method for merchant credit assessment based on multimodal data fusion, which can improve the ability to identify hidden risks and sudden risks, while reducing invalid and duplicate assessments, and improving the resource utilization efficiency and response efficiency of the credit assessment system.

[0006] To achieve the above objectives, the inventors provide a triggering method for merchant credit assessment based on multimodal data fusion, comprising:

[0007] Acquire multimodal data of the target merchants and extract multimodal features;

[0008] Based on the aforementioned multimodal features, a merchant relationship network is constructed, and the risk propagation index of the target merchants is calculated.

[0009] The risk propagation indicators are analyzed for changes based on continuous time windows to generate a risk propagation change sequence, and a risk propagation evolution diagram is constructed based on the risk propagation change sequence.

[0010] Based on the distribution differences of multimodal characteristics at different time stages of the risk propagation evolution diagram, the degree of distribution migration is calculated using the optimal transmission distance or energy distance;

[0011] When the degree of distribution migration meets the preset structural mutation judgment condition, the credit assessment process for the target merchant is triggered.

[0012] Unlike existing technologies, the above-mentioned technical solution acquires multimodal data of target merchants and extracts multimodal features. It then constructs a merchant relationship network based on these features and generates a risk propagation change sequence and a risk propagation evolution diagram by combining risk propagation indicators. Finally, it analyzes the differences in multimodal feature distribution at different time stages using optimal transmission distance or energy distance, thereby determining whether credit assessment processing should be triggered based on the degree of distribution migration. Unlike traditional methods that perform credit assessments according to fixed time periods, this application does not rely solely on a single static credit indicator for risk judgment. Instead, it analyzes the evolution of the risk propagation structure over time to identify whether the merchant's risk propagation structure undergoes abrupt changes, thus dynamically triggering credit assessment processing when the risk propagation structure changes abnormally. Because this application introduces a risk propagation evolution diagram and a multimodal feature distribution migration analysis mechanism, it can not only identify changes in transaction data but also comprehensively consider the risk diffusion problem caused by factors such as textual sentiment, performance behavior, and changes in relationships, thereby improving the ability to identify hidden and sudden risks. At the same time, by triggering credit assessment processing only when structural mutations occur, a large number of invalid and duplicate assessments can be reduced, thereby improving the resource utilization efficiency and response efficiency of the credit assessment system.

[0013] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0014] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0015] In the accompanying drawings of the instruction manual:

[0016] Figure 1 This is a flowchart of a triggering method for merchant credit assessment based on multimodal data fusion. Detailed Implementation

[0017] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0018] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0019] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0020] Please refer to Figure 1 A triggering method for merchant credit assessment based on multimodal data fusion includes:

[0021] Acquire multimodal data of the target merchants and extract multimodal features;

[0022] Based on the aforementioned multimodal features, a merchant relationship network is constructed, and the risk propagation index of the target merchants is calculated.

[0023] The risk propagation indicators are analyzed for changes based on continuous time windows to generate a risk propagation change sequence, and a risk propagation evolution diagram is constructed based on the risk propagation change sequence.

[0024] Based on the distribution differences of multimodal characteristics at different time stages of the risk propagation evolution diagram, the degree of distribution migration is calculated using the optimal transmission distance or energy distance;

[0025] When the degree of distribution migration meets the preset structural mutation judgment condition, the credit assessment process for the target merchant is triggered.

[0026] As described above, by acquiring multimodal data of target merchants and extracting multimodal features, a merchant relationship network is constructed based on these features. A risk propagation change sequence and a risk propagation evolution diagram are generated by combining risk propagation indicators. The optimal transmission distance or energy distance is then used to analyze the differences in the distribution of multimodal features at different time stages, thereby determining whether credit assessment processing should be triggered based on the degree of distribution migration. Unlike traditional credit assessments performed at fixed time periods, this application does not rely solely on a single static credit indicator for risk judgment. Instead, it analyzes the evolution of the risk propagation structure over time to identify whether the merchant's risk propagation structure undergoes abrupt changes, thus dynamically triggering credit assessment processing when the risk propagation structure changes abnormally. Because this application introduces a risk propagation evolution diagram and a multimodal feature distribution migration analysis mechanism, it can not only identify changes in transaction data but also comprehensively consider the risk diffusion problem caused by factors such as public opinion, performance behavior, and changes in relationships, thereby improving the ability to identify hidden and sudden risks. At the same time, by triggering credit assessment processing only when structural mutations occur, a large number of invalid and duplicate assessments can be reduced, thereby improving the resource utilization efficiency and response efficiency of the credit assessment system.

[0027] In some implementations, based on the distribution differences of multimodal characteristics at different time stages of the risk propagation evolution map, the degree of distribution migration is calculated using the optimal transmission distance or energy distance, including:

[0028] Based on the multimodal features associated with the nodes or edges corresponding to each time window in the risk propagation evolution diagram, construct the multimodal feature distribution corresponding to each time window;

[0029] The difference measure is calculated using the optimal transmission distance or energy distance for the multimodal feature distribution corresponding to adjacent time windows to obtain the distribution difference results between adjacent time windows;

[0030] Based on the distribution difference results, the degree of distribution migration of the risk propagation evolution map between adjacent time stages is determined; the degree of distribution migration is used to characterize the degree of change in the risk propagation structure caused by the changes in the multimodal features between adjacent time windows.

[0031] As described above, by constructing a multimodal feature distribution based on the multimodal features associated with nodes or edges corresponding to each time window in the risk propagation evolution diagram, and measuring the differences in the multimodal feature distributions between adjacent time windows, the degree of distribution migration in the risk propagation evolution diagram between adjacent time stages can be determined. In this way, this application no longer only analyzes the changes in individual risk indicators, but models the risk propagation structure from the overall distribution level, enabling the system to identify changes in the risk propagation structure caused by the combined effects of multimodal features.

[0032] For example, in some scenarios, a merchant's own transaction metrics may not change significantly, but the performance of its associated merchants, complaint texts, or transaction relationships may show an abnormal diffusion trend. In this case, the degree of distribution migration can identify hidden risks that are difficult to detect using traditional static risk assessment methods. Furthermore, this application uses the degree of distribution migration to characterize the degree of change in the risk propagation structure, enabling subsequent structural mutation determinations to be based on overall structural changes rather than fluctuations in a single indicator, thereby improving the accuracy of risk identification.

[0033] In this embodiment, nodes in the risk propagation evolution graph represent the risk propagation state within a corresponding time window, and edges represent the transition relationship of risk propagation states between different time windows. Specifically, the system can divide the target merchant's business behavior according to a preset time interval, such as dividing it into multiple time windows by day, week, or month, and generate corresponding risk propagation states based on risk propagation indicators within each time window. Each risk propagation state can serve as a node in the risk propagation evolution graph. For example, in the first time window, if the target merchant's transaction fluctuations are small, the complaint rate is low, and the risks of associated merchants are stable, the system can generate a low-risk propagation state node; while in a later time window, if the target merchant experiences an increase in refund rate, an abnormal increase in associated merchants, or an increase in negative review texts, the system can generate a medium-risk or high-risk propagation state node. Subsequently, the system establishes edges between nodes based on the changing relationship of risk propagation states between adjacent time windows. For example, when the target merchant changes from a low-risk propagation state to a medium-risk propagation state, a connecting edge is established between the corresponding two nodes to represent the transition process of risk propagation states in the time dimension.

[0034] Specifically, both nodes and edges can be associated with multimodal feature information within a corresponding time window. The multimodal features associated with nodes can represent the overall operational risk characteristics of the target merchant within the corresponding time window, such as transaction amount fluctuations, order cancellation rates, fulfillment delay rates, textual sentiment characteristics, complaint sentiment characteristics, and the proportion of abnormal associated merchants. The multimodal features associated with edges can represent the changing characteristics during the risk propagation process, such as the rate of risk growth, the degree of change in relationships, the degree of change in the transaction chain, and the degree of abnormal diffusion among upstream and downstream merchants. In this way, nodes in the risk propagation evolution graph not only represent the risk state at a certain time stage, but edges can also reflect the evolution of the risk state between different time stages.

[0035] In some implementations, the difference metric for the multimodal feature distributions corresponding to adjacent time windows is calculated using the optimal transmission distance or energy distance, including:

[0036] The multimodal feature distribution corresponding to the first time window is determined as the source distribution, and the multimodal feature distribution corresponding to the second time window is determined as the target distribution.

[0037] Based on the feature matching relationship between the source distribution and the target distribution, calculate the optimal transmission distance required for the source distribution to migrate to the target distribution;

[0038] Alternatively, the energy distance between the source distribution and the target distribution can be calculated based on the distance between samples in the source distribution and the target distribution;

[0039] The optimal transmission distance or the energy distance is determined as a measure of the difference between adjacent time windows.

[0040] As described above, by determining the multimodal feature distribution corresponding to the previous time window as the source distribution and the multimodal feature distribution corresponding to the subsequent time window as the target distribution, and using the optimal transmission distance or energy distance to measure the difference between the source and target distributions, a quantitative analysis of the degree of change in the risk propagation structure between different time stages can be achieved. Unlike traditional analysis methods based on fixed thresholds or simple mean differences, this application employs a distribution migration analysis method, which can model risk changes at the overall distribution structure level, thus enabling the identification of complex risk diffusion patterns. For example, when risk spreads from a few high-risk nodes to multiple low-risk nodes, even if the overall risk mean does not change significantly, the degree of distribution migration can still reflect the changing trend of the risk structure, thereby improving the detection capability of potential risk diffusion processes. Furthermore, since the optimal transmission distance can measure the overall migration cost between different distributions, and the energy distance can measure the statistical difference between distributions, this application can adapt to the risk analysis needs of different types of merchants, improving the applicability and stability of the system.

[0041] In some implementations, the optimal transmission distance is calculated as follows:

[0042] The multimodal feature distribution of the t-th time window is represented as: ;

[0043] The multimodal feature distribution of the (t+1)th time window is represented as: ;

[0044] in, and These represent the multimodal feature vectors of nodes or edges within the corresponding time window;

[0045] The optimal transmission distance is calculated based on the following formula:

[0046]

[0047] in, This represents the multimodal feature vector from the t-th time window. Multimodal feature vectors in the (t+1)th time window Transmission weight;

[0048] This represents the set of transfer matrices that satisfy the marginal distribution constraints.

[0049] It represents the distance metric between two multimodal feature vectors;

[0050] The optimal transmission distance As a measure of the difference between adjacent time windows.

[0051] As described above, by employing the optimal transmission distance to measure the differences in multimodal feature distributions between adjacent time windows, and calculating the distribution migration cost based on the distance metric between the transmission matrix and multimodal feature vectors, the overall migration cost required for the risk propagation structure to change from one state to another can be described more accurately. Compared to traditional risk analysis methods based on single-point differences, this application can analyze the risk propagation process from the perspective of overall structural migration, thus identifying complex structural changes such as changes in risk propagation paths, node influence, and relationships within the merchant risk relationship network. Furthermore, by introducing the optimal transmission model, a more reasonable matching relationship can be established between multimodal features, thereby improving the accuracy of comparing risk propagation structures across different time stages.

[0052] In some implementations, the energy distance is calculated as follows:

[0053] The multimodal feature distribution of the t-th time window is represented as: ;

[0054] The multimodal feature distribution of the (t+1)th time window is represented as: ;

[0055] The energy distance is calculated based on the following formula:

[0056]

[0057] in, They represent the distributions respectively. The multimodal feature vectors obtained by independent sampling are mutually independent and follow the same distribution;

[0058] They represent the distributions respectively. The multimodal feature vectors obtained by independent sampling are mutually independent and follow the same distribution;

[0059] This represents the expectation operation;

[0060] The distance norm between vectors;

[0061] The energy distance As a measure of the difference between adjacent time windows.

[0062] As described above, by using energy distance to calculate the inter-sample distance between the source and target distributions, it is possible to analyze the differences in multimodal characteristics across different time windows from a statistical distribution perspective. Since energy distance considers not only the sample distance between different distributions but also the distance between samples within each distribution, it can more stably reflect changes in the risk propagation structure. Compared to analysis using only simple statistical indicators such as mean and variance, this application can more accurately identify risk changes caused by localized abnormal behavior, sudden shifts in public opinion, or abnormal diffusion of correlations. Furthermore, energy distance has strong adaptability to different types of characteristics, thus making it suitable for credit risk analysis in complex multimodal data environments.

[0063] In some implementations, constructing the multimodal feature distribution corresponding to each time window includes:

[0064] Extract the multimodal features corresponding to the target merchant and its associated merchants within each time window;

[0065] The multimodal features are weighted according to the node positions or connections of the target merchant and its associated merchants in the risk propagation evolution diagram.

[0066] The multimodal feature distribution corresponding to the time window is generated based on the weighted multimodal features.

[0067] As described above, by extracting multimodal features corresponding to the target merchant and its associated merchants within each time window, and weighting these features based on their node positions or connections in the risk propagation evolution diagram, a multimodal feature distribution reflecting the risk propagation structure is generated. In this way, this application considers not only the merchant's own risk status but also the risk impact of associated merchants on the target merchant, thus more accurately describing the risk diffusion process in the risk propagation network. Furthermore, by weighting node positions and connections, the impact of key nodes and high-risk paths on the overall risk propagation structure can be highlighted, improving the relevance of the risk propagation structure analysis.

[0068] In some embodiments, the system constructs a multimodal feature distribution corresponding to each time window based on the multimodal features associated with the nodes or edges corresponding to each time window in the risk propagation evolution graph. Preferably, the system first extracts the multimodal feature vectors of the target merchant and its associated merchants within the corresponding time window. For example, for the target merchant, features such as transaction frequency, transaction amount, refund ratio, user rating, sentiment value of complaint text, and logistics fulfillment rate can be extracted; for associated merchants, features such as their corresponding abnormal transaction ratio, negative evaluation ratio, and correlation strength with the target merchant can be extracted. Then, the system can perform weighted processing on each multimodal feature vector according to the position of the node in the risk propagation evolution graph, the node risk level, and the connection weight of the edge. For example, associated merchants with high-frequency transaction relationships with the target merchant and higher risk levels can be assigned higher weights; while ordinary merchants with lower correlation are assigned lower weights.

[0069] After weighting, the system aggregates multiple multimodal feature vectors within the same time window to form a multimodal feature distribution for that time window. For example, multiple multimodal feature vectors can be mapped to a unified feature space, and their distribution in different risk regions can be statistically analyzed, thus forming a multimodal feature distribution that reflects the risk propagation structure of the current time window. In this way, the system constructs not a single risk value, but a distributed structural representation that reflects the overall risk propagation structure of the merchant, changes in relationships, and changes in multimodal behavior. Therefore, it can more accurately reflect the evolutionary differences in the risk propagation status of the target merchant across different time stages.

[0070] In some implementations, when the degree of distribution migration includes distribution difference results corresponding to multiple adjacent time windows, the multiple distribution difference results are aggregated to obtain the overall degree of distribution migration of the target merchant within a preset evaluation period, and the overall degree of distribution migration is updated to the new degree of distribution migration.

[0071] As described above, by aggregating the distribution difference results corresponding to multiple adjacent time windows, the overall distribution migration degree of the target merchant within a preset evaluation period is obtained, and the overall distribution migration degree is updated to a new distribution migration degree. This avoids the problem of false triggering caused by occasional fluctuations based on a single time window. In this way, this application can identify the risk diffusion trend that has been continuously accumulated, and improve the ability to identify long-term abnormal propagation behavior. At the same time, by comprehensively analyzing the distribution difference results of multiple time windows, the stability and robustness of the structural change judgment results can also be improved.

[0072] In some implementations, the step of analyzing the changes in the risk propagation indicators based on continuous time windows includes:

[0073] The risk propagation index is divided into multiple continuous time windows according to a preset time interval, and the corresponding risk propagation index value within each time window is extracted.

[0074] The risk transmission change sequence is generated based on the differences or trends in the risk transmission indicator values ​​between adjacent time windows.

[0075] As described above, by dividing the risk propagation indicators into multiple continuous time windows according to preset time intervals, and generating a risk propagation change sequence based on the differences or trends in risk propagation indicator values ​​between adjacent time windows, the changes in the risk propagation status of merchants can be continuously tracked over time. Compared to traditional methods that only analyze the current risk status, this application can identify the dynamic process of risk accumulation, diffusion, or transfer, thus enabling earlier detection of potential risk propagation trends and improving risk early warning capabilities.

[0076] In some implementations, the risk propagation change sequence is used to characterize the evolution of the risk propagation status of the target merchant over time.

[0077] As described above, by utilizing the risk propagation change sequence to characterize the evolution of the target merchant's risk propagation status over time, the system can analyze not only the static risk status but also the evolutionary patterns of the risk propagation structure. In this way, the system can identify dynamic risk characteristics such as the expansion of risk propagation paths, changes in associated nodes, and changes in the speed of risk diffusion, thereby improving the accuracy and real-time nature of credit assessment triggering.

[0078] In some implementations, constructing a risk propagation evolution diagram based on the risk propagation change sequence includes:

[0079] The risk propagation status corresponding to each time window is determined based on the risk propagation change sequence.

[0080] The risk propagation status corresponding to each time window is used as a graph node;

[0081] Based on the changing relationship of the risk propagation status between adjacent time windows, a connection relationship is established between adjacent graph nodes. This connection relationship is used to characterize the transition relationship of the risk propagation status between different time stages.

[0082] By associating the multimodal feature information within the corresponding time window with the corresponding graph nodes or connections, a risk propagation evolution graph is obtained.

[0083] As described above, by determining the risk propagation state corresponding to each time window based on the risk propagation change sequence, and using the risk propagation state corresponding to each time window as graph nodes, and then establishing connection relationships between graph nodes based on the risk propagation state change relationships between adjacent time windows, while simultaneously associating multimodal feature information within the corresponding time window to the corresponding graph nodes or connection relationships, a risk propagation evolution graph is constructed. In this way, this application can unify the temporal change process, multimodal feature information, and risk propagation relationships into a single graph structure for modeling, thus providing a more intuitive description of the evolution process of the risk propagation structure. Furthermore, since the risk propagation evolution graph simultaneously includes changes in node states and changes in connection relationships, it can more accurately identify complex risk propagation behaviors such as changes in risk diffusion paths, abnormal changes in key nodes, and changes in risk propagation direction, improving the accuracy of subsequent structural mutation determination.

[0084] In this embodiment, the structural mutation determination condition includes: the degree of distribution migration is greater than a baseline migration range determined based on historical distribution migration levels; the baseline migration range is determined based on the average and fluctuation range of the target merchant's distribution migration levels within historical time windows. Specifically, the system first obtains the distribution migration levels of the target merchant within multiple historical time windows and determines the migration level under normal operating conditions based on these historical distribution migration levels. For example, if the distribution migration level of a target merchant has been in a low fluctuation range over the past 30 time windows, this range can be used as the baseline migration range for the merchant. When the distribution migration level calculated between current adjacent time windows is greater than the baseline migration range, it indicates that the current change in the multimodal feature distribution has significantly exceeded the merchant's historical normal fluctuation level, and it can be determined that an abnormal mutation has occurred in the risk propagation structure, thereby triggering credit assessment processing.

[0085] In another embodiment, the structural mutation determination criteria include:

[0086] The distribution migration level shows a continuous upward trend across multiple consecutive time windows, and the distribution migration level at the end of the time window exceeds the preset abnormal migration threshold. If the distribution migration level of a target merchant continues to increase over three or more consecutive time windows, and the distribution migration level at the end of the time window exceeds the preset abnormal migration threshold, it indicates that the risk propagation structure of the target merchant is not an occasional fluctuation, but rather exhibits a continuous diffusion or deterioration trend. For example, if the complaint rate of a target merchant's associated merchants gradually increases, the number of abnormal performance nodes gradually increases, and negative text features continuously strengthen, then even if the triggering criteria are not met in the first few time windows, the system can trigger credit assessment processing when the continuous upward trend reaches the end abnormal threshold.

[0087] In another embodiment, the structural mutation determination criteria include:

[0088] The distribution migration between adjacent time windows exceeds a first threshold, and the change in node connections in the risk propagation evolution graph exceeds a second threshold; the change in node connections is determined based on at least one of the following: the number of new connections, the number of lost connections, or the change in connection weights. When the distribution migration between adjacent time windows exceeds the first threshold, and the change in node connections in the risk propagation evolution graph exceeds the second threshold, the system determines that the risk propagation structure of the target merchant has undergone a sudden change. For example, if the target merchant adds multiple high-risk associated merchants in a later time window, or a large number of existing stable transaction relationships disappear, and the multimodal feature distribution also shows a significant migration, it indicates that the risk change is reflected not only in the feature distribution but also in the relationship structure. By using dual-condition judgment, the risk of false triggering caused by fluctuations in a single indicator can be reduced. Specifically, the number of new connections can represent the number of new associations established by the target merchant in the current time window; the number of lost connections can represent the number of associations broken compared to the previous time window; and the change in connection weights can represent the degree of increase or decrease in the strength of existing associations. For example, when a target merchant establishes new high-weight connections with multiple merchants with high abnormal refund rates, or when connections with multiple long-term stable suppliers disappear, the system can consider the degree of change in node connection relationships to have increased. Combining this degree of change with the degree of distribution migration allows for more accurate identification of risk propagation path reconstruction.

[0089] In another embodiment, the structural mutation determination criteria include:

[0090] The direction of change corresponding to the distribution migration degree is inconsistent with the direction of change corresponding to the risk propagation indicator, and is used to identify hidden risk propagation structural changes caused by changes in the multimodal feature distribution. The risk propagation indicator may only reflect changes in the overall risk value, while the distribution migration degree reflects the overall migration of the multimodal feature distribution and structural relationships. If the risk propagation indicator does not increase significantly, or even remains stable, but the distribution migration degree continues to increase, it indicates that the risk is not directly manifested as an increase in numerical risk, but is hidden in the changes in the multimodal feature structure. For example, if the transaction amount of the target merchant remains stable, but its associated merchants migrate from a low-risk group to a group of merchants with high complaints, high refunds, or negative public opinion, the system can identify this as a hidden risk propagation structural change and trigger credit assessment processing.

[0091] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A triggering method for merchant credit assessment based on multimodal data fusion, characterized in that, include: Acquire multimodal data of the target merchants and extract multimodal features; Based on the aforementioned multimodal features, a merchant relationship network is constructed, and the risk propagation index of the target merchants is calculated. The risk propagation indicators are analyzed for changes based on continuous time windows to generate a risk propagation change sequence, and a risk propagation evolution diagram is constructed based on the risk propagation change sequence. Based on the distribution differences of multimodal characteristics at different time stages of the risk propagation evolution diagram, the degree of distribution migration is calculated using the optimal transmission distance or energy distance; When the degree of distribution migration meets the preset structural mutation judgment condition, the credit assessment process for the target merchant is triggered.

2. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 1, characterized in that, Based on the distribution differences of multimodal characteristics at different time stages of the risk propagation evolution map, the degree of distribution migration is calculated using the optimal transmission distance or energy distance, including: Based on the multimodal features associated with the nodes or edges corresponding to each time window in the risk propagation evolution diagram, construct the multimodal feature distribution corresponding to each time window; The difference measure is calculated using the optimal transmission distance or energy distance for the multimodal feature distribution corresponding to adjacent time windows to obtain the distribution difference results between adjacent time windows; Based on the distribution difference results, the degree of distribution migration of the risk propagation evolution map between adjacent time stages is determined; the degree of distribution migration is used to characterize the degree of change in the risk propagation structure caused by the changes in the multimodal features between adjacent time windows.

3. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 2, characterized in that, The difference metric for the multimodal feature distributions corresponding to adjacent time windows is calculated using the optimal transmission distance or energy distance, including: The multimodal feature distribution corresponding to the first time window is determined as the source distribution, and the multimodal feature distribution corresponding to the second time window is determined as the target distribution. Based on the feature matching relationship between the source distribution and the target distribution, calculate the optimal transmission distance required for the source distribution to migrate to the target distribution; Alternatively, the energy distance between the source distribution and the target distribution can be calculated based on the distance between samples in the source distribution and the target distribution; The optimal transmission distance or the energy distance is determined as a measure of the difference between adjacent time windows.

4. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 3, characterized in that, The optimal transmission distance is calculated as follows: The multimodal feature distribution of the t-th time window is represented as: ; The multimodal feature distribution of the (t+1)th time window is represented as: ; in, and These represent the multimodal feature vectors of nodes or edges within the corresponding time window; The optimal transmission distance is calculated based on the following formula: in, This represents the multimodal feature vector from the t-th time window. Multimodal feature vectors in the (t+1)th time window Transmission weight; This represents the set of transfer matrices that satisfy the marginal distribution constraints. It represents the distance metric between two multimodal feature vectors; The optimal transmission distance As a measure of the difference between adjacent time windows.

5. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 3, characterized in that, The energy distance is calculated as follows: The multimodal feature distribution of the t-th time window is represented as: ; The multimodal feature distribution of the (t+1)th time window is represented as: ; The energy distance is calculated based on the following formula: in, They represent the distributions respectively. The multimodal feature vectors obtained by independent sampling are mutually independent and follow the same distribution; They represent the distributions respectively. The multimodal feature vectors obtained by independent sampling are mutually independent and follow the same distribution; This represents the expectation operation; The distance norm between vectors; The energy distance As a measure of the difference between adjacent time windows.

6. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 2, characterized in that, The construction of the multimodal feature distribution corresponding to each time window includes: Extract the multimodal features corresponding to the target merchant and its associated merchants within each time window; The multimodal features are weighted according to the node positions or connections of the target merchant and its associated merchants in the risk propagation evolution diagram. The multimodal feature distribution corresponding to the time window is generated based on the weighted multimodal features.

7. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 2, characterized in that, When the distribution migration degree includes distribution difference results corresponding to multiple adjacent time windows, the multiple distribution difference results are aggregated to obtain the overall distribution migration degree of the target merchant within the preset evaluation period, and the overall distribution migration degree is updated to the new distribution migration degree.

8. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 1, characterized in that, The analysis of changes in the risk transmission indicators based on continuous time windows includes: The risk propagation index is divided into multiple continuous time windows according to a preset time interval, and the corresponding risk propagation index value within each time window is extracted. The risk transmission change sequence is generated based on the differences or trends in the risk transmission indicator values ​​between adjacent time windows.

9. A triggering method for merchant credit assessment based on multimodal data fusion according to claim 1 or 2, characterized in that, The risk propagation change sequence is used to characterize the evolution of the risk propagation status of the target merchant over time.

10. The triggering method for merchant credit assessment based on multimodal data fusion according to claim 1, characterized in that, The step of constructing a risk propagation evolution diagram based on the risk propagation change sequence includes: The risk propagation status corresponding to each time window is determined based on the risk propagation change sequence. The risk propagation status corresponding to each time window is used as a graph node; Based on the changing relationship of the risk propagation status between adjacent time windows, a connection relationship is established between adjacent graph nodes. This connection relationship is used to characterize the transition relationship of the risk propagation status between different time stages. By associating the multimodal feature information within the corresponding time window with the corresponding graph nodes or connections, a risk propagation evolution graph is obtained.

Images