Method for generating interpretable rural credit assessment model

Abstract

The invention relates to the technical field of financial risk control, and discloses an interpretable rural credit evaluation model generation method, which comprises the following steps: constructing a credit relation graph containing entity nodes and relation vectors, calibrating basic credit potential energy values of the nodes, and calculating the credit potential energy values of the nodes according to a credit conduction rule quantified with social relation influence; according to the method, the calculation of isolated individual data is converted into the structured analysis of the credit source of the evaluation object in the social relation network, so that the distortion of the evaluation result can be avoided when the individual data is sparse, and the evaluation accuracy is improved. The objective of the invention is to improve the transparency and traceability of an evaluation process and result, thereby improving the overall reliability and interaction friendliness of a credit system.

Description

Technical Field

[0001] This invention relates to a method for generating an interpretable rural credit assessment model, belonging to the field of financial risk control technology. Background Technology

[0002] In current financial lending practices, a mainstream credit assessment method is based on the applicant's individual data, using statistical or machine learning models to quantify their default risk. This method has formed a mature application system in scenarios where individual economic data is relatively complete, such as cities. Its basic technical logic is to construct a scoring model that can characterize the applicant's personal repayment ability and willingness by collecting financial and behavioral data that can be directly quantified, such as historical income, debt level and consumption records.

[0003] However, when this assessment method is applied to the specific environment of rural society, its core assumptions, which are based on individual data, become structurally mismatched with the economic reality of rural society, which is based on social relationships. In rural areas, an individual's credit status does not exist in isolation, but is deeply embedded in a diverse social network of relationships, such as family, neighborhood, and production cooperation. Their willingness and ability to repay are largely influenced by the reputation and resources of other members in this network. This influence usually manifests as informal but actually binding support or associated risk. Due to the limitations of its data collection dimensions and model construction concepts, the existing assessment method cannot effectively measure this relational capital based on social relationships and its transmission effect among individuals, thus creating a relational blind spot.

[0004] Specifically, existing technologies suffer from the following shortcomings: 1. Existing assessment methods neglect the core dimension of social networks, which plays a crucial role in rural credit, resulting in an incomplete foundation for the assessment model; 2. Given the sparse nature of individual credit data, the primary basis for assessment, the lack of a relationship dimension further amplifies the uncertainty of the assessment results, making it difficult to accurately reflect the true credit level; 3. Credit scores generated by complex algorithms are difficult to interpret, hindering effective communication between financial institutions and users, and preventing users from obtaining clear guidance on improving their credit from the assessment results. Therefore, how to construct a structured and quantifiable assessment model that can accurately reflect the credit level of entities by structuring the objectively existing social networks in rural society, and generating a transparent assessment process with interpretable results, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides an interpretable rural credit assessment model generation method, the main purpose of which is to solve the problems of distorted assessment results and opaque process caused by the neglect of social relationship networks in existing credit assessment methods.

[0006] To achieve the above objectives, this invention provides a method for generating an interpretable rural credit assessment model, comprising the following steps:

[0007] Step a: Obtain and construct a credit relationship graph containing multiple nodes, where nodes represent entities in rural society. Use a relationship vector with a predetermined data structure to record the social relationships between nodes. The data structure of the relationship vector includes fields for defining the relationship type and fields for defining the relationship strength.

[0008] Step b: For each entity node in the credit relationship graph, based on the individual basic credit information of the entity node and according to a calibration rule stored in the memory, determine an initial basic credit potential value for it.

[0009] Step c: Set and apply credit transmission rules. The credit transmission rules are based on the relationship type and relationship strength in the relationship vector, and are specifically quantified in the credit relationship graph as the weight of credit potential transmitted between different nodes.

[0010] Step d: Perform iterative calculation. In each iteration, based on the credit transmission rule and combined with the credit potential values ​​of one or more neighboring nodes directly connected to a target node, update the final credit potential value of the target node. The final credit potential value obtained after at least one round of iterative calculation is used as the credit assessment result of the entity node.

[0011] Preferably, the data structure of the relation vector also includes fields for defining the nature of social relations; the field value of the relation type, whose value set consists of core family member relations, cooperative member relations, and guarantee relations; the field value of the relation strength is selected from a predetermined set of discrete levels; and the field value of the relation nature is distinguished as positive or negative.

[0012] Preferably, the individual basic credit information includes information on whether the entity node has a bad credit record and whether it has a stable source of income; the calibration rule stored in the memory is a lookup rule, which sets a benchmark score for the basic credit potential value of the entity node, and adjusts the benchmark score up or down according to the individual basic credit information.

[0013] Preferably, in the iterative calculation, the credit transmission rule also includes an adaptive damping mechanism to dynamically adjust the transmission weights. Specifically, the adaptive damping mechanism includes: before the start of each iteration, comparing the target node's basic credit potential value with a pressure state threshold stored in memory; when the target node's basic credit potential value is lower than the pressure state threshold, generating a damping coefficient using a nonlinear damping function with S-shaped characteristics stored in memory, and using this damping coefficient to reduce the effective transmission weight when transmitting credit potential energy from neighboring nodes to the target node. This effective transmission weight w eff Determined by the following formula: w eff =w base ·f(P base ), where w base It is the static transmission weight, P, defined in the credit transmission rule. base It is the basic credit potential value of the target node, f(P) base ) is a nonlinear damping function, the value of which changes with P base It decreases as the amount decreases.

[0014] Preferably, after determining the basic credit potential value and before performing iterative calculation, the method further includes: actively sending an information incentive instruction containing a response time limit to the mobile communication terminal bound to the entity node; measuring and obtaining the response delay of the mobile communication terminal to the information incentive instruction; generating a node activity index characterizing the operational activity of the entity node based on the response delay and a mapping relationship stored in memory, wherein the shorter the response delay, the higher the node activity index; adjusting the basic credit potential value using the node activity index to generate an effective basic credit potential value, and using the effective basic credit potential value for iterative calculation.

[0015] Preferably, the method further includes: when an update trigger condition stored in memory is met, sending query instructions for external macroeconomic events in batches to entity nodes in a specific area of ​​the credit relationship graph via a short message channel; the update trigger condition includes at least one of the following: the data update time exceeds a predetermined duration and an early warning of an external macroeconomic event is detected; receiving and parsing responses to the query instructions; determining a temporary credit potential adjustment factor based on the parsed responses and a mapping table stored in memory; and temporarily adjusting the basic credit potential value of the affected entity nodes using the credit potential adjustment factor before the start of the next round of iteration calculation.

[0016] Preferably, the recording of relationship vectors and the quantification of credit transmission weights further include the following steps: the combination of the relationship vector and the corresponding credit transmission weight is used as a structured rule template and recorded on a distributed shared ledger based on consortium blockchain technology, which is shared among predetermined member institutions; when it is necessary to determine the relationship vector or transmission weight for a new social relationship, the system first searches the distributed shared ledger to see if there is a stored rule template that matches the new social relationship; if a matching rule template is found, the matching rule template is reused for the recording and quantification of the new social relationship.

[0017] Preferably, the method further includes: capturing and storing intermediate state data of the final credit potential value of each entity node after each iteration during the iterative calculation process; obtaining the geographical location information of each entity node; performing spatial aggregation processing on the intermediate state data based on the geographical location information to generate aggregated statistical values; and generating a heat map data containing risk level zoning to present the regional credit risk status based on the aggregated statistical values.

[0018] Preferably, the iterative calculation continues until the change in the final credit potential value of all or a predetermined proportion of entity nodes in the credit relationship graph is less than a convergence threshold stored in memory, at which point the process ends.

[0019] Preferably, it further includes: generating a credit composition dataset for visualization, wherein the credit composition dataset is used to characterize the contribution of the final credit potential value from the basic credit potential value, and the contribution of the credit transmission influence from each of its connected neighboring nodes.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. By constructing a credit relationship graph encompassing social relationship information between entities, and combining it with the final generated specific dataset representing credit composition, the credit assessment process is transformed from a closed calculation of isolated individual economic data to a structured analysis of the credit sources of the assessed object. As a result, the assessment results not only reflect the final credit level, but also clearly reveal that this level is composed of the individual's own basic qualities and the positive or negative impacts it receives in the social relationship network. This assessment method avoids the distortion of assessment results due to insufficient information when individual credit data is sparse, and at the same time transforms a credit review into an interactive process with clear guidance and educational significance for users, improving the transparency of the entire credit system.

[0022] 2. By introducing an adaptive damping mechanism based on the node's own basic credit potential state in the iterative calculation of credit transmission, the positive transmission effect of social relationship networks is no longer a static, idealized constant, but can be adjusted nonlinearly and prudently according to the node's own financial health. This mechanism enables the evaluation model to more realistically reflect the dynamics of social relationship support under economic pressure, avoiding overly optimistic estimates of the ability of distressed entity nodes to obtain external support, thus providing a technical information to quantify the stability of customer credit structure under stress.

[0023] 3. This invention synergizes a credit relationship graph, which serves as the basis for static credit assessment, with two different low-cost dynamic information detection mechanisms. Specifically, through proactive information incentives and response delay analysis, it can detect and quantify the implicit state of the assessed entity's daily operational activity. Meanwhile, through querying and analyzing responses to external macroeconomic events, it can instantly capture the impact of external environmental shocks on the entity. The combination of these two mechanisms allows the assessment model to transcend its reliance on historical data, gaining a dual dynamic perception capability of both the assessed entity's internal operational status and changes in the external environment. Furthermore, it transforms the risk identification model from post-hoc attribution to a more dynamic approach. This invention transforms the model into a proactive early warning system based on behavioral processes and environmental changes. Furthermore, by converting the definitions and quantifications of non-standardized social relationships by different operators into structured rule templates that can be stored on the blockchain and retrieved and reused, and by leveraging a distributed shared ledger for collaboration among predetermined institutions, it solves the bottlenecks in the efficiency of manual knowledge input and the consistency of evaluation standards faced in the large-scale application of evaluation models. This approach not only reduces the reliance on the professional experience of frontline operators and the burden of repetitive work, but also ensures the fairness and stability of the model in cross-regional and cross-time applications by building a continuously accumulating and reliably shared experience knowledge base.

[0024] 4. The method of this invention also couples the intermediate state data generated during the iterative calculation process with the geographical location information of the entity nodes. Through spatial aggregation processing, heat map data for presenting the regional credit risk status is generated. This design reuses existing information in the assessment process and extends the function of the assessment method from the credit measurement of micro-individuals to the situational awareness of the distribution and aggregation of credit risk in a specific region without increasing any new data collection costs. This provides a decision-making basis for financial institutions to carry out regional credit policy regulation and early identification of systemic risks. Attached Figure Description

[0025] Figure 1 This is an overall flowchart of the method for generating an interpretable rural credit assessment model according to the present invention;

[0026] Figure 2This is a schematic diagram illustrating the convergence process of credit potential energy values ​​at nodes with different initial states during the iterative calculation of this invention.

[0027] Figure 3 This is a schematic diagram of the on-chain consensus procedure for rule templates based on consortium blockchains according to the present invention.

[0028] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0030] The present invention discloses an interpretable rural credit assessment model generation method, the overall technical process of which is constructed as follows: First, a credit relationship graph construction step is performed to structurally represent rural social entities and their complex social connections; second, a basic credit potential calibration step is performed to assign an initial score representing the basic qualifications of each entity node in the graph; third, based on a set of preset credit transmission rules, the transmission weight of credit influence between different social relationships is quantified; finally, through a cyclically executed iterative calculation step, the transmission and evolution of credit potential in the relationship network are simulated until the credit potential values ​​of each node converge to a stable state, and the final stable credit potential value is used as the credit assessment result for the corresponding entity node. In specific rural financial application scenarios, a common objective challenge is that the structured financial data of individuals is relatively sparse, while the social relationship capital that has a significant impact on their credit status is unstructured. The social network is difficult to measure effectively by traditional risk control models due to its complex characteristics. To address this challenge, the present invention is configured to acquire and construct a credit relationship graph containing multiple nodes. Each node represents an entity in a rural society, and the connections between nodes are recorded through a relationship vector with a predetermined data structure. This relationship vector data structure includes fields for defining relationship types, relationship strengths, and the nature of social relationships. Specifically, the relationship type field value is limited to a set of values ​​consisting of core family member relationships, cooperative member relationships, and guarantee relationships. The relationship strength field value is selected from a predetermined set of discrete levels. The relationship nature field value is distinguished as positive or negative to indicate the basic role of the relationship in credit transmission. Through this data organization method, the system can translate the originally abstract social network into a credit topology data structure that can be accurately parsed and calculated.

[0031] After the topological structure of the credit relationship graph is established, in order to ensure that each evaluated entity obtains an initial credit benchmark unaffected by the relationship network, the system faces the problem of how to objectively calibrate based on limited individual data. Therefore, the system employs a lookup-based rule stored in memory to determine the initial basic credit potential value P for each entity node. base The execution logic of this rule is as follows: First, a unified baseline score is set for all entity nodes. Then, the system retrieves the individual basic credit data of that node. This data includes at least information on whether there are any adverse credit records and whether there is a stable source of income. The baseline score is then adjusted upwards or downwards according to a lookup-based rule. For example, when an entity node is found to have an adverse credit record, its basic credit potential value is lowered by a specific score; conversely, when it is found to have a verified stable source of income, its basic credit potential value is raised by a specific score. In this way, the basic credit potential value P of each node can be ensured. base All data is generated based on objective individual information, providing an initial anchor point representing an individual's creditworthiness for the entire evaluation model. To transform objectively existing credit support or associated risks within social networks into calculable quantitative indicators, the system sets and applies credit transmission rules. The core of these rules lies in quantifying a basic transmission weight w for the transmission process of credit potential energy between different nodes, based on the relationship type and strength recorded in the relationship vector. base For example, the system can pre-configure a weight mapping table in which combinations of core family member relationships and first-level strong ties are assigned a relatively high basic transmission weight w. base The combination of cooperative member relationships and third-level ordinary associations is assigned a relatively low w. base This rule-based quantification mechanism makes the transmission effect of credit in the network deterministic and traceable, providing a clear physical basis for subsequent iterative calculations. Considering that in the real economic environment, the financial health of a node directly affects its effectiveness in absorbing external credit support, a static transmission weight may overestimate the creditworthiness of a node in financial distress. To simulate this nonlinear dynamic, this invention further introduces an adaptive damping mechanism into the credit transmission rule. The specific procedure of this mechanism is that before the start of each round of iterative calculation, the basic credit potential value P of the target node is... base Compared with a pressure state threshold stored in memory, when P base When the value falls below this threshold, the system determines that the node is under financial stress and activates a nonlinear damping function f(P) with S-shaped characteristics. base The function is designed to change with the input variable P. baseAs the value decreases monotonically, its output value serves as a damping coefficient between 0 and 1, used to adjust the effective transmission weight w. eff The effective transmission weight is determined by the following formula: w eff =w base ·f(P base By introducing this damping mechanism, the model can prudently adjust the positive transmission effect of social networks, thereby avoiding an overly optimistic estimate of the ability of distressed entities to obtain external support.

[0032] After completing the credit graph construction, initial value calibration, and rule setting, the system performs iterative calculations. In each iteration, for any target node within the graph, its updated final credit potential value is calculated by weighting its own credit potential value and the credit potential values ​​of all its directly connected neighboring nodes, combined with credit transmission rules. This iterative process continues until the change in the final credit potential value of all or a predetermined proportion of entity nodes in the credit relationship graph is less than a convergence threshold stored in memory. At this point, the final credit potential value of each node after stabilization is... As a result of the credit assessment of this entity node, and further, to achieve complete transparency and traceability of the assessment process and results, the system generates a credit composition dataset along with the final assessment result. This dataset represents the contribution of the final credit potential value from the basic credit potential value, as well as the contribution from the credit transmission influence of each connected neighboring node. To enable the model to perceive the dynamic changes in the current operating status of the assessed object, this invention also embeds a node activity detection mechanism after determining the basic credit potential value and before performing iterative calculations. This mechanism aims to obtain low-cost... To obtain real-time information reflecting the current operational activity of an entity, the procedure involves proactively sending an information incentive command, including a response time limit, to the mobile communication terminal bound to the entity node through the user interface of the business application. The system accurately measures and acquires the response latency of the mobile communication terminal to the information incentive command. To ensure reliable delivery of the command, the system is configured with a multi-channel interaction scheme. Push notifications from the mobile application are used as the primary interaction path, while SMS channels serve as a backup. When the primary path is blocked due to obstacles such as the user disabling notification permissions or the terminal lacking a data network connection, the system will seamlessly switch to the backup path. The system uses a path to ensure the completion of core tasks. Then, based on the measured response latency and a mapping relationship stored in memory, it generates a node activity index representing the operational activity of entity nodes. The shorter the response latency, the higher the node activity index. Finally, the system uses this node activity index to adjust the basic credit potential value to generate an effective basic credit potential value, which is then used for subsequent iterative calculations. Furthermore, to address the systemic impact of external macroeconomic events on the credit status of entities in specific regions, this invention also provides a dynamic temporary adjustment mechanism for credit potential.When update trigger conditions stored in memory are met, such as when the data update cycle exceeds a predetermined duration or an early warning of an external macroeconomic event is detected, the system will send batch query commands regarding the impact of the event to entity nodes in a specific geographical area of ​​the credit relationship graph via SMS channel. After receiving and parsing the user's response to the query command, the system determines a temporary credit potential adjustment factor based on the mapping table stored in memory. Before the start of the next iteration calculation, the system uses this credit potential adjustment factor to temporarily adjust the basic credit potential value of the corresponding entity nodes in the affected area. This enables the model to obtain an immediate response capability to changes in the external environment.

[0033] To address the bottleneck issues of efficiency in manual knowledge input and consistency of evaluation standards in the large-scale application of evaluation models, this invention introduces a rule template management mechanism based on consortium blockchain technology. Under this mechanism, a combination of a relationship vector and its corresponding credit transmission weight serves as a structured rule template, recorded on a distributed shared ledger based on consortium blockchain technology. This ledger is shared among predetermined member institutions such as banks and insurance companies. When it is necessary to determine a relationship vector or transmission weight for a new social relationship, the system first searches the distributed shared ledger for a matching, already stored rule template. If found, the matching rule template is reused for recording and quantifying the new social relationship. This approach, by constructing a continuously accumulating and reliably shared experience knowledge base, ensures the model's reliability in cross-regional and cross-temporal applications. Fairness and stability; finally, to extend the model's micro-level individual assessment capabilities to macro-level situational awareness of the overall regional credit risk distribution, this invention also designs a method for generating risk visualization data. During iterative calculations, this method captures and stores intermediate state data of the final credit potential value of each entity node after each iteration. Simultaneously, the system acquires the geographical location information of each entity node and, based on this information, performs spatial aggregation processing on the intermediate state data within a specific spatial range to generate aggregated statistical values. For example, it calculates the average credit potential of nodes within the region or the proportion of risky nodes. Based on these aggregated statistical values, the system ultimately generates a heatmap data that presents the regional credit risk status and includes risk level zoning, providing decision support for regional credit policy regulation and systemic risk identification.

[0034] Example 1: In a financial service scenario providing large-scale seasonal credit to a local agricultural cooperative, a credit relationship graph consisting of hundreds of farmer entities has been constructed, and the initial basic credit potential value has been calibrated. When assessing a credit line granted to the cooperative for the purchase of spring farming equipment, a sudden regional external macroeconomic event warning was detected by the system. This event was a hailstorm, the impact of which covered the townships where some members of the cooperative resided. This sudden situation rendered the initial assessment results, which relied solely on historical data, immediately ineffective in terms of timeliness. The financial institution faced a decision-making dilemma: it could not accurately quantify the impact of this shock on the cooperative's overall repayment ability in a short period. If the credit line was interrupted, it might miss the opportunity to support an economic community that was originally healthy but encountered temporary difficulties; if the original credit line was maintained, it might overlook the worsening local risks. This could lead to future asset losses. In this situation, the dynamic credit potential temporary adjustment mechanism of the present invention is activated by the updated triggering conditions. The system sends query instructions for external macroeconomic events in batches to entity nodes in the credit relationship graph whose geographical location information is within the affected townships via short message channel. Based on the received and parsed replies, a temporary credit potential adjustment factor is determined for those entity nodes that confirm that their production and operation have been impacted through a pre-set mapping table, and their basic credit potential value is temporarily reduced using this adjustment factor. This operation injects real-time dynamic data about the scope and degree of the impact of this external shock into the model. This set of real-time updated and reduced effective basic credit potential values ​​constitutes the key prerequisite for the operation of the subsequent mechanism. That is, when the graph carrying this dynamic information enters a new round of iterative calculation, the adaptive damping mechanism begins to play its role.

[0035] Before the start of the new round of iterative calculations, for each basic credit potential value P affected by the disaster... base For entity nodes that have been reduced to below the pressure state threshold, the system passes through a nonlinear damping function f(P) with sigmoid characteristics. base A damping coefficient less than 1 was generated, and according to the formula w eff =w base ·f(P base Accordingly, the effective transmission weight w when transmitting positive credit potential energy from its neighboring nodes to it was lowered. effThis mechanism resolves the inherent rigidity and fragility inherent in traditional assessment models. The model avoids both global and unstable score collapse due to sudden negative information from a few nodes, and maintaining a fragile high score by ignoring local risks. Instead, through the synergy of dynamic information detection and adaptive damping, it selectively suppresses risk transmission paths within the credit relationship graph, thus ensuring that the assessment results balance sensitivity to new risks with stability of the overall structure. After this round of iterative calculations, the converged final credit potential value output by the system exhibits a structurally non-uniform change compared to before the disaster. Simultaneously, the system generates a heatmap to represent the regional credit risk status. On this heatmap, the geographical area of ​​the affected townships is marked with a specific color representing risk clustering, while unaffected areas retain their original colors. Furthermore, a credit composition dataset for the entire cooperative is also generated. The decline in the overall credit potential of the cooperative was primarily attributed to entities within specific color-coded areas in the heatmap data, while nodes in the vast unaffected areas continued to contribute positively. This series of outputs allowed financial institutions to restructure their decision-making process. The question shifted from whether to approve loans to how to design a more flexible credit scheme based on the identified risk distribution. Ultimately, based on this assessment, financial institutions adjusted their original credit granting schemes, designing specific credit terms with deferred repayment options for affected farmers marked in the heatmap data while maintaining overall support. This application example demonstrates a risk assessment method that combines static topological analysis with dynamic, targeted information detection. The system's assessment logic shifted from fitting isolated historical data to a structured analysis of the stress response and inherent stability of the relationship network when encountering external disturbances.

[0036] Example 2: To objectively verify the role of the adaptive damping mechanism of the present invention in suppressing credit risk overestimation, this example designed and executed a numerical simulation experiment. The purpose of the experiment was to quantitatively compare the differences in the final credit potential value assessment results of entity nodes under financial stress under the two conditions of introducing and not introducing the mechanism. The experiment was conducted on a server configured with a numerical computing environment. In this environment, a synthetic credit relationship graph containing 1,000 entity nodes and about 5,000 social relationships was pre-generated. The distribution of node connectivity, relationship type and relationship strength in the graph was generated with reference to statistical data collected from real rural communities to simulate a representative application scenario.

[0037] In the experimental design, the control group and the experimental group used the exact same synthetic credit relationship graph and initial parameters. The decision-making logic for setting the stress state threshold was to balance the assessment model's sensitivity to risk and its stability. If the threshold was set too high, normal credit fluctuations might be judged as stress; if it was set too low, distressed entities might not be effectively identified. Therefore, the threshold was set at a value equal to the initial basic credit potential value P of all nodes. base The statistical distribution is correlated with the values ​​of all nodes P, and these values ​​are determined to be the values ​​of all nodes P. base The position of the mean to the next standard deviation below, in the synthetic dataset of this experiment, P base The mean was 512 and the standard deviation was 65, therefore the pressure state threshold was determined to be 447. The control group used the iterative calculation method described in the specific implementation but without the adaptive damping mechanism, while the experimental group used a calculation method that included a complete adaptive damping mechanism, which is defined by formula w. eff =w base ·f(P base As defined by ).

[0038] After the experiment was started, 100 entity nodes were randomly selected from the two sets of graphs as stress test objects, and their initial basic credit potential values ​​P were determined. base Using a step size of 10, the score was gradually reduced from 440 to 300 to simulate different levels of financial stress. Subsequently, iterative calculations were performed on both sets of graphs until the final credit potential value converged. During data acquisition, it was observed that for the same stress test node, as its P... base As the value decreases, the deviation between the final credit potential value calculated by the experimental group and the result of the control group increases non-linearly, at P base When the value is close to the pressure state threshold, the difference between the two groups of results is small, but as P approaches the threshold, the difference becomes more pronounced. base Further reductions led to a final result for the experimental group that correlated with its own P. base The results of the control group converged, while the results of the control group remained at a relatively high level, still strongly supported by the credit transmission of its neighboring nodes. Table 1 shows the results of 8 nodes randomly selected from 100 stress test subjects at different P values. base Comparison of final credit potential values ​​at different levels.

[0039] Table 1: Comparison of final credit potential energy of stress state nodes under the two models.

[0040]

[0041] Referring to Table 1, the underlying mechanism for this data trend is that, in the experimental group, when the target node's P... baseWhen the pressure state threshold is below 447, the adaptive damping mechanism is activated, and the nonlinear damping function f(P) exhibits S-shaped characteristics. base Output a damping coefficient less than 1, and this coefficient varies with P. base The decrease in weight w leads to a decrease in the credit potential energy transmitted from neighboring nodes. eff The effective transmission of the signal is weakened, so the final credit potential value of the node is closer to its own basic credit level; while in the control group, due to the lack of this damping mechanism, the transmission weight is always static w. base This causes the positive effects of neighboring nodes to be transmitted to the target node indiscriminately, resulting in an overestimation of its trustworthiness under stress.

[0042] Example 3: This example combines Figures 1 to 3 This section describes a method for generating an interpretable rural credit assessment model, such as... Figure 1 As shown, this process uses two core data sources as initial inputs: individual basic credit data containing information such as adverse credit records and stable income sources, and social relationship information containing the types and strengths of relationships such as family, cooperatives, and guarantees. Based on this, the first step is to construct a credit relationship graph to structurally represent rural social entities and their complex social connections. Subsequently, in the second step, an initial basic credit potential benchmark P is determined for each node in the graph based on the individual basic credit data. base Before entering the core computing phase, the system executes two key modules in parallel: a node activity detection mechanism and an adaptive damping mechanism. The node activity detection mechanism sends information stimulus commands to the mobile terminal, measures and acquires the response latency, and generates a node activity index based on the mapping relationship that shorter latency corresponds to a higher index. Then, in step 2.5, this index is used to adjust the basic credit potential energy to generate effective basic credit potential energy. The adaptive damping mechanism uses whether the target node's basic credit potential energy is lower than a pressure state threshold as a judgment condition. If it is lower than this threshold, nonlinearity is activated. The damping function dynamically adjusts the effective weight of credit transmission based on the sigmoid function. The outputs of these two mechanisms work together in step three, which is to perform iterative calculations to simulate the transmission and evolution of credit potential energy in the relational network until the change in potential energy value of all or a predetermined proportion of nodes is less than the convergence threshold. The calculation then converges and finally outputs the evaluation results and data products, including the final credit potential energy value evaluation results of entity nodes, a credit composition dataset to reveal the composition of credit sources and achieve transparent and traceable processes, and a regional credit risk heat map to present the regional credit risk clustering status and distribution.

[0043] like Figure 2As shown in the figure, this graph graphically illustrates the convergence process of the iterative calculation. The horizontal axis represents the number of iterations, and the vertical axis represents the credit potential value. In the initial stage of the iterative calculation, the credit potential values ​​of each node change significantly. As the number of iterations increases, the high-credit nodes, represented by the solid line, show a slight and gradual decrease in their credit potential value, starting from a relatively high initial value. The medium-credit nodes, represented by the long dashed line, show a slight and gradual increase in their credit potential value. The low-credit nodes, represented by the short dashed line, show a more obvious upward trend in their credit potential value, starting from a relatively low initial value. Finally, after about 10 iterations, the rate of change of all three curves decreases significantly and gradually approaches a horizontal line.

[0044] like Figure 3 As shown, the initial state of the process is "pending submission." At this point, the operator has defined a new rule template but has not yet submitted it to the consortium blockchain. After the consensus operation is performed on the blockchain, the proposal enters the "proposal awaiting consensus" state, meaning the proposal is placed in the consensus area of ​​the shared ledger, waiting for consortium members to verify it. At this time, consortium members will obtain the proposal and the corresponding verification dataset according to the verification activity guidelines, and conduct independent analysis locally to verify its validity. The verification result leads to two paths: if the preset number of members pass the verification, the proposal obtains consensus and is stored, the template is officially written into the main chain, and the process ends; otherwise, if the preset number of members pass the verification, the proposal is rejected, and the proposal will be sent back for modification, returning to the initial "pending submission" state, waiting for the operator to modify it and re-initiate the process. This closed-loop process ensures that any rule template included in the shared knowledge base has undergone multi-party verification and consensus, thereby guaranteeing the fairness and stability of the model in cross-institutional and large-scale applications.

[0045] Example 4: In a specific deployment scenario, the credit assessment model of this invention needs to be applied to multiple rural areas with different economic characteristics and user behavior patterns. A core engineering problem here is how to ensure that the internal parameters of the key nonlinear components in the model—namely, the adaptive damping mechanism and the node activity index generation logic—can be systematically and reproducibly calibrated to adapt to the data distribution characteristics of different regions, thereby avoiding performance degradation due to the use of fixed parameters. This requirement for deterministic calibration of the model is a prerequisite for maintaining consistency in assessment standards during large-scale cross-regional applications. To address this problem, this invention provides a set of offline calibration and parameter determination procedures for key model components. In a specific implementation of this embodiment, the nonlinear damping function f(P) with S-shaped function characteristics... base Its mathematical form employs a generalized logical stith function, expressed as follows: In this function, there are two core parameters that need to be calibrated: the gain coefficient k, which determines the steepness of the function curve, and the pressure state threshold P0, which serves as the center of symmetry of the function. The calibration procedure for these two parameters is as follows: The system first acquires a localized dataset containing historical samples. Each sample in this dataset contains the basic data of the entity node before the credit granting occurs, and its final basic credit potential value P. base And the system identifies the true outcome label of whether a default occurred after the credit was granted. Then, aiming to optimize the model's discriminative ability, the system employs a gradient descent iterative search algorithm to find the optimal combination of k and P0, such that under this parameter combination, all samples pass through f(P... base The damping coefficient value calculated by the function can separate the default sample group from the normal performance sample group to the greatest extent. Through this optimization process driven by historical data, the S-shaped function in the model is an empirically verified function that can reflect the evolution characteristics of credit risk in a specific region.

[0046] Secondly, the mapping relationship used to convert the response latency of a mobile communication terminal into a node activity index is also defined by a mathematical model with explicit parameters; in a specific implementation of this embodiment, the node activity index I... activity With response delay t delay The relationship was determined to be an exponentially decaying function, with the form: Among them, I max λ is the preset maximum activity index, while λ is the decay coefficient that determines the decay rate. The calibration procedure for the decay coefficient λ is as follows: within one observation period, the system statistically analyzes the response delay distribution of a specific user group to information incentive commands and calculates a statistic of this distribution, namely the delay value t at the 80th percentile. 80 Then, an engineering goal is set: when the user's response latency reaches t... 80 At that time, its activity index decayed to 0.1. max Based on this objective, the equations are solved. By solving the inverse problem to obtain the value of the attenuation coefficient λ, the generation logic of the node activity index is anchored to the statistical analysis of real user behavior data. Through the aforementioned deterministic procedure for the adaptive damping function and the generation logic of the node activity index, this invention transforms two core, nonlinear components in the model from externally dependent links into links that can be internally calibrated and self-adapted through standardized processes. When this evaluation model needs to be deployed to a new region, the operating agency can use local historical credit data and user behavior data to execute an offline calibration procedure once to generate a set of parameters that match the local environment. Thus, without changing the main architecture of the model, it achieves its adaptability to different data ecosystems.

[0047] Example 5: In a credit consortium application scenario composed of multiple independent financial institutions, to establish a unified and fair evaluation basis among member institutions, this invention sets up a standardized pre-on-chain verification and consensus procedure before storing a new rule template on the distributed shared ledger. When any member institution proposes a rule template proposal containing a relationship vector and corresponding credit transmission weights for a new social relationship, the proposal and its anonymized historical credit dataset used to prove its validity will be submitted to a consensus area of ​​the consortium blockchain. A predetermined proportion of other member institutions in the consortium need to obtain the verification dataset and perform independent regression analysis locally to verify the correlation between the credit transmission weights and the actual repayment results. Only when a predetermined number of member institutions return confirmation information that the verification has passed will the rule template be formally written into the main chain of the distributed shared ledger by the consensus mechanism, becoming an evaluation standard that all members can follow.

[0048] Correspondingly, to ensure that the lookup rules for calibrating the initial basic credit potential values ​​of each node in the model have an objective statistical basis, this invention also provides a parameter calibration process based on shared data from the alliance. In the initial stage of model deployment in the credit alliance, each member institution collects its anonymized historical credit data to form a shared dataset with a wider coverage. Based on this shared dataset, the system uses a multivariate logistic regression analysis method, taking individual basic credit data such as the existence of bad credit records and the existence of stable income sources as independent variables, and whether or not a default occurs as the dependent variable, to fit the model. The weight coefficients of each independent variable obtained from this regression analysis, after standardization, are used to determine the specific scores for adjusting the benchmark score in the lookup rules. In this way, the setting of the initial parameters of the model is transformed from a potentially subjective judgment process into a traceable statistical modeling process driven by the historical data of all alliance members.

[0049] Example 6: In a specific system initial deployment and calibration scenario, in order to match the core parameters of the evaluation model of this invention with the business data and risk preferences of a specific financial institution, a standardized pre-calibration and online operation procedure needs to be executed. This procedure first targets the basic transmission weight w in the credit transmission rule. base Offline calibration involves extracting all pairs of entity nodes with social relationships from the institution's historical credit data, grouping them according to relationship type, and using the credit status of one party as the independent variable and the final repayment result of the other party as the dependent variable. A regression model is then established to determine the weight coefficient that maximizes prediction accuracy; this coefficient is then identified as the basic transmission weight w for the corresponding relationship type. base .

[0050] Building upon the aforementioned, this procedure further sets the termination condition for iterative calculations, namely the convergence threshold. The process involves the system performing multiple rounds of repeated calculations on the extracted sample dataset using convergence thresholds decreasing by several orders of magnitude. The average rate of change of the final credit potential value of all nodes under different thresholds is recorded. By plotting the relationship curve between the average rate of change and the convergence threshold, the inflection point where the curve's slope flattens is found, and the threshold corresponding to this inflection point is used as the formal convergence threshold for the system's operation. Simultaneously, to quantify the triggering conditions for external macroeconomic event query commands, the procedure requires the system to be configured to handle one or more external macroeconomic events. Multiple external economic data sources are monitored, and the triggering conditions are defined as a logical combination. Specifically, when the volatility of a monitored economic index exceeds a specific percentage within a preset time window, or when a disaster warning level reaches a specific standard, this logical judgment is true, thus initiating the subsequent query process. To address potential boundary and anomaly situations that may arise during continuous system operation, the procedure also defines corresponding processing logic. When a new entity node is added to the credit relationship graph, before the first complete iteration calculation is finished, the final credit potential value of the new node is determined to be equal to its own basic credit potential value P. base The system ensures that the initial values ​​are equal and do not transmit credit information externally to avoid interference with the network due to non-converged initial values. When processing responses to query commands, the system's built-in data cleaning and verification module is responsible for checking the format compliance and logical consistency of the responses. Any responses with format errors or logical contradictions will be marked as invalid data and discarded by the system. At the same time, an alarm requiring manual verification will be sent to the system administrator. This measure can prevent abnormal or malicious data from affecting the generation process of the temporary credit potential adjustment factor.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for generating an interpretable rural credit assessment model, characterized in that, Includes the following steps: Step a: Obtain and construct a credit relationship graph containing multiple nodes, where nodes represent entities in rural society. Use a relationship vector with a predetermined data structure to record the social relationships between nodes. The data structure of the relationship vector includes fields for defining the relationship type and fields for defining the relationship strength. Step b: For each entity node in the credit relationship graph, based on the individual basic credit information of the entity node and according to a calibration rule stored in the memory, determine an initial basic credit potential value for it. Step c: Set and apply credit transmission rules. The credit transmission rules are based on the relationship type and relationship strength in the relationship vector, and are specifically quantified in the credit relationship graph as the weight of credit potential transmitted between different nodes. Step d: Perform iterative calculation. In each iteration, based on the credit transmission rule and combined with the credit potential values ​​of one or more neighboring nodes directly connected to a target node, update the final credit potential value of the target node. The final credit potential value obtained after at least one round of iterative calculation will be used as the credit assessment result for the entity node.

2. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, The relation vector data structure also includes fields for defining the nature of social relationships; the field value for relation type, whose value set consists of core family member relationships, cooperative member relationships, and guarantee relationships; the field value for relation strength is selected from a predetermined set of discrete levels; and the field value for relation nature is distinguished as positive or negative.

3. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, Individual basic credit data includes information on whether the entity node has a bad credit record and whether it has a stable source of income; the calibration rule stored in the memory is a lookup rule, which sets a benchmark score for the basic credit potential value of the entity node, and adjusts the benchmark score up or down according to the individual basic credit data.

4. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, In the iterative calculation, the credit transmission rule also includes an adaptive damping mechanism to dynamically adjust the transmission weights; The adaptive damping mechanism specifically includes: before the start of each iteration, comparing the target node's basic credit potential value with a pressure state threshold stored in memory; when the target node's basic credit potential value is lower than the pressure state threshold, generating a damping coefficient using a nonlinear damping function with S-shaped characteristics stored in memory, and using this damping coefficient to reduce the effective transmission weight when transmitting credit potential energy from neighboring nodes to the target node. This effective transmission weight w eff Determined by the following formula: w eff =w base ·f(P base ), where w base It is the static transmission weight, P, defined in the credit transmission rule. base It is the basic credit potential value of the target node, f(P) base ) is a nonlinear damping function, the value of which changes with P base It decreases as the quantity decreases.

5. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, After determining the basic credit potential value and before performing iterative calculations, the process includes: actively sending an information incentive command containing a response time limit to the mobile communication terminal bound to the entity node; measuring and acquiring the response delay of the mobile communication terminal to the information incentive command; generating a node activity index characterizing the operational activity of the entity node based on the response delay and a mapping relationship stored in memory, with a higher node activity index for shorter response delays; adjusting the basic credit potential value using the node activity index to generate an effective basic credit potential value, and using the effective basic credit potential value for iterative calculations.

6. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, Also includes: When an update trigger condition stored in memory is met, query instructions for external macroeconomic events are sent in batches to entity nodes in a specific area of ​​the credit relationship graph via a short message channel. The update trigger condition includes at least one of the following: the data update time exceeds a predetermined duration and an early warning of an external macroeconomic event is detected. Responses to the query instructions are received and parsed. Based on the parsed responses and a mapping table stored in memory, a temporary credit potential adjustment factor is determined. Before the start of the next round of iteration calculation, the basic credit potential value of the affected entity nodes is temporarily adjusted using the credit potential adjustment factor.

7. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, The recording of relationship vectors and the quantification of credit transmission weights also include the following steps: the combination of the relationship vector and the corresponding credit transmission weight is used as a structured rule template and recorded on a distributed shared ledger based on consortium blockchain technology. The distributed shared ledger is shared among predetermined member institutions. When it is necessary to determine the relationship vector or transmission weight for a new social relationship, the system first searches the distributed shared ledger to see if there is a pre-existing rule template that matches the new social relationship. If a matching rule template is found, it is reused for the recording and quantification of the new social relationship.

8. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, Also includes: During the iterative calculation process, intermediate state data of the final credit potential value of each entity node after each iteration are captured and stored. Obtain the geographical location information of each entity node; Based on geographic location information, spatial aggregation processing is performed on intermediate state data to generate aggregated statistical values; Based on the aggregated statistical values, a heat map data containing risk level zoning is generated to present the regional credit risk status.

9. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, Iterative calculation continues until the change in the final credit potential value of all or a predetermined proportion of entity nodes in the credit relationship graph is less than a convergence threshold stored in memory, at which point the process ends.

10. The method for generating an interpretable rural credit assessment model according to claim 1, characterized in that, Also includes: Generate a credit composition dataset for visualization. The credit composition dataset is used to represent the contribution of the final credit potential value from the basic credit potential value, as well as the contribution of the credit transmission effect from each of its connected neighboring nodes.

Images