Simulated credit granting marginal effect calibration analysis system and method based on artificial intelligence

Abstract

The invention discloses a simulation credit granting marginal effect calibration analysis system and method based on artificial intelligence, and belongs to the technical field of artificial intelligence. Extracting typical behavior nodes of the target enterprise in the credit granting process, and constructing a credit granting node set; obtaining a historical behavior attribute data set of each typical behavior node, and constructing a unified behavior feature space; calculating a behavior difference value based on the node behavior vector, and constructing a path parameter and a historical credit atlas in combination with an event time interval; a marginal effect modeling function is used for calculating a marginal score value, and the influence of behavior changes on the credit granting marginal effect is accurately described; a simulated credit extension scene generator is further designed, multiple groups of simulated credit extension scenes are generated through dynamic disturbance, the global deviation ratio of simulated marginal score values and marginal score values is calculated, the stability and effectiveness of the model are verified, and deep modeling and scoring mechanism optimization of behavior changes in the credit extension process are achieved; and the dynamic response capability and risk control accuracy of the credit extension model are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a simulated credit marginal effect calibration and analysis system and method based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, traditional financial lending is gradually transforming towards intelligent and automated operations. In the field of corporate credit management, more and more financial institutions are attempting to introduce big data analytics and AI models to improve the accuracy and efficiency of credit decisions. Currently, mainstream corporate credit assessment systems are mostly based on corporate financial statements, credit information, and historical performance behavior for risk modeling, supplemented by scoring card models, logistic regression models, or decision tree algorithms to determine credit limits and boundaries. Building on this, some research further attempts to model behavioral nodes such as credit application, approval, loan disbursement, and repayment from a process perspective to identify potential risks and efficiency bottlenecks in the credit granting path. These methods emphasize structural modeling of behavioral sequences, initially forming a credit process graph analysis approach centered on "node-path," providing a data foundation for subsequent dynamic credit optimization.

[0003] However, existing technical methods still have significant limitations. First, traditional modeling methods typically employ static indicators or linear scoring approaches, making it difficult to accurately depict the dynamic marginal effects of corporate behavior in the credit granting process. Second, existing models generally lack a comprehensive consideration of the temporal correlations and attribute differences between behavioral nodes, resulting in weak responsiveness to node disturbances or behavioral anomalies. Furthermore, in terms of credit granting simulation and scenario testing, existing methods are mostly based on rule-based simulations, lacking effective model validation mechanisms and making it difficult to assess the model's stability and adaptability under real-world disturbance scenarios. Therefore, how to integrate behavioral characteristic changes, time interval differences, and path graph structures to establish a dynamic marginal scoring model has become a key challenge in current credit granting process modeling research. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based simulated credit marginal effect calibration and analysis system and method to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] This paper presents an AI-based method for calibrating and analyzing the marginal effect of simulated credit granting. The method includes the following steps: Step S1: Extract typical behavioral nodes of the target enterprise in the credit granting process and construct a set of credit granting nodes; obtain historical behavioral attribute data sets for each typical behavioral node and construct a unified behavioral feature space; Step S2: Based on the unified behavioral feature space, construct node behavior vectors for typical behavioral nodes within the unified behavioral feature space, and calculate the difference in node behavioral attribute data between two adjacent typical behavioral nodes; Step S3: Obtain the event occurrence time interval between two adjacent typical behavioral nodes, and combine the difference in node behavioral attribute data to construct path parameters and a historical credit granting map; construct a marginal effect modeling function and calculate the marginal score value; Step S4: Construct a simulated credit granting scenario generator, and based on the credit granting process, sequentially simulate dynamic perturbations on a total of I typical behavioral nodes; calculate the simulated marginal score value and global deviation rate under each simulated credit granting scenario; preset thresholds, and analyze and verify the model stability.

[0007] As a preferred embodiment of the AI-based simulated credit granting marginal effect calibration and analysis method described in this invention, after user authorization, typical behavioral nodes of the target enterprise in the credit granting process are extracted from a big data platform. These typical behavioral nodes include application nodes, approval nodes, loan disbursement nodes, repayment nodes, and re-credit granting nodes. These typical behavioral nodes are then uniformly numbered, and a credit granting node set is constructed, denoted as Y = {y...} i |i∈[1,I]}, where y i Let I represent the i-th typical behavior node in the set of authorized nodes, and let I represent the total number of typical behavior nodes in the set of authorized nodes.

[0008] Historical behavioral attribute data sets for each typical behavioral node are obtained from a big data platform. An artificial intelligence feature fusion model is used to align the feature spaces of all historical behavioral attribute data sets, and a unified behavioral feature space is constructed, denoted as T = {t}. a |a∈[1,A]}, where t a Let A represent the a-th historical behavior feature after feature space alignment, and let A represent the total number of historical behavior features after feature space alignment.

[0009] It should be noted that by extracting typical behavioral nodes of the target enterprise in the credit granting process, a set of credit granting nodes is constructed. Based on the data acquisition method authorized by the user, the historical behavioral attribute data set of each node is obtained. Then, the feature space alignment processing is completed by using an artificial intelligence feature fusion model, thereby constructing a unified behavioral feature space and realizing the standardized and structured expression of the enterprise's credit granting behavior characteristics.

[0010] The key technological advantage of this step lies in resolving the issues of heterogeneity and incomparability of traditional credit granting behavior characteristics. By constructing a unified behavioral feature space, behavioral data from different enterprises, at different times, and at different nodes can be quantitatively compared and analyzed within the same space, laying a unified benchmark for subsequent vector modeling and path parameter calculation. Furthermore, this structured transformation not only improves feature extraction efficiency but also enhances the model's generalization ability, providing a foundation for algorithm interpretability and demonstrating significant value in data governance and pre-modeling.

[0011] As a preferred embodiment of the simulated credit marginal effect calibration and analysis method based on artificial intelligence described in this invention, based on the unified behavioral feature space, an artificial intelligence feature processing model is used to map the historical behavioral attribute data set of typical behavioral nodes to the unified behavioral feature space, construct the node behavioral vector of typical behavioral nodes in the unified behavioral feature space, and calculate the difference in node behavioral attribute data between two adjacent typical behavioral nodes, as detailed below:

[0012]

[0013] Where, Δx i,i+1 Represents a typical behavior node y i With typical behavior node y i+1 The difference in node behavior attribute data between them, ω a This represents the a-th historical behavior feature t after pre-defined feature space alignment processing. a The weights, X i,T [t a ] represents a typical behavior node y i In the node behavior vector under the unified behavior feature space T, the historical behavior feature t a The standardized value, X i+1,T [t a ] represents a typical behavior node y i+1 Historical behavior features t in the node behavior vectors under a unified behavior feature space T a Standardized values.

[0014] It should be noted that by utilizing an artificial intelligence feature processing model to map the behavioral attributes of each node to a unified behavioral feature space, constructing node behavior vectors, and calculating the attribute differences between any two adjacent typical behavior nodes, the quantification of behavioral change trends and the extraction of differences are achieved. Through vectorization, the multidimensional attributes of behavioral changes are compressed into a computable numerical form, thereby enabling quantitative comparisons between behaviors. Node behavior differences not only reflect micro-level changes in enterprise behavioral patterns but also provide data support for constructing path parameters. By setting the weights of each feature, the saliency of key behavioral features can be identified, enhancing the model's ability to identify credit risk-sensitive factors and thus improving the accuracy of subsequent marginal effect assessments.

[0015] As a preferred embodiment of the simulated credit marginal effect calibration analysis method based on artificial intelligence described in this invention, a marginal effect calibration model is constructed as follows:

[0016] Obtain typical behavior node y i With typical behavior node y i+1 The time interval between the events is denoted as ΔTE. i,i+1 Combined with typical behavior node y i With typical behavior node y i+1 The difference in node behavior attributes Δx i,i+1 Construct typical behavior nodes y i With typical behavior node y i+1 The path parameter between them is denoted as λ. i,i+1 =(ΔTE) i,i+1 ,Δx i,i+1 );

[0017] Based on typical behavior node y i With typical behavior node y i+1 Path parameter λ between i,i+1 =(ΔTE) i,i+1 ,Δx i,i+1 A set of credit granting process path parameters is constructed, and combined with the set of credit granting nodes Y, a historical credit granting graph is constructed, as follows:

[0018] The set of credit-granting nodes is denoted as the set of graph nodes in the historical credit-granting graph.

[0019] Connect all typical behavior nodes in the set of credit granting nodes according to the credit granting process sequence to construct connection edges between typical behavior nodes;

[0020] Based on the set of credit granting process path parameters, the path parameters between each pair of adjacent typical behavior nodes are denoted as the edge weight of the connection edge between each pair of adjacent typical behavior nodes.

[0021] A marginal effect modeling function is constructed, and the edge weight of each connecting edge in the historical credit graph is input into the marginal effect modeling function to calculate the marginal score value. The marginal effect modeling function is as follows:

[0022]

[0023] Where, μ i,i+1 Represents a typical behavior node y i With typical behavior node y i+1 The marginal score values ​​between them, where α represents the preset control adjustment weight, and β and γ represent the preset event occurrence time interval and the exponential decay coefficient of the difference in node behavior attribute data, respectively.

[0024] It should be noted that by obtaining the difference between the time interval of event occurrence and the node behavior attribute, path parameters are constructed, and on this basis, a historical credit granting map and a marginal effect modeling function are constructed, realizing the mapping modeling of micro-behavioral changes in the credit granting process on macro-scoring results.

[0025] This step introduces a graph modeling approach, systematically representing each node in the credit granting process and its evolutionary paths in a graph structure. By using path parameters as edge weights, the model is endowed with a combined explanatory power encompassing both the temporal dimension and behavioral differences. The marginal effect modeling function is constructed based on an exponential decay function, ensuring that the impact of time intervals and behavioral differences on the scoring results reflects the characteristic of "recent behavior being more important." This helps improve the model's sensitivity to current behavioral anomalies, thus constructing a credit granting scoring system with strong dynamic response capabilities and good interpretability.

[0026] As a preferred embodiment of the simulated credit marginal effect calibration analysis method based on artificial intelligence described in this invention, the marginal effect calibration model is verified as follows:

[0027] Based on historical credit granting graphs and marginal scoring values, a simulated credit granting scenario generator is constructed. Based on the credit granting process, dynamic disturbances are simulated sequentially for I typical behavioral nodes (such as adjusting the qualification data of the application node and simulating different decision preferences of the approval node) to generate I sets of simulated credit granting scenarios.

[0028] Recalculate the difference in node behavior attribute data for each simulated credit granting scenario, input the new path parameters into the marginal effect modeling function, and calculate the simulated marginal score for each simulated credit granting scenario; calculate the absolute error between the simulated marginal score and the marginal score for each simulated credit granting scenario, and then sum and average them to obtain the global deviation rate;

[0029] A global deviation rate threshold is preset. If the global deviation rate is less than or equal to the global deviation rate threshold, it is determined that the credit granting process is stable under simulated dynamic disturbances, and the marginal effect calibration model is effective.

[0030] It should be noted that by constructing a simulated credit granting scenario generator, dynamic perturbations are simulated on typical behavioral nodes based on the credit granting process, and the global deviation rate is calculated by comparing the simulated marginal score value with the original score value, thus realizing the stability verification of the marginal effect modeling function and the robustness evaluation of the model.

[0031] This step introduces a simulated perturbation verification mechanism, examining the scoring model's scoring deviation under behavioral perturbations through multiple simulated scenarios, thereby determining the model's adaptability to "non-ideal behavior" or "abnormal fluctuations." The introduction of the global bias rate serves as a metric for model effectiveness, providing a quantitative basis for subsequent model iteration and optimization, further enhancing the model's practicality and reliability. In financial risk control scenarios, this mechanism effectively avoids the risk of scoring models overfitting historical data and lacking practical fault tolerance, making it highly valuable for application and promotion.

[0032] This artificial intelligence-based simulated credit marginal effect calibration and analysis system includes: a data acquisition and spatial construction module, a vector construction and difference calculation module, a graph construction and scoring calculation module, and a dynamic simulation and analysis module.

[0033] The data acquisition and space construction module extracts typical behavioral nodes of the target enterprise in the credit granting process and constructs a set of credit granting nodes; it also acquires a set of historical behavioral attribute data for each typical behavioral node and constructs a unified behavioral feature space.

[0034] The vector construction and difference calculation module: Based on the unified behavior feature space, constructs the node behavior vector of typical behavior nodes in the unified behavior feature space, and calculates the difference of node behavior attribute data between two adjacent typical behavior nodes;

[0035] The graph construction and scoring calculation module: obtains the time interval between events between two adjacent typical behavior nodes, combines the difference in node behavior attribute data, constructs path parameters and historical credit graphs; constructs a marginal effect modeling function, and calculates the marginal score value;

[0036] The dynamic simulation and analysis module: constructs a simulated credit granting scenario generator, and based on the credit granting process, sequentially simulates dynamic perturbations on a total of I typical behavioral nodes; calculates the simulated marginal score and global deviation rate under each simulated credit granting scenario; presets thresholds, and analyzes and verifies the model stability.

[0037] Furthermore, the data acquisition and spatial construction module includes a data acquisition unit and a spatial construction unit;

[0038] The data acquisition unit: after user authorization, extracts typical behavioral nodes of the target enterprise in the credit granting process from the big data platform. The typical behavioral nodes include application nodes, approval nodes, loan disbursement nodes, repayment nodes, and re-credit granting nodes; assigns a unified number to the typical behavioral nodes and constructs a credit granting node set.

[0039] The spatial construction unit: obtains the historical behavioral attribute data set of each typical behavioral node from the big data platform, uses an artificial intelligence feature fusion model to perform feature space alignment processing on all historical behavioral attribute data sets, and constructs a unified behavioral feature space.

[0040] Furthermore, the vector construction and difference calculation module includes a vector construction unit and a difference calculation unit;

[0041] The vector construction unit: Based on the unified behavior feature space, it uses an artificial intelligence feature processing model to map the historical behavior attribute data set of typical behavior nodes to the unified behavior feature space, and constructs the node behavior vector of typical behavior nodes in the unified behavior feature space.

[0042] The difference calculation unit calculates the difference in node behavior attribute data between two adjacent typical behavior nodes.

[0043] Furthermore, the map construction and scoring calculation module includes a map construction unit and a scoring calculation unit;

[0044] The graph construction unit: obtains the event occurrence time interval between two adjacent typical behavior nodes, and constructs path parameters between two adjacent typical behavior nodes by combining the difference in node behavior attributes; based on the path parameters, constructs a set of credit granting process path parameters, and constructs a historical credit granting graph by combining the set of credit granting nodes, specifically as follows: the set of credit granting nodes is denoted as the graph node set in the historical credit granting graph; all typical behavior nodes in the set of credit granting nodes are connected in the order of the credit granting process to construct the connection edges between typical behavior nodes; based on the set of credit granting process path parameters, the path parameters between each pair of adjacent typical behavior nodes are denoted as the edge weight of the connection edge between each pair of adjacent typical behavior nodes;

[0045] The scoring calculation unit: constructs a marginal effect modeling function, inputs the edge weight of each connecting edge in the historical credit map into the marginal effect modeling function, and calculates the marginal score value.

[0046] Furthermore, the dynamic simulation and analysis module includes a dynamic simulation unit and an analysis unit;

[0047] The dynamic simulation unit: Based on historical credit granting graphs and marginal scoring values, it constructs a simulated credit granting scenario generator, and based on the credit granting process, it sequentially simulates dynamic disturbances on typical behavioral nodes to generate several sets of simulated credit granting scenarios;

[0048] Recalculate the difference in node behavior attribute data for each simulated credit granting scenario, input the new path parameters into the marginal effect modeling function, and calculate the simulated marginal score for each simulated credit granting scenario; calculate the absolute error between the simulated marginal score and the marginal score for each simulated credit granting scenario, and then sum and average them to obtain the global deviation rate;

[0049] The analysis unit: presets a global deviation rate threshold. If the global deviation rate is less than or equal to the global deviation rate threshold, it is determined that the credit granting process is stable under simulated dynamic disturbances, and the marginal effect calibration model is effective.

[0050] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: The AI-based simulated credit granting marginal effect calibration and analysis system and method provided by this invention, by constructing a unified behavioral feature space, achieves the standardization and structured expression of historical behavioral data of typical behavioral nodes in the credit granting process, providing a unified benchmark for subsequent modeling; further, it utilizes a feature processing model to generate node behavioral vectors and calculates the behavioral attribute differences between adjacent nodes, realizing a quantitative expression of the trend of enterprise behavioral changes and providing a basis for identifying key behavioral variations; subsequently, it combines the event occurrence time interval and behavioral differences to construct path parameters, establishes a historical credit granting map, and constructs a marginal effect modeling function, realizing the mapping of micro-behavioral changes in the credit granting process to a macro-scoring system, and employs an exponential decay mechanism to enhance the model's sensitivity to recent behavior; finally, it constructs a simulated credit granting scenario generator, calculates the global deviation rate and evaluates the model's stability by simulating perturbations in the credit granting process, thereby verifying the robustness and adaptability of the marginal scoring model under various behavioral scenarios. Overall, this method enhances the interpretability of the scoring model while improving the accuracy and dynamic response capability of credit risk assessment, making it suitable for intelligent credit management systems driven by big data. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0052] Figure 1 This is a schematic diagram illustrating the steps of the simulated credit marginal effect calibration and analysis method based on artificial intelligence in this invention;

[0053] Figure 2 This is a schematic diagram of the structure of the simulated credit marginal effect calibration and analysis system based on artificial intelligence of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 In this first embodiment: an artificial intelligence-based simulated credit marginal effect calibration analysis method is provided, which includes the following steps:

[0056] Step S1: Extract typical behavioral nodes of the target enterprise in the credit granting process and construct a set of credit granting nodes; obtain the historical behavioral attribute data set of each typical behavioral node and construct a unified behavioral feature space.

[0057] Specifically, after user authorization, typical behavioral nodes of the target enterprise in the credit granting process are extracted from the big data platform. These typical behavioral nodes include application nodes, approval nodes, loan disbursement nodes, repayment nodes, and re-credit granting nodes. These typical behavioral nodes are uniformly numbered, and a credit granting node set is constructed, denoted as Y = {y}. i |i∈[1,I]}, where y i Let I represent the i-th typical behavior node in the set of authorized nodes, and let I represent the total number of typical behavior nodes in the set of authorized nodes.

[0058] Historical behavioral attribute data sets for each typical behavioral node are obtained from a big data platform. An artificial intelligence feature fusion model is used to align the feature spaces of all historical behavioral attribute data sets, and a unified behavioral feature space is constructed, denoted as T = {t}. a |a∈[1,A]}, where t a Let A represent the a-th historical behavior feature after feature space alignment, and let A represent the total number of historical behavior features after feature space alignment.

[0059] The artificial intelligence feature fusion model performs feature space alignment processing on the entire historical behavioral attribute data set as follows:

[0060] For example, suppose the application nodes include numerical types (annual revenue of 10 million, debt-to-asset ratio of 30%), text types (business scope of the enterprise is "manufacturing"), and category types (enterprise size is "medium");

[0061] The approval process includes numerical (approval amount of 8 million), textual (approval opinion "revenue authenticity needs to be verified"), and rating (credit rating "AA+");

[0062] Loan disbursement milestones include numerical (repayment amount 8 million, overdue days 0) and time-based (repayment date "2023-10-01");

[0063] Taking the improved autoencoder + attention mechanism fusion model as an example:

[0064] Standardize numerical data (map to the [0, 1] interval), such as annual revenue of 10 million, which is a standardized value of 0.8 (assuming the industry maximum is 12 million), and debt-to-asset ratio of 30%, which is a standardized value of 0.3;

[0065] Embedding encoding operations are performed on categorical and graded data (e.g., credit rating "AA+" → [0.8, 0.1, 0.1], "A" → [0.6, 0.2, 0.2], which can be transformed using a pre-trained embedding matrix);

[0066] For text-based data, perform word vector transformation (e.g., "manufacturing" → Word2Vec vector [0.2, 0.5, -0.1, ...], "revenue authenticity needs to be verified" → extract sentence vector using BERT [0.3, 0.1, 0.4, ...]);

[0067] For time-based data, perform time feature extraction (e.g., repayment date 180 days from loan disbursement date → standardized value 0.6, assuming the longest period is 300 days);

[0068] The input layer receives the heterogeneous features preprocessed by each node. The encoding layer compresses the inputs of different dimensions into 5-dimensional intermediate features through a 3-layer neural network (hidden layer dimensions 20→10→5). The decoding layer reconstructs the input features and ensures that key information is preserved during the compression process by minimizing the reconstruction error (such as MSE loss).

[0069] Step S2: Based on the unified behavior feature space, construct the node behavior vector of the typical behavior node in the unified behavior feature space, and calculate the difference in node behavior attribute data between two adjacent typical behavior nodes.

[0070] Specifically, based on the unified behavioral feature space, an artificial intelligence feature processing model is used to map the historical behavioral attribute data set of typical behavioral nodes to the unified behavioral feature space, construct the node behavioral vector of typical behavioral nodes in the unified behavioral feature space, and calculate the difference in node behavioral attribute data between two adjacent typical behavioral nodes, as follows:

[0071]

[0072] Where, Δx i,i+1 Represents a typical behavior node y i With typical behavior node y i+1 The difference in node behavior attribute data between them, ω a This represents the a-th historical behavior feature t after pre-defined feature space alignment processing. a The weights, X i,T [t a ] represents a typical behavior node y i Historical behavior features t in the node behavior vectors under a unified behavior feature space T a The standardized value, X i+1,T [t a ] represents a typical behavior node y i+1 Historical behavior features t in the node behavior vectors under a unified behavior feature space T a Standardized values.

[0073] In this invention, changes in the behavior of adjacent nodes in the credit granting process (such as the difference between "the revenue claimed by the company at the time of application" and "the revenue verified at the time of approval") directly affect risk assessment. This formula can transform the multi-dimensional behavioral differences (such as amount, qualifications, time requirements, etc.) in the "application-approval" and "approval-loan" stages into a single value, solving the problem of "feature heterogeneity and inability to be directly compared" in traditional methods.

[0074] Behavioral data at different stages of the credit granting process often exhibits "dimensional heterogeneity" (e.g., "corporate asset proof" at the application stage is monetary, while "credit rating" at the approval stage is graded), making direct comparison difficult using traditional methods. This formula addresses this by mapping a unified behavioral feature space (X). i,T [t a The system converts all features into standardized values ​​and then uses absolute difference calculation to achieve "cross-dimensional difference quantification," enabling direct comparison of behavioral changes in processes such as "application-approval" and "loan disbursement-repayment."

[0075] For example, if a company declares "annual revenue of 10 million" at the application stage, but the approval stage verifies it as "8 million", this difference can be quantified using a formula to unify the difference between "the industry classification was 'manufacturing' at the time of application and 'service' at the time of approval", thus avoiding analytical gaps caused by different feature types.

[0076] Step S3: Obtain the time interval between events between two adjacent typical behavior nodes, combine the differences in node behavior attribute data to construct path parameters and historical credit map; construct a marginal effect modeling function to calculate the marginal score value.

[0077] Specifically, a marginal effect calibration model is constructed as follows:

[0078] Obtain typical behavior node y i With typical behavior node y i+1 The time interval between the events is denoted as ΔTE. i,i+1 Combined with typical behavior node y i With typical behavior node y i+1 The difference in node behavior attributes Δx i,i+1 Construct typical behavior nodes y i With typical behavior node y i+1 The path parameter between them is denoted as λ. i,i+1 =(ΔTE) i,i+1 ,Δx i,i+1 );

[0079] Based on typical behavior node y i With typical behavior node y i+1 Path parameter λ between i,i+1 =(ΔTE) i,i+1 ,Δx i,i+1 A set of credit granting process path parameters is constructed, and combined with the set of credit granting nodes Y, a historical credit granting graph is constructed, as follows:

[0080] The set of credit-granting nodes is denoted as the set of graph nodes in the historical credit-granting graph.

[0081] Connect all typical behavior nodes in the set of credit granting nodes according to the credit granting process sequence to construct connection edges between typical behavior nodes;

[0082] Based on the set of credit granting process path parameters, the path parameters between each pair of adjacent typical behavior nodes are denoted as the edge weight of the connection edge between each pair of adjacent typical behavior nodes.

[0083] A marginal effect modeling function is constructed, and the edge weight of each connecting edge in the historical credit graph is input into the marginal effect modeling function to calculate the marginal score value. The marginal effect modeling function is as follows:

[0084]

[0085] Where, μ i,i+1 Represents a typical behavior node y i With typical behavior node y i+1 The marginal score values ​​between them, where α represents the preset control adjustment weight, and β and γ represent the preset event occurrence time interval and the exponential decay coefficient of the difference in node behavior attribute data, respectively.

[0086] In this invention, the marginal score is a weighted sum of the "time efficiency score" and the "behavioral consistency score." Due to the properties of the exponential function, as the time interval increases ΔTE... i,i+1(Process delays) or increased behavioral discrepancy Δx i,i+1 When there is a contradiction between the preceding and following actions, the corresponding index item will decay rapidly, leading to a decrease in the marginal score and reflecting an increased risk; for example, if the difference between the company's "repayment period promised at the time of application" and "the actual repayment period after approval" is small (Δx) i,i+1 Small), and fast approval speed (ΔTE) i,i+1 (small), then μ i,i+1 A high score indicates that the marginal negative impact of this step on credit risk is small; conversely, a low score suggests that attention should be paid to the risk.

[0087] Formula used and When the "longer the time interval" or the "greater the behavioral difference", the marginal score μ i,i+1 The lower the value (the higher the risk), and the impact exhibits a "non-linear decay"—small differences in the near term may have a more significant impact than large differences in the long term, which aligns with the "timeliness first" risk assessment logic in real-world financial scenarios.

[0088] For example, a company's "payment due 3 days late this month" (ΔTE) i,i+1 Smaller but closer in time) compared to "10 days of delayed repayment six months ago" (ΔTE) i,i+1 The negative impact of large but long-term effects on the current marginal effect of credit is greater, and the exponential function can accurately capture this "time decay" characteristic;

[0089] Marginal score μ i,i+1 It is an intuitive numerical value (the range of which can be controlled between 0 and 1 through parameter adjustment), and its level directly reflects the "marginal impact of changes in behavior between nodes on credit granting". Compared with the traditional black box model, this design allows risk control personnel to clearly understand "whether the reason for a low score is a slow process or inconsistent behavior", thus improving the acceptability of the model in actual business.

[0090] Step S4: Construct a simulated credit granting scenario generator. Based on the credit granting process, simulate dynamic perturbations on I typical behavioral nodes in sequence; calculate the simulated marginal score and global deviation rate under each simulated credit granting scenario; preset thresholds, analyze and verify the model stability.

[0091] Specifically, the marginal effect calibration model is verified as follows:

[0092] Based on historical credit granting graphs and marginal scoring values, a simulated credit granting scenario generator is constructed. Based on the credit granting process, dynamic disturbances are simulated sequentially for I typical behavioral nodes (such as adjusting the qualification data of the application node and simulating different decision preferences of the approval node) to generate I sets of simulated credit granting scenarios.

[0093] Recalculate the difference in node behavior attribute data for each simulated credit granting scenario, input the new path parameters into the marginal effect modeling function, and calculate the simulated marginal score for each simulated credit granting scenario; calculate the absolute error between the simulated marginal score and the marginal score for each simulated credit granting scenario, and then sum and average them to obtain the global deviation rate;

[0094] A global deviation rate threshold is preset. If the global deviation rate is less than or equal to the global deviation rate threshold, it is determined that the credit granting process is stable under simulated dynamic disturbances, and the marginal effect calibration model is effective.

[0095] Please see Figure 2 In this second embodiment: an artificial intelligence-based simulated credit marginal effect calibration and analysis system is provided, which includes: a data acquisition and spatial construction module, a vector construction and difference calculation module, a graph construction and scoring calculation module, and a dynamic simulation and analysis module;

[0096] The data acquisition and space construction module extracts typical behavioral nodes of the target enterprise in the credit granting process and constructs a set of credit granting nodes; it also acquires a set of historical behavioral attribute data for each typical behavioral node and constructs a unified behavioral feature space.

[0097] The vector construction and difference calculation module: Based on the unified behavior feature space, constructs the node behavior vector of typical behavior nodes in the unified behavior feature space, and calculates the difference of node behavior attribute data between two adjacent typical behavior nodes;

[0098] The graph construction and scoring calculation module: obtains the time interval between events between two adjacent typical behavior nodes, combines the difference in node behavior attribute data, constructs path parameters and historical credit graphs; constructs a marginal effect modeling function, and calculates the marginal score value;

[0099] The dynamic simulation and analysis module: constructs a simulated credit granting scenario generator, and based on the credit granting process, sequentially simulates dynamic perturbations on a total of I typical behavioral nodes; calculates the simulated marginal score and global deviation rate under each simulated credit granting scenario; presets thresholds, and analyzes and verifies the model stability.

[0100] Furthermore, the data acquisition and spatial construction module includes a data acquisition unit and a spatial construction unit;

[0101] The data acquisition unit: after user authorization, extracts typical behavioral nodes of the target enterprise in the credit granting process from the big data platform. The typical behavioral nodes include application nodes, approval nodes, loan disbursement nodes, repayment nodes, and re-credit granting nodes; assigns a unified number to the typical behavioral nodes and constructs a credit granting node set.

[0102] The spatial construction unit: obtains the historical behavioral attribute data set of each typical behavioral node from the big data platform, uses an artificial intelligence feature fusion model to perform feature space alignment processing on all historical behavioral attribute data sets, and constructs a unified behavioral feature space.

[0103] Furthermore, the vector construction and difference calculation module includes a vector construction unit and a difference calculation unit;

[0104] The vector construction unit: Based on the unified behavior feature space, it uses an artificial intelligence feature processing model to map the historical behavior attribute data set of typical behavior nodes to the unified behavior feature space, and constructs the node behavior vector of typical behavior nodes in the unified behavior feature space.

[0105] The difference calculation unit calculates the difference in node behavior attribute data between two adjacent typical behavior nodes.

[0106] Furthermore, the map construction and scoring calculation module includes a map construction unit and a scoring calculation unit;

[0107] The graph construction unit: obtains the event occurrence time interval between two adjacent typical behavior nodes, and constructs path parameters between two adjacent typical behavior nodes by combining the difference in node behavior attributes; based on the path parameters, constructs a set of credit granting process path parameters, and constructs a historical credit granting graph by combining the set of credit granting nodes, specifically as follows: the set of credit granting nodes is denoted as the graph node set in the historical credit granting graph; all typical behavior nodes in the set of credit granting nodes are connected in the order of the credit granting process to construct the connection edges between typical behavior nodes; based on the set of credit granting process path parameters, the path parameters between each pair of adjacent typical behavior nodes are denoted as the edge weight of the connection edge between each pair of adjacent typical behavior nodes;

[0108] The scoring calculation unit: constructs a marginal effect modeling function, inputs the edge weight of each connecting edge in the historical credit map into the marginal effect modeling function, and calculates the marginal score value.

[0109] Furthermore, the dynamic simulation and analysis module includes a dynamic simulation unit and an analysis unit;

[0110] The dynamic simulation unit: Based on historical credit granting graphs and marginal scoring values, it constructs a simulated credit granting scenario generator, and based on the credit granting process, it sequentially simulates dynamic disturbances on typical behavioral nodes to generate several sets of simulated credit granting scenarios;

[0111] Recalculate the difference in node behavior attribute data for each simulated credit granting scenario, input the new path parameters into the marginal effect modeling function, and calculate the simulated marginal score for each simulated credit granting scenario; calculate the absolute error between the simulated marginal score and the marginal score for each simulated credit granting scenario, and then sum and average them to obtain the global deviation rate;

[0112] The analysis unit: presets a global deviation rate threshold. If the global deviation rate is less than or equal to the global deviation rate threshold, it is determined that the credit granting process is stable under simulated dynamic disturbances, and the marginal effect calibration model is effective.

[0113] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0114] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A simulated credit marginal effect calibration and analysis method based on artificial intelligence, characterized in that, The method includes the following steps: Step S1: Extract typical behavioral nodes of the target enterprise in the credit granting process and construct a set of credit granting nodes; obtain the historical behavioral attribute data set of each typical behavioral node and construct a unified behavioral feature space; Step S2: Based on the unified behavior feature space, construct the node behavior vector of the typical behavior node in the unified behavior feature space, and calculate the difference of node behavior attribute data between two adjacent typical behavior nodes. Step S3: Obtain the time interval between events between two adjacent typical behavior nodes, combine the difference in node behavior attribute data to construct path parameters and historical credit map; construct the marginal effect modeling function and calculate the marginal score value; Step S4: Construct a simulated credit granting scenario generator. Based on the credit granting process, simulate dynamic perturbations on I typical behavioral nodes in sequence; calculate the simulated marginal score and global deviation rate under each simulated credit granting scenario; preset thresholds, analyze and verify the model stability.

2. The simulated credit marginal effect calibration and analysis method based on artificial intelligence according to claim 1, characterized in that, The specific implementation process of step S1 includes: After user authorization, typical behavioral nodes of the target enterprise in the credit granting process are extracted from the big data platform. These typical behavioral nodes include application nodes, approval nodes, loan disbursement nodes, repayment nodes, and re-credit granting nodes. These typical behavioral nodes are uniformly numbered, and a credit granting node set is constructed, denoted as Y = {y}. i |i∈[1,I]}, where y i Let I represent the i-th typical behavior node in the set of authorized nodes, and let I represent the total number of typical behavior nodes in the set of authorized nodes. Historical behavioral attribute data sets for each typical behavioral node are obtained from a big data platform. An artificial intelligence feature fusion model is used to align the feature spaces of all historical behavioral attribute data sets, and a unified behavioral feature space is constructed, denoted as T = {t}. a |a∈[1,A]}, where t a Let A represent the a-th historical behavior feature after feature space alignment, and let A represent the total number of historical behavior features after feature space alignment.

3. The simulated credit marginal effect calibration and analysis method based on artificial intelligence according to claim 2, characterized in that, The specific implementation process of step S2 includes: Based on the unified behavioral feature space, using an artificial intelligence feature processing model, the historical behavioral attribute data set of typical behavioral nodes is mapped to the unified behavioral feature space, constructing the node behavioral vector of typical behavioral nodes in the unified behavioral feature space, and calculating the difference in node behavioral attribute data between two adjacent typical behavioral nodes, as follows: Where, Δx i,i+1 Represents a typical behavior node y i With typical behavior node y i+1 The difference in node behavior attribute data between them, ω a This represents the a-th historical behavior feature t after pre-defined feature space alignment processing. a The weights, X i,T [t a ] represents a typical behavior node y i Historical behavior features t in the node behavior vectors under a unified behavior feature space T a The standardized value, X i+1,T [t a ] represents a typical behavior node y i+1 Historical behavior features t in the node behavior vectors under a unified behavior feature space T a Standardized values.

4. The simulated credit marginal effect calibration and analysis method based on artificial intelligence according to claim 3, characterized in that, The specific implementation process of step S3 includes: The marginal effect calibration model is constructed as follows: Obtain typical behavior node y i With typical behavior node y i+1 The time interval between the occurrences of the events is denoted as ΔTE. i,i+1 Combined with typical behavior node y i With typical behavior node y i+1 The difference in node behavior attributes Δx i,i+1 Construct typical behavior nodes y i With typical behavior node y i+1 The path parameter between them is denoted as λ. i,i+1 =(ΔTE) i,i+1 ,Δx i,i+1 ); Based on typical behavior node y i With typical behavior node y i+1 Path parameters between ,i+1 =(ΔTE) i,i+1 ,Δx i,i+1 A set of credit granting process path parameters is constructed, and combined with the set of credit granting nodes Y, a historical credit granting graph is constructed, as follows: The set of credit-granting nodes is denoted as the set of graph nodes in the historical credit-granting graph. Connect all typical behavior nodes in the set of credit granting nodes according to the credit granting process sequence to construct connection edges between typical behavior nodes; Based on the set of credit granting process path parameters, the path parameters between each pair of adjacent typical behavior nodes are denoted as the edge weight of the connection edge between each pair of adjacent typical behavior nodes. A marginal effect modeling function is constructed, and the edge weight of each connecting edge in the historical credit graph is input into the marginal effect modeling function to calculate the marginal score value. The marginal effect modeling function is as follows: Where, μ i,i+1 Represents a typical behavior node y i With typical behavior node y i+1 The marginal score values ​​between them, where α represents the preset control adjustment weight, and β and γ represent the preset event occurrence time interval and the exponential decay coefficient of the difference in node behavior attribute data, respectively.

5. The simulated credit marginal effect calibration and analysis method based on artificial intelligence according to claim 4, characterized in that, The specific implementation process of step S4 includes: The marginal effect calibration model was validated as follows: Based on historical credit granting graphs and marginal scoring values, a simulated credit granting scenario generator is constructed. Based on the credit granting process, I typical behavioral nodes are dynamically disturbed in sequence to generate I sets of simulated credit granting scenarios. Recalculate the difference in node behavior attribute data for each simulated credit granting scenario, input the new path parameters into the marginal effect modeling function, and calculate the simulated marginal score for each simulated credit granting scenario; calculate the absolute error between the simulated marginal score and the marginal score for each simulated credit granting scenario, and then sum and average them to obtain the global deviation rate; A global deviation rate threshold is preset. If the global deviation rate is less than or equal to the global deviation rate threshold, it is determined that the credit granting process is stable under simulated dynamic disturbances, and the marginal effect calibration model is effective.

6. An AI-based simulated credit marginal effect calibration and analysis system, executing the AI-based simulated credit marginal effect calibration and analysis method as described in any one of claims 1-5, characterized in that, The system includes: a data acquisition and spatial construction module, a vector construction and difference calculation module, a map construction and scoring calculation module, and a dynamic simulation and analysis module; The data acquisition and space construction module extracts typical behavioral nodes of the target enterprise in the credit granting process and constructs a set of credit granting nodes; it also acquires a set of historical behavioral attribute data for each typical behavioral node and constructs a unified behavioral feature space. The vector construction and difference calculation module: Based on the unified behavior feature space, constructs the node behavior vector of typical behavior nodes in the unified behavior feature space, and calculates the difference of node behavior attribute data between two adjacent typical behavior nodes; The graph construction and scoring calculation module: obtains the time interval between events between two adjacent typical behavior nodes, combines the difference in node behavior attribute data, constructs path parameters and historical credit graphs; constructs a marginal effect modeling function, and calculates the marginal score value; The dynamic simulation and analysis module: constructs a simulated credit granting scenario generator, and based on the credit granting process, sequentially simulates dynamic perturbations on a total of I typical behavioral nodes; calculates the simulated marginal score and global deviation rate under each simulated credit granting scenario; presets thresholds, and analyzes and verifies the model stability.

7. The simulated credit marginal effect calibration and analysis system based on artificial intelligence according to claim 6, characterized in that: The data acquisition and spatial construction module includes a data acquisition unit and a spatial construction unit; The data acquisition unit: after user authorization, extracts typical behavioral nodes of the target enterprise in the credit granting process from the big data platform. The typical behavioral nodes include application nodes, approval nodes, loan disbursement nodes, repayment nodes, and re-credit granting nodes; assigns a unified number to the typical behavioral nodes and constructs a credit granting node set. The spatial construction unit: obtains the historical behavioral attribute data set of each typical behavioral node from the big data platform, uses an artificial intelligence feature fusion model to perform feature space alignment processing on all historical behavioral attribute data sets, and constructs a unified behavioral feature space.

8. The simulated credit marginal effect calibration and analysis system based on artificial intelligence according to claim 7, characterized in that: The vector construction and difference calculation module includes a vector construction unit and a difference calculation unit; The vector construction unit: Based on the unified behavior feature space, it uses an artificial intelligence feature processing model to map the historical behavior attribute data set of typical behavior nodes to the unified behavior feature space, and constructs the node behavior vector of typical behavior nodes in the unified behavior feature space. The difference calculation unit calculates the difference in node behavior attribute data between two adjacent typical behavior nodes.

9. The simulated credit marginal effect calibration and analysis system based on artificial intelligence according to claim 8, characterized in that: The map construction and scoring calculation module includes a map construction unit and a scoring calculation unit; The graph construction unit: obtains the event occurrence time interval between two adjacent typical behavior nodes, and constructs path parameters between two adjacent typical behavior nodes by combining the difference in node behavior attributes; based on the path parameters, constructs a set of credit granting process path parameters, and constructs a historical credit granting graph by combining the set of credit granting nodes, specifically as follows: the set of credit granting nodes is denoted as the graph node set in the historical credit granting graph; all typical behavior nodes in the set of credit granting nodes are connected in the order of the credit granting process to construct connection edges between typical behavior nodes; Based on the set of credit granting process path parameters, the path parameters between each pair of adjacent typical behavior nodes are denoted as the edge weight of the connection edge between each pair of adjacent typical behavior nodes. The scoring calculation unit constructs a marginal effect modeling function, inputs the edge weight of each connecting edge in the historical credit map into the marginal effect modeling function, and calculates the marginal score value.

10. The simulated credit marginal effect calibration and analysis system based on artificial intelligence according to claim 9, characterized in that: The dynamic simulation and analysis module includes a dynamic simulation unit and an analysis unit; The dynamic simulation unit: Based on historical credit granting graphs and marginal scoring values, it constructs a simulated credit granting scenario generator, and based on the credit granting process, it sequentially simulates dynamic disturbances on typical behavioral nodes to generate several sets of simulated credit granting scenarios; Recalculate the difference in node behavior attribute data for each simulated credit granting scenario, input the new path parameters into the marginal effect modeling function, and calculate the simulated marginal score for each simulated credit granting scenario; calculate the absolute error between the simulated marginal score and the marginal score for each simulated credit granting scenario, and then sum and average them to obtain the global deviation rate; The analysis unit: presets a global deviation rate threshold. If the global deviation rate is less than or equal to the global deviation rate threshold, it is determined that the credit granting process is stable under simulated dynamic disturbances, and the marginal effect calibration model is effective.

Images