A financial risk control causal topology structure discovery method and system

Abstract

This application provides a method and system for discovering causal topology structures in financial risk control, including: constructing a multi-level feature variable hierarchy based on sample space initialization and feature extraction; constructing a topology graph using a causal algorithm; verifying the fitting accuracy of the causal structure and obtaining a Bayesian information criterion score; based on the Bayesian information criterion score, employing effect decomposition to quantify the transmission mechanism of financial risk and obtain the direct contribution of causal paths; calculating the topological differences between the sampled random graph and the baseline graph and evaluating the identification results. In this application, the time sequence and causal order in the financial logic are solidified into hard constraints, reducing the reverse identification error rate, such as "default leads to low income," commonly found in traditional algorithms, to a lower level, ensuring the business compliance of the graph structure. The directed acyclic graph processed by this method improves the BIC score by approximately 0.5%–1.5% compared to purely data-driven algorithms.

Description

Technical Field

[0001] This application belongs to the field of electronic information technology, specifically relating to a method and system for discovering causal topology structures in financial risk control. Background Technology

[0002] Against the backdrop of rapid development in fintech, machine learning models (such as ensemble tree models XGBoost and deep neural networks) have become core tools for credit default prediction due to their powerful nonlinear fitting capabilities. However, these models are essentially "black box attribution," capable of capturing only the statistical correlation between features but failing to reveal the intrinsic causal logic between feature variables and default risk. This leads to significant challenges in practical scenarios such as risk pricing, credit granting decisions, and regulatory compliance (e.g., the EU GDPR and the US FCRA's requirements for decision transparency).

[0003] Currently, academia and industry primarily use explanatory tools such as SHAP for post-hoc analysis. However, in real-world financial scenarios, existing technologies face the following significant bottlenecks: First, financial characteristics (such as income, asset, and liability indicators) generally exhibit high levels of multicollinearity, making traditional correlation-based attribution methods prone to capturing spurious associations lacking physical meaning, leading to attribution bias. Second, purely data-driven causal discovery algorithms (such as the standard PC algorithm) heavily rely on conditional independence tests during the orientation phase, making them highly susceptible to statistical noise and often outputting reverse causal paths that contradict financial common sense (e.g., misjudging "default resulted in lower annual income"). Third, high-dimensional credit data lacks an explicit temporal sequence, resulting in extremely poor robustness in causal orientation identification.

[0004] Therefore, developing a technical solution that can transform the knowledge of financial experts into hard constraints on the underlying topology of algorithms, thereby decoupling risk mechanisms and improving identification accuracy, is a key bottleneck that urgently needs to be overcome in the field of responsible artificial intelligence. Summary of the Invention

[0005] This application provides a method and system for discovering causal topology structures in financial risk control, in order to solve the technical problem that existing technologies are unable to decouple risk mechanisms and improve identification accuracy.

[0006] To address the aforementioned technical problems, this application adopts the following technical solution: a method for discovering causal topology structures in financial risk control, comprising:

[0007] S1. Based on sample space initialization and feature extraction, construct a multi-level feature variable hierarchy;

[0008] S2. Based on injection direction constraints and multi-level feature variable hierarchy, a topology graph is constructed using a causal algorithm;

[0009] S3. Based on the topological structure graph, verify the fitting accuracy of the causal structure and obtain the Bayesian information criterion score;

[0010] S4. Based on Bayesian information criterion scoring, effect decomposition is adopted to quantify the transmission mechanism of financial risk and obtain the direct contribution of causal path;

[0011] S5. Based on the direct contribution of causal paths, calculate the topological differences between the sampled random graph and the baseline graph and evaluate the identification results.

[0012] Furthermore, the method in step S1 includes:

[0013] S11. Collect multiple sets of historical credit data samples and extract feature vectors containing user background characteristics, static financial indicators, dynamic lending behavior, and risk decision-making outputs.

[0014] S12. Based on the key dimensions in the feature vector set, obtain the core risk indicators;

[0015] S13. Construct a hierarchical mapping function based on core risk indicators, credit lifecycle, and business logic;

[0016] S14. Based on the hierarchical mapping function, establish the partial order topological constraints between hierarchical levels.

[0017] Furthermore, the method in step S2 includes:

[0018] S21. Inject Gaussian noise into the original data matrix and perform perturbation processing to obtain the processed matrix; x i

[0019] S22. Based on control set Obtain based on variable x i With variable x j Partial correlation coefficients and their conversion into statistics;

[0020] S23. Based on topological constraints and statistics, obtain the topological structure graph and the direction of hierarchical knowledge injection.

[0021] Furthermore, the method in step S3 includes:

[0022] S31. Accumulate the joint distribution of the entire sample and calculate the global likelihood function value of the topological structure graph;

[0023] S32. Obtain the directed acyclic graph with causal topological structure based on the global likelihood function value. Bayesian Information Criterion ( )score.

[0024] Furthermore, the method in step S4 includes:

[0025] S41. Extract the path chain with loan amount as the intervention source, debt-to-income ratio as the mediating variable, and risk label as the target, and obtain the risk transmission mediating path;

[0026] S42. Based on the risk transmission mediation path, calculate the direct contribution of the causal path caused by the change of the mediating variable.

[0027] Furthermore, the method in step S5 includes:

[0028] S51. Perform sampling with replacement from the original data to generate a B set of random subsets and construct a bootstrap resampling space;

[0029] S52. Based on the self-service resampling space, calculate the topological difference between the sampled random graph and the baseline graph;

[0030] S53. Based on topological differences, calculate the frequency of occurrence of core causal edges in B experiments and evaluate the robustness of the identification results.

[0031] Another technical solution adopted in this application is: a financial risk control causal topology discovery system, comprising:

[0032] The feature mapping module is used to define the physical attributes of financial variables and perform hierarchical system construction.

[0033] The constraint search module is used to execute a causal discovery algorithm that injects hierarchical directional constraints and outputs a directed acyclic graph.

[0034] The statistical validation module is used to calculate the Bayesian information criterion score and perform path effect decomposition.

[0035] The robustness analysis module is used to perform self-sampling experiments and output structural stability indices.

[0036] The beneficial effects of this application are: by injecting topological mapping function constraints into the algorithm orientation layer, the temporal and causal orders in the financial logic of this application are solidified into hard constraints, thereby reducing the error rate of reverse identification such as "default leads to low income" commonly found in traditional algorithms to 0. This ensures the business compliance of the graph structure; the directed acyclic graph processed by the method in this application is... The score improved by approximately 0.5 points compared to a purely data-driven algorithm. -1.5 This invention demonstrates that domain knowledge injection can effectively filter high-dimensional random noise in credit data, making the model more closely resemble the real underlying data generation mechanism (DGP). Furthermore, this invention accurately quantifies the marginal contribution of intermediate paths through NED / NIE decomposition technology. Experiments confirm the proportion of indirect effects. Da The above provides a scientific quantitative basis for financial institutions to shift from "single quota control" to "asset and liability health assessment". Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating an embodiment of the financial risk control causal topology discovery method of this application;

[0038] Figure 2 yes Figure 1 A flowchart illustrating step S1 of an embodiment;

[0039] Figure 3 yes Figure 1 A flowchart illustrating step S2 of an embodiment;

[0040] Figure 4 yes Figure 1 A flowchart illustrating step S3 in one embodiment;

[0041] Figure 5 yes Figure 1 A flowchart illustrating step S4 in one embodiment;

[0042] Figure 6 yes Figure 1 A flowchart illustrating step S5 of an embodiment;

[0043] Figure 7 This is a multicollinearity heatmap of an embodiment of the financial risk control causal topology discovery method of this application;

[0044] Figure 8 This is a schematic diagram of a causal topology based on a five-level hierarchical constraint of domain knowledge, representing an embodiment of the financial risk control causal topology discovery method of this application.

[0045] Figure 9 This is a scatter plot of the structural equation model used in an embodiment of the financial risk control causal topology discovery method of this application to verify the negative causal effect between credit score and interest rate.

[0046] Figure 10 This is a correlation attribution of an embodiment of the financial risk control causal topology discovery method of this application. ) and causal attribution ( A butterfly diagram comparing biases in risk factor identification;

[0047] Figure 11 This is a performance evaluation diagram of the traceability suggestions of an embodiment of the financial risk control causal topology discovery method of this application on three-dimensional indicators of feasibility, effectiveness and intervention cost;

[0048] Figure 12This is the Pareto front curve of the intervention cost versus target effectiveness trade-off of an embodiment of the financial risk control causal topology discovery method of this application;

[0049] Figure 13 This is a comparative analysis diagram of action suggestions generated for specific rejection samples and benchmark methods in one embodiment of the financial risk control causal topology discovery method of this application.

[0050] Figure 14 This is a bar chart comparing Bayesian Information Criterion (BIC) scores under the same level of constraint strength in an embodiment of the financial risk control causal topology discovery method of this application.

[0051] Figure 15 This is a mediating effect decomposition diagram of the "loan amount → debt-to-income ratio → default risk" path in an embodiment of the financial risk control causal topology discovery method of this application;

[0052] Figure 16 This is a line graph showing the robustness evolution of the structural Hamming distance (SHD) under self-sampling perturbation in an embodiment of the financial risk control causal topology discovery method of this application.

[0053] Figure 17 This is a stability distribution diagram of the survival frequency of key causal edges in the topological space of an embodiment of the financial risk control causal topology discovery method of this application;

[0054] Figure 18 This is a verification diagram showing the necessity of introducing causal constraints and monotonicity constraints to improve logical consistency in an ablation experiment of an embodiment of the financial risk control causal topology discovery method of this application.

[0055] Figure 19 This is a performance evaluation ROC curve of the risk intervention decision model under various constraint combinations of an embodiment of the financial risk control causal topology discovery method of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0057] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0058] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the financial risk control causal topology discovery method of this application; the method includes:

[0059] S1. Based on sample space initialization and feature extraction, construct a multi-level feature variable hierarchy.

[0060] Specifically, the method in step S1 includes:

[0061] S11. Sample Space Initialization and Feature Extraction: Collect multiple sets of historical credit data samples and extract feature vectors containing user background characteristics, static financial indicators, dynamic lending behavior, and risk decision-making outcomes. The original credit feature variable sample space is as follows: ;

[0062] S12. Based on the key dimensions in the feature vector set, establish a closed-loop mathematical operator, namely the core risk-causing indicator; among which, the core risk-causing indicator includes the debt-to-income ratio indicator and the monthly repayment amount indicator.

[0063] Among them, the characteristic variables are defined. Representative debt-to-income ratio ( Its specific representation is as follows:

[0064] (1);

[0065] Define feature variables Indicator of monthly repayment amount ( Its definition based on the equal principal and interest repayment model is as follows:

[0066] (2);

[0067] Where L is the loan principal, r is the monthly interest rate, and m is the number of repayment periods.

[0068] In one embodiment, for applicant A (Sample ID:4829), its debt-to-income ratio (DTI) is extracted according to formula (1):

[0069] ;

[0070] The monthly installment is calculated according to formula (2), with the principal set at US$15,000, the annual interest rate converted to a monthly interest rate r = 0.01, and the repayment period m = 36.

[0071] ;

[0072] S13. Multi-level ordered business hierarchy mapping: Constructing a hierarchy mapping function based on the credit lifecycle and business logic. Among them, the ordered hierarchical set of business logic is: .

[0073] in, The layer includes prior attributes such as years of service, property ownership, and geographical location;

[0074] Layer contains Annual income and credit card utilization rate are indicators of debt repayment ability;

[0075] The tier includes indicators for loan application amount and term;

[0076] The layer includes intermediate indicators of approval interest rates generated by institutional decisions;

[0077] Layer is defined as risk default label ;

[0078] As shown in Figure 8, the system divides the feature set into five ordered levels:

[0079] (Background): Length of service, housing ownership;

[0080] (Finance): Annual income, Credit card utilization rate;

[0081] (Contract): Loan amount and term;

[0082] (Pricing): Loan interest rate;

[0083] (Result): Default Risk Label. Hierarchical Relationship Found. Figure 7 The heatmap analysis provides empirical support, demonstrating the necessity of hierarchical processing in the context of multicollinearity.

[0084] S14. Based on the hierarchical mapping function, establish the partial order topological constraints between hierarchical levels.

[0085] Specifically, establish partial order topological constraints between levels: define variables and The rules for allowing pointers between them are:

[0086] (3);

[0087] S2. Based on injection direction constraints and multi-level feature variable hierarchy, a topology graph is constructed using a causal algorithm.

[0088] S21. Data Non-Singularization Preprocessing: Inject Gaussian noise into the original data matrix M. Perform perturbation processing to obtain the processed matrix. The specific formula is as follows:

[0089] (4);

[0090] Among them, execution Test: Calculate the independence statistic between the variable pairs according to formula (5). Construct an undirected skeleton at a significance level of α=0.001.

[0091] S22. Based on Conditional independence scan of transformation: calculating variables and In a given control set Partial correlation coefficient under the condition And convert it into statistics. ;

[0092] S23. Direction-based forced orientation based on hierarchical knowledge injection: Invoke the topological constraints in step S14, and for any connecting edge in the skeleton graph... its direction The determination rule is as follows:

[0093] (6);

[0094] This eliminates the reverse causal logic caused by statistical noise.

[0095] Specifically, for discovering connected edges The system retrieved the "DTI" and "Loan Interest Rate" values ​​and found that Tier(DTI) = 1 and Tier(Interest Rate) = 3.

[0096] Based on the judgment logic of formula (6): since 1 < 3, the system forcibly cuts off the feedback loop of "interest rate → DTI", fixing the direction to DTI → interest rate. The generated final causal topology is as follows: Figure 8 As shown. This step ensures that the generated cause-effect graph does not contain reverse logic (such as "default risk affects annual revenue"), achieving a technical effect of zero error rate in logic identification.

[0097] S3. Based on the topological structure graph, verify the fitting accuracy of the causal structure and obtain the Bayesian information criterion score.

[0098] The method of step S3 includes:

[0099] S31. Global Log-Likelihood Calculation: Accumulate the joint distributions of the entire sample n and calculate the global likelihood function value of the whole graph. The details are as follows:

[0100] (7);

[0101] in, For nodes The set of parent nodes, These are the regression fitting parameters;

[0102] S32. Obtain the directed acyclic graph with causal topological structure based on the global likelihood function value. Bayesian Information Criterion ( The scoring formula is as follows:

[0103] (8);

[0104] Where k is the number of edge parameters in the graph structure.

[0105] This embodiment verifies the scientific validity of the structure through comparative experiments. Experimental design: Comparing a "hierarchical constrained" T-DAG with an "unconstrained native PC graph". Experimental results ( Figure 14 The BIC score of the baseline unconstrained structure is 127,085,847.0; the BIC score of the T-DAG structure in this application is 126,592,07.3. The score reduction is calculated according to formula (8).

[0106] Experiments have shown that by incorporating domain knowledge guidance, the model achieves a higher accuracy fit to financial data within a simpler parameter space.

[0107] S4. Based on the Bayesian Information Criterion (BIC) scoring, effect decomposition is used to quantify the transmission mechanism of financial risk and obtain the direct contribution of causal paths.

[0108] Specifically, the method in step S4 includes:

[0109] S41. Extract the path chain with loan amount as the intervention source, debt-to-income ratio as the mediating variable, and risk label as the target to obtain the risk transmission mediating path.

[0110] S42. Based on the risk transmission mediation path, calculate the direct contribution of the causal path caused by the change of the mediating variable.

[0111] Specifically, mediation effect quantification: calculating the direct contribution (NIE) of the causal path caused by changes in the mediating variable, i.e., the direct contribution of the causal path, as follows:

[0112] (9);

[0113] in, The effect coefficient of the intervention source on the mediating variable. The coefficient representing the direct impact of the mediating variable on the risk objective.

[0114] S5. Based on the direct contribution of causal paths, calculate the topological differences between the sampled random graph and the baseline graph and evaluate the identification results.

[0115] Specifically, the method in step S5 includes:

[0116] S51. Perform sampling with replacement from the original data to generate a B-group random subset, and construct a bootstrap resampling space, in which... B represents self-service sampling. The number of iteration rounds;

[0117] S5.2. Based on the self-service resampling space, calculate the topological difference between the sampled random graph and the baseline graph. The specific formula is as follows:

[0118] (10);

[0119] Where A, M, and R represent the number of edges added, missing, and reversed, respectively;

[0120] S53. Based on the aforementioned topological differences, calculate the frequency of occurrence of the core causal edge in B experiments, and evaluate the anti-disturbance capability of the identification results.

[0121] To verify the system's reliability in industrial scenarios, stress tests were performed.

[0122] Test design: Perform 10 rounds of bootstrapping with a sampling rate of 80%. ).

[0123] Test results (Figures 16 and 17):

[0124] SHD test: such as Figure 16 As shown, the mean Structural Hamming Distance (SHD) between the graph structure generated in each sampling round and the baseline graph is less than 1.0, indicating no significant topological jumps.

[0125] Survival rate: such as Figure 17 As shown, the frequency of occurrence of the critical risk-causing path "DTI→Risk" is F=1.0. This verification ensures that the causal transmission map generated in this application has high robustness.

[0126] Example 1

[0127] To verify the effectiveness of the implementation of this invention, the application Numerical simulation was performed for verification.

[0128] This embodiment selects four different dimensions of credit subsets (sample set A, sample set B, sample set C, and sample set D), each containing 20,000 observations. Based on the above data, 14 feature variables are extracted, including annual income, debt-to-income ratio (DTI), FICO credit score, loan amount, and approval interest rate. Figure 9 This invention is based on structural equation modeling (SEM). The causal effect diagram of credit score on interest rate verified by the model shows that the model accurately captures the linear negative correlation mechanism of risk pricing.

[0129] The first round of feature causal association screening was conducted using the conditional independence test, as shown in Table 1.

[0130] Table 1 shows the hypothesis testing results for the characteristic variables of the four credit sample sets.

[0131]

[0132] As shown in Table 1, the p-values ​​for annual income, DTI, credit card utilization rate, FICO score, and loan amount in all four sample sets are less than 0.05, indicating a significant statistical association between these characteristic variables and default risk. Meanwhile, the p-values ​​for years of employment and homeownership... A value greater than 0.05 in some samples indicates strong independence in risk assessment. After the first round of screening, samples with significant associations were retained. Each feature variable enters the second round of causal topology construction.

[0133] The second round of validation employed the Bayesian Information Criterion (BIC) scoring method to compare the goodness of fit between the proposed hierarchical constrained T-DAG and the unconstrained native PC graph. The original BIC values ​​for the four sample sets were calculated, and the specific data are shown in Table 2. Table 2 shows that the native algorithm, lacking domain knowledge guidance, is prone to overfitting on redundant edges, resulting in higher scores. After injecting hierarchical constraints according to the method described in this invention, the BIC values ​​of all sample sets significantly decreased, and the final retained BIC results are shown in Table 3.

[0134] Table 2 shows the original causal structures of the four sample sets. Rating (unrestricted)

[0135]

[0136] Table 3 shows the samples retained after filtering the four sample sets. Rating (of this invention)

[0137]

[0138] The survival frequency of causal edges identified from the four sample sets was statistically analyzed. Figure 16, 17 The stability score diagrams for the four materials under random resampling are shown in Table 4. The top three critical paths with the highest risk contribution are finally identified.

[0139] Table 4 shows the core risk transmission paths identified in the four sample sets.

[0140]

[0141] The subsequent risk intervention plan will be generated using the path logic in Table 4. Through the analysis of... By tracing back each rejected sample, this invention achieves a 100% success rate in the actionability metric, significantly outperforming benchmark methods.

[0142] Figure 19 To compare the performance of this invention on sample set A using ROC curves, the intervention success rates of the four ablation experiments were observed. Because the complete framework of this invention simultaneously locks in the causal transmission path and logical monotonicity, its curve is closest to the upper left corner of the coordinate system, achieving an AUC value of 0.99, significantly outperforming scheme A which lacks domain knowledge guidance. This verifies the robustness of this invention in identifying risk transmission mechanisms in complex credit environments.

[0143] Example 2

[0144] In this embodiment, the "unconstrained data-driven" mode was selected as the control group for sample set A; the "monotonicity constraint only" mode was selected as the control group for sample set B; and the "hierarchical constraint only" mode was selected as the control group for sample set C. The performance benchmarking results are as follows: Figure 18 As shown.

[0145] As can be seen from the figure, due to the lack of the dual closed-loop mechanism of "hierarchical constraints + monotonicity verification" described in this invention, the control group often exhibits logical paradoxes such as "suggesting users reduce their age" or "suggesting increasing debt to reduce risk" when generating retrospective suggestions, leading to a significant decrease in executability. In contrast, the method of this invention maintains optimal logical consistency while ensuring high effectiveness, demonstrating that the domain knowledge guidance mechanism described in this invention is crucial for financial risk control scenarios.

[0146] Example 3

[0147] Comprehensive application cases and ablation experiment analysis

[0148] A retrospective recommendation was generated for an applicant who was rejected by the system due to high debt (DTI=33.4%). Figure 13 ).

[0149] Solution Comparison: Traditional Method (SHAP): such as Figure 13As shown, due to collinearity interference, the incorrect attribution is "excessively high interest rates," leading to ineffective suggestions to lower interest rates. The method of this invention accurately identifies the root cause of the risk as... Indicators. System basis. Figure 12 The Pareto front seeks the cost-optimal path, outputting the solution: "Repay existing debt to..." "Reduced by 5.2%".

[0150] Verification result: After implementing the suggestion, the user's probability of default successfully decreased below the approval threshold. For example... Figure 18 As shown, ablation experiments demonstrate that the proposed feasibility is only realized when both causal and monotonic constraints are simultaneously injected. Only then can it reach 100%.

[0151] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0152] (1) Completely eliminates the fallacy of reverse causality: by injecting topology mapping functions into the algorithm orientation layer. The application solidifies the time sequence and causal order in the financial logic into hard constraints, reducing the error rate of reverse identification such as "default leads to low income" which is common in traditional algorithms to 0%, thus ensuring the business compliance of the graph structure.

[0153] (2) Significantly improved statistical fit: Experimental data show that the directed acyclic graph processed by the method of this invention improves the BIC score by about 0.5%-1.5% compared with the pure data-driven algorithm, proving that domain knowledge injection can effectively filter high-dimensional random noise in credit data and make the model more in line with the real underlying data generation mechanism (DGP).

[0154] (3) Achieved in-depth analysis and quantification of risk mechanisms: This invention uses (NDE / NIE) decomposition technology to accurately quantify the marginal contribution of intermediary paths (such as loan amounts being transmitted to risk through DTI). Experiments have shown that the indirect effect accounts for more than 70%, providing a scientific quantitative basis for financial institutions to shift from "single quota control" to "asset and liability health assessment";

[0155] (4) It has extremely high topological robustness and industrial reliability: Through self-resampled verification, the survival frequency F of the core causal edge is stable above 0.98, and the structural Hamming distance SHD remains at an extremely low level, proving that the structure recognition result can effectively resist the sample disturbance caused by financial market fluctuations.

[0156] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for discovering causal topological structures in financial risk control, characterized in that, include: S1. Based on sample space initialization and feature extraction, construct a multi-level feature variable hierarchy; S2. Based on the injection direction constraint and the multi-level feature variable hierarchy, a topology graph is constructed using a causal algorithm; S3. Based on the aforementioned topological structure diagram, verify the fitting accuracy of the causal structure and obtain the Bayesian information criterion score; S4. Based on the Bayesian information criterion scoring, the transmission mechanism of financial risk is decomposed and quantified to obtain the direct contribution of the causal path; S5. Based on the direct contribution of the causal path, calculate the topological difference between the sampled random graph and the baseline graph and evaluate the identification results.

2. The method according to claim 1, characterized in that, The method of step S1 includes: S11. Collect multiple sets of historical credit data samples and extract feature vectors containing user background characteristics, static financial indicators, dynamic lending behavior, and risk decision-making outputs. S12. Based on the key dimensions in the feature vector set, obtain the core risk indicators; S13. Construct a hierarchical mapping function based on core risk indicators, credit lifecycle, and business logic; S14. Based on the hierarchical mapping function, establish the partial order topological constraints between hierarchical levels.

3. The method according to claim 2, characterized in that, The method of step S2 includes: S21. Inject Gaussian noise into the original data matrix and perform perturbation processing to obtain the processed matrix; S22. Based on control set , obtain based on variables With variables Partial correlation coefficients and their conversion into statistics; S23. Based on the topological constraints and the statistics, obtain the topological structure diagram and the direction of hierarchical knowledge injection.

4. The method according to claim 3, characterized in that, The method of step S3 includes: S31. Accumulate the joint distribution of all samples and calculate the global likelihood function value of the topological structure graph; S32. Based on the global likelihood function value, obtain the causal topological directed acyclic graph. The Bayesian information criterion score.

5. The method according to claim 1, characterized in that, The method of step S4 includes: S41. Extract the path chain with loan amount as the intervention source, debt-to-income ratio as the mediating variable, and risk label as the target, and obtain the risk transmission mediating path; S42. Based on the aforementioned risk transmission mediation path, calculate the direct contribution of the causal path caused by the change of the mediating variable.

6. The method according to claim 1, characterized in that, The method of step S5 includes: S51. Perform sampling with replacement from the raw data to generate... A set of random subsets is used to construct a self-sampling space; S5.

2. Based on the self-service resampling space, calculate the topological difference between the sampled random graph and the baseline graph; S53. Based on the aforementioned topological differences, calculate the core causal edges in... The frequency of occurrence in this experiment is used to evaluate the robustness of the identification results against disturbances.

7. A financial risk control causal structure discovery system applying the method of any one of claims 1-6, characterized in that, include: The feature mapping module is used to define the physical attributes of financial variables and perform hierarchical system construction. The constraint search module is used to execute a causal discovery algorithm that injects hierarchical directional constraints and outputs a directed acyclic graph. The statistical validation module is used to calculate the Bayesian information criterion score and perform path effect decomposition. The robustness analysis module is used to perform self-sampling experiments and output structural stability indices.

Images