Merchant subject risk assessment method and system based on machine learning

Abstract

The invention discloses a commercial subject risk assessment method and system based on machine learning, and belongs to the technical field of knowledge maps. Basic features of an enterprise are obtained, a knowledge graph is constructed, and risk features are identified through graph mining; carrying out qualitative analysis and quantitative analysis on the enterprise to obtain a feature tag description; constructing a sample set, setting a misjudgment cost and a class weight based on the sample set, performing parameter tuning and verification set feedback, adjusting a decision threshold, performing binning on each feature, and obtaining a final risk score; and threshold selection and probability calibration are carried out to obtain risk grade division and decision suggestions of the enterprise, and risk label interpretation and large model auxiliary interpretation are carried out. The problems that manual collection and analysis of supervision policy documents consumes time and labor, careless omissions easily occur, and final risk judgment is prone to misalignment, misjudgment and frequent missed judgment are solved. According to the method, key rules can be accurately extracted, and hidden risk conduction links and potential risk points can be mined.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and specifically to a commercial entity risk assessment method and system based on machine learning. Background Art

[0002] In the current field of enterprise risk assessment, traditional technical means have many limitations. On the one hand, the use of regulatory policies often remains at the level of manual interpretation. Regulatory policy documents are complex and contain a large amount of content. Manual collection and analysis are not only time-consuming and laborious, but also prone to omissions, resulting in the inability to accurately and comprehensively extract key rules. For example, prohibition and recommended prohibition rules are often misjudged or omitted, which makes risk assessment lack a solid basis.

[0003] The risk identification process also has shortcomings, lacking a systematic dimensionality reduction and screening mechanism. Faced with a large number of initial risk features, not only does it increase the computational cost, but it also introduces a lot of noise interference, which masks key risks and leads to inaccurate final risk judgments, frequent misjudgments and missed judgments. Summary of the invention

[0004] The purpose of the present invention is to provide a commercial entity risk assessment method and system based on machine learning to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions: In a first aspect, a commercial entity risk assessment method based on machine learning comprises the following steps: Collect and parse multi-level regulatory policy documents, extract rules, and store them in the database; these rules are an important basis and foundation for subsequent risk assessments, and provide compliance and risk judgment criteria for subsequent operations. Based on these rules, we can more accurately identify the risk points that companies may face during operations.

[0006] Obtain the basic characteristics of the enterprise, map the basic characteristics into a graph data structure, and construct a knowledge graph; by constructing a knowledge graph, various basic characteristics of the enterprise are linked in a graphical manner, so that the originally scattered data form an organic whole, providing a structured framework for subsequent risk feature mining, making it easier to discover potential risk clues from a complex relationship network.

[0007] Based on the knowledge graph, graph mining is used to identify risk features, and the risk features are structured to form a first risk feature set; the first risk feature set includes various risk features mined from the knowledge graph, which are a preliminary presentation of the enterprise's risk status; and the next step is to further screen and optimize these features in order to more accurately assess enterprise risks.

[0008] The first risk feature set is subjected to dimensionality reduction and filtering, and is verified by business experts to obtain a second risk feature set. The second risk feature set obtained through this step is more refined and accurate than the first risk feature set, providing more reliable data support for subsequent risk analysis.

[0009] A qualitative rule system is established to conduct qualitative analysis on enterprises. For enterprises whose risk levels cannot be determined through qualitative analysis, a quantitative analysis is conducted based on the second risk feature set to obtain feature label descriptions and a third risk feature set. Qualitative analysis preliminarily determines the risk level of the enterprise, and for those enterprises that cannot be determined through qualitative analysis, quantitative analysis based on the second risk feature set can more accurately characterize the risk status of the enterprise. The feature label descriptions obtained through quantitative analysis further clarify the meaning and impact of each risk feature, thereby obtaining a more accurate third risk feature set, providing a more convincing data basis for risk assessment.

[0010] Construct a sample set, set the misjudgment cost and class weight based on the sample set, perform parameter tuning and validation set feedback, and adjust the decision threshold; constructing a sample set provides a data basis for model training. Setting the misjudgment cost and class weight can balance the impact of different types of samples on the model. Through parameter tuning and validation set feedback, the performance of the model can be continuously optimized and the most appropriate decision threshold can be found; the decision threshold will directly affect the accuracy of subsequent risk assessment and provide a reasonable judgment standard for scoring and grading based on the risk feature set.

[0011] Based on the third risk feature set, the risk is divided into boxes to obtain the original score, and the fit test and business indicator assessment are performed to obtain the final risk score; threshold selection and probability calibration are performed, and based on the final risk score and feature label description, the enterprise's risk level classification and decision-making suggestions are obtained, and risk label interpretation and large-scale model-assisted interpretation are performed; the original score obtained by binning based on the third risk feature set is subjected to a fit test and business indicator assessment to ensure the rationality and effectiveness of the score, thereby obtaining the final risk score; based on this final risk score and the feature label description obtained previously, through threshold selection and probability calibration, the enterprise can be accurately classified into risk levels and decision-making suggestions can be given; finally, the large model is used to assist in the interpretation of risk labels, making the risk assessment results easier to understand and apply, and providing valuable reference for enterprise managers and relevant decision makers.

[0012] In combination with the first aspect, in a first implementation of the first aspect of the present application, the collecting and parsing of multi-level regulatory policy documents, extracting rules, and storing them in a database includes: The rules include prohibition rules, recommended prohibition rules and key focus rules; When the multi-level regulatory policy document is a scanned copy or unstructured text, it is converted into a parsable text format using an OCR tool, where OCR stands for optical character recognition; the Chinese word segmentation tool is used to segment and tag the text; An entity recognition model is used to identify key entity types from the text, wherein the entity recognition model is BERT+CRF or a named entity recognition model based on deep learning, BERT represents a bidirectional encoder representation, and CRF represents a conditional random field; a syntactic analysis tool is used to perform dependency syntactic analysis on the text to obtain dependency relationships; based on the dependency relationships, risk rule expressions are identified; the risk rule expressions include conditional constraint rules, principle requirement rules, and operational requirement rules; based on the risk rule expressions, they are mapped into qualitative rules and stored in a database; repeated, similar, or conflicting rules in the rules are merged and prioritized.

[0013] In combination with the first aspect, in a second implementation of the first aspect of the present application, the step of obtaining basic features of an enterprise, mapping the basic features to a graph data structure, and constructing a knowledge graph includes: Obtain the basic characteristics of the enterprise and construct a directed weighted graph G=(V,E); where V represents the node set. For the risk of "one person in multiple enterprises" and the risk of "cross-appointment of senior executives", the node set V includes the first enterprise node and the first natural person node , i is the index of the first enterprise node, j is the index of the first natural person node; for the risk of "registration anomaly", the node set V includes the first enterprise node , Address Node and industry nodes ; E represents the edge set. For the risk of "one person with multiple enterprises", the equity edge is ,express of enterprises The proportion of equity in the enterprise , shareholders are ,express The natural person is The shareholders of the enterprise have a shareholding ratio of ; The control right of each path is the product of the equity ratios of all edges in the path; For the risk of "cross-appointment of senior executives", the appointment edge is ,express A natural person in For the risk of "registration anomaly", the registered address is ,express of companies registered in The address of the industry classification is ,express of enterprises belong to industry.

[0014] In combination with the first aspect, in a third implementation of the first aspect of the present application, the risk features are identified based on the knowledge graph using graph mining, and the risk features are structured to form a first risk feature set, including: The graph mining includes community detection, path analysis and similarity calculation; By tracing the equity structure of enterprises, the ultimate controller is found. When the enterprises in the first enterprise set are finally gathered together by the same natural person, and the enterprises in the first enterprise set are defined as shell companies, it is marked as a risk of "one person in multiple enterprises"; when it is found that the same natural person serves as the legal representative or senior executive of multiple enterprises within the user-defined time period, the multiple enterprises are marked as the second enterprise set. When the frequency of the enterprises in the second enterprise set in the multi-level regulatory policy documents reaches the set threshold, it is marked as a risk of "cross-appointment of senior executives"; by analyzing the registered address attributes of the enterprise node, when it is found that multiple enterprises are registered at the same address, the multiple enterprises are marked as the third enterprise set. When the third enterprise set is indicated in the multi-level regulatory policy documents as involving illegal operations, it is marked as a risk of "registration abnormality"; The extracted risk features are output in a structured form to form a first risk feature set.

[0015] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, the first risk feature set is subjected to dimensionality reduction and filtering, and business expert verification to obtain a second risk feature set, including: For the risk features in the first risk feature set, abnormal risk features are eliminated using feature statistical attributes, and the feature statistical attributes include variance, missing rate and information entropy; correlation analysis is performed on the abnormal risk features after elimination, key risk patterns are screened, and the patterns are submitted to business experts for qualitative evaluation to obtain a second risk feature set.

[0016] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, establishing a qualitative rule system to conduct a qualitative analysis of the enterprise includes: Construct basic characteristic indicators and green enterprise database indicators for enterprises, wherein the basic characteristic indicators include basic information of the enterprises, and the enterprises in the green enterprise database indicators are judged to be low-risk types; based on the basic characteristic indicators and the green enterprise database indicators, determine whether the enterprises are high-risk or low-risk.

[0017] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, for an enterprise whose risk level cannot be determined by qualitative analysis, a quantitative analysis is performed based on the second risk feature set to obtain a feature label description, and a third risk feature set is obtained, including: For enterprises whose risk level cannot be determined through qualitative analysis, the risk characteristics therein are divided into boxes based on the second risk characteristic set, each box includes a number of samples, and each sample is marked as a positive sample or a negative sample; the proportion of positive samples and negative samples in each box is counted, the WOE in each box is calculated, and the IV value is accumulated to obtain the IV value; when there is a risk characteristic that has information gain for the final risk judgment result and its IV value is greater than the set threshold, it is retained; when there is a risk characteristic whose IV value is judged to be close to 0 within the set threshold, it is eliminated; SHAP value analysis is used for risk features that have passed IV value screening. Risk features with positive SHAP values ​​in high-risk entities and values ​​exceeding the set threshold are retained, otherwise they are eliminated or downgraded. Inter-feature correlation test is performed on risk features after SHAP value analysis. When the correlation between two risk features exceeds the set threshold and one of the risk features is considered to be able to represent this type of risk information after review by business experts, redundant features are deleted. The third risk feature set is determined to obtain feature label descriptions.

[0018] In combination with the first aspect, in a seventh implementation of the first aspect of the present application, constructing a sample set, setting a misjudgment cost and a class weight based on the sample set, performing parameter tuning and verification set feedback, and adjusting a decision threshold include: Construct a sample set, including whitelist samples and blacklist samples. The whitelist samples are a list of low-risk enterprises, and the blacklist samples are a list of high-risk enterprises. Use CatBoost to adjust the class weights. After feature engineering, oversample the blacklist samples and undersample the whitelist samples. Set the misjudgment cost and class weights, measure the blacklist recognition rate and whitelist false positive rate under different parameter combinations on the validation set, and gradually iterate and adjust the class weights and parameters to meet the target requirements. When the final output is still a probability score after using CatBoost and adding class weights, the decision threshold is adjusted using the following methods: use different prediction thresholds in the validation set to obtain changes in the false positive rate; find a balance point that can minimize the probability of false positives for low-risk companies while maximizing the recall rate; and perform continuous monitoring and iterative optimization.

[0019] In combination with the first aspect, in an eighth implementation of the first aspect of the present application, the binning is performed based on the third risk feature set to obtain an original score, a fit test and a business indicator assessment are performed to obtain a final risk score; threshold selection and probability calibration are performed, and based on the final risk score and feature label description, the risk level classification and decision-making suggestions of the enterprise are obtained, and risk label interpretation and large model-assisted interpretation are performed, including: Binning each feature in the third risk feature set, analyzing the distribution of feature values, testing and judging the risk feature differences between feature values ​​in different binning intervals, and adjusting the number of bins and binning boundaries; performing iterative optimization and feedback loops during model training and validation to obtain the original score of each feature; Assume the third risk feature set is ,in, represents the raw score of the independent feature parameter, represents the composite score of the combined feature parameters, An index representing the third risk feature set; In the feature parameters, the formula of the comprehensive scoring model is as follows: ; Among them, Score is the comprehensive score of feature parameters. For the The weights of the independent parameters, is a set of combined feature indexes. Assign appropriate weights ; Score is scaled to the range of 0-100 through linear transformation to obtain the original score; The original score is tested for goodness of fit using the business validation data set, evaluated based on AUC, KS value and F1 Score, and fine-tuned to obtain the final risk score; Perform threshold selection and probability calibration, obtain the enterprise's risk level classification and decision-making recommendations based on the final risk score and feature label description, and perform risk label interpretation and large model-assisted interpretation.

[0020] Second, a commercial entity risk assessment system based on machine learning, including: The regulatory policy analysis and knowledge graph construction module includes: a rule extraction unit, a graph mapping unit and a graph mining unit; wherein the rule extraction unit collects and analyzes multi-level regulatory policy documents, extracts rules, and stores them in a database; the graph mapping unit obtains the basic features of the enterprise, maps the basic features to a graph data structure, and constructs a knowledge graph; the graph mining unit uses graph mining based on the knowledge graph to identify risk features, structures the risk features, and forms a first risk feature set; The risk feature screening module includes: a dimensionality reduction and filtering unit and an expert verification unit; wherein the dimensionality reduction and filtering unit performs dimensionality reduction and filtering on the first risk feature set, and the expert verification unit performs business expert verification to obtain a second risk feature set; The qualitative and quantitative analysis module includes: a qualitative analysis unit and a quantitative analysis unit; wherein the qualitative analysis unit establishes a qualitative rule system to perform qualitative analysis on the enterprise; for the enterprise whose risk level cannot be determined by qualitative analysis, the quantitative analysis unit performs quantitative analysis based on the second risk feature set to obtain a feature label description and obtain a third risk feature set; The sample set and decision optimization module includes: a sample construction unit and a parameter tuning unit; wherein the sample construction unit constructs a sample set, sets the misjudgment cost and class weight based on the sample set, and the parameter tuning unit performs parameter tuning and verification set feedback to adjust the decision threshold; The risk scoring and decision-making module includes: a binning scoring unit, a scoring verification unit and a grading unit; wherein the binning scoring unit performs binning based on the third risk feature set to obtain an original score, and the scoring verification unit performs a fit test and a business indicator assessment to obtain a final risk score; the grading unit performs threshold selection and probability calibration, obtains the enterprise's risk grading and decision-making recommendations based on the final risk score and feature label description, and performs risk label interpretation and large model-assisted interpretation.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention collects and parses multi-level regulatory policy documents, performs word segmentation and part-of-speech tagging on the text to obtain a more granular vocabulary sequence, uses an entity recognition model to identify key entity types from the text, performs dependency syntax analysis on the text to obtain dependency relationships, identifies risk rule expressions, and maps them into qualitative risk management rules, including prohibition rules, recommended prohibition rules, and key focus rules; thereby being able to quickly traverse multi-level and massive regulatory policy content, accurately extract key rules, and accurately store them in the database.

[0022] 2. The present invention maps basic features into a graph data structure, constructs a knowledge graph in the form of a directed weighted graph, and uses graph mining to identify risk features. Graph mining includes community detection, path analysis, and similarity calculation. Risk features are structured, dimension reduction and filtering are performed, and business expert verification is performed, thus breaking the limitations of the previous isolated analysis of basic enterprise features, structuring information such as enterprise scale, industry, and business scope, establishing semantic associations between entities, and mining hidden risk transmission links and potential risk points.

[0023] 3. The present invention establishes a qualitative rule system to conduct qualitative analysis on enterprises; conducts quantitative analysis on enterprises whose risk levels cannot be judged through qualitative analysis; constructs a sample set, sets the misjudgment cost and class weight based on the sample set, performs parameter tuning and verification set feedback, and adjusts the decision threshold; performs binning to obtain the final risk score, performs threshold selection and probability calibration, so as to efficiently screen key risks, remove redundant interference, and combine sample optimization with parameter tuning; based on the final risk score and feature label description, obtains the enterprise's risk level classification and decision-making suggestions, performs risk label interpretation and large model-assisted interpretation, and can convert obscure risk labels and complex feature label descriptions into easy-to-understand natural language descriptions, greatly saving the time cost of information interpretation and communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the steps of a commercial subject risk assessment method based on machine learning of the present invention; Figure 2 It is a system structure diagram of a commercial entity risk assessment system based on machine learning in the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0027] According to one embodiment of the present invention, Figure 1 A method for assessing the risk of a commercial subject based on machine learning is shown in a schematic diagram of the steps of the method. The method for assessing the risk of a commercial subject based on machine learning comprises the following steps: Step S100: Collect and parse multi-level regulatory policy documents, extract rules, and store them in a database; obtain basic features of the enterprise, map the basic features to a graph data structure, and construct a knowledge graph; based on the knowledge graph, use graph mining to identify risk features, structure the risk features, and form a first risk feature set; Specifically, collect and parse multi-level regulatory policy documents, and store the contents of the multi-level regulatory policy documents in plain text, PDF or machine-readable format; if the multi-level regulatory policy documents are scanned copies or unstructured texts, use OCR tools to convert them into parsable text formats; use Chinese word segmentation tools to segment and tag the text to obtain a more granular vocabulary sequence; Further, an entity recognition model is used to identify key entity types from the text, wherein the entity recognition model is BERT+CRF or a named entity recognition model based on deep learning, BERT represents a bidirectional encoder representation, and CRF represents a conditional random field; a syntactic analysis tool is used to perform dependency syntactic analysis on the text, and the subject-predicate-object and attributive-adverbial-complement structures in the sentence are identified to obtain dependency relationships; based on the dependency relationships, risk rule expressions are identified; the risk rule expressions include conditional constraint rules, principle requirement rules, and operational requirement rules; based on the risk rule expressions, they are mapped into qualitative risk management rules, including prohibition rules, recommended prohibition rules, and key attention rules; and stored in a database; In a specific embodiment, prohibition rules explicitly prohibit certain types of behaviors or entities from enjoying simple account opening services, such as "simple account opening services are not applicable to those with obvious suspicious characteristics", which can be used as prohibition risk rules; it is recommended that prohibition rules require banks to strengthen audits or restrict functions under certain conditions, such as "non-counter services should be restricted to strengthen risk prevention and control", which can be regarded as restrictive measures for higher-risk accounts; key attention rules emphasize giving special attention and assessment to specific entities, such as small and micro enterprises, or specific characteristics, such as the degree of customer identity verification and account risk level.

[0028] Some examples of rules that can be extracted in this guideline are: Prohibition rules: "If obvious suspicious characteristics are found during the account opening process, the simplified account opening service shall not be used" → If the subject is found to have obvious suspicious characteristics during the account opening process, the simplified account opening service shall be prohibited.

[0029] Prohibition or restrictive rules are recommended: "Banks should establish and improve an account classification and grading management system, focusing on strengthening account risk prevention and control during and after the event by restricting non-counter services" → For entities with higher risks, banks can be advised to restrict non-counter services and increase supervision.

[0030] Key focus rules: "Banks providing simple account opening services should follow the risk-based principle" → emphasize the risk-based monitoring concept and increase due diligence efforts on entities with certain risk characteristics.

[0031] Merge and prioritize duplicate, similar or conflicting risk management rules.

[0032] Specifically, the basic characteristics of the enterprise are obtained to construct a directed weighted graph G=(V,E); where V represents a node set. For the risk of "one person in multiple enterprises" and the risk of "cross-appointment of senior executives", the node set V includes the first enterprise node and the first natural person node , i is the index of the first enterprise node, j is the index of the first natural person node; for the risk of "registration anomaly", the node set V includes the first enterprise node , Address Node and industry nodes ; E represents the edge set. For the risk of "one person with multiple enterprises", the equity edge is ,express of enterprises The proportion of equity in the enterprise , shareholders are ,express The natural person is The shareholders of the enterprise have a shareholding ratio of ; The control right of each path is the product of the equity ratios of all edges in the path; For the risk of "cross-appointment of senior executives", the appointment edge is ,express A natural person in For the risk of "registration anomaly", the registered address is ,express of companies registered in The address of the industry classification is ,express of enterprises belong to industry.

[0033] Specifically, based on the knowledge graph, risk features are identified using graph mining, wherein the graph mining includes community detection, path analysis, and similarity calculation; By tracing the corporate equity structure, the ultimate controller is found. If the companies in the first corporate set are ultimately brought together by the same natural person, and the companies in the first corporate set are defined as shell companies, it is marked as a risk of "one person in multiple companies"; if it is found that the same natural person serves as the legal representative or senior executive of multiple companies within a user-defined time period, the multiple companies are marked as a second corporate set. When the frequency of the companies in the second corporate set in the multi-level regulatory policy documents reaches the set threshold, it is marked as a risk of "cross-appointment of senior executives"; by analyzing the registered address attributes of the corporate node, if it is found that multiple companies are registered at the same address, the multiple companies are marked as a third corporate set. When the third corporate set is indicated in the multi-level regulatory policy documents as involving illegal operations, it is marked as a risk of "registration anomaly"; The extracted risk features are output in a structured form to form a first risk feature set.

[0034] In one embodiment, from the target enterprise node Starting from the target enterprise node, recursive or iterative search is performed along the equity chain; To the target natural person Path , calculate the control right on the path, the calculation formula of the control right is:

[0035] in, represents the equity ratio of the ith edge in the path, Indicates the path The number of hops in Target natural person Through multiple paths and target enterprise nodes The total control over the enterprise is:

[0036] Among them, m is arrive The total number of paths; Set a threshold If the target natural person Overall control over the business , then determine For target companies the actual controller of the company; If the same target natural person When a person controls multiple enterprises and the number of controlled enterprises exceeds the set threshold, it will be marked as a "one person, multiple enterprises" risk.

[0037] Conduct equity penetration analysis, path analysis and control rights calculation: E1: P→E (direct shareholding), control rights C(P,E)=100.

[0038] E2: Path 1: E1→E2, control right C1(P,E2)=100%×80%=80%.

[0039] E3: P→E3 (direct shareholding), control rights C(P,E3)=60%.

[0040] E4: Path 1: E2→E4, control right C1(P,E4)=80%×50%=40%.

[0041] Path 2: P→E4 (direct shareholding), control rights C2(P,E4)=20%.

[0042] Cumulative control: C(P,E4)=40%+20%=60%.

[0043] Cumulative number of controlled enterprises: N(P)=4, and the number of controlled enterprises exceeds 50%.

[0044] Extract features of “one person, multiple companies”: Low registered capital: The registered capital of all enterprises is <200,000 yuan.

[0045] The registered addresses are highly concentrated: all companies are registered at the same address: A in a certain province and city.

[0046] Similar industry attributes: E1, E2, and E4 are all consulting service industries, and E3 is educational services, which is highly related to consulting service business.

[0047] Potential shell company: The registered capital of the enterprise is low and the operating status is unclear.

[0048] According to the above analysis, E1, E2, E3, and E4 controlled by P have the following high-risk characteristics: 1. “One person owning multiple enterprises” risk: P actually controls four enterprises, which meets the identification criteria of “one person owning multiple enterprises”.

[0049] 2. “Abnormal address concentration” risk: the registered addresses of the four companies are exactly the same.

[0050] 3. “Industry homogeneity” risk: The industry attributes of enterprises are highly correlated, and there may be false transactions or interest transfers.

[0051] In another specific embodiment, NLP technology is used to parse documents to find, for example, "Zhang is the legal representative and actual controller of Company A, Company B, Company C, Company D, and Company E", or "Li and Zhang appear in the shareholder lists of multiple companies at the same time". This information will serve as the attributes of the basic nodes and edges for constructing the knowledge graph.

[0052] There are two natural person nodes (Zhang and Li), and they are associated with multiple enterprise nodes (Company A, Company B, Company C, Company D, Company E, Company F). Through graph analysis, we found that: Zhang serves as the legal representative or actual controller of companies A, B, C, D, and E. These companies have similar registration models (established at similar locations during the same period), unclear main businesses, and lack of real operating traces.

[0053] Li and Zhang appear on the shareholder or executive lists of some companies at the same time, forming a complex network relationship in which executives hold cross-appointments.

[0054] If case analysis confirms that these companies are shell companies used to open public accounts in batches and resell them at a marked-up price, the entire cluster of companies controlled by Zhang and Li will be marked as a high-risk cluster in the map, and the characteristics will be refined: "obvious cross-appointments", "one person in multiple companies", "concentrated registration time and industry", etc.

[0055] In addition, by associating with the behavioral pattern labels in the cases involved in the lawsuit (such as "using shell companies to resell corporate accounts"), it can be further confirmed that the network structure represented by these nodes and edges belongs to a typical high-risk business model, thereby providing high-value risk characteristic parameters for subsequent model construction.

[0056] Step S200: Dimension reduction and filtering are performed on the first risk feature set, and business expert verification is performed to obtain a second risk feature set; Specifically, for the risk features in the first risk feature set, abnormal risk features are preliminarily eliminated using feature statistical attributes, and the feature statistical attributes include variance, missing rate and information entropy; correlation analysis is performed on the eliminated abnormal risk features, key risk patterns are screened, and the features are submitted to business experts for qualitative evaluation to obtain a second risk feature set.

[0057] In a specific embodiment, feature statistical properties, such as variance, missing rate, and information, are used to preliminarily eliminate features that are constant or severely missing in most subjects; Focusing on key risk patterns identified in case studies, such as "one person in multiple companies", "cross-appointments" and "industry and time concentration", we screened indicators that have significant differentiation for these patterns in the feature set through association analysis, such as the number of companies with the same actual controller, the registration concentration index, and the cross-appointment rate of shareholders; Submit the shortlisted candidate features to business experts, such as compliance reviewers and risk control analysts, for qualitative evaluation, and retain features that are closely related to actual risk assessments and have strong business interpretability.

[0058] After the above process, 200 core parameters were finally refined from 700 features. These parameters have both data stability, business relevance and model availability, providing a high-quality input foundation for the construction and training of subsequent machine learning models.

[0059] Step S300: Establishing a qualitative rule system to conduct qualitative analysis on the enterprise; for the enterprise whose risk level cannot be determined by qualitative analysis, quantitative analysis is conducted based on the second risk feature set to obtain feature label descriptions and obtain a third risk feature set; Specifically, basic characteristic indicators and green enterprise database indicators are constructed for enterprises, wherein the basic characteristic indicators include basic information of the enterprises, and the enterprises in the green enterprise database indicators are directly judged as low-risk types; based on the basic characteristic indicators and the green enterprise database indicators, it is judged whether the enterprises belong to high risk or low risk.

[0060] For enterprises whose risk level cannot be determined through qualitative analysis, the risk characteristics therein are divided into boxes based on the second risk characteristic set, each box includes a number of samples, and each sample is marked as a positive sample or a negative sample; the proportion of positive samples and negative samples in each box is counted, and the WOE in each box is calculated. The WOE reflects the degree of deviation of the box from the overall positive and negative sample distribution, and the IV value is accumulated; when there is a risk characteristic that has information gain for the final risk judgment result, and its IV value is greater than the set threshold, it is retained; when there is a risk characteristic whose IV value is judged to be close to 0 within the set threshold, it is eliminated; SHAP value analysis is used for risk features that have passed IV value screening. If there is a risk feature with a positive SHAP value in high-risk entities and the value exceeds the set threshold, it will be retained; otherwise, it will be eliminated or downgraded. The risk features analyzed by SHAP value are tested for correlation between features, and the Pearson correlation index, Spearman correlation coefficient or mutual information index is used to measure the correlation between features. If the correlation between two features exceeds the set threshold, and one of the features is considered to be able to represent this type of risk information after review by business experts, the redundant features will be deleted. After verification by business experts, the third risk feature set is determined to obtain feature label descriptions.

[0061] In a specific embodiment, during the qualitative analysis phase, the system first focuses on whether the subject has obvious high-risk characteristics, such as prohibited activity associations, or low-risk characteristics, such as green enterprise library characteristics, so as to quickly classify some subjects as high-risk or low-risk, thereby reducing the workload of subsequent quantitative analysis.

[0062] Basic feature indicators provide a basic profile of the target entity. These features themselves do not directly represent risks, but provide context for subsequent rule judgments. For example: Entity type B01: Distinguish between entities such as enterprises, individual businesses, public institutions, and financial institutions. For public institutions, party and government agencies, or specific types of institutions, the Green Enterprise Database indicators may be more easily applicable and directly reduce their initial risk assessment level.

[0063] Industry attribute B02: Identify whether the industry in which the subject is located is cash-intensive (such as catering, entertainment), trade distribution, traditional manufacturing, or emerging technology industry. Industry attributes can provide a reference for qualitative judgment: in the green enterprise database or the list of high-quality enterprises, some industries or types of entities naturally have a higher degree of information disclosure and a lower risk tendency.

[0064] Enterprise size B03, economic type B05, address characteristics B06: For enterprises of economic types such as state-owned enterprises and central enterprises or enterprises located in special areas, if they have relevant records in the green enterprise database or are officially certified, qualitative analysis can quickly assess them as low-risk entities.

[0065] Basic characteristic indicators help the qualitative rule system to make more accurate judgments. The same rule may have different risk implications for a large central enterprise and a newly established small and micro enterprise. By identifying the scale, nature and industry attributes of the enterprise, conditions can be provided for qualitative classification. For example, for the central enterprises and their wholly-owned subsidiaries (L03 indicator hits) in the green enterprise database, if they are also the leaders of the national key industries, they can be directly regarded as low-risk without entering the subsequent complicated quantitative analysis link.

[0066] In the qualitative analysis rule base, the following example rules can be set: Rule 1 (low risk priority rule): Conditions: The entity meets the L04 indicator (listed company or public debt-issuing enterprise) and B03 shows the scale of a large or medium-sized enterprise, and is a state-controlled enterprise in terms of B05 economic type.

[0067] Action: Directly mark the subject as low risk without entering into quantitative analysis.

[0068] Rule 2 (Green Enterprise Fast Track): Condition: The subject hits L01 (Party and government organs, public institutions) or L05 (collective economic organizations).

[0069] Action: Directly classify it as low risk and provide simplified due diligence advice.

[0070] Rule 3 (Basic characteristics correct risk judgment): Conditions: The subject may be included in the observation due to the characteristics of the high-risk industry (policy prohibition or high-risk indicators), but if the subject also hits L08 (licensed financial institution) and B10 taxpayer credit rating is high.

[0071] Action: Reassess and reduce the subject risk level from high risk to medium-low risk, or conduct a small amount of targeted additional quantitative analysis.

[0072] Through the above expansion, the qualitative analysis stage makes full use of basic characteristic indicators to establish an objective portrait of the subject, and combines the green enterprise database indicators to quickly identify high-quality subjects that do not require in-depth analysis, and concentrates analysis resources on subjects in the middle gray area, providing a basis for improving efficiency and accuracy in the subsequent quantitative analysis stage.

[0073] In the quantitative analysis stage, the system conducts a more in-depth risk quantitative assessment of entities that are difficult to judge qualitatively. At this time, risk indicators (including 100 risk indicators in 11 categories, including policy prohibition, high risk, abnormal entity, abnormal contact information, abnormal address, abnormal decision maker, abnormal legal representative, abnormal equity governance structure, abnormal corporate history, abnormal related parties, and shell company characteristics) become the core focus. These risk indicators not only describe the potential problems of the subject itself, but also reflect the potential correlation and transmission risks between subjects through knowledge graphs and graph mining technology.

[0074] In the early stage of comprehensive feature analysis and parameter extraction, 700 potential features have been refined into 200 core parameters. These 200 parameters include three major sources of features: Business risk indicators: such as Class Z policy prohibition indicators, Class G high-risk indicators, Class Q entity abnormal indicators, Class M shell company characteristic indicators, etc.

[0075] Graph structure features: relationship features extracted from the knowledge graph (such as path length between enterprises, community risk concentration, and controller association index).

[0076] Basic and auxiliary characteristics: may include derivative parameters, combined characteristics and statistical variables of basic characteristic indicators that are not directly used for judgment in the qualitative stage (such as the rate of change of registered capital, address concentration score, abnormal proportion of affiliated enterprises, etc.).

[0077] First, the IV value is calculated for these 200 parameters. The IV value is used to measure the discrimination and predictive power of a single feature on the target variable (such as whether to reject account opening and enhance due diligence).

[0078] If there is a risk indicator that has a significant information gain for the final risk judgment result, its IV value is high and can be retained first. For features with too low IV values ​​(such as close to 0), it means that the contribution to the result prediction is extremely low and can be eliminated in this round of screening. The SHAP value analysis is used for the parameter set that has passed the initial screening of the IV value. The SHAP value provides the marginal contribution and direction of each feature in the model decision, which can help identify the features that have the most obvious impact on the decision results, and at the same time reflects the nonlinear relationship and interaction between the features and the prediction results. If a risk indicator (such as "the proportion of multiple companies involved in the same address") has a positive SHAP value and a large value in high-risk subjects, it means that the indicator has a key contribution to distinguishing high-risk and low-risk subjects. If the SHAP value of an indicator fluctuates greatly or has a low contribution value in various scenarios, its predictive effect is unstable and can be considered for further elimination or downgrading in this round. Through SHAP analysis, the candidate parameter set can be further reduced from about 150-160 to about 120-130 parameters.

[0079] The remaining parameter sets are tested for correlation between features. For example, if there are two risk indicators that are statistically highly positively correlated and overlap in business meaning (such as "legal representative controls many enterprises" and "suspected actual controller controls many enterprises" may have similar information in some scenarios), the feature with clearer business meaning, higher data stability or more interpretability is retained. The Pearson correlation coefficient, Spearman correlation coefficient or mutual information index can be used to measure the correlation between features. If the correlation between two features is too strong (such as correlation coefficient>0.8), and after review by business experts, it is considered that one of the features is sufficient to represent this type of risk information, the redundant feature is deleted. Through the correlation test, 120-130 parameters can be reduced to about 108 independent parameters, and 36 combined parameters can be further constructed (such as combining "inaccurate registered address" (T10) and "high contact information duplication rate" (E06) in the address anomaly class into a comprehensive indicator reflecting the confusion of corporate information disclosure), and finally 108+36=144 final model scoring parameters are formed.

[0080] After algorithm screening, the list of 144 pre-selected parameters will be submitted to senior compliance experts, risk managers and business analysts for manual review. Experts will verify the parameters based on the following factors: Business logic rationality: confirm that the features are explanatory for real risk management scenarios; for example, "one person, multiple companies" type features can indeed constitute a risk association with shell or abnormal account groups. Data availability and stability: If a feature data is not updated in a timely manner or is difficult to obtain continuously in the actual production environment, it can be replaced or adjusted. Regulatory compliance and practical experience: Experts are familiar with practical cases and regulatory requirements, and can judge whether certain features meet regulatory compliance standards and whether there is a risk of misjudgment. If business experts believe that certain indicators are not fully meaningful in actual scenarios or may lead to misjudgment, for example, a feature is difficult to define the legality in practice or is susceptible to data noise, then the feature will be downgraded or replaced with a suitable alternative indicator to ensure that the final feature set is usable in practice and the risk is controllable.

[0081] After completing the above automation and manual verification steps, 144 model scoring parameters are finally determined. Among them, 108 are independent parameters, reflecting the independent contribution of various risk characteristics; 36 are combined parameters, which increase the ability to distinguish specific risk scenarios by combining two or more characteristics. For example, some characteristics of "corporate evolution anomalies (X category)" and "shell company characteristics (M category)" are combined to characterize the abnormal changes of enterprises in the time dimension and the multiple risk characteristics of shell operations. The confirmed 144 parameters will enter the subsequent weighted score calculation, binning and model training links, laying a solid foundation for the quantitative analysis of the subject risk decision engine.

[0082] In another specific embodiment, the IV value is used to measure the amount of information a single feature provides for distinguishing a target variable (eg, high risk vs. low risk).

[0083] For each feature, the samples are grouped according to their distribution, and the differences in the proportion of high-risk and low-risk samples in each group are counted. If a feature has significant differences in the distribution of high and low risks in different bins, the IV value of the feature is high, indicating that the feature is very helpful in risk determination.

[0084] Based on the IV value, the features with obvious no discrimination or very low contribution are initially screened out. For example, the IV value threshold is set (such as IV < 0.02 is eliminated) to quickly reduce the number of features.

[0085] The SHAP value provides an individual-level contribution explanation for feature importance, showing the positive or negative impact of the feature on model decision-making under different samples.

[0086] For the feature set after IV value screening, the trained model (a relatively complete model can be preliminarily trained) is used to calculate the average SHAP value and distribution of each feature in the sample prediction decision.

[0087] Application of results: If the overall SHAP value contribution of a feature is unstable, the direction is difficult to explain, or the contribution is extremely low, it can be further eliminated or downgraded. At the same time, features that have clear business explanation requirements (such as the "one person, multiple companies" feature) are retained.

[0088] If two features are strongly correlated statistically (such as correlation coefficient > 0.8) and have overlapping or substitution relationships in business terms, the redundant features can be deleted to reduce model complexity and improve model generalization capabilities.

[0089] The Pearson or Spearman correlation coefficient is calculated for the remaining feature set. If the correlation between paired features is found to be too high, features with clearer business logic and more stable data acquisition are retained first, and the rest are discarded.

[0090] This step reduces multicollinearity and feature redundancy, and enhances the stability and interpretability of the model.

[0091] Data distribution analysis is performed on each continuous feature to observe its frequency distribution, central tendency (mean, median), dispersion (standard deviation) and tail characteristics.

[0092] The distribution of eigenvalues ​​was analyzed using methods such as histograms and kernel density estimation plots, and statistical tests (such as the Kolmogorov-Smirnov test) were used to determine the differences in risk characteristics between eigenvalues ​​in different intervals.

[0093] Select preliminary binning points based on the distribution pattern, such as more detailed binning in data-dense areas and appropriate merging of bins in data-sparse areas to improve statistical stability.

[0094] Based on the manually selected binning strategy, a machine learning method is used to automatically search for a better binning strategy. The number of bins and bin boundaries can be adjusted through goal-oriented grid search (GridSearch) or intelligent search strategies such as genetic algorithms and Bayesian optimization.

[0095] Select the solution with the strongest correlation between the characteristics and the target variable after binning, the most balanced sample distribution between bins, and the most significant difference between groups to ensure that each bin can better distinguish the risk level.

[0096] The determination of the binning strategy is not a one-time thing and requires continuous iteration as the model performs and data is updated.

[0097] Step S400: construct a sample set, set the misjudgment cost and class weight based on the sample set, perform parameter tuning and validation set feedback, and adjust the decision threshold; Specifically, a sample set is constructed, including whitelist samples and blacklist samples. The whitelist samples are a list of low-risk enterprises, and the blacklist samples are a list of high-risk enterprises. Before using CatBoost for training, the class weights are adjusted using the class_weights parameter of CatBoost. After feature engineering, the blacklist samples are oversampled and the whitelist samples are undersampled. The misjudgment cost and class weight are set, and the blacklist recognition rate and whitelist false positive rate under different parameter combinations are measured on the validation set. The class weights and parameters are adjusted iteratively step by step to meet the target requirements. When the final output is still a probability score after using CatBoost and adding class weights, the decision threshold is adjusted using the following methods: use different prediction thresholds in the validation set to obtain changes in the false positive rate; find a balance point that can minimize the probability of false positives for low-risk companies while maximizing the recall rate; and perform continuous monitoring and iterative optimization.

[0098] In a specific embodiment, a higher weight is set for blacklist samples to severely punish FN errors: assume Cost(FN)=10. Appropriate weights are also given to whitelist samples, but they are relatively low: such as Cost(FP)=1. In this way, CatBoost will be more inclined to reduce FN (judgment of high-risk entities) errors in iterative training, while striving to keep FP (accidentally harming high-quality companies) at an extremely low level. When using CatBoost to optimize parameters, business goals can be achieved through the following steps: Use Grid Search or Bayesian 0ptimization to search for parameters such as class_weights, learning_rate, depth, iterations, etc. The blacklist recognition rate and whitelist false positive rate under different parameter combinations are measured on the validation set, and the class weights and parameters are gradually iterated to approach the target requirements, specifically the blacklist recognition rate>97%, and the whitelist false positive rate<1%.

[0099] Even after using CatBoost and adding class weights, the final output is still a probability score. Further fine-tuning can be done through the following strategies: Try different prediction thresholds in the validation set, such as adjusting from the default 0.5 to a higher 0.6 or 0.7, and observe the changes in the false positive rate and missed positive rate. Find a balance point that can greatly reduce FP while maintaining a high recall rate (low FN). For specific business requirements, the threshold can be fine-tuned to allow the model to more prudently mark high-risk subjects in the prediction and ensure that the blacklist recognition rate meets the requirements.

[0100] After the model is launched, the performance indicators in real business operations should be monitored regularly, and class_weights and thresholds should be fine-tuned based on feedback data. If the actual false positive rate is found to be close to the upper limit of 1%, the threshold can be slightly lowered or the class_weights can be adjusted to relax the judgment criteria for low-risk entities, thereby reducing false positives. If the blacklist recognition rate drops slightly, the weight of high-risk class samples can be appropriately increased or the FN error cost can be increased, so that the model can capture high-risk features more rigorously in subsequent training iterations. By reasonably setting class weights and decision thresholds in the CatBoost model, selecting representative whitelist and blacklist sample sets, and combining cost-sensitive learning strategies, the model is expected to achieve the goal of whitelist false positive rate <1% and blacklist recognition rate >97% in actual operation. This method not only meets strict risk management requirements, but also takes into account the actual needs of customer experience and business expansion.

[0101] Step S500: binning is performed based on the third risk feature set to obtain an original score, and a fit test and business indicator assessment are performed to obtain a final risk score; threshold selection and probability calibration are performed, and based on the final risk score and feature label description, the enterprise's risk level classification and decision-making recommendations are obtained, and risk label interpretation and large model-assisted interpretation are performed.

[0102] Specifically, each feature in the third risk feature set is binned, the feature value distribution is analyzed, and the risk feature differences of feature values ​​in different bin intervals are determined by combining the Kolmogorov-Smirnov test; the number of bins and bin boundaries are adjusted by the GridSearch search strategy; in the model training and verification process, iterative optimization and feedback loops are performed to obtain the original score of each feature; Assume the third risk feature set is ,in, represents the raw score of the independent feature parameter, represents the composite score of the combined feature parameters, An index representing the third risk feature set; In the feature parameters, the formula of the comprehensive scoring model is as follows: ; Among them, Score is the comprehensive score of feature parameters. For the The weights of the independent parameters, is a set of combined feature indexes. Assign appropriate weights ; Score is scaled to the range of 0-100 through linear transformation to obtain the original score; The original score is tested for goodness of fit using the business validation data set, evaluated based on AUC, KS value and F1 Score, and fine-tuned to obtain the final risk score; AUC represents the area under the curve, KS value represents the Kolmogorov-Smirnov statistic, and F1 Score represents the balanced F score; Perform threshold selection and probability calibration, obtain the enterprise's risk level classification and decision-making recommendations based on the final risk score and feature label description, and perform risk label interpretation and large model-assisted interpretation.

[0103] In a specific embodiment, during the risk assessment process, the model not only provides quantitative scores and risk levels, but also identifies specific risk tags and hit features. These tags may involve abnormal equity structure of the enterprise, legal representative's record of involvement in a case, abnormal address concentration, duplicate contact information, etc. This allows users to understand the specific reasons and evidence chain for the enterprise being judged as high risk.

[0104] Users can query detailed information through specific interfaces, including: Basic information of the enterprise, such as name, registered capital, and industry type; The evidence summary and data source corresponding to each risk label, such as the judgment document number, regulatory notification document, and knowledge graph association results; Such detailed output ensures the transparency and traceability of the decision-making process, and provides solid data support for business personnel to conduct subsequent due diligence (CDD / EDD).

[0105] In actual business environments, even with rich feature descriptions and label explanations, risk assessment outputs may still be too complex for non-technical decision makers. With the fine-tuning of pre-trained language models and prompt design, complex model logic can be converted into natural language output, improving comprehensibility and interaction convenience.

[0106] During the fine-tuning stage, information such as enterprise risk characteristics, binning rules, scoring logic and historical case interpretations are injected into the big model to enable it to have the ability to understand proprietary domain knowledge.

[0107] The fine-tuned large model can access information repositories such as risk feature definitions, parameter weight explanations, and policy and regulatory constraints, so that it can give reasonable, compliant, and easy-to-understand answers when queried.

[0108] Use prompts to guide the output of the big model. The prompts should be concise and clear, and guide the style, depth, and content scope of the big model's answer. For example: Example Prompt 1: “The following are the risk characteristics of Enterprise A: the legal representative has many enterprises under his name (S02) and the registered address is concentrated (TO1). Please use natural language to explain the impact of these characteristics on the risk level of Enterprise A and explain why enhanced due diligence is needed for this enterprise.

[0109] For Prompt 1, the response is as follows: "The fact that the legal representative of Enterprise A has many companies under his name (indicator S02) indicates that the legal representative may control multiple companies at the same time. Such characteristics are usually associated with high-risk behaviors such as bulk shell companies and account resale in regulatory practice. In addition, the concentration of registered addresses of Enterprise A (indicator T01) means that there are too many companies with the same registered address, which may represent the existence of agent registration addresses and concentrated registration of virtual offices. In actual cases, this address concentration is often associated with abnormal corporate clusters. Taken together, these characteristics make Enterprise A obtain a higher risk score in the regulatory model, and thus it is judged as a medium-to-high risk entity. It is recommended to conduct enhanced due diligence on the enterprise, including verifying the on-site operating conditions, interviewing the legal representative, and verifying the legitimacy of related enterprises.

[0110] like Figure 2 As shown in the system structure diagram of a commercial subject risk assessment system based on machine learning, a commercial subject risk assessment system based on machine learning includes: The regulatory policy analysis and knowledge graph construction module includes: a rule extraction unit, a graph mapping unit and a graph mining unit; wherein the rule extraction unit collects and analyzes multi-level regulatory policy documents, extracts rules, and stores them in a database; the graph mapping unit obtains the basic features of the enterprise, maps the basic features to a graph data structure, and constructs a knowledge graph; the graph mining unit uses graph mining based on the knowledge graph to identify risk features, structures the risk features, and forms a first risk feature set; The risk feature screening module includes: a dimensionality reduction and filtering unit and an expert verification unit; wherein the dimensionality reduction and filtering unit performs dimensionality reduction and filtering on the first risk feature set, and the expert verification unit performs business expert verification to obtain a second risk feature set; The qualitative and quantitative analysis module includes: a qualitative analysis unit and a quantitative analysis unit; wherein the qualitative analysis unit establishes a qualitative rule system to perform qualitative analysis on the enterprise; for the enterprise whose risk level cannot be determined by qualitative analysis, the quantitative analysis unit performs quantitative analysis based on the second risk feature set to obtain a feature label description and obtain a third risk feature set; The sample set and decision optimization module includes: a sample construction unit and a parameter tuning unit; wherein the sample construction unit constructs a sample set, sets the misjudgment cost and class weight based on the sample set, and the parameter tuning unit performs parameter tuning and verification set feedback to adjust the decision threshold; The risk scoring and decision-making module includes: a binning scoring unit, a scoring verification unit and a grading unit; wherein the binning scoring unit performs binning based on the third risk feature set to obtain an original score, and the scoring verification unit performs a fit test and a business indicator assessment to obtain a final risk score; the grading unit performs threshold selection and probability calibration, obtains the enterprise's risk grading and decision-making recommendations based on the final risk score and feature label description, and performs risk label interpretation and large model-assisted interpretation.

[0111] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A commercial subject risk assessment method based on machine learning, characterized in that: The method comprises the following steps: Collect and parse multi-level regulatory policy documents, extract rules, and store them in a database; obtain basic characteristics of the enterprise, map the basic characteristics to a graph data structure, and construct a knowledge graph; based on the knowledge graph, use graph mining to identify risk characteristics, structure the risk characteristics, and form a first risk characteristic set; Perform dimensionality reduction and filtering on the first risk feature set, and perform business expert verification to obtain a second risk feature set; Establishing a qualitative rule system to conduct qualitative analysis on the enterprise; for enterprises whose risk level cannot be determined through qualitative analysis, conducting quantitative analysis based on the second risk feature set to obtain feature label descriptions and obtain a third risk feature set; Construct a sample set, set the misjudgment cost and class weight based on the sample set, perform parameter tuning and validation set feedback, and adjust the decision threshold; Based on the third risk feature set, the risk is divided into boxes to obtain the original score, and the fit test and business indicator assessment are performed to obtain the final risk score. Threshold selection and probability calibration are performed, and based on the final risk score and feature label description, the enterprise's risk level classification and decision-making recommendations are obtained, and risk label interpretation and large model-assisted interpretation are performed.

2. A commercial entity risk assessment method based on machine learning according to claim 1, characterized in that: The collection and analysis of multi-level regulatory policy documents, extraction of rules, and storage in the database include: The rules include prohibition rules, recommended prohibition rules and key focus rules; When the multi-level regulatory policy document is a scanned copy or unstructured text, it is converted into a parsable text format using an OCR tool, where OCR stands for optical character recognition; the Chinese word segmentation tool is used to segment and tag the text; An entity recognition model is used to identify key entity types from the text, where the entity recognition model is BERT+CRF or a named entity recognition model based on deep learning, where BERT represents a bidirectional encoder representation and CRF represents a conditional random field; a syntactic analysis tool is used to perform dependency syntactic analysis on the text to obtain dependency relationships; based on the dependency relationships, risk rule expressions are identified; based on the risk rule expressions, they are mapped into qualitative rules and stored in a database; and repeated, similar or conflicting rules in the rules are merged and prioritized.

3. The commercial subject risk assessment method based on machine learning according to claim 1 is characterized in that: The step of obtaining basic features of an enterprise, mapping the basic features to a graph data structure, and constructing a knowledge graph includes: Obtain the basic characteristics of the enterprise and construct a directed weighted graph G=(V,E); wherein V represents a node set, E represents an edge set, and the relationship between different node sets is described by relationship edges. For different types of risk characteristics, the node set contains different nodes.

4. A commercial subject risk assessment method based on machine learning according to claim 3, characterized in that: The method of identifying risk features by using graph mining based on the knowledge graph and structuring the risk features to form a first risk feature set includes: The graph mining includes community detection, path analysis and similarity calculation; By tracing the equity structure of enterprises, the ultimate controller is found. When the enterprises in the first enterprise set are finally gathered together by the same natural person and the enterprises in the first enterprise set are defined as shell companies, they are marked as the first risk; when it is found that the same natural person serves as the legal representative or senior executive of multiple enterprises within the user-defined time period, the multiple enterprises are marked as the second enterprise set, and when the frequency of the enterprises in the second enterprise set in the multi-level regulatory policy documents reaches the set threshold, they are marked as the second risk; by analyzing the registered address attributes of the enterprise nodes, when it is found that multiple enterprises are registered at the same address, the multiple enterprises are marked as the third enterprise set, and when the third enterprise set is indicated in the multi-level regulatory policy documents as involving illegal operations, it is marked as the third risk; the extracted risk features are output in a structured form to form the first risk feature set.

5. The commercial entity risk assessment method based on machine learning according to claim 1 is characterized in that: The first risk feature set is subjected to dimensionality reduction and filtering, and business expert verification to obtain a second risk feature set, including: For the risk features in the first risk feature set, abnormal risk features are eliminated using feature statistical attributes, and the feature statistical attributes include variance, missing rate and information entropy; correlation analysis is performed on the abnormal risk features after elimination, key risk patterns are screened, and the patterns are submitted to business experts for qualitative evaluation to obtain a second risk feature set.

6. A commercial subject risk assessment method based on machine learning according to claim 1, characterized in that: The above-mentioned establishment of a qualitative rule system to conduct qualitative analysis of the enterprise includes: Construct basic characteristic indicators and green enterprise database indicators for enterprises, wherein the basic characteristic indicators include basic information of the enterprises, and the enterprises in the green enterprise database indicators are judged to be low-risk types; based on the basic characteristic indicators and the green enterprise database indicators, determine whether the enterprises are high-risk or low-risk.

7. A commercial entity risk assessment method based on machine learning according to claim 6, characterized in that: For the enterprises whose risk level cannot be determined by qualitative analysis, quantitative analysis is performed based on the second risk feature set to obtain feature label descriptions, thereby obtaining a third risk feature set, including: For enterprises whose risk level cannot be determined through qualitative analysis, the risk characteristics therein are divided into boxes based on the second risk characteristic set, each box includes a number of samples, and each sample is marked as a positive sample or a negative sample; the proportion of positive samples and negative samples in each box is counted, the WOE in each box is calculated, and the IV value is accumulated to obtain the IV value, where WOE represents the weight of evidence and the IV value represents the information value; when there is a risk characteristic that has information gain for the final risk judgment result and its IV value is greater than the set threshold, it is retained; when there is a risk characteristic whose IV value is judged to be close to 0 within the set threshold, it is eliminated; The risk features after IV value screening are analyzed using SHAP value, where SHAP value represents the Shapley additive explanatory value. Risk features with positive SHAP value in high-risk entities and values ​​exceeding the set threshold are retained, otherwise they are eliminated or downgraded. The risk features after SHAP value analysis are tested for correlation between features. When the correlation between two risk features exceeds the set threshold, and one of the risk features is considered to be able to represent this type of risk information after review by business experts, the redundant features are deleted. The third risk feature set is determined to obtain feature label descriptions.

8. The commercial entity risk assessment method based on machine learning according to claim 1 is characterized in that: The sample set is constructed, misjudgment cost and class weight are set based on the sample set, parameter tuning and validation set feedback are performed, and decision thresholds are adjusted, including: Construct a sample set, including whitelist samples and blacklist samples. The whitelist samples are a list of low-risk enterprises, and the blacklist samples are a list of high-risk enterprises. Adjust the class weights. After feature engineering, oversample the blacklist samples and undersample the whitelist samples. Set the misjudgment cost and class weights, measure the blacklist recognition rate and whitelist false positive rate under different parameter combinations on the validation set, and gradually iterate and adjust the class weights and parameters to meet the target requirements. When class weights are added and the final output is still a probability score, the decision threshold is adjusted using the following methods: use different prediction thresholds in the validation set to obtain changes in the false positive rate; find a balance point that can minimize the probability of false positives for low-risk companies and maximize the recall rate; and perform continuous monitoring and iterative optimization.

9. The commercial entity risk assessment method based on machine learning according to claim 1 is characterized in that: The binning is performed based on the third risk feature set to obtain an original score, and a fit test and a business indicator assessment are performed to obtain a final risk score; Perform threshold selection and probability calibration, obtain the enterprise's risk level classification and decision-making recommendations based on the final risk score and feature label description, and interpret risk labels and assist in large-scale model interpretation, including: Binning each feature in the third risk feature set, analyzing the distribution of feature values, testing and judging the risk feature differences between feature values ​​in different binning intervals, and adjusting the number of bins and binning boundaries; performing iterative optimization and feedback loops during model training and validation to obtain the original score of each feature; Assume the third risk feature set is ,in, represents the raw score of the independent feature parameter, represents the composite score of the combined feature parameters, An index representing the third risk feature set; In the feature parameters, the formula of the comprehensive scoring model is as follows: ； Among them, Score is the comprehensive score of feature parameters. For the The weights of the independent parameters, is a set of combined feature indexes. Assign appropriate weights ; Scale the Score through linear transformation to get the original score; The original score is tested for goodness of fit using the business validation data set, and the final risk score is obtained by evaluation based on the AUC, KS value, and F1 Score; wherein AUC represents the area under the curve, KS value represents the Kolmogorov-Smirnov statistic, and F1 Score represents the balanced F score; Perform threshold selection and probability calibration, obtain the enterprise's risk level classification and decision-making recommendations based on the final risk score and feature label description, and perform risk label interpretation and large model-assisted interpretation.

10. A commercial subject risk assessment system based on machine learning, using a commercial subject risk assessment method based on machine learning as claimed in any one of claims 1 to 9, characterized in that: include: The regulatory policy analysis and knowledge graph construction module includes: a rule extraction unit, a graph mapping unit and a graph mining unit; wherein the rule extraction unit collects and analyzes multi-level regulatory policy documents, extracts rules, and stores them in a database; the graph mapping unit obtains the basic features of the enterprise, maps the basic features to a graph data structure, and constructs a knowledge graph; the graph mining unit uses graph mining based on the knowledge graph to identify risk features, structures the risk features, and forms a first risk feature set; The risk feature screening module includes: a dimensionality reduction and filtering unit and an expert verification unit; wherein the dimensionality reduction and filtering unit performs dimensionality reduction and filtering on the first risk feature set, and the expert verification unit performs business expert verification to obtain a second risk feature set; The qualitative and quantitative analysis module includes: a qualitative analysis unit and a quantitative analysis unit; wherein the qualitative analysis unit establishes a qualitative rule system to perform qualitative analysis on the enterprise; for the enterprise whose risk level cannot be determined by qualitative analysis, the quantitative analysis unit performs quantitative analysis based on the second risk feature set to obtain a feature label description and obtain a third risk feature set; The sample set and decision optimization module includes: a sample construction unit and a parameter tuning unit; wherein the sample construction unit constructs a sample set, sets the misjudgment cost and class weight based on the sample set, and the parameter tuning unit performs parameter tuning and verification set feedback to adjust the decision threshold; The risk scoring and decision-making module includes: a binning scoring unit, a scoring verification unit and a grading unit; wherein the binning scoring unit performs binning based on the third risk feature set to obtain an original score, and the scoring verification unit performs a fit test and a business indicator assessment to obtain a final risk score; the grading unit performs threshold selection and probability calibration, obtains the enterprise's risk grading and decision-making recommendations based on the final risk score and feature label description, and performs risk label interpretation and large model-assisted interpretation.

Images