A Corporate Financial Risk Early Warning Method Integrating Multi-Source Heterogeneous Knowledge Graphs and Evidence Theory

Abstract

This invention provides a corporate financial risk early warning method that integrates multi-source heterogeneous knowledge graphs and evidence theory. Based on a corporate financial risk early warning system, this invention innovatively constructs and quantifies a corporate equity and supply chain network knowledge graph, comprehensively depicting the complex relationships between enterprises from both capital and operational dimensions. It applies D-S evidence theory to integrate multi-source information from heterogeneous networks and traditional financial indicators, effectively resolving information conflicts and uncertainties, and improving model robustness. Employing an inherently interpretable EBM model for risk early warning not only ensures high prediction accuracy but also clearly reveals the specific impact mechanisms of each network feature on financial risk. It extracts deep network structure features and risk transmission features from the knowledge graph, fully mining the deep knowledge contained within the network. In addition to outputting risk levels, it also provides multi-dimensional information such as confidence level, feature contribution, and risk transmission path, providing comprehensive decision support for corporate management.

Description

Technical Field

[0001] This invention relates to the intersection of knowledge graph technology and corporate financial risk management, specifically to a method for constructing corporate equity and supply chain knowledge graphs in the financial field, and a method for early warning of corporate financial risks based on knowledge graphs and evidence theory. Background Technology

[0003] Corporate financial risk early warning is a core component of modern enterprise management. Its effectiveness not only affects the survival and development of the enterprise itself but also has a profound impact on the stable operation of the entire industrial supply chain. Currently, with increased global economic uncertainty and structural adjustments in industrial and supply chains, the identification and prevention of corporate financial risks have become particularly urgent and complex. Enterprises are no longer isolated "economic islands" but are deeply embedded in a complex, multi-layered, and multi-dimensional network formed by equity, supply chains, and guarantee chains. However, traditional financial risk early warning research, from early statistical models such as Z-scores to current machine learning models, while continuously improving in predictive accuracy, still largely treats enterprises as independent entities, relying on their own financial statement data. This "individual perspective" ignores the network connections between enterprises and fails to effectively capture external shocks and chain reactions transmitted through the network.

[0004] In recent years, complex network theory has been introduced into risk warning research. For example, Zhou et al. (2023) constructed a linkage mode transmission network to analyze the dynamic correlation between innovative financial assets and traditional assets. Wang et al. (2024) verified the contribution of network correlation to systemic risk prediction through an interpretable fingerprint model. Some scholars have also fused multi-source networks; for example, Li et al. (2025) integrated stock return networks, risk spillover networks, and market trading networks, and used an interpretable graph neural network model for risk prediction. However, these studies lack an effective knowledge representation framework when dealing with multi-source heterogeneous data, resulting in insufficient network feature extraction and difficulty in capturing the deep mechanisms of risk transmission.

[0005] Against the backdrop of the integration of fintech and artificial intelligence, knowledge graph technology provides an advanced method for corporate financial risk early warning. In recent years, literature attempting to use knowledge graph technology for corporate financial risk early warning has gradually emerged. Wang Haofen et al. (2020) pointed out that with the deepening of application scenarios, the representation and acquisition of multiple types of knowledge, the reasoning and computation of massive amounts of knowledge, and the timeliness of knowledge will become urgent problems to be solved in the construction of corporate knowledge graphs. Chen Xiaojun et al. (2020) systematically studied corporate risk knowledge graphs, realizing the retrieval and utilization of knowledge graphs through intelligent question answering. Zheng Jie et al. (2022) proposed, from a regulatory perspective, to judge the potential tax risks of corporate financial data and other related information, and used Neo4j to apply it to the construction of a knowledge graph for corporate tax compliance risk, realizing the storage and retrieval of the knowledge graph. Although knowledge graph technology has begun to be applied to corporate financial risk early warning, existing research remains at a relatively superficial level of theoretical feasibility discussion, and no early warning graph applicable to the real situation of enterprises has been developed.

[0006] Based on the above problems, this invention proposes a method for constructing equity and supply chain knowledge graphs based on multi-source heterogeneous data of enterprises, as well as a method for early warning of enterprise financial risks based on knowledge graphs and evidence theory. Summary of the Invention

[0007] This invention provides a method for early warning of corporate financial risks that integrates multi-source heterogeneous knowledge graphs and evidence theory. It constructs corporate equity network knowledge graphs and supply chain network knowledge graphs, transforming unstructured relational data into analyzable structured knowledge, comprehensively depicting the complex relationships between enterprises from both capital and operational dimensions. Deep network structure and risk features are extracted from the knowledge graphs, fully mining the deep knowledge contained within the network. An interpretable augmented machine (EBM) model is used to train each evidence source, obtaining the model's "belief" probability (BPA) for each risk level. Finally, the synthesis rules of DS evidence theory are applied to fuse the BPAs from three independent evidence sources: equity network, supply chain network, and financial indicators, effectively handling conflicts and uncertainties between evidence and outputting the final financial risk warning result.

[0008] To achieve the objective of this invention, the technical solution adopted is as follows:

[0009] A method for early warning of corporate financial risks that integrates multi-source heterogeneous knowledge graphs and evidence theory is proposed. The method is based on a corporate financial risk early warning system, which includes a data input and preprocessing module, a risk level definition module, a heterogeneous knowledge graph construction module, a multi-dimensional feature engineering extraction module, a basic credibility allocation acquisition module, and a DS evidence fusion and decision-making module.

[0010] A method for early warning of corporate financial risks that integrates multi-source heterogeneous knowledge graphs and evidence theory is proposed. The method is based on a corporate financial risk early warning system, which includes a data input and preprocessing module, a risk level definition module, a heterogeneous knowledge graph construction module, a multi-dimensional feature engineering extraction module, a basic credibility allocation acquisition module, and a DS evidence fusion and decision-making module.

[0011] A corporate financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory, comprising the following steps:

[0012] S1. Data Input and Preprocessing: Collect financial indicator data from the CNRDS database and the CNFI database, clean and preprocess the financial indicator data, perform entity identification and normalization for the enterprise domain, and perform Z-score standardization on the financial indicator data.

[0013] S2. Risk Level Definition: Select key financial indicators from the financial indicator data as model input, initialize the Gaussian mixture model using the GaussianMixture function, call the gmm.fit function to train the key financial indicator data, iteratively optimize the model parameters through the expectation-maximization (EM) algorithm, use the gmm.predict function to determine the category of each sample according to the probability density function, output the risk_label array, and use the risk_label as the target variable y for subsequent model training to construct a financial risk early warning and identification framework;

[0014] S3. Construction of Heterogeneous Knowledge Graph: Using the Neo4j graph database, create enterprise entity nodes in the equity network. For each equity investment record, connect the enterprise entity nodes of the two investing parties through network edges; extract basic information and a list of participating and holding companies from the CPCD module of the CNRS database, and extract cross-shareholding details and equity participation / control details from the listed company alliance-equity alliance module of the CSMAR database; construct the equity network knowledge graph.

[0015] The Neo4j graph database was used to create enterprise entity nodes in the supply chain network. A directed graph was constructed based on the data of the top five customers and top five suppliers of the listed company. Information on major suppliers, major customers, and shareholder information of suppliers and customers was extracted from the Supply Chain Research SCRD module of the CNRS database. Information on the top five customers and suppliers and supply chain alliances was extracted from the Supply Chain Research module of the CSMAR database. A knowledge graph of the supply chain network was constructed.

[0016] S4. Multidimensional feature extraction: Use the centrality module in the NetworkX library to extract degree centrality, proximity centrality, betweenness centrality, feature vector centrality and structural holes. Calculate the sum of squares of the shareholding ratios HHI=sum(ratio^2). Use the nx.pagerank function with the holding ratio holding_ratio as the edge weight to calculate the PageRank value of each node.

[0017] The centrality module in the NetworkX library is used to extract degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural holes. The standard deviation of supplier concentration over three consecutive years, std_concentration, is calculated, and then the coefficient of variation, cv = std_concentration / mean_concentration, is calculated.

[0018] Key financial indicators are extracted from the financial indicator data to obtain the enterprise financial status feature FIN_feature. The concat function in the pandas library is used to merge the equity network structure feature G_EQ, equity network risk feature V_EQ, supply chain network structure feature G_SC, supply chain network risk feature V_SC and enterprise financial status feature FIN_feature into a complete feature matrix X_complete, which is used as the input for subsequent model training.

[0019] S5. Basic Credibility Allocation: Divide the feature matrix X_complete into three independent evidence sources: equity network evidence source, supply chain network evidence source, and financial indicator evidence source. Use the ExplainableBoostingClassifier function in the interpreter library to initialize the EBM model, and call ebm_source.fit to train an independent EBM model for each evidence source.

[0020] Use the ebm.predict_proba function to obtain the predicted probability distribution of each model for the test set samples, and output the softmax probability distribution as the basic confidence assignment (BPA) in the DS evidence theory. Use the ebm.explain_global function to obtain the global feature importance, and use the ebm.explain_local function to obtain the feature contribution map of a single sample.

[0021] S6, DS Evidence Fusion and Decision Making: Defining the Recognition Framework ={low risk, medium risk, high risk} represents three possible financial risk states. The conflict coefficient K is calculated for the BPA of each pair of evidence sources. The BPA of the three evidence sources is fused using the DS evidence theory synthesis rule to obtain the final fused BPA m_final. Based on the fused BPA distribution m_final, the final risk warning result is output using the maximum confidence rule. The prediction result is compared with the real risk label risk_label to calculate the evaluation index and verify the effectiveness of the model.

[0022] As a preferred method, step S1 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0023] S11. Data Extraction: Collect financial indicator data from the CNRDS database and the CNFI database;

[0024] S12. Data Cleaning and Preprocessing: For duplicate records in financial indicator data, a hash function is used to generate a unique identifier for each transaction record, and the identifiers are compared to remove duplicates; for missing values ​​in financial indicator data, since filling in the transaction amount will cause discrepancies with the actual data, they are directly deleted; for missing timestamps in financial indicator data, if business logic allows, the transaction record date is used as an approximation, or the average timestamp of adjacent records is used.

[0025] S13. Entity Recognition and Normalization: Perform entity recognition and normalization on enterprise stock codes and stock abbreviations in the financial indicator data. Use a natural language processing model to calculate entity similarity. Normalize entities with similarity above the threshold and remove special characters and spaces.

[0026] S14. Data Standardization: Perform Z-score standardization on the financial indicator data. Use the standard_scaler function to convert each indicator into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the impact of differences in the dimensions of different indicators on model training.

[0027] As a preferred method, step S2 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0028] S21. Indicator Selection: Four key financial indicators—profitability, solvency, operating efficiency, and growth potential—are selected from the financial indicator data as inputs to the Gaussian mixture model. Profitability financial indicators include return on assets, return on equity, operating profit margin, cost-to-profit ratio, and main business ratio. Solvency financial indicators include current ratio, quick ratio, debt-to-equity ratio, and times interest earned. Operating efficiency financial indicators include inventory turnover, accounts receivable turnover, current asset turnover, fixed asset turnover, and total asset turnover. Growth potential financial indicators include revenue growth rate, net profit growth rate, net cash flow growth rate from operating activities, and total asset growth rate.

[0029] S22. Initialization of Gaussian Mixture Model: Use the GaussianMixture function to initialize the Gaussian mixture model, set n_components=3 to indicate that the samples are divided into three categories of financial risk: high, medium and low; set covariance_type='full' to allow each cluster to have its own covariance matrix;

[0030] S23. Model Training: Call the gmm.fit function to train the financial indicator data, and iteratively optimize the model parameters through the expectation-maximization (EM) algorithm, including the mean, covariance matrix and weight of each Gaussian component, until the log-likelihood function converges.

[0031] S24. Risk Level Classification: The gmm.predict function is used to determine the category of each sample based on the probability density function, and the risk_label array is output, where 0 represents low risk, 1 represents medium risk, and 2 represents high risk. According to the clustering results, there are 25 high-risk companies, 201 medium-risk companies, and 36 low-risk companies out of 262 companies.

[0032] S25. Identification Framework Construction: Using risk_label as the target variable y for subsequent model training, a financial risk early warning identification framework is constructed to ensure the objectivity and scientific nature of the risk level definition.

[0033] Furthermore, the specific algorithm flow for step S2 of the enterprise financial risk early warning method is as follows:

[0034] ① Input the financial indicator data matrix X_financial:

[0035] 1. X_financial is an n×m matrix, where n is the number of companies and m is the number of financial indicators;

[0036] 2. Standardize X_financial to obtain X_standardized;

[0037] ② Gaussian mixture model clustering:

[0038] 1. Initialize the GaussianMixture model and set the number of clusters k=3;

[0039] 2. Call gmm.fit(X_standardized) to train the model;

[0040] 3. Use gmm.predict(X_standardized) to obtain the risk label (risk_label) for each sample;

[0041] ③ Output the risk level classification results:

[0042] 1. Count the number of samples in each category to verify the rationality of the division;

[0043] 2. Save risk_label as the target variable for subsequent model training.

[0044] As a preferred method, step S3 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0045] S31. Equity Network Node Creation: Using the Neo4j graph database, enterprise entities are used as nodes, including listed companies and their subsidiaries, joint ventures, associates and sub-subsidiaries; the Cypher statement CREATE (n:Company {name:company_name, code: company_code, industry: industry_type}) is used to create enterprise entity nodes, and a unique identifier node_id is assigned to each entity;

[0046] S32. Equity Network Edge Creation: For each equity investment record, connect the two investing companies' nodes through a network edge; use the Cypher statement MATCH (a:Company {code: investor_code}), (b:Company {code: investee_code}) CREATE (a)-[r:EquityInvest {ratio: holding_ratio, date: invest_date}]->(b) to create an equity investment relationship edge, with the edge weight being the shareholding ratio (holding_ratio);

[0047] S33. Data sources for equity network: Basic information and a list of participating and holding companies are extracted from the CPCD module of the CNRS database; Cross-shareholding details and equity participation details are extracted from the listed company alliance-equity alliance module of the CSMAR database; Finally, an equity network knowledge graph with 97,140 nodes and 417,487 edges is constructed.

[0048] S34. Supply Chain Network Node Creation: Using the Neo4j graph database, with enterprise entities as nodes, including listed companies and their top five suppliers and top five customers; using the Cypher statement CREATE (n:SupplyChainNode {name:company_name, code: company_code, is_listed: is_listed_flag, industry:industry_type}) to create supply chain nodes, adding four attributes: company stock code, company stock abbreviation, company sub-industry, and company node size;

[0049] S35. Supply Chain Network Edge Creation: Construct a directed graph based on the purchase amount of the top five suppliers and the revenue of the top five customers of the listed company; use the Cypher statement MATCH (a:SupplyChainNode {code: supplier_code}),(b:SupplyChainNode {code: buyer_code}) CREATE (a)-[r:SupplyRelation {amount:transaction_amount, ratio: transaction_ratio, year: transaction_year}]->(b) to create supply or purchase relationship edges, with the edge weight being the ratio of sales / purchase amount to the total amount, transaction_ratio;

[0050] S36. Supply chain network data sources: Extract information on major suppliers, major customers, and shareholder information of suppliers and customers from the Supply Chain Research SCRD module of the CNRS database; extract information on the top five customers and suppliers and supply chain alliances from the Supply Chain Research module of the CSMAR database; and finally construct a supply chain network knowledge graph with 12,965 nodes and 44,201 edges.

[0051] Furthermore, the specific algorithm flow for step S3 of the enterprise financial risk early warning method is as follows:

[0052] ① Construction of equity network graph:

[0053] 1. Initialize the Neo4j graph database connection using the Graph function of the py2neo library;

[0054] 2. Traverse the equity data table `equity_data` and perform node creation and edge creation operations for each record;

[0055] 3. Use the merge statement to create each enterprise entity, ensuring the exhaustiveness and uniqueness of each enterprise entity;

[0056] 4. Assign a weight to each edge as the holding ratio;

[0057] ② Supply chain network mapping:

[0058] 1. Initialize the Neo4j graph database connection using the Graph function of the py2neo library;

[0059] 2. Traverse the supply chain data table supply_chain_data and perform node creation and edge creation operations for each record;

[0060] 3. Use the merge statement to create each enterprise entity, ensuring the exhaustiveness and uniqueness of each enterprise entity;

[0061] 4. Assign a weight to each edge as the proportion of sales or purchases to the total amount (transaction_ratio);

[0062] ③Graphical verification:

[0063] 1. Use the Cypher statement `MATCH (n) RETURN count(n)` to verify the number of nodes;

[0064] 2. Use the Cypher statement `MATCH ()-[r]->() RETURN count(r)` to verify the number of edges;

[0065] 3. Check the connectivity of the graph to ensure there are no isolated nodes.

[0066] As a preferred method, step S4 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0067] S41. Equity Network Structure Feature G_EQ Extraction: The following features are extracted using the centrality module of the NetworkX library:

[0068] 1. Degree centrality: Calculated using the nx.degree_centrality function; a higher value indicates greater importance in the network.

[0069] 2. Proximity Centrality: Calculated using the nx.closeness_centrality function, it reflects the degree of proximity between a company and other companies in a network;

[0070] 3. Betweenness centrality: Calculated using the nx.betweenness_centrality function, it represents the number of times a firm acts as a bridge between two other firms on the shortest path.

[0071] 4. Eigenvector centrality: Calculated using the nx.eigenvector_centrality function, it measures the overall influence and centrality of a listed company or shareholder within its network of connections.

[0072] 5. Structural holes: Calculated using the networkx.algorithms.structural_holes function, this represents the number of structural holes occupied. The larger the value, the greater the information advantage.

[0073] S42. Extraction of V_EQ characteristics of equity network risk:

[0074] 1. Equity Concentration: The sum of squares of the shareholding ratios of shareholders, HHI=sum(ratio^2), reflects the degree of concentration of the equity structure;

[0075] 2. PageRank centrality based on shareholding ratio: Using the nx.pagerank function, the PageRank value of each node is calculated with the shareholding ratio as the edge weight. Companies with higher values ​​are in a core position in the entire network.

[0076] S43. Extraction of Supply Chain Network Structural Features G_SC: Similar to the structural features of equity networks, the centrality module of the NetworkX library is used to extract features, including degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural holes. These features reflect the degree of association between the focal enterprise and its supply chain partners and other highly interconnected enterprises in the network, and measure whether the enterprise is connected to other unconnected supply chain clusters.

[0077] S44. Supply Chain Network Risk Characteristics V_SC Extraction:

[0078] 1. Supplier instability: Calculate the standard deviation of supplier concentration (std_concentration) over three consecutive years, and then calculate the coefficient of variation (cv) = std_concentration / mean_concentration. The higher the value, the more unstable the supplier dependency structure.

[0079] 2. Customer instability: The calculation method is symmetrical with that of supplier instability, reflecting the degree of change in a company's dependence on its major customers;

[0080] S45. Extraction of Enterprise Financial Status Characteristics (FIN_feature): Extract the following four categories of key financial indicators from the financial statements:

[0081] 1. Profitability financial indicators: Return on assets, Return on equity, Operating profit margin, Cost-to-profit ratio, and main business ratio;

[0082] 2. Financial indicators of solvency: current ratio, quick ratio, debt-to-equity ratio, and times interest earned;

[0083] 3. Financial indicators of operational capabilities: inventory turnover, accounts receivable turnover, current asset turnover, fixed asset turnover, and total asset turnover;

[0084] 4. Financial indicators of growth potential: Revenue growth rate, net profit growth rate, net cash flow growth rate from operating activities, and total asset growth rate;

[0085] S46. Feature Integration: Use the concat function of the pandas library to merge the equity network structure feature G_EQ, equity network risk feature V_EQ, supply chain network structure feature G_SC, supply chain network risk feature V_SC, and enterprise financial status feature FIN_feature into a complete feature matrix X_complete, which will be used as the input for subsequent model training.

[0086] Furthermore, the specific algorithm flow for step S4 of the enterprise financial risk early warning method is as follows:

[0087] ① Network feature extraction:

[0088] Load the equity network graph (equity_graph) and the supply chain network graph (supply_graph). For each node, call the function to calculate the degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural hole constraint.

[0089] ② Risk characteristic calculation:

[0090] 1. Traverse the equity data table and calculate the equity concentration for each company;

[0091] 2. Call nx.pagerank(equity_graph, weight='holding_ratio') to calculate PageRank centrality;

[0092] 3. Traverse the supply chain data table and calculate the supplier and customer instability for each enterprise;

[0093] ③ Construction of the feature matrix:

[0094] 1. Initialize an empty list feature_list;

[0095] 2. For each enterprise, add network characteristics, risk characteristics, and financial indicators to the feature_list;

[0096] 3. Use pd.DataFrame(feature_list) to create the feature matrix X_complete;

[0097] 4. Use StandardScaler to standardize the feature matrix.

[0098] As a preferred method, step S5 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0099] S51. Evidence Source Division: Divide the feature matrix X_complete into three independent evidence sources:

[0100] Evidence sources for equity networks include structural characteristics and risk characteristics of equity networks;

[0101] Sources of evidence for supply chain networks include structural characteristics and risk characteristics of supply chain networks;

[0102] Financial indicator evidence sources include indicators of profitability, solvency, operational efficiency, and growth potential.

[0103] S52. EBM Model Initialization: Initialize the EBM model using the ExplainableBoostingClassifier function from the interpreter library. Set random_state to ensure reproducible results, set the number of weak learners n_estimators=300, and set the learning rate learning_rate=0.01.

[0104] S53. Model Training: Call ebm_source.fit to train an independent EBM model for each evidence source;

[0105] S54, BPA Acquisition: Use the ebm.predict_proba function to obtain the predicted probability distribution of each model for the test set samples, and output the softmax probability distribution as the basic confidence assignment (BPA) in the DS evidence theory;

[0106] S55. Interpretability Analysis: Use the ebm.explain_global function to obtain the global feature importance, and use the ebm.explain_local function to obtain the feature contribution map of a single sample, clearly showing how each feature affects the final prediction result.

[0107] Furthermore, the specific algorithm flow for step S5 of the enterprise financial risk early warning method is as follows:

[0108] ①Evidence source model training:

[0109] 1. Initialize three EBM model instances: ebm_equity, ebm_supply, and ebm_financial;

[0110] 2. For each source of evidence, call model.fit(X_source, y) to train the model;

[0111] 3. Use cross-validation to evaluate the performance of each model and record the three evaluation metrics: accuracy, recall, and F1 score.

[0112] ②BPA generation: For the test set samples, call ebm_source.predict_proba(X_source_test) to obtain the equity network BPA, supply chain network BPA, and financial indicator BPA respectively;

[0113] Each BPA is an n×3 matrix, where n is the number of samples and 3 is the number of risk levels, such as high, medium, and low.

[0114] ③ Interpretable output:

[0115] Call ebm.explain_global() to generate a global feature importance ranking graph;

[0116] Call ebm.explain_local(X_sample) to generate a feature contribution map for a single sample.

[0117] As a preferred method, step S6 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0118] S61. Definition of Recognition Framework: Define the recognition framework. ={Low Risk, Medium Risk, High Risk} represents three possible financial risk levels; each BPA is... The fundamental probability assignment function on the power set;

[0119] S62. Conflict Coefficient Calculation: For each pair of evidence sources' BPAs, calculate the conflict coefficient K; use the NumPy library to iterate through all focal element combinations, and for those satisfying... The focal element combination is accumulated as m1(B)×m2(C); the conflict coefficient K measures the degree of conflict between two sources of evidence, and the larger the K value, the more serious the conflict.

[0120] S63, Application of DS Synthesis Rules: The synthesis rules of DS evidence theory are used to fuse the BPAs of three evidence sources to obtain the final fused BPA m_final;

[0121] S64. Decision rule: Based on the fused BPA distribution m_final, the final risk warning result is output using the maximum confidence rule; for each sample, the risk level A with the largest m_final(A) is selected as the prediction result, and the confidence level confidence=m_final(A) corresponding to the risk level is output at the same time.

[0122] S65. Early Warning Result Output: Compare the predicted results with the true risk label (risk_label) and calculate eight evaluation metrics: accuracy, precision, recall, F1 score (f1_score), area under the ROC curve (AUC), area under the PR curve (AUC-PR), geometric mean (Gmean), and false negative rate (FNR) to verify the effectiveness of the model.

[0123] Furthermore, the specific algorithm flow for step S6 of the enterprise financial risk early warning method is as follows:

[0124] ① Calculation of conflict coefficient:

[0125] 1. Initialize the conflict coefficient K=0;

[0126] 2. Traverse all focal elements B of evidence source 1 and all focal elements C of evidence source 2;

[0127] 3. If Then K = K + m1(B) × m2(C);

[0128] 4. Return the conflict coefficient K, which measures the degree of conflict between sources of evidence;

[0129] ②D-S evidence fusion:

[0130] 1. Initialize the fusion result m_fused as the identity matrix;

[0131] 2. For every two evidence sources BPA m1 and m2, call the ds_combine(m1, m2) function to fuse them;

[0132] 3. In the ds_combine function, iterate through all focal element combinations B and C;

[0133] 4. For the focal element combination that satisfies B∩C=A, accumulate m1(B)×m2(C) to m_fused(A);

[0134] 5. Normalize the fusion results using a normalization factor of 1 / (1-K);

[0135] 6. Return the fused BPA m_fused;

[0136] ③Final decision:

[0137] 1. For the fused BPA m_final, call np.argmax(m_final, axis=1) to obtain the predicted risk level of each sample;

[0138] 2. Call np.max(m_final, axis=1) to get the prediction confidence for each sample;

[0139] 3. Compare the predicted results with the actual labels and calculate the evaluation indicators;

[0140] 4. Output the final financial risk warning results, including risk level and confidence level;

[0141] ④ Model Evaluation:

[0142] The model's overall performance is verified by calling functions to calculate eight evaluation metrics: accuracy, precision, recall, F1 score (f1_score), area under the ROC curve (AUC), area under the PR curve (AUC-PR), geometric mean (Gmean), and false negative rate (FNR).

[0143] The beneficial effects of this invention are as follows:

[0144] (1) Construction of multi-source heterogeneous networks and comprehensive risk characterization: This invention innovatively constructs and quantifies the knowledge graph of corporate equity and supply chain networks, comprehensively characterizing the complex relationships between enterprises from the two dimensions of capital and operation; most existing technologies regard enterprises as independent individuals or only focus on a single network dimension, which cannot fully capture the complex relationships between enterprises formed through equity and supply chain; This invention, through a two-layer knowledge graph structure, transforms unstructured relational data into analyzable structured knowledge, solves the problem of constructing and quantifying the characteristics of multi-source heterogeneous networks, and lays the foundation for comprehensively capturing the risk transmission between enterprises; Experimental results show that equity networks and supply chain networks provide complementary risk information, neither of which can be omitted, and the integration of the two can significantly improve the accuracy of risk identification;

[0145] (2) Fusion of DS Evidence Theory and Improvement of Robustness: This invention applies DS evidence theory to fuse multi-source information from heterogeneous networks and traditional financial indicators, effectively solving the problems of information conflict and uncertainty, and improving the robustness of the model. Existing technologies lack effective fusion mechanisms to handle conflicts between multi-source information. This invention uses DS synthesis rules to measure the degree of conflict between evidence sources, effectively handling conflicts and uncertainties between evidence, and finally generating a BPA that fuses all information. Experimental results show that even though EBM itself is very strong, the fusion model surpasses EBM in indicators that focus more on imbalanced samples, indicating that the DS rule effectively handles the conflict of multi-source information and improves decision accuracy.

[0146] (3) EBM Interpretable Model and Risk Mechanism Revelation: This invention uses an inherently interpretable EBM model for risk warning, which not only ensures high prediction accuracy but also clearly reveals the specific impact mechanism of each network feature on financial risk. Although deep learning-based graph models have strong performance, their "black box" characteristics limit their application and trust in risk management decisions. This invention provides accurate single-feature contribution maps through the "glass box" property of EBM, clearly showing how each feature affects the final prediction result. This makes the model not only a prediction tool but also an analysis tool, helping users to deeply understand the mechanism of network features on financial risk and enhancing the credibility of decisions.

[0147] (4) Deep knowledge extraction and feature engineering optimization; This invention extracts deep network structure features and risk transmission features from the knowledge graph, and fully explores the deep knowledge contained in the network; Most existing technologies use simple network indicators and fail to fully explore the deep knowledge contained in the network; This invention comprehensively captures the positional advantages, information advantages and dependency structure risks of enterprises in the network by extracting multi-dimensional features such as centrality, structural holes, equity concentration, and supplier instability; Experimental results show that there is a certain conflict between financial indicators and network features, which indicates that network features have captured the "forward-looking" or "structural" risks that have not yet been reflected in the financial statements, proving the effectiveness of deep knowledge extraction;

[0148] (5) Multi-dimensional risk assessment and decision support; This invention not only outputs risk level, but also provides multi-dimensional information such as confidence level, feature contribution, and risk transmission path, providing comprehensive decision support for enterprise management; Most existing technologies only output a single risk score or level, lacking in-depth risk cause analysis; This invention helps users understand the source and transmission mechanism of risk through interpretability analysis, formulate targeted risk management strategies, and improve the level and effectiveness of enterprise risk management. Attached Figure Description

[0149] Figure 1This is the overall technical flowchart of the enterprise financial risk early warning method that integrates multi-source heterogeneous knowledge graphs and evidence theory. Detailed Implementation

[0150] The present invention will be further described and illustrated below with reference to specific embodiments.

[0151] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0152] In the description of this invention, it should be understood that the terms "upper", "lower", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship of the technical solution, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0153] like Figure 1 As shown, a corporate financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory is presented. This method is based on a corporate financial risk early warning system, which includes:

[0154] Data Input and Preprocessing Module: This module is responsible for collecting corporate equity, supply chain, and financial data, and performing cleaning, standardization, and missing value processing. To demonstrate the effectiveness of this invention, it is validated based on the CNRS and CSMAR databases. The data includes corporate equity data, supply chain data, and financial statement data. The data spans from 2015 to 2024, and a total of 262 Chinese A-share listed manufacturing companies were selected as the research sample.

[0155] Risk Level Definition Module: This module uses Gaussian mixture model clustering method to objectively classify the financial risk level of enterprises based on preprocessed financial data and construct an identification framework.

[0156] Heterogeneous Knowledge Graph Construction Module: This module constructs enterprise equity network knowledge graphs and supply chain network knowledge graphs respectively, transforming unstructured relational data into structured knowledge;

[0157] Multidimensional Feature Engineering Extraction Module: This module extracts deep network structure features and risk transmission features from knowledge graphs, and extracts corporate financial indicator features from reports.

[0158] Basic Credibility Assignment Acquisition Module: This module uses the Interpretable Augmentation Machine (EBM) model, taking each feature dimension as an independent source of evidence to obtain the model's "belief" probability for each risk level;

[0159] DS Evidence Fusion and Decision Module: This module applies the synthesis rules of DS evidence theory to fuse BPA from three independent evidence sources, handle information conflicts, and output the final financial risk warning result.

[0160] like Figure 1 As shown, a corporate financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory is proposed. The corporate financial risk early warning method includes the following steps:

[0161] S1. Data Input and Preprocessing: Collect financial indicator data from the CNRDS database and the CNFI database, clean and preprocess the financial indicator data, perform entity identification and normalization for the enterprise domain, and perform Z-score standardization on the financial indicator data.

[0162] S2. Risk Level Definition: Select key financial indicators from the financial indicator data as model input, initialize the Gaussian mixture model using the GaussianMixture function, call the gmm.fit function to train the key financial indicator data, iteratively optimize the model parameters through the expectation-maximization (EM) algorithm, use the gmm.predict function to determine the category of each sample according to the probability density function, output the risk_label array, and use the risk_label as the target variable y for subsequent model training to construct a financial risk early warning and identification framework;

[0163] S3. Construction of Heterogeneous Knowledge Graph: Using the Neo4j graph database, create enterprise entity nodes in the equity network. For each equity investment record, connect the enterprise entity nodes of the two investing parties through network edges; extract basic information and a list of participating and holding companies from the CPCD module of the CNRS database, and extract cross-shareholding details and equity participation / control details from the listed company alliance-equity alliance module of the CSMAR database; construct the equity network knowledge graph.

[0164] The Neo4j graph database was used to create enterprise entity nodes in the supply chain network. A directed graph was constructed based on the data of the top five customers and top five suppliers of the listed company. Information on major suppliers, major customers, and shareholder information of suppliers and customers was extracted from the Supply Chain Research SCRD module of the CNRS database. Information on the top five customers and suppliers and supply chain alliances was extracted from the Supply Chain Research module of the CSMAR database. A knowledge graph of the supply chain network was constructed.

[0165] S4. Multidimensional feature extraction: Use the centrality module in the NetworkX library to extract degree centrality, proximity centrality, betweenness centrality, feature vector centrality and structural holes. Calculate the sum of squares of the shareholding ratios HHI=sum(ratio^2). Use the nx.pagerank function with the holding ratio holding_ratio as the edge weight to calculate the PageRank value of each node.

[0166] The centrality module in the NetworkX library is used to extract degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural holes. The standard deviation of supplier concentration over three consecutive years, std_concentration, is calculated, and then the coefficient of variation, cv = std_concentration / mean_concentration, is calculated.

[0167] Key financial indicators are extracted from the financial indicator data to obtain the enterprise financial status feature FIN_feature. The concat function in the pandas library is used to merge the equity network structure feature G_EQ, equity network risk feature V_EQ, supply chain network structure feature G_SC, supply chain network risk feature V_SC and enterprise financial status feature FIN_feature into a complete feature matrix X_complete, which is used as the input for subsequent model training.

[0168] S5. Basic Credibility Allocation: Divide the feature matrix X_complete into three independent evidence sources: equity network evidence source, supply chain network evidence source, and financial indicator evidence source. Use the ExplainableBoostingClassifier function in the interpreter library to initialize the EBM model, and call ebm_source.fit to train an independent EBM model for each evidence source.

[0169] Use the ebm.predict_proba function to obtain the predicted probability distribution of each model for the test set samples, and output the softmax probability distribution as the basic confidence assignment (BPA) in the DS evidence theory. Use the ebm.explain_global function to obtain the global feature importance, and use the ebm.explain_local function to obtain the feature contribution map of a single sample.

[0170] S6, DS Evidence Fusion and Decision Making: Define an identification framework Θ={low risk, medium risk, high risk} to represent three possible financial risk states. Calculate the conflict coefficient K for the BPA of each pair of evidence sources. Use the DS evidence theory synthesis rules to fuse the BPAs of the three evidence sources to obtain the final fused BPA m_final. Based on the fused BPA distribution m_final, use the maximum confidence rule to output the final risk warning result. Compare the prediction result with the true risk label risk_label, calculate evaluation indicators, and verify the model's effectiveness.

[0171] Furthermore, step S1 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0172] S11. Data Extraction: Collect financial indicator data from the CNRDS database and the CNFI database;

[0173] S12. Data Cleaning and Preprocessing: For duplicate records in financial indicator data, a hash function is used to generate a unique identifier for each transaction record, and the identifiers are compared to remove duplicates; for missing values ​​in financial indicator data, since filling in the transaction amount will cause discrepancies with the actual data, they are directly deleted; for missing timestamps in financial indicator data, if business logic allows, the transaction record date is used as an approximation, or the average timestamp of adjacent records is used.

[0174] S13. Entity Recognition and Normalization: Perform entity recognition and normalization on enterprise stock codes and stock abbreviations in the financial indicator data. Use a natural language processing model to calculate entity similarity. Normalize entities with similarity above the threshold and remove special characters and spaces.

[0175] S14. Data Standardization: Perform Z-score standardization on the financial indicator data. Use the standard_scaler function to convert each indicator into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the impact of differences in the dimensions of different indicators on model training.

[0176] Furthermore, step S2 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0177] S21. Indicator Selection: Four key financial indicators—profitability, solvency, operating efficiency, and growth potential—are selected from the financial indicator data as inputs to the Gaussian mixture model. Profitability financial indicators include return on assets, return on equity, operating profit margin, cost-to-profit ratio, and main business ratio. Solvency financial indicators include current ratio, quick ratio, debt-to-equity ratio, and times interest earned. Operating efficiency financial indicators include inventory turnover, accounts receivable turnover, current asset turnover, fixed asset turnover, and total asset turnover. Growth potential financial indicators include revenue growth rate, net profit growth rate, net cash flow growth rate from operating activities, and total asset growth rate.

[0178] S22. Initialization of Gaussian Mixture Model: Use the GaussianMixture function to initialize the Gaussian mixture model, set n_components=3 to indicate that the samples are divided into three categories of financial risk: high, medium and low; set covariance_type='full' to allow each cluster to have its own covariance matrix;

[0179] S23. Model Training: Call the gmm.fit function to train the financial indicator data, and iteratively optimize the model parameters through the expectation-maximization (EM) algorithm, including the mean, covariance matrix and weight of each Gaussian component, until the log-likelihood function converges.

[0180] S24. Risk Level Classification: The gmm.predict function is used to determine the category of each sample based on the probability density function, and the risk_label array is output, where 0 represents low risk, 1 represents medium risk, and 2 represents high risk. According to the clustering results, there are 25 high-risk companies, 201 medium-risk companies, and 36 low-risk companies out of 262 companies.

[0181] S25. Identification Framework Construction: Using risk_label as the target variable y for subsequent model training, a financial risk early warning identification framework is constructed to ensure the objectivity and scientific nature of the risk level definition.

[0182] Furthermore, the specific algorithm flow for step S2 of the enterprise financial risk early warning method is as follows:

[0183] ① Input the financial indicator data matrix X_financial:

[0184] 1. X_financial is an n×m matrix, where n is the number of companies and m is the number of financial indicators;

[0185] 2. Standardize X_financial to obtain X_standardized;

[0186] ② Gaussian mixture model clustering:

[0187] 1. Initialize the GaussianMixture model and set the number of clusters k=3;

[0188] 2. Call gmm.fit(X_standardized) to train the model;

[0189] 3. Use gmm.predict(X_standardized) to obtain the risk label (risk_label) for each sample;

[0190] ③ Output the risk level classification results:

[0191] 1. Count the number of samples in each category to verify the rationality of the division;

[0192] 2. Save risk_label as the target variable for subsequent model training.

[0193] Furthermore, step S3 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0194] S31. Equity Network Node Creation: Using the Neo4j graph database, enterprise entities are used as nodes, including listed companies and their subsidiaries, joint ventures, associates and sub-subsidiaries; the Cypher statement CREATE (n:Company {name:company_name, code: company_code, industry: industry_type}) is used to create enterprise entity nodes, and a unique identifier node_id is assigned to each entity;

[0195] S32. Equity Network Edge Creation: For each equity investment record, connect the two investing companies' nodes through a network edge; use the Cypher statement MATCH (a:Company {code: investor_code}), (b:Company {code: investee_code}) CREATE (a)-[r:EquityInvest {ratio: holding_ratio, date: invest_date}]->(b) to create an equity investment relationship edge, with the edge weight being the shareholding ratio (holding_ratio);

[0196] S33. Data sources for equity network: Basic information and a list of participating and holding companies are extracted from the CPCD module of the CNRS database; Cross-shareholding details and equity participation details are extracted from the listed company alliance-equity alliance module of the CSMAR database; Finally, an equity network knowledge graph with 97,140 nodes and 417,487 edges is constructed.

[0197] S34. Supply Chain Network Node Creation: Using the Neo4j graph database, with enterprise entities as nodes, including listed companies and their top five suppliers and top five customers; using the Cypher statement CREATE (n:SupplyChainNode {name:company_name, code: company_code, is_listed: is_listed_flag, industry:industry_type}) to create supply chain nodes, adding four attributes: company stock code, company stock abbreviation, company sub-industry, and company node size;

[0198] S35. Supply Chain Network Edge Creation: Construct a directed graph based on the purchase amount of the top five suppliers and the revenue of the top five customers of the listed company; use the Cypher statement MATCH (a:SupplyChainNode {code: supplier_code}),(b:SupplyChainNode {code: buyer_code}) CREATE (a)-[r:SupplyRelation {amount:transaction_amount, ratio: transaction_ratio, year: transaction_year}]->(b) to create supply or purchase relationship edges, with the edge weight being the ratio of sales / purchase amount to the total amount, transaction_ratio;

[0199] S36. Supply chain network data sources: Extract information on major suppliers, major customers, and shareholder information of suppliers and customers from the Supply Chain Research SCRD module of the CNRS database; extract information on the top five customers and suppliers and supply chain alliances from the Supply Chain Research module of the CSMAR database; and finally construct a supply chain network knowledge graph with 12,965 nodes and 44,201 edges.

[0200] Furthermore, the specific algorithm flow for step S3 of the enterprise financial risk early warning method is as follows:

[0201] ① Construction of equity network graph:

[0202] 1. Initialize the Neo4j graph database connection using the Graph function of the py2neo library;

[0203] 2. Traverse the equity data table `equity_data` and perform node creation and edge creation operations for each record;

[0204] 3. Use the merge statement to create each enterprise entity, ensuring the exhaustiveness and uniqueness of each enterprise entity;

[0205] 4. Assign a weight to each edge as the holding ratio;

[0206] ② Supply chain network mapping:

[0207] 1. Initialize the Neo4j graph database connection using the Graph function of the py2neo library;

[0208] 2. Traverse the supply chain data table supply_chain_data and perform node creation and edge creation operations for each record;

[0209] 3. Use the merge statement to create each enterprise entity, ensuring the exhaustiveness and uniqueness of each enterprise entity;

[0210] 4. Assign a weight to each edge as the proportion of sales or purchases to the total amount (transaction_ratio);

[0211] ③Graphical verification:

[0212] 1. Use the Cypher statement `MATCH (n) RETURN count(n)` to verify the number of nodes;

[0213] 2. Use the Cypher statement `MATCH ()-[r]->() RETURN count(r)` to verify the number of edges;

[0214] 3. Check the connectivity of the graph to ensure there are no isolated nodes.

[0215] Furthermore, step S4 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0216] S41. Equity Network Structure Feature G_EQ Extraction: The following features are extracted using the centrality module of the NetworkX library:

[0217] 1. Degree centrality: Calculated using the nx.degree_centrality function; a higher value indicates greater importance in the network.

[0218] 2. Proximity Centrality: Calculated using the nx.closeness_centrality function, it reflects the degree of proximity between a company and other companies in a network;

[0219] 3. Betweenness centrality: Calculated using the nx.betweenness_centrality function, it represents the number of times a firm acts as a bridge between two other firms on the shortest path.

[0220] 4. Eigenvector centrality: Calculated using the nx.eigenvector_centrality function, it measures the overall influence and centrality of a listed company or shareholder within its network of connections.

[0221] 5. Structural holes: Calculated using the networkx.algorithms.structural_holes function, this represents the number of structural holes occupied. The larger the value, the greater the information advantage.

[0222] S42. Extraction of V_EQ characteristics of equity network risk:

[0223] 1. Equity Concentration: The sum of squares of the shareholding ratios of shareholders, HHI=sum(ratio^2), reflects the degree of concentration of the equity structure;

[0224] 2. PageRank centrality based on shareholding ratio: Using the nx.pagerank function, the PageRank value of each node is calculated with the shareholding ratio as the edge weight. Companies with higher values ​​are in a core position in the entire network.

[0225] S43. Extraction of Supply Chain Network Structural Features G_SC: Similar to the structural features of equity networks, the centrality module of the NetworkX library is used to extract features, including degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural holes. These features reflect the degree of association between the focal enterprise and its supply chain partners and other highly interconnected enterprises in the network, and measure whether the enterprise is connected to other unconnected supply chain clusters.

[0226] S44. Supply Chain Network Risk Characteristics V_SC Extraction:

[0227] 1. Supplier instability: Calculate the standard deviation of supplier concentration (std_concentration) over three consecutive years, and then calculate the coefficient of variation (cv) = std_concentration / mean_concentration. The higher the value, the more unstable the supplier dependency structure.

[0228] 2. Customer instability: The calculation method is symmetrical with that of supplier instability, reflecting the degree of change in a company's dependence on its major customers;

[0229] S45. Extraction of Enterprise Financial Status Characteristics (FIN_feature): Extract the following four categories of key financial indicators from the financial statements:

[0230] 1. Profitability financial indicators: Return on assets, Return on equity, Operating profit margin, Cost-to-profit ratio, and main business ratio;

[0231] 2. Financial indicators of solvency: current ratio, quick ratio, debt-to-equity ratio, and times interest earned;

[0232] 3. Financial indicators of operational capabilities: inventory turnover, accounts receivable turnover, current asset turnover, fixed asset turnover, and total asset turnover;

[0233] 4. Financial indicators of growth potential: Revenue growth rate, net profit growth rate, net cash flow growth rate from operating activities, and total asset growth rate;

[0234] S46. Feature Integration: Use the concat function of the pandas library to merge the equity network structure feature G_EQ, equity network risk feature V_EQ, supply chain network structure feature G_SC, supply chain network risk feature V_SC, and enterprise financial status feature FIN_feature into a complete feature matrix X_complete, which will be used as the input for subsequent model training.

[0235] Furthermore, the specific algorithm flow for step S4 of the enterprise financial risk early warning method is as follows:

[0236] ① Network feature extraction:

[0237] Load the equity network graph (equity_graph) and the supply chain network graph (supply_graph). For each node, call the function to calculate the degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural hole constraint.

[0238] ② Risk characteristic calculation:

[0239] 1. Traverse the equity data table and calculate the equity concentration for each company;

[0240] 2. Call nx.pagerank(equity_graph, weight='holding_ratio') to calculate PageRank centrality;

[0241] 3. Traverse the supply chain data table and calculate the supplier and customer instability for each enterprise;

[0242] ③ Construction of the feature matrix:

[0243] 1. Initialize an empty list feature_list;

[0244] 2. For each enterprise, add network characteristics, risk characteristics, and financial indicators to the feature_list;

[0245] 3. Use pd.DataFrame(feature_list) to create the feature matrix X_complete;

[0246] 4. Use StandardScaler to standardize the feature matrix.

[0247] Furthermore, step S5 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0248] S51. Evidence Source Division: Divide the feature matrix X_complete into three independent evidence sources:

[0249] Evidence sources for equity networks include structural characteristics and risk characteristics of equity networks;

[0250] Sources of evidence for supply chain networks include structural characteristics and risk characteristics of supply chain networks;

[0251] Financial indicator evidence sources include indicators of profitability, solvency, operational efficiency, and growth potential.

[0252] S52. EBM Model Initialization: Initialize the EBM model using the ExplainableBoostingClassifier function from the interpreter library. Set random_state to ensure reproducible results, set the number of weak learners n_estimators=300, and set the learning rate learning_rate=0.01.

[0253] S53. Model Training: Call ebm_source.fit to train an independent EBM model for each evidence source;

[0254] S54, BPA Acquisition: Use the ebm.predict_proba function to obtain the predicted probability distribution of each model for the test set samples, and output the softmax probability distribution as the basic confidence assignment (BPA) in the DS evidence theory;

[0255] S55. Interpretability Analysis: Use the ebm.explain_global function to obtain the global feature importance, and use the ebm.explain_local function to obtain the feature contribution map of a single sample, clearly showing how each feature affects the final prediction result.

[0256] Furthermore, the specific algorithm flow for step S5 of the enterprise financial risk early warning method is as follows:

[0257] ①Evidence source model training:

[0258] 1. Initialize three EBM model instances: ebm_equity, ebm_supply, and ebm_financial;

[0259] 2. For each source of evidence, call model.fit(X_source, y) to train the model;

[0260] 3. Use cross-validation to evaluate the performance of each model and record the three evaluation metrics: accuracy, recall, and F1 score.

[0261] ②BPA generation: For the test set samples, call ebm_source.predict_proba(X_source_test) to obtain the equity network BPA, supply chain network BPA, and financial indicator BPA respectively;

[0262] Each BPA is an n×3 matrix, where n is the number of samples and 3 is the number of risk levels, such as high, medium, and low.

[0263] ③ Interpretable output:

[0264] Call ebm.explain_global() to generate a global feature importance ranking graph;

[0265] Call ebm.explain_local(X_sample) to generate a feature contribution map for a single sample.

[0266] Furthermore, step S6 of the enterprise financial risk early warning method specifically includes the following sub-steps:

[0267] S61. Definition of Recognition Framework: Define the recognition framework. ={Low Risk, Medium Risk, High Risk} represents three possible financial risk levels; each BPA is... The fundamental probability assignment function on the power set;

[0268] S62. Conflict Coefficient Calculation: For each pair of evidence sources' BPAs, calculate the conflict coefficient K; use the NumPy library to iterate through all focal element combinations, and for those satisfying... The focal element combination is accumulated as m1(B)×m2(C); the conflict coefficient K measures the degree of conflict between two sources of evidence, and the larger the K value, the more serious the conflict.

[0269] S63, Application of DS Synthesis Rules: The synthesis rules of DS evidence theory are used to fuse the BPAs of three evidence sources to obtain the final fused BPA m_final;

[0270] S64. Decision rule: Based on the fused BPA distribution m_final, the final risk warning result is output using the maximum confidence rule; for each sample, the risk level A with the largest m_final(A) is selected as the prediction result, and the confidence level confidence=m_final(A) corresponding to the risk level is output at the same time.

[0271] S65. Early Warning Result Output: Compare the predicted results with the true risk label (risk_label) and calculate eight evaluation metrics: accuracy, precision, recall, F1 score (f1_score), area under the ROC curve (AUC), area under the PR curve (AUC-PR), geometric mean (Gmean), and false negative rate (FNR) to verify the effectiveness of the model.

[0272] Furthermore, the specific algorithm flow for step S6 of the enterprise financial risk early warning method is as follows:

[0273] ① Calculation of conflict coefficient:

[0274] 1. Initialize the conflict coefficient K=0;

[0275] 2. Traverse all focal elements B of evidence source 1 and all focal elements C of evidence source 2;

[0276] 3. If Then K = K + m1(B) × m2(C);

[0277] 4. Return the conflict coefficient K, which measures the degree of conflict between sources of evidence;

[0278] ②D-S evidence fusion:

[0279] 1. Initialize the fusion result m_fused as the identity matrix;

[0280] 2. For every two evidence sources BPA m1 and m2, call the ds_combine(m1, m2) function to fuse them;

[0281] 3. In the ds_combine function, iterate through all focal element combinations B and C;

[0282] 4. For the focal element combination that satisfies B∩C=A, accumulate m1(B)×m2(C) to m_fused(A);

[0283] 5. Normalize the fusion results using a normalization factor of 1 / (1-K);

[0284] 6. Return the fused BPA m_fused;

[0285] ③Final decision:

[0286] 1. For the fused BPA m_final, call np.argmax(m_final, axis=1) to obtain the predicted risk level of each sample;

[0287] 2. Call np.max(m_final, axis=1) to get the prediction confidence for each sample;

[0288] 3. Compare the predicted results with the actual labels and calculate the evaluation indicators;

[0289] 4. Output the final financial risk warning results, including risk level and confidence level;

[0290] ④ Model Evaluation:

[0291] The model's overall performance is verified by calling functions to calculate eight evaluation metrics: accuracy, precision, recall, F1 score (f1_score), area under the ROC curve (AUC), area under the PR curve (AUC-PR), geometric mean (Gmean), and false negative rate (FNR).

[0292] like Figure 1 The diagram illustrates the operational process of a corporate financial risk early warning method that integrates multi-source heterogeneous knowledge graphs and evidence theory.

[0293] First, the system collects enterprise equity, supply chain, and financial statement data from CNRS and CSMAR databases, and performs cleaning and preprocessing to remove duplicate records and outliers, followed by entity identification and normalization. Next, the system uses Gaussian mixture model clustering to analyze the enterprise's financial situation, objectively classifying it into high, medium, and low financial risk levels, and constructing an identification framework. Then, the system constructs enterprise equity network knowledge graphs and supply chain network knowledge graphs respectively, transforming unstructured relational data into analyzable structured knowledge. Based on this, the system extracts deep network structure features and risk transmission characteristics from the knowledge graphs, and simultaneously extracts enterprise financial indicator features from the financial statements.

[0294] The system treats equity network characteristics, supply chain network characteristics, and financial indicators as independent sources of evidence, inputting them into an Interpretable Augmentation Machine (EBM) model to obtain the model's basic credibility assignment for each source of evidence within the risk framework. Subsequently, the system applies the synthesis rules of the DS evidence theory to fuse the Basic Perceptions (BPAs) from the three independent sources, effectively handling conflicts and uncertainties between evidence, and ultimately generating a unified BPA that incorporates all information. Based on the distribution of the fused BPA, the system outputs the final financial risk warning result.

[0295] Users can view the enterprise's risk warning results through the system interface, including risk level, confidence level, and analysis of the contribution of each feature to the risk. The system also supports comparative model evaluation, comparing the proposed model with benchmark models (such as LR, DT, SVM, RF, etc.) to verify the overall performance of the proposed model. Experimental results show that the model integrating all indicators has the best overall performance, proving that integrating "equity network, supply chain network, and financial indicators" through DS evidence theory is indeed more discriminative than a single dimension or pairwise integration. The system also provides interpretability analysis, demonstrating the specific impact mechanism of each network feature on financial risk, helping users to deeply understand the risk transmission path.

[0296] This invention is based on the principle of structured representation of enterprise relationship networks using knowledge graphs. It constructs equity relationships and supply chain relationships between enterprises into a knowledge graph, forming a structured network representation. The knowledge graph is defined as... ,in For enterprise node set, For enterprise relationship edge set, For an equity network, the edge weights are a set of edge weights. Indicates enterprise For enterprises The shareholding ratio meets the requirements. For supply chain networks, marginal rights Indicates enterprise With enterprises The proportion of transactions between them to the total amount satisfies Through the adjacency matrix Represents the network structure, where If an edge exists ,otherwise This structured representation transforms unstructured, interconnected data into a computable network, laying the mathematical foundation for subsequent feature extraction and risk quantification. This enables the precise characterization and quantification of capital ties and operational dependencies between enterprises.

[0297] This invention utilizes the principles of multidimensional feature extraction and risk quantification based on complex network theory. It extracts network structure features and risk transmission features from knowledge graphs, and quantifies a company's positional advantage and risk exposure within the network using mathematical formulas. Structural features include: degree centrality. Reflecting nodes Number of direct connections; proximity centrality ,in For nodes arrive The shortest path length reflects the proximity of a node to other nodes; betweenness centrality. ,in For nodes arrive The total number of shortest paths, For the nodes The number of shortest paths reflects the intermediary control ability of a node; eigenvector centrality ,in The largest eigenvalue of the adjacency matrix reflects the overall influence of the nodes. Risk characteristics include: equity concentration. ,in For the first The shareholding ratio of each shareholder reflects the degree of concentration of the equity structure; PageRank centrality based on shareholding ratio. ,in The damping coefficient is... Pointing to a node The set of nodes reflects the core position of each node in the network; supplier instability ,in For the first Supplier concentration in 2018 The average concentration reflects the degree of volatility in the supplier dependence structure. These characteristics, expressed through mathematical formulas, precisely quantify a firm's structural position and risk exposure within the network, providing calculable feature inputs for risk warning.

[0298] This invention is based on the objective definition principle of risk levels using Gaussian Mixture Models (GMMs). It employs GMMs to perform cluster analysis on corporate financial situations, objectively classifying financial risk levels. Let the corporate financial feature vector be... ,in For financial indicators, the probability density function of the Gaussian mixture model is... ,in Number of risk level categories (as defined in this invention) (corresponding to high, medium, and low risk levels) For the first The mixing coefficients of the Gaussian components satisfy the following conditions: and , For the first A Gaussian distribution, It is the mean vector. Let be the covariance matrix. The parameters are iteratively optimized using the Expectation-Maximization (EM) algorithm. Maximize the log-likelihood function .sample Belongs to the The posterior probability of the class is The risk level of the sample is determined based on the maximum a posteriori probability principle. This method avoids subjective thresholding and achieves an objective definition of risk level through probability distribution fitting, providing accurate target variables for subsequent model training.

[0299] Based on the fundamental credibility assignment principle of the Interpretable Augmentation Machine (EBM), this invention employs the Interpretable Augmentation Machine (EBM) model to obtain the risk probability assignment for each feature dimension, serving as the Basic Credibility Assignment (BPA) for the DS evidence theory. EBM is a generalized additive model, and its mathematical form is... ,in This is the connection function (the softmax function is used for multi-class classification problems). For the sample Belongs to the The probability of a risk level. For the first The intercept term of the class, For the first The feature is related to the first The contribution function for each class of risk is learned using the Boosting algorithm. For multi-class classification problems, the softmax function is... This invention uses equity network characteristics, supply chain network characteristics, and financial indicators as independent sources of evidence, and trains an independent EBM model for each source of evidence, the output probability distribution of which... This is considered the basic credibility allocation in the DS evidence theory, where This represents the evidence source index. The interpretability of EBM lies in the fact that the learning of each feature is independent, and the feature contribution function... It can be visualized as a single feature contribution map, clearly showing how each feature affects the final prediction result, achieving "glass box" interpretability.

[0300] This invention utilizes the multi-source information fusion and decision-making principle based on the DS evidence theory. It applies the synthesis rules of the DS evidence theory to fuse the basic credibility allocations from multiple independent evidence sources, effectively handling information conflicts and uncertainties. A recognition framework is established. For the set of all possible risk levels, For the first The basic credibility allocation function for each source of evidence satisfies and For two independent sources of evidence and Its DS synthesis rule is defined as follows: ,in Conflict coefficient To measure the degree of conflict between two sources of evidence, For multiple sources of evidence, an iterative synthesis approach is employed. Basic credibility allocation after fusion Indicates the proposition The overall level of trust is considered. The final decision-making rule is to select the risk level with the highest level of trust. The advantage of the DS evidence theory lies in its ability to quantify and handle conflicts between pieces of evidence, when the conflict coefficient... When the value is large, use a normalization factor. Adjusting the fusion results ensures the robustness of decision-making. Based on this principle, this invention effectively integrates evidence from three dimensions—equity networks, supply chain networks, and financial indicators—to obtain a comprehensive risk assessment, significantly improving the accuracy and credibility of risk warnings.

[0301] The technical solutions disclosed in the embodiments of the present invention have been described in detail above. Specific embodiments have been used to illustrate the principles and implementation methods of the embodiments of the present invention. The description of the above embodiments is only for helping to understand the principles of the embodiments of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for fusing multi-source heterogeneous knowledge graph and evidence theory for enterprise financial risk early warning, characterized in that, The enterprise financial risk early warning method is based on an enterprise financial risk early warning system, which includes a data input and preprocessing module, a risk level definition module, a heterogeneous knowledge graph construction module, a multi-dimensional feature engineering extraction module, a basic credibility allocation acquisition module, and a DS evidence fusion and decision-making module. The enterprise financial risk early warning method includes the following steps: S1. Data Input and Preprocessing: Collect financial indicator data from the CNRDS database and the CNFI database, clean and preprocess the financial indicator data, perform entity identification and normalization for the enterprise domain, and perform Z-score standardization on the financial indicator data. S2. Risk Level Definition: Select key financial indicators from the financial indicator data as model input, initialize the Gaussian mixture model using the GaussianMixture function, call the gmm.fit function to train the key financial indicator data, iteratively optimize the model parameters through the expectation-maximization (EM) algorithm, use the gmm.predict function to determine the category of each sample according to the probability density function, output the risk_label array, and use the risk_label as the target variable y for subsequent model training to construct a financial risk early warning and identification framework; S3. Construction of Heterogeneous Knowledge Graph: Using the Neo4j graph database, create enterprise entity nodes in the equity network. For each equity investment record, connect the enterprise entity nodes of the two investing parties through network edges; extract basic information and a list of participating and holding companies from the CPCD module of the CNRS database, and extract cross-shareholding details and equity participation / control details from the listed company alliance-equity alliance module of the CSMAR database; construct the equity network knowledge graph. The Neo4j graph database was used to create enterprise entity nodes in the supply chain network. A directed graph was constructed based on the data of the top five customers and top five suppliers of the listed company. Information on major suppliers, major customers, and shareholder information of suppliers and customers was extracted from the Supply Chain Research SCRD module of the CNRS database. Information on the top five customers and suppliers and supply chain alliances was extracted from the Supply Chain Research module of the CSMAR database. A knowledge graph of the supply chain network was constructed. S4. Multidimensional feature extraction: Use the centrality module in the NetworkX library to extract degree centrality, proximity centrality, betweenness centrality, feature vector centrality and structural holes. Calculate the sum of squares of the shareholding ratios HHI=sum(ratio^2). Use the nx.pagerank function with the holding ratio holding_ratio as the edge weight to calculate the PageRank value of each node. The centrality module in the NetworkX library is used to extract degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural holes. The standard deviation of supplier concentration over three consecutive years, std_concentration, is calculated, and then the coefficient of variation, cv = std_concentration / mean_concentration, is calculated. Key financial indicators are extracted from the financial indicator data to obtain the enterprise financial status feature FIN_feature. The concat function in the pandas library is used to merge the equity network structure feature G_EQ, equity network risk feature V_EQ, supply chain network structure feature G_SC, supply chain network risk feature V_SC and enterprise financial status feature FIN_feature into a complete feature matrix X_complete, which is used as the input for subsequent model training. S5. Basic Credibility Allocation: Divide the feature matrix X_complete into three independent evidence sources: equity network evidence source, supply chain network evidence source, and financial indicator evidence source. Use the ExplainableBoostingClassifier function in the interpreter library to initialize the EBM model, and call ebm_source.fit to train an independent EBM model for each evidence source. Use the ebm.predict_proba function to obtain the predicted probability distribution of each model for the test set samples, and output the softmax probability distribution as the basic confidence assignment (BPA) in the DS evidence theory. Use the ebm.explain_global function to obtain the global feature importance, and use the ebm.explain_local function to obtain the feature contribution map of a single sample. S6, D-S evidence fusion and decision: define the recognition framework {low risk, medium risk, high risk} represents three possible financial risk states, the conflict coefficient K is calculated for the BPA of each two evidence sources, the BPA of the three evidence sources is fused using the D-S evidence theory synthesis rule to obtain the final fusion BPA m_final, the final risk warning result is output according to the fused BPA distribution m_final using the maximum confidence rule, the prediction result is compared with the true risk label risk_label, the evaluation index is calculated and the model validity is verified.

2. The method of claim 1, wherein the method comprises: Step S1 of the aforementioned enterprise financial risk early warning method specifically includes the following sub-steps: S11. Data Extraction: Collect financial indicator data from the CNRDS database and the CNFI database; S12. Data Cleaning and Preprocessing: For duplicate records in financial indicator data, a hash function is used to generate a unique identifier for each transaction record, and the identifiers are compared to remove duplicates; for missing values ​​in financial indicator data, since filling in the transaction amount will cause discrepancies with the actual data, they are directly deleted; for missing timestamps in financial indicator data, if business logic allows, the transaction record date is used as an approximation, or the average timestamp of adjacent records is used. S13. Entity Recognition and Normalization: Perform entity recognition and normalization on enterprise stock codes and stock abbreviations in the financial indicator data. Use a natural language processing model to calculate entity similarity. Normalize entities with similarity above the threshold and remove special characters and spaces. S14. Data Standardization: Perform Z-score standardization on the financial indicator data. Use the standard_scaler function to convert each indicator into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the impact of differences in the dimensions of different indicators on model training.

3. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 1, characterized in that, Step S2 of the enterprise financial risk early warning method specifically includes the following sub-steps: S21. Indicator Selection: Four key financial indicators—profitability, solvency, operating efficiency, and growth potential—are selected from the financial indicator data as inputs to the Gaussian mixture model. Profitability financial indicators include return on assets, return on equity, operating profit margin, cost-to-profit ratio, and main business ratio. Solvency financial indicators include current ratio, quick ratio, debt-to-equity ratio, and times interest earned. Operating efficiency financial indicators include inventory turnover, accounts receivable turnover, current asset turnover, fixed asset turnover, and total asset turnover. Growth potential financial indicators include revenue growth rate, net profit growth rate, net cash flow growth rate from operating activities, and total asset growth rate. S22. Initialization of Gaussian Mixture Model: Use the GaussianMixture function to initialize the Gaussian mixture model, set n_components=3 to indicate that the samples are divided into three categories of financial risk: high, medium and low; set covariance_type='full' to allow each cluster to have its own covariance matrix; S23. Model Training: Call the gmm.fit function to train the financial indicator data, and iteratively optimize the model parameters through the expectation-maximization (EM) algorithm, including the mean, covariance matrix and weight of each Gaussian component, until the log-likelihood function converges. S24. Risk Level Classification: The gmm.predict function is used to determine the category of each sample based on the probability density function, and the risk_label array is output, where 0 represents low risk, 1 represents medium risk, and 2 represents high risk. According to the clustering results, there are 25 high-risk companies, 201 medium-risk companies, and 36 low-risk companies out of 262 companies. S25. Identification Framework Construction: Using risk_label as the target variable y for subsequent model training, a financial risk early warning identification framework is constructed to ensure the objectivity and scientific nature of the risk level definition.

4. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 3, characterized in that, The specific algorithm flow for step S2 of the enterprise financial risk early warning method is as follows: ① Input the financial indicator data matrix X_financial:

1. X_financial is an n×m matrix, where n is the number of companies and m is the number of financial indicators; 2. Standardize X_financial to obtain X_standardized; ② Gaussian mixture model clustering:

1. Initialize the GaussianMixture model and set the number of clusters k=3; 2. Call gmm.fit(X_standardized) to train the model; 3. Use gmm.predict(X_standardized) to obtain the risk label (risk_label) for each sample; ③ Output the risk level classification results:

1. Count the number of samples in each category to verify the rationality of the division; 2. Save risk_label as the target variable for subsequent model training.

5. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 1, characterized in that, Step S3 of the enterprise financial risk early warning method specifically includes the following sub-steps: S31. Equity Network Node Creation: Using the Neo4j graph database, enterprise entities are used as nodes, including listed companies and their subsidiaries, joint ventures, associates and sub-subsidiaries; the Cypher statement CREATE (n:Company {name:company_name, code: company_code, industry: industry_type}) is used to create enterprise entity nodes, and a unique identifier node_id is assigned to each entity; S32. Equity Network Edge Creation: For each equity investment record, connect the two investing companies' nodes through a network edge; use the Cypher statement MATCH (a:Company {code: investor_code}), (b:Company{code: investee_code}) CREATE (a)-[r:EquityInvest {ratio: holding_ratio,date: invest_date}]->(b) to create an equity investment relationship edge, with the edge weight being the shareholding ratio (holding_ratio); S33. Data sources for equity network: Basic information and a list of participating and holding companies are extracted from the CPCD module of the CNRS database; Cross-shareholding details and equity participation details are extracted from the listed company alliance-equity alliance module of the CSMAR database; Finally, an equity network knowledge graph with 97,140 nodes and 417,487 edges is constructed. S34. Supply Chain Network Node Creation: Using the Neo4j graph database, with enterprise entities as nodes, including listed companies and their top five suppliers and top five customers; using the Cypher statement CREATE (n:SupplyChainNode {name:company_name, code: company_code, is_listed: is_listed_flag, industry:industry_type}) to create supply chain nodes, adding four attributes: company stock code, company stock abbreviation, company sub-industry, and company node size; S35. Supply Chain Network Edge Creation: Construct a directed graph based on the purchase amount of the top five suppliers and the revenue of the top five customers of the listed company; use the Cypher statement MATCH (a:SupplyChainNode {code: supplier_code}), (b:SupplyChainNode {code: buyer_code}) CREATE (a)-[r:SupplyRelation {amount:transaction_amount, ratio: transaction_ratio, year: transaction_year}]->(b) to create supply or purchase relationship edges, with the edge weight being the ratio of sales / purchase amount to the total amount, transaction_ratio; S36. Supply chain network data sources: Extract information on major suppliers, major customers, and shareholder information of suppliers and customers from the Supply Chain Research SCRD module of the CNRS database; extract information on the top five customers and suppliers and supply chain alliances from the Supply Chain Research module of the CSMAR database; and finally construct a supply chain network knowledge graph with 12,965 nodes and 44,201 edges.

6. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 5, characterized in that, The specific algorithm flow for step S3 of the enterprise financial risk early warning method is as follows: ① Construction of equity network graph:

1. Initialize the Neo4j graph database connection using the Graph function of the py2neo library; 2. Traverse the equity data table `equity_data` and perform node creation and edge creation operations for each record; 3. Use the merge statement to create each enterprise entity, ensuring the exhaustiveness and uniqueness of each enterprise entity; 4. Assign a weight to each edge as the holding ratio; ② Supply chain network mapping:

1. Initialize the Neo4j graph database connection using the Graph function of the py2neo library; 2. Traverse the supply chain data table supply_chain_data and perform node creation and edge creation operations for each record; 3. Use the merge statement to create each enterprise entity, ensuring the exhaustiveness and uniqueness of each enterprise entity; 4. Assign a weight to each edge as the proportion of sales or purchases to the total amount (transaction_ratio); ③Graphical verification:

1. Use the Cypher statement `MATCH (n) RETURN count(n)` to verify the number of nodes; 2. Use the Cypher statement `MATCH ()-[r]->() RETURN count(r)` to verify the number of edges; 3. Check the connectivity of the graph to ensure there are no isolated nodes.

7. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 1, characterized in that, Step S4 of the enterprise financial risk early warning method specifically includes the following sub-steps: S41. Equity Network Structure Feature G_EQ Extraction: The following features are extracted using the centrality module of the NetworkX library:

1. Degree centrality: Calculated using the nx.degree_centrality function; a higher value indicates greater importance in the network.

2. Proximity Centrality: Calculated using the nx.closeness_centrality function, it reflects the degree of proximity between a company and other companies in a network; 3. Betweenness centrality: Calculated using the nx.betweenness_centrality function, it represents the number of times a firm acts as a bridge between two other firms on the shortest path.

4. Eigenvector centrality: Calculated using the nx.eigenvector_centrality function, it measures the overall influence and centrality of a listed company or shareholder within its network of connections.

5. Structural holes: Calculated using the networkx.algorithms.structural_holes function, this represents the number of structural holes occupied. The larger the value, the greater the information advantage. S42. Extraction of V_EQ characteristics of equity network risk:

1. Equity Concentration: The sum of squares of the shareholding ratios of shareholders, HHI=sum(ratio^2), reflects the degree of concentration of the equity structure; 2. PageRank centrality based on shareholding ratio: Using the nx.pagerank function, the PageRank value of each node is calculated with the shareholding ratio as the edge weight. Companies with higher values ​​are in a core position in the entire network. S43. Extraction of Supply Chain Network Structural Features G_SC: Similar to the structural features of equity networks, the centrality module of the NetworkX library is used to extract features, including degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural holes. These features reflect the degree of association between the focal enterprise and its supply chain partners and other highly interconnected enterprises in the network, and measure whether the enterprise is connected to other unconnected supply chain clusters. S44. Supply Chain Network Risk Characteristics V_SC Extraction:

1. Supplier instability: Calculate the standard deviation of supplier concentration (std_concentration) over three consecutive years, and then calculate the coefficient of variation (cv) = std_concentration / mean_concentration. The higher the value, the more unstable the supplier dependency structure.

2. Customer instability: The calculation method is symmetrical with that of supplier instability, reflecting the degree of change in a company's dependence on its major customers; S45. Extraction of Enterprise Financial Status Characteristics (FIN_feature): Extract the following four categories of key financial indicators from the financial statements:

1. Profitability financial indicators: Return on assets, Return on equity, Operating profit margin, Cost-to-profit ratio, and main business ratio; 2. Financial indicators of solvency: current ratio, quick ratio, debt-to-equity ratio, and times interest earned; 3. Financial indicators of operational capabilities: inventory turnover, accounts receivable turnover, current asset turnover, fixed asset turnover, and total asset turnover; 4. Financial indicators of growth potential: Revenue growth rate, net profit growth rate, net cash flow growth rate from operating activities, and total asset growth rate; S46. Feature Integration: Use the concat function of the pandas library to merge the equity network structure feature G_EQ, equity network risk feature V_EQ, supply chain network structure feature G_SC, supply chain network risk feature V_SC, and enterprise financial status feature FIN_feature into a complete feature matrix X_complete, which will be used as the input for subsequent model training.

8. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 7, characterized in that, The specific algorithm flow for step S4 of the enterprise financial risk early warning method is as follows: ① Network feature extraction: Load the equity network graph (equity_graph) and the supply chain network graph (supply_graph). For each node, call the function to calculate the degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, and structural hole constraint. ② Risk characteristic calculation:

1. Traverse the equity data table and calculate the equity concentration for each company; 2. Call nx.pagerank(equity_graph, weight='holding_ratio') to calculate PageRank centrality; 3. Traverse the supply chain data table and calculate the supplier and customer instability for each enterprise; ③ Construction of the feature matrix:

1. Initialize an empty list feature_list; 2. For each enterprise, add network characteristics, risk characteristics, and financial indicators to the feature_list; 3. Use pd.DataFrame(feature_list) to create the feature matrix X_complete; 4. Use StandardScaler to standardize the feature matrix.

9. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 1, characterized in that, Step S5 of the enterprise financial risk early warning method specifically includes the following sub-steps: S51. Evidence Source Division: Divide the feature matrix X_complete into three independent evidence sources: Evidence sources for equity networks include structural characteristics and risk characteristics of equity networks; Sources of evidence for supply chain networks include structural characteristics and risk characteristics of supply chain networks; Financial indicator evidence sources include indicators of profitability, solvency, operational efficiency, and growth potential. S52. EBM Model Initialization: Initialize the EBM model using the ExplainableBoostingClassifier function from the interpreter library. Set random_state to ensure reproducible results, set the number of weak learners n_estimators=300, and set the learning rate learning_rate=0.

01. S53. Model Training: Call ebm_source.fit to train an independent EBM model for each evidence source; S54, BPA Acquisition: Use the ebm.predict_proba function to obtain the predicted probability distribution of each model for the test set samples, and output the softmax probability distribution as the basic confidence assignment (BPA) in the DS evidence theory; S55. Interpretability Analysis: Use the ebm.explain_global function to obtain the global feature importance, and use the ebm.explain_local function to obtain the feature contribution map of a single sample, clearly showing how each feature affects the final prediction result.

10. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 9, characterized in that, The specific algorithm flow for step S5 of the enterprise financial risk early warning method is as follows: ①Evidence source model training:

1. Initialize three EBM model instances: ebm_equity, ebm_supply, and ebm_financial; 2. For each source of evidence, call model.fit(X_source, y) to train the model; 3. Use cross-validation to evaluate the performance of each model and record the three evaluation metrics: accuracy, recall, and F1 score. ②BPA generation: For the test set samples, call ebm_source.predict_proba(X_source_test) to obtain the equity network BPA, supply chain network BPA, and financial indicator BPA respectively; Each BPA is an n×3 matrix, where n is the number of samples and 3 is the number of risk levels, such as high, medium, and low. ③ Interpretable output: Call ebm.explain_global() to generate a global feature importance ranking graph; Call ebm.explain_local(X_sample) to generate a feature contribution map for a single sample.

11. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 1, characterized in that, Step S6 of the enterprise financial risk early warning method specifically includes the following sub-steps: S61. Definition of Recognition Framework: Define the recognition framework. ={Low Risk, Medium Risk, High Risk} represents three possible financial risk levels; each BPA is... The fundamental probability assignment function on the power set; S62. Conflict Coefficient Calculation: For each pair of evidence sources' BPAs, calculate the conflict coefficient K; use the NumPy library to iterate through all focal element combinations, and for those satisfying... The focal element combination is accumulated as m1(B)×m2(C); the conflict coefficient K measures the degree of conflict between two sources of evidence, and the larger the K value, the more serious the conflict. S63, Application of DS Synthesis Rules: The synthesis rules of DS evidence theory are used to fuse the BPAs of three evidence sources to obtain the final fused BPA m_final; S64. Decision rule: Based on the fused BPA distribution m_final, the maximum confidence rule is used to output the final risk warning result; For each sample, select the risk level A with the highest m_final(A) as the prediction result, and output the confidence level corresponding to the risk level: confidence=m_final(A); S65. Early Warning Result Output: Compare the predicted results with the true risk label (risk_label) and calculate eight evaluation metrics: accuracy, precision, recall, F1 score (f1_score), area under the ROC curve (AUC), area under the PR curve (AUC-PR), geometric mean (Gmean), and false negative rate (FNR) to verify the effectiveness of the model.

12. The enterprise financial risk early warning method integrating multi-source heterogeneous knowledge graphs and evidence theory according to claim 11, characterized in that, The specific algorithm flow for step S6 of the enterprise financial risk early warning method is as follows: ① Calculation of conflict coefficient:

1. Initialize the conflict coefficient K=0; 2. Traverse all focal elements B of evidence source 1 and all focal elements C of evidence source 2; 3. If Then K = K + m1(B) × m2(C); 4. Return the conflict coefficient K, which measures the degree of conflict between sources of evidence; ②D-S evidence fusion:

1. Initialize the fusion result m_fused as the identity matrix; 2. For every two evidence sources BPA m1 and m2, call the ds_combine(m1, m2) function to fuse them; 3. In the ds_combine function, iterate through all focal element combinations B and C; 4. For the focal element combination that satisfies B∩C=A, accumulate m1(B)×m2(C) to m_fused(A); 5. Normalize the fusion results using a normalization factor of 1 / (1-K); 6. Return the fused BPA m_fused; ③Final decision:

1. For the fused BPA m_final, call np.argmax(m_final, axis=1) to obtain the predicted risk level of each sample; 2. Call np.max(m_final, axis=1) to get the prediction confidence for each sample; 3. Compare the predicted results with the actual labels and calculate the evaluation indicators; 4. Output the final financial risk warning results, including risk level and confidence level; ④ Model evaluation: The model's overall performance is verified by calling functions to calculate eight evaluation metrics: accuracy, precision, recall, F1 score (f1_score), area under the ROC curve (AUC), area under the PR curve (AUC-PR), geometric mean (Gmean), and false negative rate (FNR).

Images