Method for generating dynamic graph of financing risk report based on multi-source invoice data

Abstract

The present application relates to the technical field of financing risk analysis, in particular to a financing risk report dynamic graph generation method based on multi-source invoice data; obtaining input invoice data and output invoice data, extracting first business features and second business features; taking financing events as time anchor points, for each financing event, applying an operating risk analysis model to process the first business features and the second business features, generating a pre-financing risk set and a post-financing risk set including risk types; performing first-order differential analysis on the pre-financing risk set and the post-financing risk set, identifying risk state parameters; for multiple financing events, performing second-order differential analysis, identifying the cross-cycle change trend of the target enterprise risk mode; based on the first-order and second-order differential analysis, generating a multi-layer time sequence knowledge graph to constitute a financing risk dynamic graph. The present application realizes dynamic analysis of financing risk through multi-source invoice data recognition.

Description

Technical Field

[0001] This invention relates to the field of financing risk analysis technology, specifically a method for generating dynamic graphs of financing risk reports based on multi-source invoice data. Background Technology

[0002] In modern finance, especially in venture capital, equity financing, and credit approval, risk assessment of enterprises is a core element in determining the direction and safety of capital investment. In particular, financial institutions urgently need a precise, dynamic, and forward-looking risk assessment mechanism when conducting post-investment management, post-loan monitoring, and subsequent decision support.

[0003] Accurately identifying the true changes in a company's operating behavior after obtaining funding is crucial for measuring capital efficiency and predicting future risks. Current mainstream risk assessment methods primarily rely on analyzing regularly published financial statements and conducting qualitative due diligence. Financial statement analysis assesses a company's historical operating performance through various financial ratios, while qualitative investigations use interviews, site visits, and other methods to make subjective judgments about non-quantitative factors such as management and market conditions.

[0004] While these methods provide a basic risk perspective, their inherent flaws are becoming increasingly serious in today's market environment. They generally lack the ability to analyze the dynamic correlation and causal relationship of key events; instead, they tend to conduct isolated risk assessments at a certain point in time, rather than using the major business event of "obtaining financing" as an analytical anchor to systematically and fully examine the substantive changes in the company's core business behavior before and after the event.

[0005] To address this, a method for generating dynamic graphs of financing risk reports based on multi-source invoice data is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for generating dynamic graphs of financing risk reports based on multi-source invoice data, which enables dynamic analysis of financing risks through multi-source invoice data identification.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for generating dynamic graphs of financing risk reports based on multi-source invoice data includes:

[0009] Obtain the input invoice data and output invoice data of the target company; extract the first business feature based on the input invoice data, and extract the second business feature based on the output invoice data;

[0010] Anchoring the target company's financing event as a time anchor, for each financing event, a pre-trained business risk analysis model is applied to process the first and second business characteristics, generating a pre-financing risk set and a post-financing risk set including risk types;

[0011] A first-order differential analysis was conducted on the pre-financing risk set and the post-financing risk set to identify the risk status parameters before and after a single financing event.

[0012] For two consecutive financing events, a second-order differential analysis is further performed to identify the cross-cycle trend of the target company's risk pattern by comparing the risk status parameters of different financing events.

[0013] Based on the first-order and second-order differential analysis, a multi-layer time-series knowledge graph is generated; the multi-layer time-series knowledge graph structures and correlates risk status parameters, cross-cycle change trends and financing events to form a dynamic graph of financing risk.

[0014] The first business characteristic includes at least customer concentration data and active customer growth data, which characterize the customer structure.

[0015] The second business characteristic includes at least supplier concentration data that characterizes the supply chain structure and core material procurement price fluctuation data.

[0016] Cross-validate the first and second business features to generate a third feature vector characterizing the authenticity of the business transaction; the cross-validation includes identifying whether there is a circular transaction flow between the target enterprise and the transaction entity that is highly matched in terms of amount and time.

[0017] The risk types included in the pre-financing risk set and the post-financing risk set are identified by the operational risk analysis model. The risk types include at least: customer structure risk, supply chain structure risk and transaction authenticity risk.

[0018] The customer structure risk is identified based on the first business characteristic, specifically including the risk of excessive customer concentration and the risk of fluctuations in the number of active customers;

[0019] The supply chain structure risks are identified based on the second business characteristic, specifically including core supplier dependence risk and key material procurement price fluctuation risk;

[0020] The risk to the authenticity of the transaction is identified based on a third feature vector, specifically including the risk of circular transactions.

[0021] The implementation process of the aforementioned operational risk analysis model includes:

[0022] Training phase: In offline mode, a multi-class classifier is trained using a sample set containing labeled risk types through a supervised learning algorithm;

[0023] Diagnostic phase: The feature vector of the target company is input into a trained multi-class classifier, which outputs the predicted probability of each type of risk. Based on a confidence threshold determined by statistical analysis of historical data, the final pre-financing risk set and post-financing risk set are determined.

[0024] The first-order differentiation analysis identifies risk status parameters as: new risk, increased risk, and improved risk by comparing the relationship between the risk types of the pre-financing risk set and the post-financing risk set and the preset risk threshold.

[0025] The newly added risk refers to the risk type appearing for the first time after financing and the risk indicator exceeding the preset risk threshold.

[0026] The aforementioned risk exacerbation refers to the fact that the risk type exists both before and after financing, but its risk indicators further deteriorate after financing.

[0027] The aforementioned risk improvement refers to a situation where a risk type existed before financing but its risk indicators recovered to within a safe threshold after financing.

[0028] The risk indicators are identified and determined based on the quantitative data corresponding to the risk type.

[0029] The second-order differential analysis is achieved through a pre-set logical rule base, which contains at least one rule to compare the changes in risk status associated with different financing events, thereby identifying cross-cycle trends.

[0030] The structure of the multi-layer temporal knowledge graph specifically includes:

[0031] One timeline, using each financing event as a chronological node;

[0032] An entity layer, with target companies and transaction entities as nodes;

[0033] A pattern layer, with risk type nodes and cross-cycle change trends as nodes;

[0034] Nodes in the three levels are connected by edges with preset semantic types and quantized weight attributes to achieve structured association.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. By quantifying customer concentration, activity growth, supplier dependence, and material price fluctuations, this invention enables technical solutions to directly address the core of an enterprise, identifying deep-seated structural risks that are difficult to detect in conventional financial data; it further defines the data foundation for risk analysis, ensuring the depth and relevance of risk analysis.

[0037] 2. This invention trains a business risk analysis model using a supervised learning algorithm, solidifying the risk identification process into a "training-diagnosis" machine learning paradigm. It transforms expert experience and knowledge into repeatable and scalable algorithmic capabilities. This not only reduces reliance on human experience and ensures the consistency of assessment standards, but also makes it possible to perform efficient and batch automated risk scanning on a large number of enterprises. It provides automated and standardized technical support for risk assessment.

[0038] 3. By comparing the changes in associated risks across multiple financing events, this invention identifies recurring patterns in a company's operations and financing activities, which helps to understand a company's financing risks from a more macro perspective and empowers the risk assessment system to gain insights into long-term trends and underlying patterns. Furthermore, by integrating the complex analysis results from multiple dimensions such as time, entity, and risk patterns into a unified and intuitive view, it achieves efficient information transmission and holistic understanding. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the process for generating a dynamic graph of financing risk reports based on multi-source invoice data according to the present invention.

[0040] Figure 2 This is a schematic diagram of the risk type identification process of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of the multi-layer temporal knowledge graph of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1:

[0044] This invention proposes a method for generating dynamic graphs of financing risk reports based on multi-source invoice data, the process of which is as follows: Figure 1 As shown, it specifically includes:

[0045] Obtain the input invoice data and output invoice data of the target company; extract the first business feature based on the input invoice data, and extract the second business feature based on the output invoice data;

[0046] Anchoring the target company's financing event as a time anchor, for each financing event, a pre-trained business risk analysis model is applied to process the first and second business characteristics, generating a pre-financing risk set and a post-financing risk set including risk types;

[0047] A first-order differential analysis was conducted on the pre-financing risk set and the post-financing risk set to identify the risk status parameters before and after a single financing event.

[0048] For two consecutive financing events, a second-order differential analysis is further performed to identify the cross-cycle trend of the target company's risk pattern by comparing the risk status parameters of different financing events.

[0049] Based on the first-order and second-order differential analysis, a multi-layer time-series knowledge graph is generated; the multi-layer time-series knowledge graph structures and correlates risk status parameters, cross-cycle change trends and financing events to form a dynamic graph of financing risk.

[0050] Preferably, structured data of all input and output invoices of the target enterprise within a specified time frame is securely and compliantly obtained from an authorized electronic invoice service platform, tax software, or ERP (Enterprise Resource Planning) system. Required key fields should include at least: invoice code, invoice number, invoice date, buyer's name and taxpayer identification number, seller's name and taxpayer identification number, product or service name, specifications, unit, quantity, unit price, amount, tax rate, and tax amount.

[0051] Preferably, the first business feature includes at least customer concentration data and active customer growth data, which characterize the customer structure; the second business feature includes at least supplier concentration data and core material procurement price fluctuation data, which characterize the supply chain structure.

[0052] The calculation method for customer concentration data is as follows: within a specific time window (such as 6 months) before / after a financing event, the percentage of the target company's cumulative sales to its top five customers (ranked by transaction amount) to the total sales during that window period; the higher this value, the greater the customer concentration risk.

[0053] The calculation method for active customer growth data is as follows: First, "active customer" is defined as a customer who has made at least one transaction with the target company in the current month; then, the number of monthly active customers is calculated, and its month-on-month growth rate is calculated as (number of active customers in the current month - number of active customers in the previous month) / number of active customers in the previous month; sharp fluctuations in this indicator indicate business instability.

[0054] The calculation method for supplier concentration data is similar to that for customer concentration, which is to calculate the percentage of the total purchase amount of the top five suppliers to the total purchase amount.

[0055] Core material procurement price fluctuation data: First, identify "core materials", which are the goods or services with the highest total procurement amount (e.g., the top three); then, calculate the coefficient of variation (standard deviation / mean) of the average monthly procurement unit price of each core material over the past specified period. The higher the value, the more unstable the cost of key raw materials and the greater the supply chain risk.

[0056] This invention defines the data foundation for risk analysis by quantifying customer concentration, activity growth, supplier dependence, and material price fluctuations. This enables the technical solution to directly address the core indicators of an enterprise, identify deep-seated structural risks that are difficult to detect in conventional financial data, and ensure the depth and relevance of risk analysis.

[0057] The first and second business features are cross-validated to generate a third feature vector representing the authenticity of the business transaction. The cross-validation includes identifying whether there is a circular transaction flow that matches in terms of amount and time between the target enterprise and the transaction entity. The matching in terms of amount and time is determined according to a preset threshold.

[0058] The process of identifying circular transaction risks includes: constructing a transaction network graph with enterprises (target enterprise, its customers, and its suppliers) as nodes and invoices as directed edges. Graph algorithms such as Depth-First Search (DFS) are applied to identify whether closed transaction loops exist in the network, especially those loops that occur within a short period (e.g., within a week) and have highly matched transaction amounts (difference less than 5%). If such loops exist, a Boolean feature indicating "circular transaction risk exists" is generated.

[0059] Preferably, in cross-validation, not only are closed loops with the same starting and ending point identified, but graph algorithms are also used to search for transaction links within a preset time window that start from the target company, pass through intermediate transaction entities, and ultimately flow funds to individuals or corporate entities that are related to the target company (such as the same legal person, actual controller, or major shareholder). These are identified as quasi-circular transaction flows. The identified quasi-circular transaction flows are used as part of the third feature vector to identify the risks of unfair off-balance-sheet fund circulation or related-party transactions through complex transaction structures.

[0060] The third feature vector is obtained by integrating information from the identified circular and quasi-circular transaction flows.

[0061] By identifying and analyzing quasi-circular transaction flows, we can reveal deep-seated risks that ordinary circular transaction detection cannot detect, thus achieving a refined upgrade in circular transaction flow identification.

[0062] This invention performs closed-loop detection on the purchase and sales invoice flow, increases the verification between invoice data, effectively filters out noise caused by fraudulent transactions, ensures that subsequent risk analysis is based on real and credible business, and guarantees the accuracy of the final assessment results.

[0063] The risk types included in the pre-financing risk set and the post-financing risk set are identified by the operational risk analysis model. The risk types include at least: customer structure risk, supply chain structure risk and transaction authenticity risk.

[0064] The risk type identification process is as follows: Figure 2 As shown:

[0065] The customer structure risk is identified based on the first business characteristic, specifically including the risk of excessive customer concentration and the risk of fluctuations in the number of active customers;

[0066] Furthermore, “payment cycle stability data” that characterizes customer payment behavior can be extracted as a supplement to the first business characteristic (customer structure).

[0067] The data extraction process includes: for each sales invoice, linking it to its corresponding payment record and calculating the "payment cycle" = (payment receipt date - invoice issuance date); using the financing time as an anchor point, statistically analyzing the stability of the payment cycle for the 6 months before and after financing; for each major customer (top 10 in terms of transaction volume in the current period), calculating the mean and standard deviation of their payment cycle; and then calculating the "payment cycle variation coefficient" = standard deviation / mean.

[0068] Finally, the coefficients of variation of all major customers are weighted and averaged according to their transaction amounts to obtain the final "payment cycle stability data" indicator.

[0069] The supply chain structure risks are identified based on the second business characteristic, specifically including core supplier dependence risk and key material procurement price fluctuation risk;

[0070] The risk to the authenticity of the transaction is identified based on a third feature vector, specifically including the risk of circular transactions.

[0071] This invention associates data features with specific risk types, so that the model output is no longer cold numbers, but a diagnostic conclusion with clear business meaning, which greatly enhances the readability and operability of risk reports, makes it easier for decision-makers to quickly locate problems, and achieves standardization and interpretability of risk assessment results.

[0072] The implementation process of the aforementioned operational risk analysis model includes:

[0073] Training phase: In offline mode, a multi-class classifier is trained using a sample set containing labeled risk types through a supervised learning algorithm;

[0074] Diagnostic phase: The feature vector of the target company is input into a trained multi-class classifier, which outputs the predicted probability of each type of risk. Based on a confidence threshold determined by statistical analysis of historical data, the final pre-financing risk set and post-financing risk set are determined.

[0075] Preferably, in this embodiment, the operational risk analysis model adopts the Gradient Boosting Decision Tree (GBDT) model, specifically implemented through the open-source LightGBM framework. This model is trained offline on an anonymized invoice dataset. The risk labels in the dataset (such as customer structure risk, supply chain structure risk, transaction authenticity risk, etc.) are verified and labeled by senior credit analysts based on hard facts such as whether the company has experienced credit defaults or been listed in the list of companies with abnormal operations after the labeled date.

[0076] The labeling follows these principles: "Transaction authenticity risk" has the highest priority, and any circular transaction flow detected will be labeled as this risk; for companies with "credit delinquency," their risk indicators for the six months prior to the delinquency will be reviewed. If only customer concentration exceeds the 80th percentile of the industry, it will be labeled as "customer structure risk." If only material procurement price fluctuations exceed the 80th percentile, it will be labeled as "supply chain structure risk." If multiple indicators exceed the limits simultaneously, it will be labeled as a composite risk, and all relevant risk types will be noted in the label.

[0077] During training, the model takes the defined feature vector as input and multi-class risk labels as output, and is optimized using five-fold cross-validation. The learning rate is set to 0.05 and the maximum tree depth is 6.

[0078] Diagnosis and Set Generation: Calculate the pre- and post-financing data of the target company separately, and combine the calculated features such as "customer concentration data", "active customer growth data", "supplier concentration data", "core material procurement price fluctuation data" and "circular transaction risk" Boolean value into a feature vector in a predetermined order. Input the vector into the trained GBDT model.

[0079] Determine the risk set: Set a confidence threshold. This threshold is determined by plotting an ROC curve on the model's validation set and selecting the point that maximizes the Youden index; in this embodiment, it is set to 0.75. When the probability of a certain risk type output by the model is greater than 0.75, that risk type is included in the set. This results in the "pre-financing risk set" and the "post-financing risk set".

[0080] This invention trains models using supervised learning algorithms and solidifies the risk identification process into a "training-diagnosis" machine learning paradigm, transforming expert experience and knowledge into repeatable and scalable algorithmic capabilities. This not only reduces reliance on human experience and ensures the consistency of assessment standards, but also makes it possible to perform efficient and automated risk scanning on a large number of enterprises. It provides automated and standardized technical support for risk assessment.

[0081] The first-order differentiation analysis identifies risk status parameters as: new risk, increased risk, and improved risk by comparing the relationship between the risk types of the pre-financing risk set and the post-financing risk set and the preset risk threshold.

[0082] The newly added risk refers to the risk type appearing for the first time after financing and the risk indicator exceeding the preset risk threshold.

[0083] The aforementioned risk exacerbation refers to the fact that the risk type exists both before and after financing, but its risk indicators further deteriorate after financing.

[0084] The aforementioned risk improvement refers to a situation where a risk type existed before financing but its risk indicators recovered to within a safe threshold after financing.

[0085] The risk indicators are identified and determined based on the quantitative data corresponding to the risk type.

[0086] The method for determining the preset risk threshold includes: acquiring a sample dataset of historical invoices from companies in the same industry that contain the target company; calculating the numerical distribution of each risk indicator (such as customer concentration) in the sample dataset; and setting the 80th percentile of the numerical distribution of this indicator as the risk threshold for the industry to ensure the objectivity and industry adaptability of the threshold. These thresholds can be dynamically adjusted according to different industry characteristics and sample data.

[0087] Preferably, the 80th percentile is determined by backtesting historical data and finding that this threshold has the highest accuracy (or F1 score) in distinguishing between subsequently defaulting and non-defaulting companies. As an alternative, for risk indicators with obvious skewed distributions, an outlier detection method based on box plots can be used, where the upper quartile Q3 plus 1.5 times the interquartile range (IQR) is used as the risk threshold.

[0088] This invention precisely categorizes the ambiguous changes before and after financing into three distinct states: "new," "aggravated," and "improved." Through this structured classification, the evaluation of financing effectiveness is no longer a qualitative description but a quantifiable conclusion, making the evaluation results clearer and more specific. This enables a structured and quantitative assessment of the impact of a single financing event.

[0089] The second-order differential analysis is implemented through a pre-defined logical rule base. This rule base includes, but is not limited to, the following exemplary rules:

[0090] Rule 1 (Identification of Continuous Deterioration Trend): If the risk status parameter for the same risk type (such as customer structure risk) is "risk aggravation" in two consecutive financing events for the target company in the Nth and N+1th consecutive financing events, then THEN will generate a cross-cycle trend node indicating that "this risk type is showing a continuous deterioration trend".

[0091] Rule 2 (Identification of Volatility Improvement Trend): If the risk status parameter of the target company in the Nth financing event is "risk aggravated", AND the status parameter of this risk type in the N+1th financing event is "risk improved", THEN generates a node "this risk type shows a volatility improvement trend".

[0092] Rule 3 (Identification of Continuous Improvement Trends): If the target company's status parameter for a certain risk type is "risk improvement" in both the Nth and N+1th financing events, then a node indicating "this risk type shows a continuous improvement trend" will be generated.

[0093] Rule 4 (Identification of the trend of deterioration after improvement): IF the risk status parameter of the target company in the Nth financing event is "risk improved", AND the risk status parameter of this risk type changes to "risk added" or "risk aggravated" in the N+1th financing event, THEN generates a node "this risk type shows a trend of deterioration after improvement".

[0094] Preferably, different weighting coefficients are assigned to the "financing round" (such as angel round, Series A, Series B) and "financing distance" (the length of time since the current analysis) of each financing event; when using the logical rule base to identify cross-cycle change trends, the risk status parameter of each financing event is multiplied by its corresponding weighting coefficient, so that financing events that occur recently and are at a critical stage of enterprise development have a greater impact on the final trend judgment.

[0095] Specifically, in this embodiment, the weight of "financing round" is set to the same weight, and the time distance is normalized according to the time length of different financing events from this analysis to obtain the weight of time distance.

[0096] By using dynamic weighting, trend analysis is no longer a simple comparison of event sequences, but incorporates considerations of the enterprise life cycle and information timeliness, making the model's judgment of risk trends more accurate.

[0097] This invention identifies recurring patterns in a company's operations and financing behavior by comparing the associated risk changes across multiple financing events. This helps to understand a company's financing risks from a more macro perspective and empowers the risk assessment system to gain insights into long-term trends and underlying patterns.

[0098] The structure of the multi-layer temporal knowledge graph is as follows: Figure 3 As shown, it specifically includes:

[0099] One timeline, using each financing event as a chronological node;

[0100] An entity layer, with target companies and transaction entities as nodes;

[0101] A pattern layer, with risk type nodes and cross-cycle change trends as nodes;

[0102] Nodes in the three levels are connected by edges with preset semantic types and quantized weight attributes to achieve structured associations. Specifically:

[0103] The connection between the time layer and the pattern layer: The "Financing Event" node is connected to the "Risk State Parameter" node through an edge with the semantic type "Exhibition". The "weight" attribute value of this edge can be associated with the core risk indicator corresponding to the risk state; specifically, it can be determined by combining the actual value of the core risk indicator corresponding to the risk state with the preset risk threshold of the indicator.

[0104] Connections within the pattern layer: The "Risk State Parameters" node is connected to the "Cross-cycle Change Trend" node via an edge with the semantic type "Evolve into".

[0105] Connection between entity layer and schema layer: The "target enterprise" node is connected to the "risk type" node through an edge with the semantic type "existence".

[0106] Preferably, in the entity layer, for transaction entities that have a direct transaction relationship with the target enterprise, the operational risk analysis model is also applied to generate their own risk set; and in the multi-layer time-series knowledge graph, a risk transmission path is established from the transaction entity node with risk to the target enterprise node according to the invoice flow, in order to analyze and quantify the possibility and degree of impact of risks from core suppliers or core customers being transmitted to the target enterprise.

[0107] This design creatively expands the perspective of risk analysis from "point" to "chain," not only focusing on the risks of the target company itself, but also on the risks of its upstream and downstream partners. It also makes this potential transmission path explicit and modeled on a knowledge graph, enabling the identification of potential risks caused by external entities.

[0108] This invention integrates complex analysis results from multiple dimensions, such as time, entities, and risk patterns, into a unified and intuitive view. It achieves efficient information transfer and holistic understanding through a visualized and structured knowledge graph.

[0109] Example 2:

[0110] This invention proposes a method for generating dynamic graphs of financing risk reports based on multi-source invoice data, including:

[0111] Obtain the input invoice data and output invoice data of the target company; extract the first business feature based on the input invoice data, and extract the second business feature based on the output invoice data;

[0112] Anchoring the target company's financing event as a time anchor, for each financing event, a pre-trained business risk analysis model is applied to process the first and second business characteristics, generating a pre-financing risk set and a post-financing risk set including risk types;

[0113] A first-order differential analysis was conducted on the pre-financing risk set and the post-financing risk set to identify the risk status parameters before and after a single financing event.

[0114] For two consecutive financing events, a second-order differential analysis is further performed to identify the cross-cycle trend of the target company's risk pattern by comparing the risk status parameters of different financing events.

[0115] Based on the first-order and second-order differential analysis, a multi-layer time-series knowledge graph is generated; the multi-layer time-series knowledge graph structures and correlates risk status parameters, cross-cycle change trends and financing events to form a dynamic graph of financing risk.

[0116] The first business characteristic includes at least customer concentration data and active customer growth data, which characterize customer structure; the second business characteristic includes at least supplier concentration data and core material procurement price fluctuation data, which characterize supply chain structure. The first and second business characteristics are cross-validated to generate a third feature vector characterizing the authenticity of business transactions; the cross-validation includes identifying whether there is a circular transaction flow between the target enterprise and the transaction entity that is highly matched in terms of amount and time.

[0117] Specifically, in this embodiment, the characteristic data of the target enterprise is identified as follows:

[0118] The risk types included in the pre-financing risk set and the post-financing risk set are identified by the operational risk analysis model. The risk types include at least: customer structure risk, supply chain structure risk and transaction authenticity risk.

[0119] The customer structure risk is identified based on the first business characteristic, specifically including the risk of excessive customer concentration and the risk of fluctuations in the number of active customers;

[0120] The supply chain structure risks are identified based on the second business characteristic, specifically including core supplier dependence risk and key material procurement price fluctuation risk;

[0121] The risk to the authenticity of the transaction is identified based on a third feature vector, specifically including the risk of circular transactions.

[0122] Furthermore, “payment cycle stability data” that characterizes customer payment behavior can be extracted as a supplement to the first business characteristic (customer structure).

[0123] The data extraction process includes: for each sales invoice, linking it to its corresponding payment record and calculating the "payment cycle" = (payment receipt date - invoice issuance date); using the financing time as an anchor point, statistically analyzing the stability of the payment cycle for the 6 months before and after financing; for each major customer (top 10 in terms of transaction volume in the current period), calculating the mean and standard deviation of their payment cycle; and then calculating the "payment cycle variation coefficient" = standard deviation / mean.

[0124] Finally, the coefficients of variation of all major customers are weighted and averaged according to their transaction amounts to obtain the final "payment cycle stability data" indicator.

[0125] The implementation process of the aforementioned operational risk analysis model includes:

[0126] Training phase: In offline mode, a multi-class classifier is trained using a sample set containing labeled risk types through a supervised learning algorithm; Diagnosis phase: The feature vector of the target company is input into the trained multi-class classifier, which outputs the predicted probability of each risk type. Based on a confidence threshold determined by statistical analysis of historical data, the final pre-financing risk set and post-financing risk set are determined.

[0127] The first-order differentiation analysis identifies risk status parameters as: new risk, increased risk, and improved risk by comparing the relationship between the risk types of the pre-financing risk set and the post-financing risk set and the preset risk threshold.

[0128] The second-order differential analysis is achieved through a pre-set logical rule base, which contains at least one rule to compare the changes in risk status associated with different financing events, thereby identifying cross-cycle trends.

[0129] The structure of the multi-layered temporal knowledge graph specifically includes: a time layer, with each financing event as a temporal node; an entity layer, with target companies and transaction entities as nodes; and a pattern layer, with risk type nodes and cross-cycle change trends as nodes. The nodes in the three layers are connected by edges with preset semantic types and quantitative weight attributes to achieve structured association.

[0130] This invention acquires input and output invoice data, extracts first and second business characteristics; using financing events as time anchors, for each financing event, it applies an operational risk analysis model to process the first and second business characteristics, generating a pre-financing risk set and a post-financing risk set including risk types; it performs a first-order differential analysis on the pre-financing and post-financing risk sets to identify risk status parameters; for two consecutive financing events, it performs a second-order differential analysis to identify the cross-cycle trend of the target company's risk pattern; based on the first and second-order differential analyses, it generates a multi-layer time-series knowledge graph, forming a dynamic financing risk graph, realizing dynamic analysis of financing risks.

[0131] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating dynamic graphs of financing risk reports based on multi-source invoice data, characterized in that, include: Obtain the target company's input and output invoice data; Extract the first business characteristic from the input invoice data, and extract the second business characteristic from the output invoice data; The first business characteristic includes at least customer concentration data and active customer growth data, which characterize the customer structure. The second business characteristic includes at least supplier concentration data characterizing the supply chain structure and core material procurement price fluctuation data; Anchoring the target company's financing event as a time anchor, for each financing event, a pre-trained business risk analysis model is applied to process the first and second business characteristics, generating a pre-financing risk set and a post-financing risk set including risk types; The implementation process of the aforementioned operational risk analysis model includes: Training phase: In offline mode, a multi-class classifier is trained using a sample set containing labeled risk types through a supervised learning algorithm; Diagnostic phase: The feature vector of the target company is input into a trained multi-class classifier, which outputs the predicted probability of each type of risk. Based on a confidence threshold determined by statistical analysis of historical data, the final pre-financing risk set and post-financing risk set are determined. A first-order differential analysis was conducted on the pre-financing risk set and the post-financing risk set to identify the risk status parameters before and after a single financing event. For two consecutive financing events, a second-order differential analysis is further performed to identify the cross-cycle trend of the target company's risk pattern by comparing the risk status parameters of different financing events. Based on first-order and second-order differential analysis, a multi-layer time-series knowledge graph is generated; the multi-layer time-series knowledge graph structures and correlates risk status parameters, cross-cycle change trends and financing events to form a dynamic graph of financing risk.

2. The method for generating a dynamic graph of financing risk reports based on multi-source invoice data according to claim 1, characterized in that: Cross-validate the first and second business features to generate a third feature vector characterizing the authenticity of the business transaction; the cross-validation includes identifying whether there is a circular transaction flow that matches in terms of amount and time between the target enterprise and the transaction entity.

3. The method for generating a dynamic graph of financing risk reports based on multi-source invoice data according to claim 2, characterized in that: The risk types included in the pre-financing risk set and the post-financing risk set are identified by the operational risk analysis model. The risk types include at least: customer structure risk, supply chain structure risk and transaction authenticity risk. The customer structure risk is identified based on the first business characteristic, specifically including the risk of excessive customer concentration and the risk of fluctuations in the number of active customers; The supply chain structure risks are identified based on the second business characteristic, specifically including core supplier dependence risk and key material procurement price fluctuation risk; The risk to the authenticity of the transaction is identified based on a third feature vector, specifically including the risk of circular transactions.

4. The method for generating a dynamic graph of financing risk reports based on multi-source invoice data according to claim 1, characterized in that: The first-order differentiation analysis identifies risk status parameters as: new risk, increased risk, and improved risk by comparing the relationship between the risk types of the pre-financing risk set and the post-financing risk set and the preset risk threshold. The newly added risk refers to the risk type appearing for the first time after financing and the risk indicator exceeding the preset risk threshold. The aforementioned risk exacerbation refers to the fact that the type of risk exists both before and after financing, but its risk indicators further deteriorate after financing. The aforementioned risk improvement refers to a situation where a risk type existed before financing but its risk indicators recovered to within a safe threshold after financing. The risk indicators are identified and determined based on the quantitative data corresponding to the risk type.

5. The method for generating a dynamic graph of financing risk reports based on multi-source invoice data according to claim 1, characterized in that: The second-order differential analysis is achieved through a pre-set logical rule base, which contains at least one rule to compare the changes in risk status associated with different financing events, thereby identifying cross-cycle trends.

6. The method for generating a dynamic graph of financing risk reports based on multi-source invoice data according to claim 1, characterized in that: The structure of the multi-layer temporal knowledge graph specifically includes: One timeline, using each financing event as a chronological node; An entity layer, with target companies and transaction entities as nodes; A pattern layer, with risk type nodes and cross-cycle change trends as nodes; Nodes in the three levels are connected by edges with preset semantic types and quantized weight attributes to achieve structured association.

Images