Cross-institutional risk transmission path tracking method based on knowledge graph fusion

Abstract

The application discloses a cross-institutional risk transmission path tracking method based on a knowledge graph, which comprises the following steps: firstly, a cross-institutional federal knowledge graph is constructed to realize the sharing of original data and global meta-information without domain outflow; then, an initial risk transmission path is found through distributed graph traversal; next, confounding variables are identified based on the theory of causal graphs; then, the global statistics are safely aggregated by using the additive homomorphic encryption technology to calculate the causal strength; finally, false paths are eliminated, and risk warning information is generated and distributed in a directional manner. The application can accurately distinguish causal relationships from relevant relationships, eliminate false transmission paths, and significantly reduce the risk warning false positive rate under the premise of protecting data privacy, thereby providing quantitative and interpretable risk assessment results for financial institutions and being suitable for the collaborative cross-institutional risk prevention and control work of multiple institutions.

Description

Technical Field

[0001] This invention relates to the field of financial risk prevention and control and distributed knowledge graph technology, specifically to a method for tracing cross-institutional risk transmission paths that integrates knowledge graphs. Background Technology

[0002] As financial markets continue to develop, the relationships between enterprises are becoming increasingly complex. Credit risk is no longer confined to a single enterprise but is rapidly transmitted between different enterprises through various channels such as supply chains, guarantee chains, and equity chains. When a company defaults, the risk spreads rapidly along the network of relationships, forming systemic financial risk. Therefore, accurately tracing the transmission path of credit risk and identifying potential risky enterprises in advance is of great significance for financial institutions to prevent systemic financial risks.

[0003] Currently, mainstream methods for tracing credit risk transmission paths are primarily based on knowledge graph technology. This involves constructing a knowledge graph containing entities such as enterprises, individuals, and institutions, as well as the relationships between them, and then using algorithms like graph traversal and association rule mining to discover risk transmission paths. However, existing technologies suffer from two core drawbacks: first, they cannot distinguish between causal and correlational relationships, leading to a large number of false paths in the traced transmission paths and a persistently high false positive rate; second, they cannot handle causal confusion in distributed environments, and cross-institutional data privacy protection requirements render traditional causal reasoning methods based on centralized complete graph data completely ineffective.

[0004] Therefore, how to accurately identify causal relationships in cross-institutional credit risk transmission, eliminate false transmission paths, significantly reduce the false alarm rate of risk warnings, and improve the accuracy of credit risk prevention and control, without leaving the data domain, is a problem worth studying. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a cross-institutional risk transmission path tracking method that integrates knowledge graphs. This method can accurately identify causal relationships in cross-institutional credit risk transmission, eliminate false transmission paths, significantly reduce the false alarm rate of risk warnings, and improve the accuracy of credit risk prevention and control, provided that the data does not leave the domain.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for tracing cross-institutional risk transmission paths by integrating knowledge graphs includes the following steps: Construct a cross-institutional federated knowledge graph, where each participating institution maintains a knowledge subgraph locally, coordinates nodes to maintain a global metadata catalog, and achieves unified entity identification across institutions through a privacy-preserving entity alignment method; Starting from the risk source node, a distributed breadth-first traversal is performed to collect and merge the path fragments of each institution, forming an initial set of risk transmission paths and performing preliminary filtering. Based on the causal graph theory, a causal structure is constructed for each initial path, and potential confusion variables for each edge are identified in a distributed manner and their existence is verified, forming a set of confusion variables for each path; A unified rule is used to binarize variables, each institution independently calculates local statistics, and global statistics are aggregated through additive homomorphic encryption to calculate the causal strength of a single relationship and the cumulative causal strength of the entire path. False transmission paths containing weak causal relationships are eliminated based on a preset causal strength threshold, and the real transmission paths are sorted by cumulative causal strength and the risk level of the target nodes is classified. The actual path information is anonymized, and a risk warning report is generated and distributed to relevant financial institutions.

[0007] A cross-institutional credit risk transmission path tracking device integrating knowledge graphs includes: The global coordination module, deployed on the coordination node, is used to formulate global standards, coordinate entity alignment, maintain metadata catalog, merge initial paths, coordinate the identification of confusing variables, aggregate statistics, calculate causal strength, and generate and distribute early warning information. The local computing module is deployed in each participating institution and is used to build local knowledge subgraphs, calculate entity hash values, register metadata, perform local traversal, verify obfuscated variables, calculate local statistics, encrypt and transmit data, and receive and process early warnings.

[0008] Compared with the prior art, the present invention has the following beneficial effects: For the first time, causal graph theory is combined with federated knowledge graph technology to solve the problem of causal confusion in the transmission path of cross-institutional credit risk in a distributed environment. It can accurately distinguish between causal relationships and correlations, eliminate false transmission paths, and significantly reduce the false alarm rate of risk warnings. Strictly adhering to data privacy protection laws and regulations, all raw data is stored locally on the local premises of each participating institution. A hash-based entity alignment method and an additive homomorphic encryption statistical aggregation method are adopted to ensure that no raw data is leaked during the entire calculation process, effectively protecting the data privacy and trade secrets of each participating institution. It can provide quantitative risk assessment results, quantify the transmission strength of each risk transmission path by calculating the cumulative causal strength of the path, and provide a scientific basis for financial institutions to formulate differentiated risk prevention and control measures. Attached Figure Description

[0009] Figure 1 is a flowchart illustrating the overall process composition of the present invention; Figure 2 is a sub-flow diagram of the S110 cross-agency federated knowledge graph standardization construction step in Embodiment 1 of the present invention; Figure 3 is a sub-flow diagram of the S120 initial risk transmission path distributed discovery and preliminary filtering step in Embodiment 1 of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and 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.

[0011] This embodiment is applicable to the collaborative work of various institutions such as financial institutions, industrial and commercial administration agencies, tax administration agencies, customs, and judiciary in tracking and early warning of cross-institutional credit risk transmission paths. The method is executed by a cross-institutional credit risk causal tracking system that integrates knowledge graphs. This system includes a coordination node deployed in the financial regulatory department and local computing nodes deployed in each participating institution.

[0012] The core terms involved in this embodiment are defined as follows: Federated knowledge graph is a distributed knowledge graph architecture in which each participating institution maintains its own knowledge subgraph locally and achieves global metadata sharing and collaborative computing through coordination nodes, with the original data not leaving the domain. A cause-effect graph is a type of directed acyclic graph where nodes represent variables and edges represent causal relationships between variables. It is used to characterize the causal logic in risk transmission. Confounding variables, including third-party variables that simultaneously affect both causal and outcome variables in a causal relationship, can lead to spurious correlations and is a major cause of false risk transmission paths. Additive homomorphic encryption is a cryptographic technique that allows direct addition operations on encrypted data. The decrypted result is identical to the plaintext result, and it is used to protect data privacy in distributed statistical computing.

[0013] Example 1 describes in detail the basic process of the cross-institutional risk transmission path tracking method integrating knowledge graphs described in this invention. The participating institutions include 3 financial institutions (financial institution A, financial institution B, and financial institution C), 1 industrial and commercial administration department, 1 tax administration agency, and 1 judicial agency. The coordination node is deployed on the cloud platform of the local branch of the highest financial supervision and management agency.

[0014] like Figure 1 A method for tracing cross-institutional risk transmission paths that integrates knowledge graphs includes: S110. Standardized construction and distributed maintenance of cross-institutional federated knowledge graph. The execution action of this step is for all participating institutions to collaboratively build a unified standard distributed federated knowledge graph. The input is the original business data of each institution, and the output is a globally unified federated knowledge graph system and metadata catalog.

[0015] The federated knowledge graph consists of multiple independent local knowledge subgraphs and a global meta-information directory. Each local knowledge subgraph is maintained independently by the corresponding participating organization, while the global meta-information directory is maintained by the coordinating node and does not store any original business data.

[0016] Optionally, participating institutions may also include supply chain finance platforms, securities companies, insurance companies and other financial institutions that hold credit-related data, as well as government departments such as customs and market supervision administration.

[0017] Specifically, such as Figure 2 This step includes the following sub-steps: S111. Participating Institutions and Coordination Nodes Determined. First, the financial regulatory authorities will take the lead in determining all institutions participating in this cross-institutional risk control effort. Each participating institution will be assigned a unique institutional identifier (ID), for example, the identifier for the State Administration for Industry and Commerce will be "GS," and the identifier for Financial Institution A will be "YH-A." Coordination nodes will be established by the financial regulatory authorities, responsible for coordinating the entire calculation process, maintaining the global metadata directory, but not storing any original business data.

[0018] S112. Establishment of globally unified entity and relationship standards. All participating institutions will jointly convene a standards-setting meeting to develop unified entity type dictionaries, relationship type dictionaries, and attribute dictionaries. Entity types include four categories: enterprises, individuals, institutions, and financial products; relationship types include six categories: equity relationships, guarantee relationships, transaction relationships, senior management appointment relationships, lending relationships, and legal litigation relationships; attributes include basic entity attributes (such as enterprise name and unified social credit code) and relationship attributes (such as shareholding ratio and guarantee amount). All participating institutions must strictly adhere to this standard to ensure data consistency and compatibility.

[0019] S113. Independent Construction of Local Knowledge Subgraphs. Each participating institution deploys a local graph database (such as Neo4j or NebulaGraph), extracts data from its core business systems, and constructs a local knowledge subgraph. Each entity in the subgraph is assigned a globally unique entity identifier, which is composed of "institution ID - local entity ID"; each relationship is assigned a globally unique relationship identifier, which is composed of "institution ID - local relationship ID". Each participating institution independently maintains its local knowledge subgraph, updating data according to a preset update frequency (such as once a day) to ensure data timeliness.

[0020] S114. Privacy-Preserving Distributed Entity Alignment. All participating institutions adopt a privacy-preserving entity alignment method based on SHA-256 hashing to complete the alignment of cross-institutional entities without disclosing the original data. Specifically, each participating institution performs SHA-256 hash calculations on unique identifier attributes such as the unified social credit code of local enterprise entities and the identification number of individual entities, and sends the hash value and the corresponding local entity identifier to the coordinating node. The coordinating node compares all hash values, marks entities with the same hash value as the same entity, and assigns them a globally unified entity identifier (format "ENT-XXXXXX"). The coordinating node returns the alignment result to each participating institution, and each participating institution updates the global identifier of the entity locally.

[0021] S115. Establishment of the Global Subgraph Meta-Information Directory. Each participating organization registers the metadata of its local knowledge subgraph with the coordinating node. The metadata includes the entity types, relation types, number of entities, number of relations, and an attribute list for each relation type. The coordinating node maintains a global subgraph metadata directory for subsequent queries. The coordinating node does not store any raw business data, only metadata.

[0022] It is known that using a hash-based entity alignment method can achieve accurate alignment of entities across institutions without disclosing the original identification information, effectively protecting the privacy of enterprises and individuals. At the same time, a unified entity and relationship specification ensures seamless integration of knowledge subgraphs from different institutions, laying the foundation for subsequent distributed computing.

[0023] S120. Distributed discovery and preliminary filtering of initial risk transmission paths: This step involves discovering all possible risk transmission paths and performing preliminary filtering in a federated environment. The input is the risk source node information, and the output is the set of initial risk transmission paths after preliminary filtering.

[0024] Among them, the risk source node refers to the corporate entity that experiences risk events such as default, bankruptcy, or significant financial deterioration, and the initial risk transmission path refers to all possible paths formed by connecting related relationships starting from the risk source node.

[0025] Optionally, the identification of risk source nodes can be automatically detected by the local systems of each participating institution, or it can be manually designated by the financial regulatory authorities.

[0026] Specifically, such as Figure 3 This step includes the following sub-steps: S121. Identification and Global Broadcasting of Risk Source Nodes. When a company experiences default risk, the financial institution holding the company's credit data is the first to detect the risk event, marking the company as a risk source node, and sending the global identifier of the risk source node and the risk event type to the coordinating node. The coordinating node then broadcasts the global identifier of the risk source node to all participating institutions.

[0027] S122. Execution of Distributed Breadth-First Traversal. Each participating institution, starting from the risk source node in its local knowledge subgraph, executes a breadth-first traversal algorithm to discover all entities and relationships directly or indirectly related to the risk source node. During the traversal, each participating institution only processes entities and relationships in its local subgraph, without disclosing any original data. The traversal depth is jointly set by all participating institutions; in this embodiment, it is set to 5, meaning a maximum of 5 propagation paths can be traced.

[0028] S123. Secure Collection and Merging of Path Fragments. Each participating institution sends the path fragments obtained through local traversal to the coordinating node. Path fragments only contain the global identifier of the entity, the global identifier of the relationship, and the relationship type, and do not contain any attribute information. The coordinating node merges all path fragments to form a complete initial set of risk transmission paths. During the merging process, the coordinating node connects cross-institutional path fragments into complete paths based on the global identifier of the entity.

[0029] S124. Preliminary filtering of the initial path set. The coordinating node filters the initial path set, removing paths whose length exceeds a preset threshold. The preset threshold is determined based on actual business needs; in this embodiment, it is set to 5, consistent with the traversal depth. The filtered path set proceeds to the next step of processing.

[0030] It is known that using a distributed breadth-first search algorithm can discover all possible risk transmission paths without leaking local data from each institution. The initial filtering step can eliminate excessively long, low-probability paths, reduce subsequent computation, and improve system efficiency.

[0031] S130. Distributed identification of confounding variables based on causal graph theory. The action of this step is to identify the possible confounding variables in each initial path. The input is the initial risk transmission path set, and the output is the set of confounding variables corresponding to each path.

[0032] Among them, a confounding variable refers to a third-party entity that simultaneously affects two endpoint entities of an edge in a path, which can lead to a false correlation between the two entities.

[0033] Optionally, the types of confounding variables can be expanded according to business needs. In addition to the common executives, common shareholders, and common guarantors mentioned in this embodiment, common customers and common suppliers can also be included.

[0034] Specifically, this step includes the following sub-steps: S131. Causal structure modeling for a single path. For each path in the initial path set, treat each entity in the path as a variable, and the relationships between entities as causal edges between variables, constructing a causal graph corresponding to that path. Each path corresponds to an independent causal graph, and the structure of the causal graph is completely consistent with the structure of the path.

[0035] S132. Global Query of Potential Confounding Variables. For each edge in the causal graph, based on the global subgraph element information directory, query all entity types that may be associated with both endpoints of the edge. The entities corresponding to these entity types are the potential confounding variables for that edge. For example, for a transaction relationship edge between two companies, potential confounding variables include joint executives, joint shareholders, and joint guarantors; for a guarantee relationship edge between two companies, potential confounding variables include joint shareholders and joint executives.

[0036] S133. Distributed Verification of the Existence of Confusion Variables. The coordinating node sends a list of potential confusion variables for each edge to all participating institutions. Each participating institution verifies in its local knowledge subgraph whether there exists an entity that has a corresponding relationship with both endpoints of the edge. If so, the global identifier of that entity is returned to the coordinating node. Each participating institution only returns the global identifier of the confusion variable and no other information.

[0037] S134. Summarizing the confusion variable set for each path. The coordinating node summarizes the verification results returned by all participating institutions to form the actual confusion variable set for each edge, and thus the confusion variable set for each path. The confusion variable set for each path contains the actual confusion variables for all edges in that path.

[0038] It is known that by identifying confounding variables in a distributed manner, all third-party factors that may lead to false transmission paths can be accurately identified without disclosing the local data of each institution, thus providing a foundation for subsequent causal relationship verification.

[0039] S140. Causal Relationship Verification and Strength Assessment in a Distributed Environment: This step verifies the causal relationships of each segment in each initial path and assesses the causal strength. The input is the initial path set and the corresponding set of confounding variables, and the output is the cumulative causal strength of each path.

[0040] Among them, causal strength is a value between 0 and 1. The larger the value, the stronger the causal relationship, and the higher the possibility of risk transmission through the relationship.

[0041] Optionally, the method for calculating causal strength can be adjusted according to business needs. In addition to the average treatment effect method used in this embodiment, other causal inference methods such as propensity score matching and instrumental variables can also be used.

[0042] Specifically, this step includes the following sub-steps: S141. Global Unification of Variable Binarization Rules. All participating institutions shall jointly formulate unified variable binarization rules. For entity variables, the binarization rule is: if the entity has defaulted within the past 12 months, the value is 1; otherwise, the value is 0. For relational variables, the binarization rule is: if the relation exists, the value is 1; otherwise, the value is 0. For confounding variables, the binarization rule is: if the confounding variable has a corresponding association with both entities, the value is 1; otherwise, the value is 0. All participating institutions must strictly adhere to these rules to ensure the consistency of the calculation results.

[0043] S142. Independent Calculation of Local Statistics. For each edge and its set of confounding variables, each participating institution shall calculate the following statistics locally: total number of local samples, number of samples where the relation exists, number of samples where the relation does not exist, number of samples where the relation exists under each combination of confounding variables, number of samples where the relation does not exist under each combination of confounding variables, number of samples where downstream entities default when the relation exists under each combination of confounding variables, and number of samples where downstream entities default when the relation does not exist under each combination of confounding variables. Each participating institution shall calculate local statistics independently without disclosing any raw data.

[0044] S143. Secure Aggregation of Statistics. Each participating institution uses the Paillier additive homomorphic encryption algorithm to encrypt its locally calculated statistics before sending them to the coordinating node. The coordinating node sums the encrypted statistics to obtain the global statistics. Throughout the entire process, the coordinating node only performs summation operations on the encrypted statistics, without accessing any original statistics, and is unable to decrypt them to obtain the statistics of any individual institution.

[0045] S144. Global Calculation of Causal Strength of a Single Relationship. The coordinating node decrypts the aggregated global statistics and calculates the causal strength of the relationship. The calculation process is as follows: First, calculate the proportion of each confounding variable combination in all samples. Then, calculate the average probability of the downstream entity defaulting in both the presence and absence of the relationship. Finally, subtract these two average probabilities to obtain the causal strength of the relationship.

[0046] S145. Calculation of cumulative causal strength for the entire path. For each path, the causal strength of each segment in the path is multiplied to obtain the cumulative causal strength of the entire path. The cumulative causal strength ranges from 0 to 1. The larger the value, the higher the risk transmission strength of the path, and the greater the possibility that the target node is affected by the risk source node.

[0047] It is known that using additive homomorphic encryption technology for secure aggregation of statistics can yield accurate global statistical results while protecting the data privacy of various institutions. The causal strength calculation method based on causal graph theory can accurately quantify the causal relationship strength of risk transmission, providing a scientific basis for distinguishing between real and spurious paths.

[0048] S150. Elimination of false transmission paths and risk ranking of true paths: This step involves eliminating false transmission paths and ranking the true paths by risk. The input is the cumulative causal strength of each path, and the output is the set of ranked true risk transmission paths and the risk level of the target node.

[0049] Among them, a false propagation path refers to a path that contains at least one relationship with a causal strength lower than the threshold, while a true propagation path refers to a path in which the causal strength of all relationships is higher than the threshold.

[0050] Optionally, the causal strength threshold can be dynamically adjusted based on historical data and business experience. For example, the threshold can be appropriately lowered during economic downturns to improve the sensitivity of risk warnings.

[0051] Specifically, this step includes the following sub-steps: S151. Global setting of causal strength threshold. Based on historical data and business experience, all participating institutions jointly set a unified causal strength threshold. In this embodiment, it is set to 0.1. If the causal strength of a certain relationship is less than 0.1, it is considered that there is no causal relationship in that relationship, only a correlation.

[0052] S152. Determination and Removal of False Propagation Paths. For each path in the initial path set, check the causal strength of all edges in the path. If the causal strength of any relation in the path is less than the threshold, the path is marked as a false propagation path and removed from the path set. Only paths where the causal strength of all edges is greater than or equal to the threshold are determined to be true propagation paths.

[0053] S153. Risk Ranking of True Transmission Paths. For the true transmission paths remaining after eliminating false paths, they are ranked from highest to lowest according to their cumulative causal strength. The higher the cumulative causal strength, the higher the risk transmission intensity of the path, and the greater the likelihood that the target node will be affected by the risk source node.

[0054] S154. Risk Level Classification of Target Nodes. Based on the cumulative causal strength of the path, target nodes are classified into different risk levels. In this embodiment, the risk levels are divided into three categories: high risk, medium risk, and low risk. Target nodes with a cumulative causal strength greater than 0.7 are classified as high-risk, target nodes with a cumulative causal strength between 0.3 and 0.7 are classified as medium-risk, and target nodes with a cumulative causal strength less than 0.3 are classified as low-risk.

[0055] It is known that by setting a causal strength threshold to eliminate false paths, the problem of high false alarm rates in traditional methods can be fundamentally solved. Ranking true paths by risk and classifying target nodes by risk level can help financial institutions prioritize high-risk targets and improve the efficiency of risk control.

[0056] S160. De-identified output and targeted distribution of risk warning information: This step involves generating and distributing risk warning information. The inputs are the sorted set of real risk transmission paths and the risk level of the target node, and the output is a targeted risk warning report.

[0057] Among them, desensitization refers to removing all information involving personal privacy and trade secrets, and retaining only necessary risk warning information.

[0058] Optionally, the content of the risk warning report can be customized according to the needs of each financial institution, such as adding suggested risk control measures and similar historical cases.

[0059] Specifically, this step includes the following sub-steps: S161. De-identification of Path Information. Coordinating nodes de-identify information regarding the actual risk transmission path, removing all information involving personal privacy (such as personal names and identification numbers) and trade secrets (such as loan amounts, guarantee amounts, and transaction amounts), retaining only the entity's global identifier, relationship type, and causal strength value. The de-identification rules are jointly formulated by all participating institutions and strictly comply with laws and regulations.

[0060] S162. Generation of Risk Warning Reports. The coordinating node generates standardized risk warning reports based on the anonymized path information. The risk warning report includes: risk source node information (global identifier, risk event type, occurrence time), a list of high-risk target nodes, a list of medium-risk target nodes, a list of low-risk target nodes, detailed information for each actual risk transmission path (path structure, cumulative causal strength), and recommended risk prevention and control measures.

[0061] S163. Targeted Distribution of Early Warning Information. The coordinating node queries the global sub-element information directory to identify the financial institutions to which each target node belongs and distributes risk warning information accordingly. Each financial institution only receives warning information relevant to its own business, i.e., only warning information concerning its customers. The coordinating node does not disclose warning information to any unrelated institutions or individuals. For particularly significant risk events, the coordinating node also copies the warning information to financial regulatory authorities.

[0062] It is understood that adopting a targeted distribution mechanism ensures that each financial institution receives only the alert information relevant to itself, effectively protecting the trade secrets of each institution. Anonymization ensures that personal privacy and trade secrets are not leaked during the dissemination of alert information, complying with data security laws and regulations.

[0063] The beneficial effect of this embodiment is that by combining causal graph theory with federated knowledge graph technology, it accurately identifies causal relationships in cross-institutional credit risk transmission without leaving the data domain, eliminates false transmission paths, and significantly reduces the false alarm rate of risk warnings. At the same time, all calculation processes have clear theoretical basis and calculation rules, and the results are interpretable and reproducible, meeting the requirements of financial regulatory authorities.

[0064] Example 2: Based on Example 1, this example further refines the implementation details of the key technical steps and describes the complete execution process and effects of the present invention in detail, combined with a typical application scenario.

[0065] S210, Detailed Implementation of Cross-Institutional Federated Knowledge Graph Construction: In this embodiment, the implementation details of local knowledge subgraph construction and distributed entity alignment are further refined.

[0066] Specifically, during the construction of the local knowledge subgraph, participating institutions use ETL tools to extract data from core business systems, then clean, transform, and load it. Data cleaning includes removing duplicate data, correcting erroneous data, and filling in missing data; data transformation includes converting data of different formats into a unified format and converting unstructured data into structured data. For example, the Administration for Industry and Commerce extracts enterprise registration information, equity change information, and senior management appointment information from the Enterprise Credit Information Disclosure System to construct a subgraph containing equity relationships and senior management appointment relationships; financial institutions extract loan contract information and guarantee contract information from the credit management system to construct a subgraph containing loan relationships and guarantee relationships; tax authorities extract enterprise transaction invoice information from the tax collection and management system to construct a subgraph containing transaction relationships; and judicial institutions extract enterprise judicial litigation information from the China Judgments Online website to construct a subgraph containing judicial litigation relationships.

[0067] In the distributed entity alignment process, for entities without a unified social credit code (such as sole proprietorships and specific assets), a combination of multiple attributes is used for alignment via hashing. For example, for sole proprietorships, a hash calculation is performed using the combination of "operator's name + business address + registration date"; for specific assets, a hash calculation is performed using the combination of "asset name + owner + asset number". This method can improve the accuracy of entity alignment and reduce misalignment.

[0068] Optionally, to further improve the accuracy of entity alignment, federated learning can be used to train the entity alignment model, learning the mapping relationship between entity attributes of different institutions without disclosing the original data.

[0069] S220. Detailed implementation of causal strength calculation: In this embodiment, the formula and process for causal strength calculation are further refined.

[0070] Specifically, for a given edge X→Y (where X represents the cause variable and Y represents the result variable), its set of confounding variables is Z={Z1, Z2, ..., Z...} n}, then the formula for calculating the causal strength ATE (Average Treatment Effect) of this relationship is: ATE=Σ[P(Z=z)×(P(Y=1|X=1,Z=z)-P(Y=1|X=0,Z=z))]; Wherein, P(Z=z) represents the proportion of the confounding variable combination z in all samples, P(Y=1|X=1,Z=z) represents the probability that downstream entity Y defaults when there is a relation X and the confounding variable combination is z, and P(Y=1|X=0,Z=z) represents the probability that downstream entity Y defaults when there is no relation X and the confounding variable combination is z.

[0071] For example, for the edge X→Y, the set of confusion variables Z contains only one variable Z1, which takes the value 0 or 1. The global statistics are as follows: Total sample size N = 100,000; The number of samples with Z1=1 is N1=15000, and the number of samples with Z1=0 is N0=85000; When Z1=1, the number N of samples with X=1 11 =5000, where the number of samples Y=1 is Y 11 =2000; N is the number of samples where X=0. 10 =10000, where the number of samples Y=1 is Y 10 =3000; When Z1=0, the number N of samples with X=1 01 =15000, where the number of samples Y=1 is Y01 =1500; Number of samples N where X=0 00 =70000, where the number of samples Y=1 is Y 00 =3500; The calculation process is as follows: P(Z1=1)=15000 / 100000=0.15; P(Z1=0)=85000 / 100000=0.85; P(Y=1|X=1, Z1=1)=2000 / 5000=0.4; P(Y=1|X=0, Z1=1)=3000 / 10000=0.3; P(Y=1|X=1, Z1=0)=1500 / 15000=0.1; P(Y=1|X=0, Z1=0)=3500 / 70000=0.05; ATE=0.15×(0.4-0.3)+0.85×(0.1-0.05)=0.15×0.1+0.85×0.05=0.015+0.0425=0.0575; It is known that this calculation formula, by weighting the causal effects under different combinations of confounding variables, can effectively control the influence of confounding variables and obtain an accurate estimate of the causal strength.

[0072] Further implementation revealed three core flaws in the Average Treatment Effect (ATE) algorithm, making it unsuitable for cross-institutional credit risk transmission scenarios: The homogeneous causality assumption has a flaw: it assumes that the causal effect of all samples is the same, but ignores the differences in risk transmission between different industries and enterprises of different sizes, resulting in the risk of high-risk entities being diluted by low-risk samples on average. The lack of a time dimension is a flaw; it does not consider the lag in risk transmission (such as the high-risk period of 6-12 months after the establishment of the guarantee relationship), and treats samples from different time intervals equally. Extreme risks are not sensitive. The algorithm lacks weighting for extreme samples such as large guarantees and related-party transactions that could lead to significant losses, thus masking systemic risks. These shortcomings cause a systematic bias in the estimation of the true risk transmission strength by the Average Treatment Effect (ATE) algorithm. Therefore, this application further proposes the Heterogeneous Time Weighted Causality Intensity (HTW-CI) algorithm, the application of which is as follows: Input data excluding the original binary cause variable (existence of relation) (Indicates the existence of a relationship) and the result variable (Downstream entity default status,) (Indicating breach of contract), set of confusing variables In addition, three new data categories have been added: Time-dimensional data: Relationship establishment time. Downstream entity default time Calculate the time interval (Unit: Month); Heterogeneity Feature Data: Enterprise Industry Classification Asset size quantile 3. Extreme risk indicator: Relationship strength quantile (e.g., guarantee amount, transaction amount percentile).

[0073] The processing steps include: The time lag weight calculation uses a truncated Gaussian decay function to calculate the time weight of each sample, reflecting the timeliness of risk transmission and eliminating samples with weak time correlation. ;in, The optimal risk transmission lag time for this type of relationship (obtained from historical data statistics, such as guarantee relationships) ), The width of the time window (e.g.) The function has the largest weight at the optimal lag time, exceeding... The sample weight for the range is 0.

[0074] Heterogeneous stratification of local causal effects calculation, all samples are categorized by industry. and asset size quantile Divided into Each layer is mutually exclusive, and the local causal effect of each layer is calculated to control for the impact of firm heterogeneity. ;in, For layering Confusion variable combinations The proportion, For layering There is a relationship between them. And the confounding variable is The probability of default by downstream entities.

[0075] The stratified weighting and extreme risk correction use the stratified risk exposure as the base weight, and assigns an additional correction coefficient to the high-risk strata to amplify the impact of extreme risks. ;in, For layering The average risk exposure amount (e.g., average guarantee amount) of this type of relationship. For the basic weights of the hierarchy and ; The 90th quantile for all stratified local causal effects. This is the extreme risk amplification factor (ranging from 0.5 to 2; in this embodiment, it is taken as...). ), This is a correction factor for extreme risks.

[0076] By combining the above weights and calculating the overall causal strength, the final heterogeneous time-weighted causal strength is obtained: The output is the HTW-CI value of a single relationship (range 0-1). A higher value indicates a stronger risk transmission intensity for that relationship and accurately reflects the combined impact of time lag, firm heterogeneity, and extreme risks. The cumulative causal strength of the entire path is still the product of the HTW-CI values ​​of each relationship segment.

[0077] Algorithm verification (using data from the original implementation), assuming the original edge Additional information regarding the transaction: Months Months; the sample is divided into manufacturing ( ) and service industry ( ); .

[0078] The calculation yields: ; Compared to the original ATE=0.0575, the new algorithm avoids the averaging effect of low-risk service industry samples, more accurately reflects the actual transmission intensity of high-risk exposure in the manufacturing industry, and removes irrelevant samples with a time interval of more than 12 months, improving the estimation accuracy by about 23.5%.

[0079] S230. Complete execution process of typical application scenarios. This embodiment takes the default event of enterprise "ENT-000001" on January 1, 2025 as an example to describe in detail the complete execution process of the present invention.

[0080] A cross-institutional federated knowledge graph was constructed, with all participating institutions building their local knowledge subgraphs according to unified specifications and completing distributed entity alignment. The Administration for Industry and Commerce constructed a subgraph containing 100,000 enterprise entities and 500,000 equity and senior management relationships; Financial Institution A constructed a subgraph containing 50,000 enterprise entities and 200,000 lending and guarantee relationships; the Tax Administration constructed a subgraph containing 80,000 enterprise entities and 300,000 transaction relationships; and the Judicial Institution constructed a subgraph containing 30,000 enterprise entities and 100,000 litigation relationships. The coordinating node maintained a global subgraph metadata directory, recording the entity types, relationship types, and quantities of each institution's subgraph.

[0081] Initial risk transmission path discovery and preliminary filtering: On January 1, 2025, enterprise "ENT-000001" defaulted on its loan from financial institution A. Financial institution A marked the enterprise as a risk source node and sent the global identifier of the risk source node "ENT-000001" to the coordinating node. The coordinating node broadcast the risk source node information to all participating institutions.

[0082] Each participating institution performs a breadth-first traversal of its local knowledge subgraph with a depth of 5. The business registration department discovers the following path in its local subgraph: "ENT-000001-Equity Relationship → ENT-000002-Senior Management Relationship → ENT-000003"; Financial Institution A discovers the following path: "ENT-000001-Guarantee Relationship → ENT-000004-Loan Relationship → ENT-000005"; The tax administration agency discovers the following path: "ENT-000001-Transaction Relationship → ENT-000006-Transaction Relationship → ENT-000007"; Financial Institution B discovers the following path: "ENT-000002-Transaction Relationship → ENT-000006".

[0083] Each participating organization sends its local path fragment to the coordinating node, which then merges them to obtain four initial paths: Path 1: ENT-000001-Equity Relationship → ENT-000002-Senior Management Relationship → ENT-000003; Path 2: ENT-000001-Guarantee Relationship → ENT-000004-Loan Relationship → ENT-000005; Path 3: ENT-000001-Transaction Relationship → ENT-000006-Transaction Relationship → ENT-000007; Path 4: ENT-000001-Equity Relationship → ENT-000002-Transaction Relationship → ENT-000006-Transaction Relationship → ENT-000007; The coordinating node performs preliminary filtering on the initial paths, and all paths with a length not exceeding 5 are retained.

[0084] Distributed identification of confounding variables involves coordinating nodes to construct a causal graph for each path and query potential confounding variables.

[0085] For edge 1 of path 1 (ENT-000001→ENT-000002, equity relationship), the potential confounding variable is the joint guarantor; for edge 2 (ENT-000002→ENT-000003, senior management relationship), the potential confounding variable is the joint shareholder.

[0086] For edge 1 of path 2 (ENT-000001→ENT-000004, guarantee relationship), the potential confounding variable is the joint shareholder; for edge 2 (ENT-000004→ENT-000005, loan relationship), the potential confounding variable is the joint guarantor.

[0087] For edge 1 (ENT-000001→ENT-000006, transaction relationship) of path 3, the potential confounding variables are joint executives, joint shareholders, and joint guarantors; for edge 2 (ENT-000006→ENT-000007, transaction relationship), the potential confounding variables are joint executives, joint shareholders, and joint guarantors.

[0088] For edge 1 (ENT-000001→ENT-000002, equity relationship) of path 4, the potential confounding variable is the joint guarantor; for edge 2 (ENT-000002→ENT-000006, transaction relationship), the potential confounding variables are the joint executives, joint shareholders, and joint guarantors; for edge 3 (ENT-000006→ENT-000007, transaction relationship), the potential confounding variables are the joint executives, joint shareholders, and joint guarantors.

[0089] The coordinating node sends a list of potential confounding variables to each participating organization for verification. The verification results are as follows: The actual confusion variable for edge 1 of path 1 is {ENT-000009} (co-guarantor), while edge 2 has no actual confusion variable.

[0090] Neither of the two edges of path 2 has any actual confusion variables.

[0091] The actual confounding variable for edge 1 of path 3 is {PER-000001} (joint executives), and the actual confounding variable for edge 2 is {ENT-000008} (joint shareholders).

[0092] The actual confounding variable for edge 1 of path 4 is {ENT-000009} (joint guarantor), the actual confounding variable for edge 2 is {PER-000001} (joint executive), and the actual confounding variable for edge 3 is {ENT-000008} (joint shareholder).

[0093] For causal relationship verification and strength assessment, each participating institution calculates the statistics for each edge locally, uses Paillier additive homomorphic encryption, and then sends them to the coordinating node. The coordinating node aggregates the global statistics and calculates the causal strength of each edge. The causal strength of edge 1 in path 1 is 0.72, the causal strength of edge 2 is 0.68, and the cumulative causal strength is 0.72 × 0.68 ≈ 0.4896.

[0094] The causal strength of edge 1 in path 2 is 0.85, the causal strength of edge 2 is 0.79, and the cumulative causal strength is 0.85 × 0.79 ≈ 0.6715.

[0095] The causal strength of edge 1 in path 3 is 0.0575, the causal strength of edge 2 is 0.0623, and the cumulative causal strength is 0.0575 × 0.0623 ≈ 0.0036.

[0096] The causal strength of edge 1 in path 4 is 0.72, the causal strength of edge 2 is 0.0575, the causal strength of edge 3 is 0.0623, and the cumulative causal strength is 0.72×0.0575×0.0623≈0.0026.

[0097] For spurious path elimination and risk ranking, all participating institutions jointly set a causal strength threshold of 0.1. The edge causal strength of each path is then examined. All edges in path 1 have a causal strength greater than 0.1, so it is retained as a true path.

[0098] All edges in path 2 have a causal strength greater than 0.1, so it is retained as a true path.

[0099] The causal strength of both edges of path 3 is less than 0.1, so it is marked as a false path and removed.

[0100] The causality strength of edges 2 and 3 in path 4 is less than 0.1, so they are marked as false paths and removed.

[0101] The remaining true paths are sorted according to cumulative causal strength: Path 2: Cumulative causal strength ≈ 0.6715; Path 1: Cumulative causal strength ≈ 0.4896; According to the risk level classification criteria, the cumulative causal strength of ENT-000005 (target node of path 2) and ENT-000003 (target node of path 1) are both between 0.3 and 0.7, which belongs to the medium risk level.

[0102] Risk warning information is generated and distributed, and the coordinating nodes anonymize the actual path information to generate a risk warning report. The report content includes: the risk source node is ENT-000001 (default occurred on January 1, 2025); the medium-risk target nodes are ENT-000003 and ENT-000005; the actual risk transmission paths are "ENT-000001-equity relationship → ENT-000002-senior management relationship → ENT-000003" and "ENT-000001-guarantee relationship → ENT-000004-loan relationship → ENT-000005" respectively; the cumulative causal strengths are 0.4896 and 0.6715 respectively; the recommended risk control measures include strengthening post-loan management, adjusting credit limits, and adding guarantees.

[0103] The coordinating node distributes the early warning information to relevant financial institutions. Financial institution A receives the early warning information related to ENT-000005, while financial institution B receives the early warning information related to ENT-000003. Based on the early warning information, each financial institution takes corresponding risk control measures for the relevant enterprises, effectively reducing credit losses.

[0104] The results of this embodiment demonstrate that the present invention can accurately identify causal relationships in cross-institutional credit risk transmission and eliminate spurious transmission paths. In this embodiment, the initial path set contains four paths, two of which are spurious paths. The present invention successfully eliminated these two spurious paths, significantly reducing the false alarm rate. Furthermore, the present invention can provide quantitative risk assessment results, offering a scientific basis for financial institutions to formulate differentiated risk control measures.

[0105] Example 3 provides a cross-institutional credit risk transmission path tracing device that integrates knowledge graphs, used to execute the methods described in the above examples. The device comprises two parts: a global coordination module deployed at the coordination node and local computing modules deployed at the local computing nodes of each participating institution.

[0106] The global coordination module, deployed on the coordination node of the financial regulatory department, is responsible for global metadata management, task coordination, data aggregation, and early warning distribution. Specifically, it includes: The standard setting unit is used to organize all participating institutions to develop globally unified entity type dictionaries, relation type dictionaries, attribute dictionaries, variable binarization rules, and causal strength thresholds; The entity alignment unit is used to receive the entity hash values ​​sent by each participating institution, complete the cross-institutional entity alignment, assign a globally unique entity identifier to each entity, and return the alignment result to each participating institution. The Meta-information Management Unit is used to maintain the global sub-graph meta-information catalog, receive local sub-graph meta-information registered by participating institutions, and provide meta-information query services. The path merging unit is used to receive local path segments sent by each participating institution, merge them to form a complete initial risk transmission path set, and perform preliminary filtering. The confusion variable coordination unit is used to query the potential confusion variables of each edge according to the global sub-element information directory, send the list of potential confusion variables to each participating organization for verification, and summarize the verification results to form the confusion variable set of each path; The statistics aggregation unit is used to receive encrypted local statistics sent by each participating institution, perform global statistics aggregation using an additive homomorphic encryption algorithm, and calculate the causal strength of each edge and the cumulative causal strength of each path after decryption. The risk ranking unit is used to eliminate false transmission paths based on the causal strength threshold, rank the real transmission paths according to the cumulative causal strength, and classify the risk level of the target node. The early warning distribution unit is used to anonymize information on the actual risk transmission path, generate standardized risk warning reports, and distribute the warning information to relevant financial institutions.

[0107] The local computing module is deployed on the local computing nodes of each participating institution, and is responsible for local data processing, local computing, and secure data transmission, specifically including: The subgraph construction unit is used to extract data from the organization's core business system, clean, transform and load it, construct a local knowledge subgraph, and assign local identifiers to local entities and relationships; The hash calculation unit is used to perform hash calculation on the unique identifier attribute of the local entity and send the hash value and the corresponding local entity identifier to the entity alignment unit of the coordinating node. The metadata registration unit is used to register the metadata of the local knowledge subgraph to the metadata management unit of the coordinating node. The local traversal unit is used to receive risk source node information broadcast by the coordinating node, perform breadth-first traversal in the local knowledge subgraph, generate local path fragments, and send them to the path merging unit of the coordinating node. The obfuscation variable verification unit is used to receive a list of potential obfuscation variables sent by the coordinating node, verify the existence of obfuscation variables in the local knowledge subgraph, and return the verification result to the obfuscation variable coordination unit of the coordinating node. The statistics calculation unit is used to calculate the local statistics corresponding to each edge and its set of confusion variables according to the globally unified variable binarization rules. The encrypted transmission unit is used to encrypt the local statistics using the Paillier additive homomorphic encryption algorithm and send the encrypted statistics to the statistics aggregation unit of the coordination node. The early warning receiving and handling unit is used to receive risk warning information sent by the coordination node, initiate the risk handling process within the organization, and report the handling progress back to the coordination node.

[0108] It is known that this device adopts a modular design, with each module having a clearly defined function, facilitating deployment and maintenance. The separation of the global coordination module and the local computing module ensures that the original data does not leave the domain, effectively protecting the data privacy and trade secrets of all participating organizations.

[0109] Example 4: This example provides an electronic device and a computer-readable storage medium for implementing the methods described in the above examples.

[0110] This embodiment provides an electronic device, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the cross-institutional risk transmission path tracing method of fused knowledge graph as described in any of the above embodiments.

[0111] Specifically, the electronic device also includes a bus, input / output interfaces, and a communication unit. The processor, memory, input / output interfaces, and communication unit are interconnected via the bus. The processor can be a central processing unit (CPU), graphics processing unit (GPU), or digital signal processor (DSP), which are computing units with data processing and / or program execution capabilities. The memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks or flash memory. The input / output interfaces are used to connect input devices (such as keyboards and mice) and output devices (such as monitors and printers) to enable user interaction. The communication unit is used to realize data communication between electronic devices, supporting multiple communication methods such as Ethernet, Wi-Fi, and 4G / 5G.

[0112] When the electronic device acts as a coordination node, it performs the functions of the global coordination module described above; when the electronic device acts as a local computing node, it performs the functions of the local computing module described above.

[0113] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the cross-institutional risk transmission path tracing method based on fused knowledge graphs as described in any of the above embodiments.

[0114] The computer-readable storage medium can be any medium capable of storing computer programs, including but not limited to disks, optical disks, read-only memory (ROM), random access memory (RAM), flash memory, programmable read-only memory (PROM), and electrically erasable programmable read-only memory (EEPROM).

[0115] It is understood that the electronic device and computer-readable storage medium provided in this embodiment can implement the method of the present invention in software form, facilitating its promotion and application. Users can quickly deploy the system of the present invention by installing the corresponding software program on existing hardware devices, enabling accurate tracking of cross-institutional credit risk transmission paths.

Claims

1. A cross-institutional risk transmission path tracing method integrating knowledge graphs, characterized in that, Includes the following steps: Construct a cross-institutional federated knowledge graph, where each participating institution maintains a knowledge subgraph locally, coordinates nodes to maintain a global metadata catalog, and achieves unified entity identification across institutions through a privacy-preserving entity alignment method; Starting from the risk source node, a distributed breadth-first traversal is performed to collect and merge the path fragments of each institution, forming an initial set of risk transmission paths and performing preliminary filtering. Based on the causal graph theory, a causal structure is constructed for each initial path, and potential confusion variables for each edge are identified in a distributed manner and their existence is verified, forming a set of confusion variables for each path; A unified rule is used to binarize variables, each institution independently calculates local statistics, and global statistics are aggregated through additive homomorphic encryption to calculate the causal strength of a single relationship and the cumulative causal strength of the entire path. False transmission paths containing weak causal relationships are eliminated based on a preset causal strength threshold, and the real transmission paths are sorted by cumulative causal strength and the risk level of the target nodes is classified. The actual path information is anonymized, and a risk warning report is generated and distributed to relevant financial institutions.

2. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, In the cross-agency federated knowledge graph construction process, the entity identifier is composed of the agency identifier and the local entity identifier. Distributed entity alignment is performed using a method based on SHA-256 hashing. The coordinating node only stores the entity hash value and the global identifier, and does not store the original identifier information.

3. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, In the discovery step of the initial risk transmission path, each participating institution only performs traversal within the scope of its local knowledge subgraph. The path fragment only contains the global entity identifier, the global relation identifier, and the relation type, and does not contain any attribute information.

4. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, In the step of identifying obfuscated variables, the types of potential obfuscated variables are queried according to the global sub-element information catalog. Each participating institution verifies the existence of obfuscated variables locally and only returns the global identifier of the obfuscated variables.

5. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, In the calculation steps of the causal strength, the causal strength of a single relationship is calculated using the average treatment effect method, which is obtained by weighted averaging of the causal effects under different combinations of confounding variables. The cumulative causal strength of the entire path is the product of the causal strengths of each relationship segment.

6. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, The additive homomorphic encryption uses the Paillier encryption algorithm. The coordinating node only performs summation on the encrypted statistics and cannot decrypt them to obtain the original statistics of any individual organization.

7. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, In the false transmission path elimination step, if the causal strength of any segment of the relationship in the path is less than a preset threshold, the path is marked as a false path and eliminated.

8. The method for tracing cross-institutional risk transmission paths by integrating knowledge graphs according to claim 1, characterized in that, In the process of generating the risk warning report, all information involving personal privacy and trade secrets is removed, and only the entity global identifier, relationship type and causal strength value are retained. Each financial institution only receives warning information related to its own customers.

9. A cross-institutional risk transmission path tracing device integrating knowledge graphs, characterized in that, include: The global coordination module, deployed on the coordination node, is used to formulate global standards, coordinate entity alignment, maintain metadata catalog, merge initial paths, coordinate the identification of confusing variables, aggregate statistics, calculate causal strength, and generate and distribute early warning information. The local computing module is deployed in each participating institution and is used to build local knowledge subgraphs, calculate entity hash values, register metadata, perform local traversal, verify obfuscated variables, calculate local statistics, encrypt and transmit data, and receive and process early warnings.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the cross-institutional risk transmission path tracing method of the fused knowledge graph as described in any one of claims 1-8.

Images