Industrial chain data analysis method and system based on knowledge graph and intelligent risk control

By constructing an explicit knowledge graph of the industrial chain and identifying implicit relationships, and dynamically adjusting the risk propagation weights, the problem of insufficient identification of implicit relationships in existing technologies is solved, enabling more accurate risk assessment and timely early warning.

CN121437153AInactive Publication Date: 2026-01-30BEIJING YUNRONG TIANXIA TECHNOLOGY CO LTD

Patent Information

Application Number
CN202511614090.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-06
Publication Date
2026-01-30
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing technologies struggle to identify implicit connections such as cyclical fund flows, closed-loop fund systems, and joint controlling shareholders in supply chain data analysis, leading to biased risk assessments, a lack of monitoring capabilities for the dynamic evolution of the supply chain structure, and an inability to capture risk signals in a timely manner.

Method used

Construct an explicit knowledge graph of the industrial chain based on knowledge graphs, identify periodic capital flows and capital loops through temporal co-occurrence patterns, identify common actual controllers by combining equity and personnel relationship graphs, dynamically adjust risk propagation weights, and generate comprehensive risk assessment results by combining topological structure change analysis.

Benefits of technology

It improves the accuracy and comprehensiveness of risk association identification, enables more precise risk quantification assessment, timely captures abnormal change patterns, and enhances the foresight and sensitivity of risk prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121437153A_ABST
    Figure CN121437153A_ABST
Patent Text Reader

Abstract

The invention provides an industrial chain data analysis method and system based on a knowledge graph and intelligent risk control, and relates to an intelligent risk control technology, and the method comprises the steps: constructing an industrial chain dominant knowledge graph which comprises a main body node and a dominant association edge, recognizing implicit association, forming an enhanced knowledge graph, dividing levels according to a centrality metric value and a control path number, and obtaining an enhanced knowledge graph; and a differential risk propagation coefficient is set, an abnormal change mode is identified in combination with time sequence correlation analysis, and finally a comprehensive risk assessment result is generated. The risk association hidden in the industrial chain can be effectively found, and the risk identification accuracy and the early warning timeliness are improved.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to intelligent risk control technology, and in particular to an industry chain data analysis method and system based on a knowledge graph and intelligent risk control. BACKGROUND

[0002] With the rapid development of digital economy, industry chain data analysis and risk management have become increasingly important. The transaction relationships and fund flows between enterprises constitute a complex industrial ecological network, and these relationships not only include direct supplier-customer relationships, but also various implicit associations. The existing industry chain data analysis and risk control technology has the following defects and deficiencies:

[0003] The existing technology mainly focuses on explicit transaction relationships, and has limited ability to identify implicit associations such as periodic fund flows, fund closed loops, and the same actual controller, which makes it impossible to fully grasp the real structure of the industry chain and the risk transmission path. These implicit associations are often important channels for risk concealment and transmission, and ignoring these associations will cause significant deviations in risk assessment.

[0004] The existing risk assessment methods usually use static risk weights, and fail to differentiate the risk transmission coefficients according to the hierarchical status and control ability of different subjects in the industry chain, and also fail to consider the differentiated influence of different types of implicit associations on risk transmission, resulting in a simplified risk transmission model that cannot accurately reflect the risk diffusion law in the actual industry chain.

[0005] Most of the existing technology uses static analysis methods, and lacks the ability to continuously monitor and analyze the dynamic evolution of the structure of the industry chain, and cannot timely capture abnormal changes in the attributes and relationship structure of the subjects in the industry chain. These changes are often early manifestations of risk signals, and static analysis cannot discover the behavior patterns of enterprises reconfiguring through association relationships to evade risk monitoring, thereby reducing the timeliness and accuracy of risk early warning. SUMMARY

[0006] The embodiments of the present application provide an industry chain data analysis method and system based on a knowledge graph and intelligent risk control, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiments of the present application provide an industry chain data analysis method based on a knowledge graph and intelligent risk control, comprising:

[0008] Obtaining transaction data and relationship data between subjects in an industry chain, and constructing an explicit knowledge graph of the industry chain containing subject nodes and explicit association edges;

[0009] Periodic fund flow is identified based on a time sequence co-occurrence pattern of the transaction data, fund closed loop is detected based on a fund flow direction tracking, common actual controller is identified based on a penetration analysis of the inter-subject relationship data, an implicit association edge is generated and an implicit association type is labeled, the industry chain explicit knowledge graph is joined to form an enhanced knowledge graph;

[0010] The levels are divided according to a centrality measure value of the subject node in the enhanced knowledge graph and a control path number, a differentiated risk propagation coefficient is set for different levels, a risk propagation weight is dynamically adjusted in combination with the implicit association type, and a multi-hop weighted propagation is performed along the association edge to obtain a propagated risk value;

[0011] A snapshot sequence of the enhanced knowledge graph is constructed, attribute evolution tracks and topological structure change characteristics of the subject node are extracted, abnormal change patterns and association relationship reconstruction patterns are identified through time sequence association analysis, and a comprehensive risk assessment result is generated through fusion calculation with the propagated risk value.

[0012] The steps of constructing an industry chain explicit knowledge graph including subject nodes and explicit association edges include:

[0013] Entity recognition and attribute extraction are performed on the transaction data and the inter-subject relationship data, the identified industry chain subjects are taken as subject nodes, and the extracted attributes are taken as node attributes;

[0014] Based on transaction counterpart information and transaction direction information in the transaction data, a transaction relationship edge is established between subject nodes having transaction flow; based on stock holding information, personnel appointment information and qualification sharing information in the inter-subject relationship data, a personnel relationship edge is established between subject nodes having stock relationship, personnel association or qualification association; the transaction relationship edge and the personnel relationship edge are taken as explicit association edges;

[0015] Edge attributes of the explicit association edges are labeled, the edge attributes including association strength and association direction;

[0016] Based on the subject nodes, node attributes, explicit association edges and edge attributes, an industry chain explicit knowledge graph is constructed.

[0017] The steps of forming an enhanced knowledge graph include:

[0018] Periodic characteristic vectors of the transaction data are extracted through frequency domain transformation, a periodic stability index is calculated based on energy distribution of the periodic characteristic vectors, and subject node pairs with a periodic stability index exceeding a preset stability threshold are labeled as periodic fund flow associations;

[0019] constructing a directed graph for the fund flow direction in the transaction data and performing loop detection to identify a subject node sequence of a closed path; extracting a fund inflow and outflow timestamp of each node in the subject node sequence and calculating a time interval sequence; based on the time interval sequence, marking a node sequence completing fund circulation as a fund closed loop association;

[0020] constructing an equity relationship graph and a personnel relationship graph for the equity holding information and the personnel appointment information in the inter-subject relationship data respectively, and performing reverse tracing from a target subject to be identified to a root node; cross- verifying the tracing results to identify multiple target subjects traced to the same root node; marking an association between the multiple target subjects as a common actual controller association;

[0021] adding the periodic fund flow association, the fund closed loop association and the common actual controller association as implicit association edges to the industrial chain explicit knowledge graph to form an enhanced knowledge graph.

[0022] The step of identifying the common actual controller association comprises:

[0023] constructing an equity relationship graph based on the equity holding information and constructing a personnel relationship graph based on the personnel appointment information; setting path weight decay rules for the equity relationship graph and the personnel relationship graph respectively, with the equity path weight being multiplied by the holding ratio and the personnel path weight being decreased by the appointment level; starting from a target subject to be identified, performing reverse tracing in the two graphs respectively, and terminating the tracing when the path weight is lower than a preset weight threshold or the tracing depth exceeds a preset depth threshold;

[0024] cross-matching candidate root nodes in the two graphs, calculating a weighted fusion value of the equity tracing path weight and the personnel tracing path weight of a node appearing in the candidate root node set of both graphs as a penetration path confidence, and selecting a candidate root node with the highest confidence as an actual controller root node for each target subject according to the penetration path confidence;

[0025] counting all homologous target subjects traced to the same actual controller root node, and marking a two-by-two association between the homologous target subjects as a common actual controller association.

[0026] The step of calculating the propagated risk value comprises:

[0027] calculating a centrality measure value and a control path number, constructing a two-dimensional feature space based on the centrality measure value and the control path number, clustering and dividing the subject nodes in the two-dimensional feature space, dividing the subject nodes into a core level, an intermediate level and a peripheral level, and setting a decreasing differentiated risk propagation coefficient for each level respectively;

[0028] obtaining implicit association types of each association edge in the enhanced knowledge graph, setting a corresponding basic propagation weight for different implicit association types; identifying the levels of two end subject nodes connected by each association edge, adjusting the basic propagation weight to obtain a dynamic risk propagation weight based on the level difference of the two end subject nodes and the risk propagation coefficient of the levels to which the two end subject nodes belong;

[0029] setting an initial risk node and assigning an initial risk value; propagating risk from the initial risk node to adjacent subject nodes along the association edge; multiplying the risk value of the current node by the dynamic risk propagation weight of the connected association edge to obtain a propagation risk increment and accumulating it to the adjacent subject nodes, and iteratively performing multi-hop propagation to obtain the post-propagation risk value of each subject node.

[0030] The step of calculating the comprehensive risk assessment result comprises:

[0031] sampling the enhanced knowledge graph at multiple time points to construct a snapshot sequence; extracting attribute information of each subject node in the snapshot sequence and concatenating them in time sequence to form an attribute evolution track, calculating the topological feature vector of the connected association edge of each subject node and concatenating them in time sequence to form a topological structure change feature;

[0032] calculating the attribute change rate of the attribute evolution track, identifying the time period exceeding the preset change rate threshold as an attribute abnormal change time period; calculating the topological similarity of the topological structure change feature, identifying the time period below the preset similarity threshold as a topological abnormal change time period; performing time window overlap analysis on the attribute abnormal change time period and the topological abnormal change time period, identifying the subject node that simultaneously satisfies attribute abnormality and topological abnormality as an abnormal change mode; and identifying the subject node combination that occurs association edge reconstruction as an association relationship reconstruction mode by counting the addition, deletion and type change of association edges in the snapshot sequence;

[0033] assigning a risk enhancement factor to the abnormal change mode and the association relationship reconstruction mode respectively; and calculating the comprehensive risk assessment result by weighted fusion of the post-propagation risk value and the corresponding risk enhancement factor.

[0034] The step of identifying the subject node combination that occurs association edge reconstruction as an association relationship reconstruction mode by counting the addition, deletion and type change of association edges in the snapshot sequence comprises:

[0035] Counting the addition, deletion and type change of the associated edges in the snapshot sequence, identifying the subject nodes with edge changes, and extracting the local subgraph centered on the subject nodes with edge changes; performing graph isomorphism detection on the local subgraphs before and after the change and calculating the topological structure similarity; for the local subgraph with topological structure similarity higher than the preset isomorphism threshold, calculating the semantic similarity of the subject node sequence on the path before and after the change; marking the change that meets the preset conditions in topological structure similarity and node sequence semantic similarity as semantic equivalent reconstruction, and recording the involved subject nodes;

[0036] Counting the edge change frequency of each subject node in the snapshot sequence, and marking the subject node with edge change frequency exceeding the preset frequency threshold as a high-frequency reconstruction node;

[0037] Marking the subject node that meets the semantic equivalent reconstruction and high-frequency reconstruction at the same time as the associated relationship reconstruction mode.

[0038] The second aspect of the embodiment of the application provides an industrial chain data analysis system based on a knowledge graph and intelligent risk control, comprising:

[0039] A first unit is configured to acquire transaction data and relationship data between subjects in an industrial chain, and construct an industrial chain explicit knowledge graph comprising subject nodes and explicit associated edges;

[0040] A second unit is configured to identify periodic fund flow based on the time sequence co-occurrence mode of the transaction data, detect fund closed loop based on fund flow direction tracking, identify common actual controllers based on penetration analysis of the relationship data between the subjects, generate implicit associated edges and mark the implicit associated type, and add the industrial chain explicit knowledge graph to form an enhanced knowledge graph;

[0041] A third unit is configured to divide levels according to the centrality measure value and the number of control paths of the subject nodes in the enhanced knowledge graph, set different risk propagation coefficients for different levels, dynamically adjust the risk propagation weight in combination with the implicit associated type, and perform multi-hop weighted propagation along the associated edges to obtain a propagated risk value;

[0042] A fourth unit is configured to construct a snapshot sequence of the enhanced knowledge graph, extract attribute evolution tracks and topological structure change characteristics of the subject nodes, identify abnormal change modes and associated relationship reconstruction modes through time sequence association analysis, fuse the propagated risk value, and generate a comprehensive risk assessment result.

[0043] The third aspect of the embodiment of the application,

[0044] An electronic device is provided, comprising:

[0045] A processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] Fourth aspect of the present invention,

[0049] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] This invention constructs an explicit knowledge graph of the industrial chain and identifies implicit relationships to form an enhanced knowledge graph, effectively revealing the complex explicit and implicit relationship network between entities. It solves the problem that traditional methods are difficult to discover hidden relationships and improves the accuracy and comprehensiveness of risk association identification.

[0051] This invention introduces a hierarchical differentiated risk propagation mechanism and a dynamic weight adjustment method, making risk propagation more consistent with the risk diffusion patterns in the actual industrial chain. It overcomes the problem of risk propagation distortion caused by treating all entities equally in traditional risk assessment, and achieves more accurate risk quantification assessment.

[0052] This invention combines the temporal evolution characteristics and topological structure change analysis of knowledge graphs to capture abnormal change patterns and relationship reconstruction behaviors in the industrial chain in a timely manner, providing early warning at the risk incubation stage, significantly improving the foresight and sensitivity of risk prediction, and providing more scientific and efficient technical support for industrial chain risk management. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the supply chain data analysis method based on knowledge graphs and intelligent risk control, as described in an embodiment of the present invention.

[0054] Figure 2 A flowchart for constructing an enhanced knowledge graph. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0056] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0057] Figure 1 A flowchart of an embodiment of the industrial chain data analysis method based on a knowledge graph and intelligent risk control of the application is shown in FIG. 1. As shown in FIG. 1, the method comprises the following steps. Figure 1

[0058] Obtaining transaction data and inter-subject relationship data of a plurality of subjects in an industrial chain, and constructing an industrial chain explicit knowledge graph comprising subject nodes and explicit association edges;

[0059] Based on the time sequence co-occurrence pattern recognition of the transaction data, periodic fund flow is identified, based on fund flow tracking, fund closed loop is detected, based on the penetration analysis of the inter-subject relationship data, common actual controllers are identified, implicit association edges are generated and implicit association types are labeled, and the industrial chain explicit knowledge graph is added to form an enhanced knowledge graph;

[0060] According to the centrality measure value and the number of control paths of the subject nodes in the enhanced knowledge graph, different levels are divided, different risk propagation coefficients are set for different levels, the risk propagation weight is dynamically adjusted in combination with the implicit association type, and multi-hop weighted propagation is performed along the association edges to obtain a propagated risk value;

[0061] A snapshot sequence of the enhanced knowledge graph is constructed, the attribute evolution track and the topological structure change feature of the subject nodes are extracted, the abnormal change mode and the association relationship reconstruction mode are identified through time sequence correlation analysis, and the propagated risk value is fused and calculated to generate a comprehensive risk assessment result.

[0062] In an optional embodiment, the step of obtaining transaction data and inter-subject relationship data of a plurality of subjects in an industrial chain, and constructing an industrial chain explicit knowledge graph comprising subject nodes and explicit association edges comprises:

[0063] Performing entity recognition and attribute extraction on the transaction data and inter-subject relationship data, taking the identified industrial chain subjects as subject nodes, and taking the extracted attributes as node attributes;

[0064] Based on the transaction counterparty information and transaction direction information in the transaction data, transaction relationship edges are established between subject nodes with transaction flow; based on the equity holding information, personnel appointment information and qualification sharing information in the inter-subject relationship data, personnel relationship edges are established between subject nodes with equity relationship, personnel association or qualification association; the transaction relationship edges and the personnel relationship edges are taken as explicit association edges;

[0065] Performing edge attribute labeling on the explicit association edges, the edge attributes comprising association strength and association direction;

[0066] Based on the subject nodes, node attributes, explicit association edges and edge attributes, an industrial chain explicit knowledge graph is constructed.​

[0067] For example, the transaction data in the industrial chain can include invoice data, fund flow data, order data, etc., and the inter-subject relationship data can include business information, equity information, personnel appointment information, etc. For example, in an electronic manufacturing industrial chain, the transaction data from chip supplier A to terminal manufacturer B includes: transaction number T20230501, transaction amount 500 million yuan, transaction date May 1, 2023, transaction product is processor chip, quantity 10000. And the inter-subject relationship data includes: the legal representative of enterprise A is Zhang, the registered capital is 1 billion yuan, and it was established in 2010; enterprise B holds 30% equity of enterprise C, etc.

[0068] When performing entity recognition and attribute extraction on transaction data and inter-subject relationship data, the named entity recognition technology can be used to identify the industrial chain subjects. The specific implementation mode is to use a pre-trained natural language processing model, combined with an industrial field dictionary, to identify enterprise names, organization entities, etc. from the text data. For example, from the text "A company supplies 1000 devices to B company", A company and B company are identified as subject nodes. Attribute extraction is to extract various attribute information of the subject from structured or unstructured data, such as the registered capital, business scope, establishment time, etc. of the enterprise. In practical applications, structured attributes can be directly extracted from enterprise business data, and attribute information can also be extracted from enterprise websites, news reports, etc. unstructured text. For example, from the business data, extract the attribute information of "A company was established in 2015, with a registered capital of 1000 million yuan, and the main business is electronic component manufacturing".

[0069] When establishing transaction relationship edges based on transaction counterpart information and transaction direction information in transaction data, it is necessary to extract the buyer and seller information from the transaction records and determine the direction of the transaction. For example, from the invoice data, the invoicer and the payee can be extracted as the seller and the buyer, respectively; from the fund flow data, the payer and the payee can be extracted to determine the flow direction. For example, in an electronic industrial chain, if there is a transaction record of A supplying goods to B, a transaction relationship edge from A to B is established between A and B. The attributes of the transaction relationship edge can include transaction frequency, transaction amount, transaction product type, etc.

[0070] When establishing personnel relationship edges based on inter-subject relationship data, it is necessary to analyze stock holding information, personnel position information, and qualification sharing information. Stock holding relationship can be extracted from enterprise shareholder list, business change records and other data, such as enterprise A holding 20% of the stock of enterprise B, then a stock holding relationship edge is established between A and B. Personnel association can be achieved by analyzing enterprise executives, board members and other information, such as Mr. Wang serving as the director of enterprise C and enterprise D, then a personnel relationship edge is established between C and D. Qualification association refers to the case where multiple enterprises share the same qualification certificate or license, such as enterprises E and F jointly holding a certain patent or production license, then a qualification relationship edge is established between E and F.

[0071] When labeling the edge attributes of explicit association edges, it is necessary to calculate the association strength and determine the association direction. Association strength can be calculated by multiple dimensions, such as transaction relationship edge, which can consider transaction amount proportion, transaction frequency, transaction duration and other factors, and when quantifying, the three indicators are normalized to the 0-1 interval and then weighted and summed according to the weight of 4:3:3; the association strength of stock holding relationship edge can consider the holding proportion and holding method, when the holding proportion is more than 50%, the association strength is set to 1.0, between 20% and 50%, it is 1.5 times of the holding proportion with a upper limit of 1.0, less than 20%, it is equal to the holding proportion, indirect holding needs to be multiplied by a link attenuation coefficient such as 0.9 for one layer indirect and 0.8 for two layers; the association strength of personnel relationship edge can consider the importance of the associated personnel's position and the number of associated personnel, by establishing a position scoring table, where the chairman or general manager is 1.0, the deputy general manager is 0.8, and the department manager is 0.5, then multiplied by the number of people coefficient, i.e. 0.5 for 1 person, 0.8 for 2-3 people, and 1.0 for more than 4 people. Association direction indicates the direction of influence or dependence between subjects, such as the transaction direction of suppliers to customers, the investment direction of the stock holding direction of the invested party, etc. For example, the annual supply amount of enterprise A to enterprise B accounts for 50% of the total procurement amount of enterprise B, i.e. the proportion score is 0.5, the annual transaction is 12 times normalized to 0.8, and the continuous transaction time is more than 3 years, i.e. the time coefficient is 1.0, then the association strength of the transaction relationship edge can be labeled as 0.5 multiplied by 0.4 plus 0.8 multiplied by 0.3 plus 1.0 multiplied by 0.3 equals 0.74, which is medium-high intensity, and the direction is A pointing to B.

[0072] Finally, when constructing the explicit knowledge graph of the industrial chain, the subject nodes, node attributes, explicit association edges, and edge attributes are integrated into a graph database. In practical applications, graph database technologies such as Neo4j can be used to store and manage the knowledge graph. By defining appropriate graph patterns, including node types, relationship types, and their attributes, the identified subject nodes and their attributes, explicit association edges and their attributes are entered into the graph database. For example, create a "enterprise" type node with attributes including name, registered capital, establishment time, etc.; create a "transaction" type relationship with attributes including transaction amount, transaction frequency, etc. After completing data entry, various analyses can be performed using the query language of the graph database, such as finding the upstream and downstream relationships of a specific enterprise, analyzing key node enterprises in the industrial chain, and discovering potential risk transmission paths. In the explicit knowledge graph of a certain electronic industrial chain, the complete industrial chain structure from raw material suppliers to component manufacturers and then to terminal product manufacturers can be clearly seen, as well as the transaction relationships, equity relationships, and personnel relationships between the subjects.

[0073] By precise entity recognition and attribute extraction, the present application constructs an explicit knowledge graph of the industrial chain, achieving the structured expression of the transaction relationships and personnel relationships of the industrial chain. This method can accurately depict the explicit connection relationships between subjects, laying a solid data foundation for subsequent implicit association mining, significantly improving the visualization degree and analysis efficiency of the industrial chain relationships, and making the risk transmission path more clear and visible.

[0074] In an optional implementation, the step of forming an enhanced knowledge graph includes:

[0075] A periodic feature vector of the transaction data is extracted through frequency domain transformation, a periodic stability index is calculated based on the energy distribution of the periodic feature vector, and a subject node pair with a periodic stability index exceeding a preset stability threshold is marked as a periodic fund flow association;

[0076] A directed graph is constructed for the fund flow direction in the transaction data, and loop detection is performed to identify a sequence of subject nodes in a closed path; the fund inflow and outflow timestamps of each node in the sequence of subject nodes are extracted, and a sequence of time intervals is calculated; based on the sequence of time intervals, a sequence of nodes completing a fund cycle is marked as a fund closed loop association;

[0077] The equity holding information and personnel employment information in the inter-subject relationship data are used to construct an equity relationship graph and a personnel relationship graph, respectively, and reverse tracing is performed from a target subject to be identified to a root node; the tracing results are cross-validated to identify multiple target subjects that trace to the same root node; and the association between the multiple target subjects is marked as a common actual controller association;

[0078] The periodic fund flow association, fund closed loop association and common actual controller association are taken as implicit association edges, and are added to the industry chain explicit knowledge graph to form an enhanced knowledge graph.

[0079] In combination Figure 2 An enhanced knowledge graph flowchart is constructed to illustrate the process of forming an enhanced knowledge graph. In the process of forming an enhanced knowledge graph, transaction data needs to be processed to identify periodic fund flow associations. The transaction records between each pair of subject nodes with transaction relationships are processed in the frequency domain. Specifically, for the transaction time series {t1, t2,..., tn} between subject A and subject B and the corresponding transaction amount series {a1, a2,..., an}, discrete Fourier transform or fast Fourier transform is used to convert them to the frequency domain. In the frequency spectrum obtained after transformation, the frequency components are sorted by energy size, and the periods corresponding to the top three energy distributions are extracted to form a period feature vector P={p1, p2, p3}. To evaluate the stability of the period, the stability of the transaction behavior under each period is calculated. For example, for the period p1, the proportion of transactions within a tolerance range such as plus or minus 3 days of each time point in the period to the total transaction volume is calculated to obtain the stability index s1. The stability indexes s2 and s3 of other periods are calculated in the same way. If max(s1, s2, s3) is greater than 0.75, i.e., a preset stability threshold, the subject pair is marked as having a periodic fund flow association. For example, a trade company and its supplier have 36 transactions within a year. After Fourier transform of the transaction time series, it is shown that 30 days is the main period, i.e., the energy highest frequency component corresponds to the period, and there are 30 transactions within the period falling within a range of plus or minus 3 days of each period point. The stability index is 30 divided by 36, which is 0.83, which exceeds the preset threshold of 0.75. Accordingly, an implicit association edge of periodic fund flow is established between the two, and the period parameter is marked as 30 days and the stability is 0.83.

[0080] To identify the fund closed loop association, a transaction directed graph G equal to the combination of V and E is constructed, where V is the set of subject nodes, E is the set of transaction edges, the direction of the edge represents the direction of the fund flow, and the weight of the edge represents the transaction amount. Perform a depth-first search algorithm or a strongly connected component detection algorithm on the directed graph G to detect closed loops. When a path {v1→v2→...→vk→v1} forms a closed loop, extract the transaction timestamps {t1, t2,..., tk} corresponding to each edge, calculate the time interval sequence {t2-t1, t3-t2,..., t1'-tk} of adjacent transactions, where t1' is the transaction time back to the starting point v1. If all time intervals are within a predefined reasonable range, for example, 1 to 10 days, and the decay rate of the transaction amount on the loop is within a reasonable range, for example, no more than 5% of each transaction, then all the subject node sequences in the loop are marked as fund closed loop association. For example, subject C transfers 500,000 yuan to subject D on May 1, 2023, subject D transfers 480,000 yuan to subject E on May 3, 2023, and subject E transfers 460,000 yuan to subject C on May 5, 2023. These three transactions constitute a complete closed loop, the time intervals are 2 days and 2 days respectively, which are within the reasonable range, and the fund decay rates are 4% and 4.2% respectively, which do not exceed the 5% threshold. Identify that the three subjects have fund closed loop association and label the implicit association edge as involving subjects {C, D, E}, closed loop period 4 days, and total decay rate 8%.

[0081] Then the common actual controller correlation is identified. The equity relationship graph Gs is constructed based on the stock holding information in the inter-subject relationship data, and the personnel relationship graph Gp is constructed based on the personnel appointment information. For the target subject set T to be identified for the actual controller equal to {t1, t2,..., tm}, reverse tracing is performed from each target subject ti in Gs, searching upwards along the stock holding relationship, and the path weight decay rule according to claim 4 is used, and when the path weight is lower than a preset weight threshold such as 0.1 or the tracing depth exceeds a preset depth threshold such as 5 layers, the tracing is terminated, and a set of equity candidate root nodes Rs is obtained. At the same time, in Gp, each target subject is traced upwards along the personnel appointment relationship to find key management personnel, and the weight is decreased according to the appointment level, and a set of personnel candidate root nodes Rp is obtained. Rs and Rp are cross-matched, the penetration path confidence of the nodes appearing in both sets is calculated, and the node with the highest confidence is selected as the actual controller root node. All target subjects traced to the same actual controller root node are counted, and a common actual controller correlation is established between these homologous target subjects. For example, the equity tracing analysis of target subject F and target subject G shows that F is held by natural person H through two layers of indirect holding, with 55% equity, i.e. path weight 0.55 multiplied by decay coefficient 0.8 equals 0.44, and G is held by H through one layer of direct holding, with 60% equity, i.e. path weight 0.6, and H appears in the equity candidate root node set of both; while the personnel relationship analysis shows that the chairman of F and G is H, and there are 3 same members in the management team of F and G, accounting for more than 60% of the respective management team, and H also appears in the personnel candidate root node set. The penetration path confidence of F and G to H is 0.72 and 0.80 respectively, both of which are the highest values, and accordingly H is determined to be the common actual controller of F and G, and a common actual controller correlation edge is established between F and G.

[0082] The identified implicit association is integrated into the industry chain knowledge graph. The original industry chain explicit knowledge graph mainly contains explicit associations between subject nodes, such as transaction relationships, equity relationships, personnel relationships, etc. The identified periodic fund flow association, fund closed loop association and common actual controller association are added to the original knowledge graph as new edge types to form an enhanced knowledge graph. In order to distinguish different types of implicit associations, the edges are labeled with implicit association types, including association type name, confidence of association discovery, discovery time, related parameters such as cycle value, closed loop path, etc. The storage of the enhanced knowledge graph uses the property graph model, and each node and edge can be attached with multiple attributes, which is convenient for subsequent risk propagation calculation and anomaly analysis. For example, in the supply chain analysis of a manufacturing enterprise, the enhanced knowledge graph not only shows the direct transaction relationship between its upstream and downstream suppliers and customers, but also reveals the implicit association between some suppliers that have no direct association on the surface, such as the periodic fund flow between two suppliers, and the implicit association between some customers and potential competitors sharing the same actual controller. These implicit association edges are identified in the graph with different colors or line types, providing a more comprehensive association perspective for enterprise risk prevention and control.

[0083] Through the implementation of the above technical means, the enhanced knowledge graph effectively reveals the implicit association between subjects that cannot be identified by traditional explicit analysis methods, significantly improving the depth and accuracy of industry chain association relationship analysis, and providing a more powerful data foundation and decision support for financial risk prevention, industry chain security evaluation, and association transaction monitoring.

[0084] In an optional implementation, the step of identifying the common actual controller association comprises:

[0085] An equity relationship graph is constructed based on equity holding information, and a personnel relationship graph is constructed based on personnel appointment information; path weight decay rules are set for the equity relationship graph and the personnel relationship graph respectively, the equity path weight is multiplied by the holding proportion, and the personnel path weight is decreased according to the appointment level; starting from the target subject to be identified actual controller, reverse tracing is performed in the two graphs, and the tracing is terminated when the path weight is lower than a preset weight threshold or the tracing depth exceeds a preset depth threshold;

[0086] The candidate root nodes in the two graphs are cross-matched, the weighted fusion value of the equity tracing path weight and the personnel tracing path weight of the nodes appearing in the candidate root node set of the two graphs at the same time is calculated as the penetration path confidence; the candidate root node with the highest confidence is selected as the actual controller root node for each target subject according to the penetration path confidence;

[0087] All homologous target subjects traced to the same actual controller root node are counted, and the pairwise association between the homologous target subjects is marked as a common actual controller association.

[0088] For example, two relationship graphs need to be constructed: a stock ownership relationship graph and a personnel relationship graph. The stock ownership relationship graph is constructed based on the shareholding information between each enterprise entity, where the node represents an enterprise or an individual, the directed edge represents the shareholding relationship, and the weight on the edge represents the shareholding proportion. For example, if A company holds 30% of the shares of B company, a directed edge from A to B is established in the graph, and the weight is 0.3. The personnel relationship graph is constructed based on the information of the positions held by individuals in each enterprise entity, where the node represents an enterprise or an individual, the directed edge represents the position relationship, and the weight on the edge represents the weight value corresponding to the position level. The weight mapping table can be set according to the control power of the position, such as a chairman weight of 0.8, a general manager weight of 0.7, a director weight of 0.6, and a supervisor weight of 0.4. For example, if Zhang holds the position of chairman of C company, a directed edge from Zhang to C company is established in the graph, and the weight is set to 0.8.

[0089] For these two relationship graphs, different path weight decay rules need to be set. In the stock ownership relationship graph, the path weight is calculated by multiplying the shareholding proportions, i.e., if A holds 30% of the shares of B, and B holds 50% of the shares of C, then the indirect shareholding path weight of A to C is 0.3 multiplied by 0.5, which equals 0.15. In the personnel relationship graph, the path weight is calculated by decreasing the position level, and a level decrease coefficient can be set, with the direct position level being 0 level and the weight coefficient being 1.0, and the weight coefficient decreasing by 0.2 for each additional level of indirect position. For example, if natural person P directly holds the position of chairman of A company, which is 0 level, then the personnel path weight of P to A company is 0.8 multiplied by 1.0, which equals 0.8; if A company controls C company, then the indirect position of P to C company is 1 level, and the personnel path weight of P to C company is 0.8 multiplied by 0.8, which equals 0.64, where the first 0.8 is the chairman position weight, and the second 0.8 is the 1 level decrease coefficient, i.e., 1.0 minus 0.2.

[0090] Starting from the target subject to be identified, reverse tracing is performed in the two relationship graphs. In the stock ownership relationship graph, the target subject is traced in the reverse direction along the shareholding relationship, and the cumulative weight of each path is calculated. When the path cumulative weight is below a preset weight threshold, such as 0.05, or the tracing depth exceeds a preset depth threshold, such as 5 levels, the tracing of the path is terminated. For example, if the weight threshold is set to 0.05 and the depth threshold is set to 5 levels, for the path A→B→C→target subject, if the cumulative weight is 0.04, which is lower than 0.05, or the number of levels has reached 6, which exceeds 5, then the tracing of the upper-level shareholders of A is terminated, and A is taken as the candidate root node of the path. Similarly, in the personnel relationship graph, the target subject is traced in the reverse direction along the position relationship, and when the path weight is below a preset threshold, such as 0.3, or the tracing depth exceeds a preset value, such as 3 levels, the tracing is terminated and the candidate root node is recorded.

[0091] By reverse tracing, a series of candidate root nodes will be obtained in the two relationship graphs. Cross matching is performed on these candidate root nodes to find out the nodes that appear in both the candidate root node set of the equity relationship graph and the candidate root node set of the personnel relationship graph. For these cross-appearing nodes, the weighted fusion value of the equity trace path weight and the personnel trace path weight is calculated as the penetration path confidence. The weighted fusion can be performed in the following way: the equity path weight and the personnel path weight are respectively assigned weight coefficients a and β, and a plus β equals 1, and then the weighted sum is calculated. For example, if the path weight of a node in the equity graph is 0.2 and the path weight in the personnel graph is 0.7, and a is set to 0.6 and β is set to 0.4, then the penetration path confidence of the node is 0.2 multiplied by 0.6 plus 0.7 multiplied by 0.4, which equals 0.4. When a node only appears in the candidate set of one graph, the path weight of the non-appearing graph is calculated as 0.

[0092] For each target subject, the candidate root node with the highest confidence is selected as the actual controller root node of the subject according to the calculated penetration path confidence. For example, for target company D, if through the above calculation, it is found that the penetration path confidence of natural person Wang is 0.65, which is higher than that of other candidate root nodes such as Li, which is 0.45, then Wang is determined as the actual controller of company D.

[0093] After completing the identification of the actual controllers of each target subject, all target subjects that trace back to the same actual controller root node are counted, and these subjects are marked as homologous subjects with common actual controller association. These homologous subjects are combined two by two, and each pair of combination is marked as a common actual controller association relationship. For example, if it is found that companies D, E and F all trace back to Wang as the actual controller, then the three pairs of relationships of D and E, D and F, and E and F are marked as common actual controller associations respectively.

[0094] Specific case explanation: Assume that there are enterprise entities A, B, C, D, E and natural persons P, Q and R. P holds 60% of the shares of company A and serves as the chairman, holds 40% of the shares of company B and serves as a director; Q holds 40% of the shares of company A and serves as the general manager; R holds 30% of the shares of company B and serves as the chairman; company A holds 70% of the shares of company C, company B holds 80% of the shares of company D and 50% of the shares of company E. The position weight is set as 0.8 for chairman, 0.7 for general manager and 0.6 for director.

[0095] After constructing the equity graph, we trace back from C and get the weight of path C←A←P as 0.7 times 0.6 equals 0.42, and the weight of path C←A←Q as 0.7 times 0.4 equals 0.28. From D, we get the weight of path D←B←P as 0.8 times 0.4 equals 0.32, and the weight of path D←B←R as 0.8 times 0.3 equals 0.24. From E, we get the weight of path E←B←P as 0.5 times 0.4 equals 0.2, and the weight of path E←B←R as 0.5 times 0.3 equals 0.15. These weights are all higher than the preset threshold 0.05 and the number of layers does not exceed 5, so P, Q, and R are all recorded as candidate root nodes.

[0096] In the personnel graph, we trace back from C and get the weight of path C←A←P as 0.8 times 0.8 equals 0.64, where the first 0.8 is the weight of the chairman position and the second 0.8 is the 1-layer decreasing coefficient. The weight of path C←A←Q as 0.7 times 0.8 equals 0.56. From D, we get the weight of path D←B←P as 0.6 times 0.8 equals 0.48, and the weight of path D←B←R as 0.8 times 0.8 equals 0.64. From E, we get the weight of path E←B←P as 0.48, and the weight of path E←B←R as 0.64.

[0097] After cross-matching, for company C, the candidate root nodes are P and Q. Assuming that a equals 0.5 and β equals 0.5, the confidence of P is 0.42 times 0.5 plus 0.64 times 0.5 equals 0.53, and the confidence of Q is 0.28 times 0.5 plus 0.56 times 0.5 equals 0.42, so P is the actual controller of C. For company D, the candidate root nodes are P and R, the confidence of P is 0.32 times 0.5 plus 0.48 times 0.5 equals 0.4, and the confidence of R is 0.24 times 0.5 plus 0.64 times 0.5 equals 0.44, so R is the actual controller of D. For company E, the candidate root nodes are P and R, the confidence of P is 0.2 times 0.5 plus 0.48 times 0.5 equals 0.34, and the confidence of R is 0.15 times 0.5 plus 0.64 times 0.5 equals 0.395, so R is the actual controller of E.

[0098] Ultimately, it was identified that D and E share a common actual controller, R. Therefore, the relationship between D and E was marked as a common actual controller relationship. An implicit association edge between D and E was added to the enhanced knowledge graph. The attributes of this edge were labeled as association type as common actual controller, actual controller as R, and confidence level as 0.44 for D to R and 0.395 for E to R.

[0099] This invention employs a dual verification mechanism of equity and personnel paths, combined with path weight decay rules, to achieve high-precision identification of the actual controller. This method can overcome the surface fog of complex shareholding structures and overlapping appointments, accurately trace actual control relationships, effectively identify seemingly unrelated entities controlled by the same actual controller, and reveal the substantive control architecture hidden behind complex relationship networks, providing crucial support for risk correlation analysis.

[0100] In one alternative implementation, the step of calculating the post-propagation risk value includes:

[0101] Calculate the centrality metric and the number of control paths; construct a two-dimensional feature space based on the centrality metric and the number of control paths; cluster the main nodes in the two-dimensional feature space, dividing the main nodes into core level, middle level and peripheral level, and setting a decreasing differentiated risk propagation coefficient for each level;

[0102] Obtain the implicit association type of each associated edge in the enhanced knowledge graph, and set corresponding basic propagation weights for different implicit association types; identify the hierarchy of the two main nodes connected by each associated edge, and adjust the basic propagation weights based on the hierarchy difference of the two main nodes and the risk propagation coefficient of the hierarchy to which the two main nodes belong to obtain dynamic risk propagation weights;

[0103] Set an initial risk node and assign an initial risk value; propagate the risk from the initial risk node along the associated edge to the adjacent subject node; multiply the risk value of the current node by the dynamic risk propagation weight of the connected associated edge to obtain the propagation risk increment and accumulate it to the adjacent subject node; iteratively execute multi-hop propagation to obtain the propagated risk value of each subject node.

[0104] For example, the centrality metric is obtained by calculating the degree centrality of nodes, which involves counting the number of edges connecting each principal node. For instance, if node A connects to 10 edges, its degree centrality is 10; if node B connects to 5 edges, its degree centrality is 5. The number of control paths is obtained by calculating the number of pairs of nodes each node connects to as an intermediary node, using the betweenness centrality algorithm to count the number of paths passing through that node in all shortest paths. For example, in a graph with shortest paths CAD, EAF, and GAH, node A acts as a bridge node, controlling the connections between C and D, E and F, and G and H respectively; therefore, node A has 3 control paths.

[0105] A two-dimensional feature space is constructed based on the calculated centrality metric values and the number of control paths. In this two-dimensional space, the horizontal coordinate represents the centrality metric value, and the vertical coordinate represents the number of control paths. For example, the coordinates of node A are (10, 3), indicating that the centrality metric value is 10 and the number of control paths is 3; the coordinates of node B are (5, 1).

[0106] In the constructed two-dimensional feature space, the subject nodes are clustered and divided using the K-means clustering algorithm, and the number of clusters K is set to 3. All nodes are divided into three categories: core level, intermediate level, and peripheral level. When clustering, the Euclidean distance is used as the similarity measure, and the cluster center is iteratively optimized until convergence. After clustering, the clusters are sorted according to the position of the cluster center in the two-dimensional space, and the cluster with the largest coordinate value is marked as the core level, the intermediate level is marked as the intermediate level, and the smallest is marked as the peripheral level. Three levels are set to different risk propagation coefficients: the core level node is set to a propagation coefficient of 0.8, the intermediate level node is set to a propagation coefficient of 0.5, and the peripheral level node is set to a propagation coefficient of 0.3, reflecting the characteristic of decreasing risk propagation ability with level.

[0107] Obtain the implicit association type of each association edge in the enhanced knowledge graph. The implicit association type includes periodic fund flow, fund closed loop association, common actual controller association, and other implicit associations, as well as existing explicit associations such as transaction relationship, equity relationship, personnel relationship, etc. Set the corresponding basic propagation weight for different association types: the periodic fund flow weight is set to 0.6, the fund closed loop association weight is set to 0.7, the common actual controller association weight is set to 0.9, the transaction relationship weight is set to 0.5, the equity relationship weight is set to 0.9, and the personnel relationship weight is set to 0.7. For example, the association edge type between node A and node B is common actual controller association, and the basic propagation weight of the edge is 0.9.

[0108] The dynamic risk propagation weight is equal to the basic propagation weight multiplied by the source node propagation coefficient multiplied by (1 + level difference coefficient). The level difference coefficient is calculated according to the level difference between the source node and the target node. When the source node level is higher than the target node, that is, the risk propagates from high level to low level, the level difference coefficient takes a positive value of 0.2; when the source node level is lower than the target node, that is, the risk propagates in the opposite direction, the level difference coefficient takes a negative value of -0.3 to reflect the attenuation of reverse propagation; when the levels of the two nodes are the same, the level difference coefficient takes 0. For example, when the core level node A (propagation coefficient 0.8) propagates risk to the intermediate level node B (propagation coefficient 0.5), the level difference coefficient takes 0.2, and if the association type between A and B is equity relationship (basic weight 0.9), the dynamic propagation weight is 0.9x0.8x(1+0.2)=0.864.

[0109] Mark the node X as the initial risk node, and assign it an initial risk value of 100. From the initial risk node X, propagate the risk to adjacent subject nodes along the associated edges. Suppose X is connected to node Y through an associated edge of periodic capital flow, X is a core level node (propagation coefficient 0.8), and Y is an intermediate level node (propagation coefficient 0.5), then the dynamic risk propagation weight of the associated edge is 0.6*0.8*(1+0.2)=0.576. Multiply the risk value 100 of X by the dynamic risk propagation weight 0.576 to obtain the propagation risk increment 57.6, which is added to the risk value of node Y to make the risk value of Y change from the initial 0 to 57.6.

[0110] Iteratively perform the multi-hop propagation process, set the maximum propagation hop number to 3, and the risk propagation increment threshold to 5. After one-hop propagation from node X, node Y obtains a risk value of 57.6; node Y continues to propagate the risk to its adjacent node Z, assuming that the association between Y and Z is a capital closed-loop association (base weight 0.7), Y is an intermediate level (propagation coefficient 0.5), and Z is a peripheral level (propagation coefficient 0.3), then the dynamic risk propagation weight is 0.7*0.5*(1+0.2)=0.42, the risk increment propagated by Y to Z is 57.6*0.42=24.19, and node Z obtains a risk value of 24.19; Z continues to propagate to its adjacent node W, assuming that the association between Z and W is a transaction relationship (base weight 0.5), and Z and W are both peripheral levels, then the level difference coefficient is 0, the dynamic risk propagation weight is 0.5*0.3*(1+0)=0.15, the propagation risk increment is 24.19*0.15=3.63, which is lower than the threshold 5, so the propagation from W is stopped. If the risk increment of a path is always higher than the threshold 5 within three hops, the propagation is forcibly stopped after the third hop.

[0111] After completing the risk propagation of all paths, the final propagated risk values of the subject nodes in the enhanced knowledge graph are obtained. For example, the risk value of node X is 100 (initial value), the risk value of node Y is 57.6, the risk value of node Z is 24.19, and the risk value of node W is 3.63. These risk values reflect the degree of influence of each node on the initial risk node in the associated network, and the higher the risk value, the greater the influence of the initial risk node on the node, providing a quantitative basis for subsequent risk management, association review, credit decision-making, and other applications.

[0112] The present application considers the risk bearing and transmission characteristics of different position subjects in the industrial chain, overcomes the homogeneity assumption of traditional risk propagation models, can more accurately simulate the actual risk transmission process, and greatly improves the accuracy of risk diffusion prediction and the pertinence of risk assessment.

[0113] In an optional implementation, the step of calculating the comprehensive risk assessment result comprises:

[0114] sampling the enhanced knowledge graph at multiple time points to construct a snapshot sequence; extracting attribute information of each subject node in the snapshot sequence and concatenating the attribute information in time sequence to form an attribute evolution track; calculating a topological feature vector of a connection association edge of each subject node and concatenating the topological feature vector in time sequence to form a topological structure change feature;

[0115] calculating an attribute change rate for the attribute evolution track and identifying a time period exceeding a preset change rate threshold as an attribute abnormal change time period; calculating a topological similarity for the topological structure change feature and identifying a time period below a preset similarity threshold as a topological abnormal change time period; performing time window overlap analysis on the attribute abnormal change time period and the topological abnormal change time period, and identifying a subject node satisfying both attribute abnormality and topological abnormality as an abnormal change mode; and identifying a subject node combination in which an association edge is reconstructed as an association relationship reconstruction mode by counting the addition, deletion, and type change of association edges in the snapshot sequence;

[0116] assigning a risk enhancement factor to the abnormal change mode and the association relationship reconstruction mode, respectively; and calculating a comprehensive risk assessment result by weighted fusion of the propagation risk value and the corresponding risk enhancement factor.

[0117] For example, sampling the enhanced knowledge graph at multiple time points is the first step of risk assessment. The enhanced knowledge graph is sampled at predetermined time intervals, such as every day, every week, or every month, to form a time sequence snapshot sequence. For example, if a financial transaction network is analyzed, the sampling can be performed once a week, and 12 consecutive weeks of sampling can be performed to obtain a snapshot sequence G equal to {G1, G2,..., G12}, where each snapshot Gi represents the knowledge graph state of the i-th week, and includes all subject nodes, association edges, and attribute information at that time point.

[0118] Based on the obtained snapshot sequence, the attribute information of each subject node is extracted, and the attribute evolution track is formed by concatenating the attribute information in time sequence. Taking an enterprise subject as an example, the attributes include registered capital, number of employees, annual transaction amount, asset-liability ratio, and other financial and operating indicators. For enterprise A, the registered capital attribute evolution track can be represented as a sequence of attribute values at 12 time points [1000 million, 1000 million, 1000 million, 1000 million, 5000 million, 5000 million, 5000 million, 5000 million, 5000 million, 5000 million, 2000 million, 2000 million], and the annual transaction amount evolution track is [50 million, 60 million, 55 million, 58 million, 28 million, 30 million, 29 million, 31 million, 30 million, 29.5 million, 8 million, 7.5 million].

[0119] Meanwhile, the topological feature vector of the connection correlation edge of each subject node is calculated, and the topological structure change feature is formed in time sequence. The topological feature can include the degree centrality of the node, i.e. the number of connection edges, the betweenness centrality, i.e. the number of control paths, the closeness centrality, i.e. the average shortest path length to other nodes, and the like. For example, the degree centrality change of enterprise A at 12 time points is [5, 5, 6, 6, 6, 12, 15, 15, 15, 4, 4, 4], and the betweenness centrality change is [2, 2, 3, 3, 3, 8, 10, 10, 10, 1, 1, 1]. The topological feature vector sequence is formed by combining these indicators.

[0120] When calculating the attribute change rate of the attribute evolution trajectory, the attribute value difference of adjacent time points is used to calculate the change rate. The specific calculation formula is: the attribute change rate is equal to (the attribute value at the later time minus the attribute value at the earlier time) divided by the attribute value at the earlier time multiplied by 100%. When the absolute value of the change rate exceeds the preset threshold, the time period is marked as an attribute abnormal change period. Taking the registered capital of enterprise A as an example, from the 4th week to the 5th week, the registered capital changes from 10 million to 50 million, and the change rate is (5000-1000) divided by 1000 multiplied by 100%, which is equal to 400%. If the preset increase threshold is 100%, the 4th to 5th week period is marked as an attribute abnormal change period. Similarly, from the 10th week to the 11th week, the registered capital decreases from 50 million to 20 million, and the change rate is (2000-5000) divided by 5000 multiplied by 100%, which is equal to negative 60%. If the decrease threshold is set to 50%, i.e. the absolute value of the change rate exceeds 50%, the time period is also marked as abnormal.

[0121] The cosine similarity measurement method is used for the analysis of the topological structure change feature. The cosine similarity of the topological feature vectors of adjacent time points is calculated, and the calculation formula is: the similarity is equal to the dot product of the two vectors divided by the product of the lengths of the two vectors. When the similarity is lower than the preset threshold, it is marked as a topological abnormal change period. Taking enterprise A as an example, the topological feature vector of the 5th week is [6, 3], i.e. the degree centrality is 6 and the betweenness centrality is 3, and the topological feature vector of the 6th week is [12, 8]. The cosine similarity is calculated as (6*12+3*8) divided by (the square root of 6 squared plus 3 squared multiplied by the square root of 12 squared plus 8 squared), which is equal to 96 divided by (the square root of 45 multiplied by the square root of 208), which is equal to 96 divided by 96.59, approximately equal to 0.994. However, if the normalized Euclidean distance is used as the dissimilarity measure, the dissimilarity is equal to the square root of [(12-6) squared plus (8-3) squared] divided by the square root of 2, and then divided by the maximum distance normalization, which is equal to about 0.5. The similarity is equal to 1 minus 0.5, which is equal to 0.5. If the preset similarity threshold is 0.7, the 5th to 6th week is marked as a topological abnormal change period. The degree centrality decreases from 15 to 4 and the betweenness centrality decreases from 10 to 1 in the 9th to 10th week, and the similarity is lower, about 0.3, which is also marked as abnormal.

[0122] By time window overlap analysis, the time period that satisfies both attribute anomaly and topology anomaly is identified. The specific method is: for each subject node, find its all attribute anomaly change time period set Ta and topology anomaly change time period set Tt, calculate the intersection Ta∩Tt, if the intersection is not empty, the subject node is marked as an abnormal change pattern in the intersection time period. Take enterprise A as an example, its attribute anomaly time period includes the 4th to 5th week and the 10th to 11th week, and the topology anomaly time period includes the 5th to 6th week and the 9th to 10th week. Since the 10th to 11th week partially overlaps or approaches the 9th to 10th week, the 9th to 11th week can be regarded as an overlapping interval by extending the time window such as a positive or negative 1-week tolerance, so enterprise A is marked as an abnormal change pattern in the 9th to 11th week.

[0123] The changes of associated edges in the snapshot sequence are also counted, including new edges that do not exist at time ti but appear at ti+1, deleted edges that exist at time ti but disappear at ti+1, and type changed edges that have changed associated type attributes. For each subject node, the number of associated edge changes in each time period is counted, and the edge change rate is calculated as the number of changed edges divided by the original number of edges. If the edge change rate exceeds a preset threshold such as 50%, the related subject combination is marked as an associated relationship reconstruction pattern. For example, enterprise A has 15 associated edges in the 9th week, only 4 associated edges in the 10th week, 11 edges are deleted, and the edge change rate is 11 divided by 15, which is 73.3%, exceeding the 50% threshold. Specifically, the transaction relationship edge between enterprise A and enterprise B disappears in the 9th week, and the relationship edges between enterprise A and multiple long-term partners such as enterprises C, D, and E also disappear in the same period, so enterprise A and these partners B, C, D, and E are marked as an associated relationship reconstruction pattern.

[0124] For the identified abnormal change patterns and associated relationship reconstruction patterns, risk enhancement factors are assigned respectively. The setting of risk enhancement factors can be determined according to historical data analysis and domain expert experience. For example, by analyzing the enterprise samples that have experienced risk events in history, the frequency of abnormal change patterns before the risk events is counted, and it is found that the risk realization probability of abnormal change pattern enterprises is 1.5 times that of normal enterprises, so the risk enhancement factor of abnormal change pattern is set to 1.5. Similarly, the risk realization probability of associated relationship reconstruction pattern enterprises is 1.3 times that of normal enterprises, so the risk enhancement factor of associated relationship reconstruction pattern is set to 1.3. If a subject satisfies multiple abnormal patterns at the same time, the risk enhancement factor is calculated in a cumulative manner such as 1.5×1.3 equals 1.95.

[0125] Finally, the post-propagation risk value is weighted and fused with the corresponding risk enhancement factor to calculate the comprehensive risk assessment result. The post-propagation risk value is the risk value of each subject node obtained by the risk propagation algorithm in claim 5. The specific fusion formula is: the comprehensive risk assessment result is equal to the post-propagation risk value multiplied by the risk enhancement factor. Taking enterprise A as an example, if its post-propagation risk value is 0.6 and it is identified as an abnormal change pattern with an enhancement factor of 1.5, then its comprehensive risk assessment result is 0.6x1.5=0.9. If the post-propagation risk value of enterprise B is 0.4 and it is identified as a correlation relationship reconstruction pattern with an enhancement factor of 1.3, then its comprehensive risk assessment result is 0.4x1.3=0.52. If the post-propagation risk value of enterprise C is 0.5 and it is identified as an abnormal change pattern and a correlation relationship reconstruction pattern, then its comprehensive risk assessment result is 0.5x1.5x1.3=0.975.

[0126] The application comprehensively considers the attribute changes, topological structure changes and correlation relationship reconstruction of the subject nodes, combines the risk propagation results, and obtains more accurate comprehensive risk assessment results, which can provide quantitative basis for risk management decisions such as credit approval, risk early warning and asset portfolio management of financial institutions.

[0127] In an optional implementation, the step of identifying the subject node combination that occurs correlation edge reconstruction and marking as a correlation relationship reconstruction pattern includes:

[0128] The step of identifying the subject node combination that occurs correlation edge reconstruction and marking as a correlation relationship reconstruction pattern includes:

[0129] The step of identifying the subject node combination that occurs correlation edge reconstruction and marking as a correlation relationship reconstruction pattern includes:

[0130] The step of identifying the subject node combination that occurs correlation edge reconstruction and marking as a correlation relationship reconstruction pattern includes:

[0131] Exemplarily, in the dynamic monitoring process of the knowledge graph, the associated edges in the snapshot sequence are scanned in full, and the edge set difference between adjacent snapshots is compared at each time. For two snapshots at time t and time t+1, edge set Et and edge set E(t+1) are constructed, the newly added edge set is the difference set of E(t+1) minus Et, and the deleted edge set is the difference set of Et minus E(t+1). For the edges existing at both times, check the edge type attribute field, and if the type attribute value changes, record it as a type changing edge. The edge type attribute includes predefined type labels such as transaction relationship, equity relationship, and guarantee relationship, which are stored in the type field in the edge attribute dictionary. Extract the end point subject nodes of all newly added edges, deleted edges and type changing edges to form the subject node set that has changed edges. Taking the enterprise relationship network as an example, if the transaction relationship edge between enterprise A and enterprise B disappears at time 5, and a guarantee relationship edge with enterprise C is newly added, then enterprise A is identified as a subject node that has changed edges.

[0132] For each subject node that has changed edges, extract the local subgraph centered on the node. The extraction range of the local subgraph is controlled by the hop parameter, and the commonly used hop is set to 1 hop or 2 hops, that is, all direct neighbor nodes of the center node and the connection edges between these neighbors are included. The 1-hop subgraph extraction is fast but the information is limited, and the 2-hop subgraph can capture more rich local topology structure but the calculation overhead increases. Extract the local subgraph at the time before the change and the time after the change, denoted as subgraph Gt and subgraph G(t+1). The subgraph data structure includes node set, edge set and attribute dictionary of nodes and edges, the node set is stored in the form of node identifier list, and the edge set is stored in the form of start-end two-tuple list. Taking enterprise A as an example, its 1-hop subgraph at time 5 includes enterprise A and its 5 directly associated enterprises B, C, D, E, F, and all 12 connection edges between the 6 enterprises. At time 6, the 1-hop subgraph of enterprise A includes enterprise A and its associated enterprises C, D, E, F, G, H, and the connection edges between the enterprises are 15.

[0133] The graph isomorphism detection is performed on the local sub-graphs before and after the change to determine whether the topological structures of the two sub-graphs are the same or similar after removing the node labels. The graph isomorphism detection adopts an algorithm based on normalized labels. First, the degree, ordered sequence of neighbor node degrees, and number of triangles of each node are calculated as topological invariants. These invariants are combined to form a node signature. Then, the node signatures are sorted in lexicographical order, and a normalized label is assigned to each node. If the number of nodes of the two sub-graphs is the same and the normalized label sequence is completely consistent, it is determined that they are completely isomorphic. In actual applications, the requirement of complete isomorphism is too strict, and the topological structure similarity is calculated to quantify the similarity of the two sub-graphs. The method for calculating the topological structure similarity is to count the number of matching nodes in the two sub-graphs, divide the number of matching nodes by the total number of nodes to obtain the node-level similarity, and at the same time count the number of matching edges, divide the number of matching edges by the total number of edges to obtain the edge-level similarity. Finally, the node similarity and edge similarity are weighted and averaged to obtain the topological structure similarity. The weight coefficients can be set to node weight 0.4 and edge weight 0.6 to emphasize the importance of edge connection mode. Taking enterprise A as an example, the sub-graph at time 5 has 6 nodes and 12 edges, and the sub-graph at time 6 has 7 nodes and 15 edges. Through normalized label matching, it is found that the topological signatures of 5 nodes are consistent, and the connection mode of 10 edges is matched, so the node similarity is 5 divided by 7, which is approximately equal to 0.714, the edge similarity is 10 divided by 15, which is approximately equal to 0.667, and the topological structure similarity is 0.714 multiplied by 0.4 plus 0.667 multiplied by 0.6, which is approximately equal to 0.686.

[0134] A preset isomorphism threshold is set for screening local subgraphs with little change in topology structure. The isomorphism threshold is usually set between 0.6 and 0.8, and below the threshold indicates that the subgraph topology has changed significantly, belonging to structural reconstruction rather than semantic equivalent reconstruction. For local subgraphs with a topology similarity higher than the isomorphism threshold, the semantic similarity of the sequence of principal nodes on the path before and after the change is further calculated. The path extraction strategy is to enumerate all simple paths with a length of no more than 3 hops in the local subgraph, and a simple path refers to a path without repeating the same node. For each path, the sequence of principal nodes on the path is extracted, and the semantic information of the node sequence is derived from the attribute features of the node, including principal type, industry classification, business scope, qualification level, and other attribute fields. A semantic vector is constructed for each principal node, and the dimension of the semantic vector is the dimension of the predefined semantic space, and the commonly used dimension is 128 or 256. The semantic vector is generated by a pre-trained entity embedding model, and the model is trained based on the entity co-occurrence relationship and attribute similarity in the knowledge graph. For the node sequence of the path, the semantic vectors of the nodes in the sequence are concatenated or average-pooled in the order of the path to form the semantic vector representation of the path. The corresponding paths in the two subgraphs before and after the change are determined by the node matching relationship, for example, the path A to B to C at time 5 corresponds to the path A to B to C at time 6, even if node C is replaced by a new associated node D at time 6, but if C and D are semantically similar, the path can still be considered as corresponding. The cosine similarity of the semantic vectors of the corresponding paths is calculated, and the semantic similarity of all corresponding paths is averaged to obtain the semantic similarity of the node sequence. Taking the subgraph of enterprise A as an example, there is a path A to B to D at time 5, and a path A to C to D at time 6, if B and C are both steel trade enterprises and have similar business scope, the cosine similarity of their semantic vectors is 0.92, and the semantic similarity of the corresponding path is 0.92. The semantic similarity of the node sequence is 0.88 by averaging all corresponding paths.

[0135] A preset condition is set for the topology similarity and the semantic similarity of the node sequence, and both similarities need to meet their respective threshold requirements to be marked as semantic equivalent reconstruction. The topology similarity threshold is set to 0.6 to 0.8 as described above, and the node sequence semantic similarity threshold is usually set to 0.75 to 0.9, and the high threshold ensures that the business semantics before and after the reconstruction are basically unchanged. Taking enterprise A as an example, its topology similarity 0.686 is higher than the isomorphism threshold of 0.6, and the node sequence semantic similarity 0.88 is higher than the semantic threshold of 0.75, so the change is marked as semantic equivalent reconstruction, and the principal nodes involved include all nodes B, C, D, E, F, G, and H in the local subgraph of enterprise A. The record data structure is a list of principal node identifiers, stored in the semantic equivalent reconstruction event table, and the table fields include time, central node, involved node list, topology similarity value, and semantic similarity value.

[0136] The edge change frequency of each subject node in the snapshot sequence is counted, and the edge change is defined as the number of times of adding, deleting or changing the type of the edge of the subject node between any adjacent time points. For a snapshot sequence containing 12 time points, there are 11 pairs of adjacent time points, and each subject node has a maximum of 11 edge change opportunities. All time point pairs are traversed, and the edge change times of each subject node are accumulated. Taking enterprise A as an example, there is no edge change between time 1 and time 2, 1 edge is added between time 2 and time 3, there is no change between time 3 and time 4, 2 edges are deleted and 3 edges are added between time 4 and time 5, but the change of the time point pair is counted as 1 change event, and 1 edge change occurs between time 5 and time 6, and so on. Enterprise A has a total of 5 edge changes in 11 time point pairs. A preset frequency threshold is set, and the threshold is determined according to the length of the snapshot sequence and the business stability requirement. For a sequence of 12 time points, the frequency threshold can be set to 3 times, that is, if the edge change frequency of the subject node exceeds 27% of the time point pairs, it is considered to be high frequency. If the edge change frequency of enterprise A is 5 times, which exceeds the threshold of 3 times, enterprise A is marked as a high-frequency reconstruction node. The high-frequency reconstruction node list is stored in the high-frequency node table, and the table fields include the node identifier, the edge change frequency, and the change time list.

[0137] The subject nodes that meet both semantic equivalence reconstruction and high-frequency reconstruction are marked as association relationship reconstruction mode. The specific operation is to perform intersection operation on the involved node list in the semantic equivalence reconstruction event table and the high-frequency reconstruction node list, and the nodes in the intersection are the nodes that meet both conditions. Taking enterprise A as an example, enterprise A appears in the semantic equivalence reconstruction event table and exists in the high-frequency reconstruction node list, so enterprise A is marked as association relationship reconstruction mode. The marking result is stored in the association relationship reconstruction mode table, and the table fields include the node identifier, the reconstruction type label set to semantic equivalence plus high frequency, the related time list, and the risk enhancement factor. The risk enhancement factor is used in subsequent risk assessment, and the nodes of the association relationship reconstruction mode are assigned a risk enhancement factor of 1.3.

[0138] The present application proposes the concept of semantic equivalence reconstruction, combines graph isomorphism detection and semantic similarity analysis, and realizes high-precision association relationship reconstruction mode recognition. This method can identify those surface structure changes but the essential relationship does not change, reveals the structural changes that deliberately avoid supervision, especially has very strong recognition ability to those behaviors that hide risks by frequent adjustment of relationship structure, and greatly improves the discovery ability of hidden risk transmission.

[0139] In a second aspect of the embodiment of the present application, an industry chain data analysis system based on a knowledge graph and intelligent risk control is provided, comprising:

[0140] A first unit is configured to acquire transaction data and inter-subject relationship data of a plurality of subjects in an industrial chain, and construct an explicit knowledge graph of the industrial chain including subject nodes and explicit association edges.

[0141] A second unit is configured to identify periodic fund flow based on a time sequence co-occurrence mode of the transaction data, detect fund closed loop based on fund flow direction, identify common actual controller based on penetration analysis of the inter-subject relationship data, generate implicit association edges and label implicit association types, and add the implicit association edges to the explicit knowledge graph of the industrial chain to form an enhanced knowledge graph.

[0142] A third unit is configured to divide levels according to a centrality measure value and a number of control paths of a subject node in the enhanced knowledge graph, set different risk propagation coefficients for different levels, dynamically adjust risk propagation weights in combination with the implicit association types, and perform multi-hop weighted propagation along the association edges to obtain a propagated risk value.

[0143] A fourth unit is configured to construct a snapshot sequence of the enhanced knowledge graph, extract attribute evolution tracks and topological structure change features of the subject nodes, identify abnormal change modes and associated relationship reconstruction modes through time sequence association analysis, and perform fusion calculation with the propagated risk value to generate a comprehensive risk assessment result.

[0144] In a third aspect, an electronic device is provided, including:

[0145] a processor;

[0146] a memory for storing processor-executable instructions;

[0147] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0148] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0149] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having stored thereon computer readable program instructions that, when executed by a computer, cause the computer to carry out various aspects of the present application.

[0150] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An industry chain data analysis method based on a knowledge graph and intelligent risk control, characterized in that, The method comprises the following steps: obtaining transaction data and relationship data between subjects in an industrial chain, and constructing an industrial chain explicit knowledge graph comprising subject nodes and explicit association edges; based on the time sequence co-occurrence mode of the transaction data, identifying periodic fund flow, based on the fund flow direction, tracking and detecting fund closed loop, based on the relationship data between the subjects, identifying common actual controllers through penetration analysis, generating implicit association edges and labeling the implicit association type, adding the industrial chain explicit knowledge graph to form an enhanced knowledge graph; According to the centrality measure value and the number of control paths of the subject node in the enhanced knowledge graph, the level is divided, the difference risk propagation coefficient is set for different levels, the risk propagation weight is dynamically adjusted combined with the implicit association type, and the multi-hop weighted propagation is performed along the association edge to obtain the propagated risk value; Constructing a snapshot sequence of the enhanced knowledge graph, extracting the attribute evolution track and topological structure change feature of the subject node, identifying abnormal change mode and association relationship reconstruction mode through time sequence correlation analysis, and fusing and calculating the propagated risk value to generate a comprehensive risk assessment result.

2. The method of claim 1, wherein, The step of obtaining transaction data and relationship data between subjects in an industrial chain, and constructing an industrial chain explicit knowledge graph comprising subject nodes and explicit association edges comprises: performing entity recognition and attribute extraction on the transaction data and relationship data between subjects, taking the identified industrial chain subjects as subject nodes, and taking the extracted attributes as node attributes; Based on the transaction counterparty information and transaction direction information in the transaction data, transaction relationship edges are established between subject nodes with transaction flow; based on the stock holding information, personnel appointment information and qualification sharing information in the relationship data between subjects, personnel relationship edges are established between subject nodes with stock relationship, personnel association or qualification association; the transaction relationship edges and personnel relationship edges are taken as explicit association edges; Perform edge attribute labeling on the explicit association edges, and the edge attributes include association strength and association direction; Based on the subject nodes, node attributes, explicit association edges and edge attributes, an industrial chain explicit knowledge graph is constructed.

3. The method of claim 1, wherein, The step of forming an enhanced knowledge graph comprises: extracting the periodic feature vector of the transaction data through frequency domain transformation, calculating the periodic stability index based on the energy distribution of the periodic feature vector, and marking the subject node pair with a periodic fund flow association when the periodic stability index exceeds a preset stability threshold; constructing a directed graph based on the fund flow direction in the transaction data and performing loop detection to identify the sequence of subject nodes in the closed path; extract the fund inflow and outflow timestamps of each node in the sequence of subject nodes and calculate the time interval sequence; based on the time interval sequence, mark the node sequence completing the fund circulation as a fund closed loop association; Based on the stock holding information and personnel appointment information in the relationship data between the subjects, a stock relationship graph and a personnel relationship graph are constructed, respectively, and the reverse tracing is performed from the target subject to be identified actual controller to the root node; cross-verify the tracing results, identify multiple target subjects traced to the same root node; and mark the association between the multiple target subjects as a common actual controller association; The periodic fund flow association, fund closed loop association and common actual controller association are taken as implicit association edges to form an enhanced knowledge graph by adding the implicit association edges to the industrial chain explicit knowledge graph.

4. The method of claim 3, wherein, The step of identifying the common actual controller association comprises: An equity relationship graph is constructed based on equity holding information, and a personnel relationship graph is constructed based on personnel appointment information; path weight decay rules are set for the equity relationship graph and the personnel relationship graph respectively, the equity path weight is multiplied by the holding ratio, and the personnel path weight is decreased by the appointment level; reverse tracing is performed in the two graphs starting from a target subject to be identified actual controller, and the tracing is terminated when the path weight is lower than a preset weight threshold or the tracing depth exceeds a preset depth threshold; candidate root nodes in the two graphs are cross-matched, the weighted fusion value of the equity tracing path weight and the personnel tracing path weight of a node appearing in the candidate root node set of the two graphs at the same time is calculated as the penetration path confidence; the candidate root node with the highest confidence is selected as the actual controller root node for each target subject according to the penetration path confidence; all homologous target subjects traced to the same actual controller root node are counted, and the pairwise association between the homologous target subjects is marked as a common actual controller association.

5. The method of claim 1, wherein, The step of calculating the propagated risk value comprises: a centrality measure value and a control path quantity are calculated; a two-dimensional feature space is constructed based on the centrality measure value and the control path quantity; the subject nodes are clustered and divided in the two-dimensional feature space, and the subject nodes are divided into a core level, an intermediate level and a periphery level, and a decreasing differentiated risk propagation coefficient is set respectively; the implicit association type of each association edge in the enhanced knowledge graph is obtained, the corresponding basic propagation weight is set for different implicit association types; the levels of the two end subject nodes connected by each association edge are identified, and the basic propagation weight is adjusted based on the level difference of the two end subject nodes and the risk propagation coefficient of the levels to which the two end subject nodes belong to obtain a dynamic risk propagation weight; an initial risk node is set and an initial risk value is assigned; the risk is propagated from the initial risk node to adjacent subject nodes along the association edges; the risk value of the current node is multiplied by the dynamic risk propagation weight of the connected association edge to obtain a propagation risk increment and accumulate to the adjacent subject nodes, and the multi-hop propagation is iteratively performed to obtain the propagated risk value of each subject node.

6. The method of claim 1, wherein, The step of calculating the comprehensive risk assessment result comprises: a snapshot sequence is constructed by sampling the enhanced knowledge graph at multiple time points; attribute information of each subject node in the snapshot sequence is extracted to form an attribute evolution track in time sequence, and a topological feature vector of the connected association edge of each subject node is calculated to form a topological structure change feature in time sequence; The attribute evolution trajectory is calculated to obtain an attribute change rate, and a time period exceeding a preset change rate threshold is identified as an attribute abnormal change time period; the topology structure change feature is calculated to obtain a topology similarity, and a time period lower than a preset similarity threshold is identified as a topology abnormal change time period; time window overlap analysis is performed on the attribute abnormal change time period and the topology abnormal change time period, and a subject node simultaneously satisfying attribute abnormality and topology abnormality is marked as an abnormal change mode; addition, deletion and type change of an associated edge in the snapshot sequence are counted, and a subject node combination in which the associated edge is reconstructed is marked as an associated relationship reconstruction mode; The abnormal change mode and the associated relationship reconstruction mode are respectively given a risk enhancement factor; the post-propagation risk value and the corresponding risk enhancement factor are weighted and fused to calculate a comprehensive risk assessment result.

7. The method of claim 6, wherein, The step of counting addition, deletion and type change of an associated edge in the snapshot sequence, and identifying a subject node in which the associated edge is reconstructed as an associated relationship reconstruction mode comprises: The addition, deletion and type change of the associated edge in the snapshot sequence are counted, the subject node in which the edge is changed is identified, and a local subgraph centered on the subject node in which the edge is changed is extracted; graph isomorphism detection is performed on the local subgraphs at the time before and after the change, and a topology structure similarity is calculated; for the local subgraph with a topology structure similarity higher than a preset isomorphism threshold, a semantic similarity of a subject node sequence on a path before and after the change is calculated; the change that simultaneously satisfies the topology structure similarity and the node sequence semantic similarity is marked as semantic equivalent reconstruction, and the involved subject nodes are recorded; The edge change frequency of each subject node in the snapshot sequence is counted, and the subject node with an edge change frequency exceeding a preset frequency threshold is marked as a high-frequency reconstruction node; The subject node that simultaneously satisfies semantic equivalent reconstruction and high-frequency reconstruction is marked as an associated relationship reconstruction mode.

8. An industry chain data analysis system based on a knowledge graph and intelligent risk control, used to implement the method of any one of the preceding claims 1-7, characterized in that, It comprises: A first unit configured to obtain transaction data and inter-subject relationship data of a plurality of subjects in an industrial chain, and construct an industrial chain explicit knowledge graph comprising subject nodes and explicit associated edges; A second unit configured to identify periodic capital flow based on a time sequence co-occurrence mode of the transaction data, detect a capital closed loop based on a capital flow direction, identify a common actual controller based on penetration analysis of the inter-subject relationship data, generate an implicit associated edge and mark an implicit associated type, and add the industrial chain explicit knowledge graph to form an enhanced knowledge graph; A third unit configured to divide levels according to a centrality measure value and a control path number of a subject node in the enhanced knowledge graph, set different risk propagation coefficients for different levels, dynamically adjust a risk propagation weight in combination with the implicit associated type, and perform multi-hop weighted propagation along associated edges to obtain a post-propagation risk value; A fourth unit configured to construct a snapshot sequence of the enhanced knowledge graph, extract an attribute evolution trajectory and a topology structure change feature of a subject node, identify an abnormal change mode and an associated relationship reconstruction mode through time sequence association analysis, and perform fusion calculation with the post-propagation risk value to generate a comprehensive risk assessment result.

9. An electronic device, comprising: It comprises: a processor; a memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Opposite-public marketing method and system for constructing knowledge graph based on multiple dimensions

    CN115186099A

  • Enterprise actual holder identification and risk conduction relation mining method and equipment

    CN120144655A

  • Risk identification method based on supplier association relationship, medium and electronic equipment

    CN120373845A

  • Invoice closed-loop transaction risk assessment method based on graph neural network

    CN120374116A

  • Stockholder penetration analysis system and method

    CN120562539A

Cited By

  • Knowledge graph dynamic construction and optimization method based on group behavior and state propagation

    CN121809640A

  • Method for dynamic construction and optimization of knowledge graph based on group behavior and state propagation

    CN121809640B