A stage-constrained enterprise risk propagation prediction method and system
By unifying the mapping of corporate self-assessment and annual report evidence, and combining life cycle conditions to complete the transmission edge, the system adopts bounded synchronous confluence propagation and path sequencing to solve the unified mapping problem of multi-source evidence processing in the enterprise risk analysis system, thus achieving the accuracy and reproducibility of the risk transmission chain.
Patent Information
- Application Number
- CN202611402875.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-10
AI Technical Summary
Existing enterprise risk analysis systems lack a unified mapping mechanism in multi-source evidence processing, resulting in the same graph outputting similar results for different enterprises. Furthermore, they lack reliable edge completion methods, and the weight calculation is unstable during propagation, making it unable to adapt to the differences in sample distribution at different life cycle stages.
By unifying the mapping of corporate self-assessment and annual report evidence, and supplementing missing transmission edges based on the life cycle conditions of complete case isolation, the risk transmission chain is reproducibly predicted and intervened by adopting bounded synchronous convergence propagation and evidence-based path ranking, combined with fixed graph snapshots and early warning benchmarks.
It enables differentiated risk transmission edge generation for different enterprises and life cycle stages, improves the accuracy of missing edge completion and the reproducibility of transmission chains, and ensures the stability and traceability of the propagation results.
Smart Images

Figure CN123070731A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of enterprise risk analysis technology, and in particular relates to a method and system for predicting the transmission of enterprise risks under stage constraints. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Enterprise risk does not refer to all business operation risks, but rather to the possibility and impact of adverse events or deviations in the operational status of science and technology innovation enterprises in areas such as algorithm security, data compliance, technological ethics, intellectual property, geopolitical risks, talent risks, and R&D risks. At the computational level, based on a risk indicator dictionary, these risks are subdivided into secondary risk indicators, and each secondary risk indicator is modeled as a risk indicator node in a knowledge graph. Enterprise self-assessment answers, annual report text evidence, and manual confirmation records are standardized and mapped to form the initial node activation quantity and evidence set, which are used for subsequent risk propagation and path prediction.
[0004] In existing technologies, enterprise risk analysis systems typically employ methods such as questionnaire scoring, annual report keyword extraction, or knowledge graph propagation. Knowledge graph propagation solutions generally calculate risk propagation probabilities based on pre-defined edge relationships or network structures. For example, they may use adjacency matrices and random walks to mine propagation paths, calculate the impact range based on hierarchical weight attenuation, or assess the overall risk value based on node connection density.
[0005] However, the aforementioned schemes lack a unified mapping mechanism for multi-source evidence. Questionnaires, annual reports, and case chains are processed separately, failing to form standardized initial node states. This leads to convergent results for different companies using the same graph. Furthermore, there is a lack of reliable methods to complete missing edges. Explicit edges may miss implicit connections, while edge completion based on global frequency is susceptible to duplicate information and ignores differences in sample distribution across different stages. During multi-hop propagation, weight multiplication leads to excessive decay of path values, while simple addition of multiple predecessors causes activations to exceed limits, resulting in insufficient numerical stability.
[0006] Therefore, there is an urgent need for a solution to complete and predict the transmission chain of multi-source risk evidence and life cycle cases applicable to enterprises. Summary of the Invention
[0007] To overcome the shortcomings of the prior art, this invention provides a stage-constrained enterprise risk transmission prediction method and system. It aims to achieve reproducible prediction of risk transmission chains and quantification of intervention effects under fixed graph snapshots and early warning benchmarks by uniformly mapping enterprise self-assessment and annual report evidence, completing missing transmission edges based on life cycle conditions isolated by complete cases, and adopting bounded synchronous convergence propagation and evidence-based path sorting.
[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: Firstly, a method for predicting the transmission of enterprise risks under stage constraints is disclosed, including: Establish a risk indicator dictionary and a lifecycle dictionary; acquire multi-source data related to enterprise risks, and generate an initial activation vector containing activation quantity and evidence identifier; Obtain the occurrence cases and parse them into a node sequence. Using complete cases as isolation units, statistically obtain the hybrid transition posterior of stage constraints. Virtual candidate edges are completed based on the hybrid transition posterior, the lifecycle sensitivity of the target node, and the independent case support, and then fixed as a graph snapshot; Within the aforementioned graph snapshot, explicit edges and virtual edges are synchronously and boundedly propagated to update node activation values; Starting from the initial activated node, candidate paths are sorted based on the updated activation amount, edge quality, endpoint importance, and length penalty, and the evidenced critical path is output. Based on the updated activation volume and the critical path, a comprehensive risk value is calculated using a fixed quantile benchmark, and a graded early warning system is generated.
[0009] Furthermore, the generation of the initial activation vector includes: mapping the enterprise's self-assessment answer, annual report evidence, and manual confirmation result to risk indicator nodes respectively, calculating the self-assessment strength, annual report evidence strength, and confirmation coefficient, and synthesizing the initial activation amount of the node according to the preset fusion rules, while writing the evidence identifiers involved in the calculation into the evidence set of the node.
[0010] Furthermore, the calculation of the hybrid transfer posterior includes: counting the number of global transfers, the number of stage transfers, the number of global occurrences of the source node, and the number of stage occurrences of the source node, respectively, and obtaining the global transfer posterior and the stage transfer posterior after smoothing; setting the reliability according to the stage sample size, and weighting and mixing the global posterior and the stage posterior.
[0011] Furthermore, the completion of virtual candidate edges includes generating virtual candidate edges only when the candidate edge score of the node pair reaches a preset completion threshold and the number of independent evidences reaches a preset lower limit.
[0012] Furthermore, the synchronous bounded propagation adopts Noisy-OR merging and adds node persistence terms; in each round, the activation amount of each node is synchronously updated based on the state of the previous round, so that the activation amount is kept in the range of [0,1].
[0013] Furthermore, the ranking of candidate paths includes: for each edge in the candidate path, its edge quality is jointly determined by the effective edge coefficient of the edge, the node persistence coefficient of the node pointed to by the edge, and the edge confidence of the edge; the path score is calculated by combining the geometric mean of the edge quality of each edge in the path, the importance weight of the path endpoint, and the length penalty coefficient; when outputting the path, the path node sequence, support case identifier, life cycle stage, and endpoint consequence information are attached.
[0014] Furthermore, the comprehensive risk value is determined by a weighted calculation of the node exposure level, the highest path score, and the range of affected nodes.
[0015] Secondly, a stage-constrained enterprise risk transmission prediction system is disclosed, including: The basic dictionary and initial activation module are used to build a risk indicator dictionary and a life cycle dictionary, and to acquire multi-source data related to enterprise risks to generate an initial activation vector containing activation volume and evidence identifiers. The Case Analysis and Posterior Statistics module is used to acquire the cases that have occurred and parse them into a sequence of nodes. Using the complete case as the isolation unit, it statistically obtains the hybrid transition posterior of the stage constraints. The virtual candidate edge completion module is used to complete virtual candidate edges based on the hybrid transition posterior, the lifecycle sensitivity of the target node, and the independent case support, and is fixed as a graph snapshot; The bounded synchronous propagation module is used to synchronously propagate explicit edges and virtual edges within the graph snapshot and update the node activation count. The path sorting module is used to sort candidate paths based on the updated activation quantity, edge quality, endpoint importance, and length penalty, starting from the initial activated node, and output the evidenced critical path. The early warning module is used to calculate a comprehensive risk value and form a graded early warning based on the updated activation amount and the critical path, combined with a fixed quantile benchmark.
[0016] Thirdly, a computer device is disclosed, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the steps of the method described above.
[0017] Fourthly, a computer-readable storage medium is disclosed having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.
[0018] The above one or more technical solutions have the following beneficial effects: This invention completes virtual candidate edges by using a hybrid transfer posterior of stage constraints and independent case evidence for dual screening. This enables the same source risk to generate differentiated candidate transmission edges at different enterprise lifecycle stages, reducing the occlusion of specific stage features by global high-frequency edges. At the same time, it suppresses false strong correlations caused by duplicate information within the same case, improving the accuracy and stage adaptability of missing edge completion.
[0019] This invention employs synchronous Noisy-OR merging and incorporates node persistence terms and feedback thresholds to perform bounded propagation, eliminating the influence of node traversal order on the propagation results and avoiding unbounded amplification caused by feedback loops. It combines edge quality, endpoint importance, and length penalty to sort critical paths, and each path simultaneously outputs edge coefficients, lifecycle multipliers, support cases, and endpoint consequences, making the risk transmission chain deduction results both verifiable and traceable.
[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0022] Figure 1 This is a flowchart of a stage-constrained enterprise risk transmission prediction method according to Embodiment 1 of the present invention; Figure 2 This is a block diagram of a stage-constrained enterprise risk transmission prediction system according to Embodiment 2 of the present invention. Detailed Implementation
[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0026] To facilitate understanding, before providing a detailed explanation of the embodiments of the present invention, the terms involved in Embodiment 1 of the present invention will be explained first: Risk indicator nodes: Graph nodes represented by risk dimensions and secondary indicator codes. In this embodiment, they include seven primary dimensions and thirty-one secondary indicators: algorithm security, data compliance, technology ethics, intellectual property, geopolitical game, talent risk, and R&D risk.
[0027] Case transmission chain: A sequence of risk indicator nodes extracted from past cases and arranged chronologically or causally.
[0028] Life cycle stages: startup, growth, maturity, or expansion phases of a business.
[0029] Virtual candidate edges: Directed edges that are not present in the original graph but are temporarily added to the inference snapshot based on evidence of phase condition transition and independent case evidence.
[0030] Initial activation vector: A non-negative intensity vector obtained by mapping the company's self-assessment answers or annual report risk evidence to each risk indicator node.
[0031] Graph snapshot: A set of nodes, edges, parameters, evidence versions, and warning benchmarks fixed during a single simulation.
[0032] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a stage-constrained enterprise risk transmission prediction method, including: Step 1: Establish a risk indicator dictionary and a life cycle dictionary as a unified coding basis for subsequent case analysis, evidence mapping, stage statistics and dissemination calculations.
[0033] The risk indicator dictionary is used to assign stable codes to the primary dimensions and secondary indicators involved in enterprise risk analysis, avoiding naming inconsistencies or ambiguities when mapping multi-source data to graph nodes.
[0034] In this embodiment, the first-level dimensions include seven dimensions: algorithm security, data compliance, technology ethics, intellectual property, geopolitical game, talent risk, and R&D risk.
[0035] The algorithm security dimension includes five secondary indicators, which correspond to the risks of leakage of core algorithm assets and disruption of computing power supply chain, failure of model reliability and robustness, risk of algorithm interpretability and compliance accountability, risk of abuse of generative artificial intelligence and content security, and risk of artificial intelligence technology iteration and business continuity.
[0036] The data compliance dimension also includes five secondary indicators, which correspond to the risks of compliant use of sensitive user data, the risks of full-chain security management of core R&D data, the risks of data security of open source components and third-party supply chains, the risks of cross-border data transmission and critical infrastructure security, and the risks of network attacks and system governance traceability.
[0037] The science and technology ethics dimension includes four secondary indicators, which correspond to the ethical risks of AI algorithm fairness and technology application, ethical risks of life science trials and protection of the rights and interests of subjects, ethical risks of technology abuse and public safety, and the risk of the absence of ethical review mechanisms and governance systems for science and technology innovation enterprises.
[0038] The intellectual property dimension includes five secondary indicators, which correspond to the risks of leakage of trade secrets and R&D results, risks of technology patent layout and barrier failure, risks of technology infringement disputes and patent litigation, risks of industry-university-research cooperation and intellectual property ownership, and risks of international intellectual property and business compliance.
[0039] The geopolitical game dimension includes four secondary indicators, which correspond to the risks of technological sanctions and barriers to product export, the risks of supply chain disruptions for core equipment and key materials, the risks of obstructed access to overseas markets and technology implementation, and the risks of overseas R&D layout and multinational team management.
[0040] The talent risk dimension includes four secondary indicators, which correspond to the risks of core talent attraction, development and reserve and talent pipeline gap, core backbone loss and R&D pipeline continuity, equity incentive and career development path matching, and rapid team expansion and organizational management.
[0041] The R&D risk dimension includes four secondary indicators, which correspond to the risks of technology R&D route selection and strategic deviation, R&D project pipeline advancement and resource allocation coordination, technology engineering verification and mass production adaptation, and R&D results industrialization and market revenue transformation.
[0042] As a result, all 31 secondary indicators are incorporated into the coding framework of the seven primary dimensions. Each risk indicator node is jointly represented by the risk dimension code and the secondary indicator code, ensuring that the node has a unique and traceable identifier in the map snapshot, case chain, and early warning results.
[0043] The lifecycle dictionary includes the startup, growth, maturity, and expansion stages of an enterprise. Each stage is coded and its weight, threshold, and version number are recorded for each risk indicator.
[0044] By managing version numbers, when the definition, weight, or threshold of indicators are adjusted according to business rules or regulatory requirements, the parameter versions used in different simulation batches can be clearly distinguished and reviewed. This ensures that the subsequent case chain sample construction, initial activation vector generation, stage condition transition a posteriori calculation, and early warning quantile benchmark division are all executed based on the same set of stable coding and version benchmarks, thereby improving the overall reproducibility and auditability of the method.
[0045] Furthermore, a knowledge graph of the fixed explicit transmission relationships used for calculation during the risk transmission process is constructed, namely, a risk transmission computation graph, the specific form of which is as follows: ,in This represents a set of nodes consisting of 31 secondary risk indicators. This represents the set of explicit risk propagation edges that were already established prior to this virtual edge patching. The risk propagation computation graph only includes risk propagation edges that have sufficient supporting evidence or have been manually verified.
[0046] edge set It consists of explicit risk transmission edges, each edge representing a confirmed transmission relationship between two risk indicators. For any edge... This means that risks can originate from the source node. Transmitted to the target node .
[0047] Step 2: Obtain multi-source data related to enterprise risk and generate an initial activation vector containing activation values and evidence identifiers. Specifically: Enterprise risk-related multi-source data includes enterprise self-assessment responses, annual report evidence, and manual confirmation results, among other heterogeneous risk evidence.
[0048] The enterprise self-assessment data comes from the risk self-assessment questionnaire filled out by the assessment subjects. It includes at least the enterprise's current selected life cycle stage, the answers of 1 to 5 points for the 31 risk indicator questions, and the corresponding assessment batch identifier and submission timestamp. This data directly reflects the enterprise's proactive assessment of its own status at each risk node.
[0049] The annual report data comes from the publicly disclosed annual report documents of enterprises. Specifically, it includes the uploaded annual report document itself, the document summary value used to verify the integrity of the document, the structured content obtained after text extraction, the page number or paragraph location information of the evidence, the list of hit keywords, and the original evidence fragments containing context. This information enables the risk descriptions in the annual report to be accurately located and mapped to the corresponding risk indicator nodes.
[0050] Manually verified data consists of records generated by risk analysts after reviewing candidate evidence for annual reports automatically extracted by the system. This includes the acceptance or rejection status of each candidate piece of evidence, the manually adjusted risk intensity value, the operation timestamp, and a complete verification record. Its purpose is to perform quality calibration and correction on the automated mapping results. Furthermore, this method supports the integration of optional business data as supplementary evidence sources. This optional business data includes, but is not limited to, internal incident records, audit issue lists, compliance check results, R&D project anomaly reports, and personnel change records. If the aforementioned business systems are not yet integrated at the initial stage of project implementation, this type of data is not a mandatory input item but is reserved as an interface for subsequent expansion of data sources.
[0051] The system receives self-assessment answers or annual report evidence from enterprises and maps the question options, text evidence, and manual confirmation results to the corresponding risk indicator nodes. Specifically, the risk indicator dictionary stores the indicator code, standard name, primary dimension, alias set, keyword list, associated question number, and version number for each indicator.
[0052] Based on the risk indicator dictionary, risk evidence from different sources is mapped using differentiated strategies according to their respective data characteristics.
[0053] For enterprise self-assessment data, since the question design itself corresponds one-to-one with risk indicators, a direct mapping method is adopted: each question in the self-assessment questionnaire is associated with a specific risk indicator code. For example, the self-assessment item with question number "2.1" is directly mapped to the risk indicator node "2.1" without the need for text matching or disambiguation processing. The life cycle stage selected by the enterprise is also recorded as the identifier of the current projection stage, and the answers from 1 to 5 are normalized and used as the original value of the self-assessment intensity of that node.
[0054] The annual report text is first segmented into sentences, then matched with keywords and aliases of each indicator. Upon a match, candidate evidence is generated, recording the sentence content, matched words, and source location. If a sentence matches multiple indicators, multiple candidates are formed separately, without forced merging during the automatic phase. All candidate evidence must be manually verified before becoming valid input; rejected evidence does not activate nodes. The final output includes node codes, initial strength, and a set of evidence identifiers used in the calculation, ensuring the mapping process is traceable.
[0055] For any risk indicator node Three types of strengths were calculated: self-assessment strength, annual report evidence strength, and confirmation coefficient.
[0056] The self-assessment strength comes from the nodes in the enterprise questionnaire or self-assessment form. The formula for calculating the self-assessment strength of the relevant question options is as follows:
[0057] in, The score is based on the question, ranging from 1 to 5. It is mapped to the indicator A set of questions; This refers to the question weight. Currently, the platform generally assigns one secondary indicator to each question, which can be simplified to:
[0058] That is, 1 point corresponds to 0, and 5 points correspond to 1.
[0059] The strength of evidence in annual reports stems from the details and key points within the annual report text. Related risk descriptions, keywords, or event extraction results:
[0060] in, It is in the annual report and indicators A set of matching candidate evidence. When candidate evidence exists, a default value of 0.5 is initially assigned, followed by user confirmation or adjustment. The seven-dimensional comprehensive score calculated on the annual report page is a separate set of dimensional rules and cannot be directly written as a second-level node strength.
[0061] The confirmation coefficient, derived from manual review of the automatically extracted results or supplementary confirmation of missing information, is used to adjust the reliability of the first two types of intensity. The manual confirmation coefficient is defined as follows:
[0062] The above strengths are synthesized according to the preset fusion rules to obtain nodes. initial activation amount .
[0063]
[0064] in: For nodes The initial activation amount; For self-assessment of intensity; For the strength of evidence in annual reports; This is a manually confirmed coefficient; This indicates whether the corresponding data source exists; it is set to 1 if it exists, and 0 otherwise. For data source fusion weights; clip indicates that the result is restricted to 0-1; when the denominator is 0, it is set to zero. .
[0065] During fusion, the identifiers of each evidence source are preserved, and the identifiers of the evidence involved in the calculation are written into the evidence set of that node, so that any activation can be traced back to a specific evidence item, avoiding the occurrence of unexplainable values.
[0066] Among them, the evidence identifier refers to the system generating a globally unique evidence identifier for each piece of evidence involved in the mapping. This identifier is not the evidence content itself, but a unique number pointing to a complete evidence record in the independent evidence table.
[0067] Each evidence record must include at least the following fields: evidenceId (unique identifier of evidence), sourceType (self-assessment / annual report / manual confirmation / historical case), sourceRecordId (assessment result ID / annual report document ID / case ID), sourceLocator (issue number / page number / paragraph offset), contentDigest (content summary), acquiredAt (acquired time), indicatorCode (mapped indicator code), mappingRuleId (mapping rule version), strength (strength value of the evidence's contribution), confirmationStatus (pending confirmation / confirmed / rejected), and validPeriod (validity period).
[0068] It should be understood that the fusion rules can be implemented in different ways, such as maximum value fusion, weighted average, DS evidence synthesis, or output of a calibrated classification model. However, regardless of the method used, the evidence source identifier and confirmation coefficient must be retained to maintain the traceability and auditability of the initial activation vector.
[0069] By generating initial activation values, data from different companies, evidence sources, and collection methods are unified into non-negative activation vectors within the same indicator space, with each dimension of the activation value carrying an evidence set identifier. This initial activation vector serves as the initial state for propagation in subsequent graph snapshots, enabling the same graph to output differentiated paths matching the evidence for different companies, avoiding convergent results based solely on a fixed network structure. Furthermore, since the evidence identifier is retained with the activation value, the activation source can be directly returned in subsequent path ranking and early warning interpretation, enhancing the verifiability of the final risk conclusions.
[0070] Step 3: Obtain the cases that have occurred and parse them into a node sequence. Using the complete case as the isolation unit, statistically obtain the hybrid transition posterior of the stage constraints.
[0071] This step constructs case chain samples using existing risk transmission cases and uses complete cases as isolation units to statistically determine the transition posterior of stage constraints. This provides a staged prior basis for subsequently completing the missing but actually existing risk transmission edges in the graph, specifically including: Step 3.1: Construct Case Chain Samples. Each occurred case is parsed into a directed indicator sequence composed of risk indicator nodes, arranged according to the chronological order of event occurrence or causal progression. Specifically, this includes: first, reading the case title, summary, lifecycle stage, transmission chain text, source link, and case number. Then, the transmission chain text is split into event fragments using arrows, semicolons, periods, and causal connectors. Each fragment is cleaned by removing format prefixes such as "positive transmission" and first-level dimension parentheses after the indicator name, and performing space, punctuation, and full / half-width character normalization. The cleaned fragments are mapped to 31 second-level indicator codes using indicator standard names, aliases, and keywords. After mapping, the node sequence is arranged according to the original arrow order, event occurrence order, or explicit causal relationship. Reverse feedback chains are saved separately and not included in the positive adjacent edge statistics. Finally, n-1 adjacent directed edge samples are generated for each sequence containing n nodes, with all samples inheriting the same caseId.
[0072] Among them, "occurring cases" refers to publicly disclosed or available records of corporate risk events. A complete record of an occurring case must include at least the following components: a unique case identifier (caseId), a case title, the name of the company or entity involved, a case summary, lifecycle stage markers and their determination criteria, the main risk dimension and main secondary indicators, a directed node sequence arranged by risk indicator nodes in causal or chronological order, a description of the transmission type, the source address of the evidence, and the time of acquisition.
[0073] For each case, save the case identifier, the company's lifecycle stage, adjacent node pairs, evidence summary, and evidence source. For a single case containing... Generate a case chain of nodes. There are 3 adjacent directed edge samples, each sample representing the transmission relationship from one node to the next.
[0074] It should be noted that when performing training, validation, or validation partitioning in the future, the complete case should be used as the smallest isolation unit. Different adjacent edge samples generated in the same case should not be assigned to the training set and validation set at the same time, so as to avoid evaluation bias caused by duplicate information within the same case.
[0075] Step 3.2, Computational phase conditional transition posterior verification. For the source node... Candidate target nodes and enterprise stage The following frequencies were counted separately: number of global transfers. This indicates the source node in all stages. The target node then appears. Number of times; number of stage transitions , indicating the stage Next source node The target node then appears. The number of times; the global occurrence count of the source node. , indicating the source node Total number of occurrences across all stages; Number of occurrences in the source node stage , indicating the source node In the stage The frequency of occurrence of the following values. After smoothing the above frequencies, the global transition posterior is obtained. and phase transition post-test :
[0076]
[0077] in: For all training cases The number of adjacent edges; For all training cases, This represents the total number of adjacent edges of the source node; For the stage The following appears The number of times; For the stage The following This represents the total number of adjacent edges of the source node; For global transfer posterior; This is for the post-transfer axiom.
[0078] Given the significant differences in sample size across different stages, directly applying the stage-transition posterior may not be robust enough if the sample size at a stage is small. Therefore, the reliability is set based on the sample size at each stage. The calculation formula is as follows:
[0079] in, This is the shrinkage constant for stage reliability, taken as a positive number, used to control the degree to which stage statistics shrink towards global statistics; For the stage In the training cases, using indicators This represents the number of adjacent transmitted samples of the source node.
[0080] Reliability Between 0 and 1, the larger the sample size at each stage, the better. The closer the result is to 1, the more reliable the stage posterior. Subsequently, the global transition posterior and the stage transition posterior are mixed to obtain the mixed posterior:
[0081] This hybrid posterior achieves an adaptive balance between global and stage statistics: when stage samples are insufficient, it relies more on global transfer patterns; when stage samples are sufficient, it better reflects the transmission characteristics unique to that life cycle stage. This approach avoids the masking of specific stage differences by global high-frequency edges and also prevents posterior fluctuations caused by sparse stage samples.
[0082] Step 4: Complete the virtual candidate edges based on the hybrid transfer posterior, the lifecycle sensitivity of the target node, and the independent case support, and fix them as a graph snapshot.
[0083] Using the stage constraint hybrid transition posterior obtained in step 3, combined with the life cycle sensitivity of the target node and the evidence support of independent cases, the node pairs that do not exist in the original graph are scored and screened, virtual candidate edges with statistical basis and evidence support are filled in, and the graph structure after completion is fixed as the graph snapshot used in this inference to ensure the consistency of subsequent propagation and comparison calculation.
[0084] In practice, for node pairs in the risk propagation computation graph that do not have direct edges... Calculate virtual candidate edge scores The calculation formula is as follows:
[0085] in, For the stage Next source node Pointing to candidate target node Hybrid transfer posterior; For the target node In the stage The normalized lifecycle sensitivity reflects the relative significance of the node being triggered or deteriorated in the current lifecycle stage; In addition to the currently pending verification case, support source nodes Point to target node The proportion of independent case evidence is used to measure whether the transmission relationship occurs repeatedly in multiple independent cases; The lifecycle mixing coefficient controls the contribution weight of lifecycle sensitivity to edge scoring; The independent evidence mixing coefficient controls the contribution weight of independent case evidence support to the edge score.
[0086] In this embodiment, within the same stage, the weights of all indicators are normalized to obtain the normalized lifecycle sensitivity. The specific calculation formula is as follows:
[0087] in: It is a node in the risk model In the stage The AHP global weights; the denominator is the stage. The maximum weight among the next 31 nodes; .
[0088] The formation of virtual candidate edges is subject to three constraints through this scoring formula: first, the statistical regularity provided by the posterior of the stage condition transition; second, the sensitivity correction of the target node under a specific life cycle stage; and third, the support strength of independent evidence across cases.
[0089] Only when A virtual candidate edge is only actually generated in the current graph snapshot when the preset edge-filling threshold is reached and the number of independent pieces of evidence supporting the node pair reaches the preset lower limit. The generated virtual candidate edge is accompanied by information such as a set of supporting cases, confidence level, and validity period. The set of supporting cases is used for subsequent evidence tracing, the confidence level is used for edge quality calculation when ranking paths, and the validity period is used to indicate that the virtual edge is only valid in the current graph snapshot.
[0090] The completed virtual candidate edges, together with the explicit edges in the original graph, are fixed as a graph snapshot.
[0091] Compared to directly supplementing edges based on global frequency, this embodiment, through the dual constraints of stage posterior and independent evidence, can suppress false strong associations caused by duplicate information within the same case, while avoiding the masking of differences in specific lifecycle stages by globally high-frequency edges. Furthermore, since virtual candidate edges are only effective in the current graph snapshot, they need to be recalculated when the graph version or evidence version changes, thus ensuring the timeliness and reproducibility of the edge supplementation results.
[0092] Step 5: Within the graph snapshot, synchronously and boundedly propagate explicit edges and virtual edges, and update the node activation count.
[0093] After obtaining the graph snapshot, the effective propagation capacity of explicit edges and virtual candidate edges is calculated uniformly, and multi-source propagation is performed synchronously and in a bounded manner to update the activation quantity of each risk indicator node. This provides the node state after propagation for subsequent path ranking and early warning calculation, including: Step 5.1: Calculate the effective edge coefficients. This applies to any explicit edge or virtual candidate edge in the graph snapshot. Obtain the basic propagation probability of this edge. Damage intensity Target stage multiplier and time delay To avoid the damage intensity becoming excessive or overflowing in continuous multiplication, the damage intensity is first bounded to obtain a bounded damage intensity:
[0094] in, As a damage intensity scale parameter, it must be greater than 0. The larger the value, the smaller the bounded response corresponding to the same damage intensity, which preserves the relative differences in damage size and prevents extreme values from being unreasonably amplified during propagation.
[0095] Subsequently, the relative stage multiplier of the target node is calculated:
[0096]
[0097] in, For the stage Next target node The stage multiplier; This represents the lower limit of the stage multiplier, ranging from 0 to 1; For the target node in the stage The global weights of AHP under; This is for normalized lifecycle sensitivity.
[0098] Finally, by combining the factors and setting an upper limit, the effective edge coefficients are obtained:
[0099] in, The basic propagation probability parameter of the edge; For bounded damage response; The stage multiplier for the target node; For the edge The time constant must be greater than 0, and the unit is simulation steps.
[0100] Step 5.2: After completing the calculation of the effective edge coefficients, execute synchronous multi-source propagation.
[0101] In the Wheel, for any node Summarize all its predecessor nodes Contributions:
[0102] in, Predecessor node In the The activation amount of the wheel, For the edge The effective edge coefficients are then determined. Subsequently, the nodes are updated using the Noisy-OR merging formula. In the Wheel activation amount:
[0103] in, For nodes The persistence coefficient represents the degree to which a node's existing activations are retained in the next round; the product term represents the number of nodes that have not been triggered by any predecessor contributions. The probability complement. This merging mechanism ensures that the activation levels do not linearly exceed the upper limit when multiple precursors act together, naturally maintaining... Within the interval. Each round of calculation uses a snapshot of the graph from the previous round to update synchronously, meaning that the new activation amount of all nodes is determined only by the activation amount from the previous round, avoiding differences in results due to different node traversal orders.
[0104] If feedback edges form a closed loop during propagation, the contribution continues to propagate along the feedback edge only when the confidence level of the feedback edge, the number of feedback rounds, and the increment threshold are all satisfied simultaneously; otherwise, the contribution of the feedback edge in the current round is truncated, but its record in the evidence tracing is retained. The propagation process continues until the preset maximum number of rounds is reached, or the maximum activation increment of the entire graph is lower than the convergence threshold. Through synchronous updates, Noisy-OR merging, and feedback thresholding, this step can maintain bounded activation and deterministic results under conditions of multiple predecessors, feedback loops, and multi-hop propagation.
[0105] Step 6: Starting from the initial activated node, sort the candidate paths based on the updated activation amount, edge quality, endpoint importance, and length penalty, and output the evidence-based critical path.
[0106] This step extracts candidate transmission paths based on the node activation state after propagation, starting from the initially activated node. It then sorts the paths by comprehensively considering edge quality, endpoint importance, and length penalty, and outputs the critical path with evidence traceability information. This allows the risk transmission chain to not only have numerical calculation results but also to be able to review the case basis and parameter contributions of each step.
[0107] In practice, the nodes with activation values greater than a preset threshold in the initial activation vector generated in step 2 are first used as the starting node set. Then, in the fixed graph snapshot from step 4, simple paths originating from these starting nodes with a length not exceeding a preset upper limit are searched. A simple path is defined as a path that does not repeatedly pass through the same node, avoiding interference from closed loops or duplicate nodes on the path score. The search method can employ depth-first search, breadth-first search, or shortest path equivalence calculation based on negative logarithms, depending on the number of nodes and computational resources.
[0108] For candidate paths Suppose it contains A directed edge. For each edge in the path... Calculate edge mass Its definition is:
[0109]
[0110]
[0111] in: The effective edge coefficients obtained in step 5; For the edge target node The percentage of risk retained after intervention measures have taken effect; For the edge Confidence level; To reduce the risk ratio; For the stage The following is from the source node Point to target node Virtual candidate edge scoring; To limit the calculation results to between 0 and 1.
[0112] Since the longer the path length, the smaller the product of edge qualities obtained by multiplying them is, directly using the product will naturally favor shorter paths, thus masking potentially critical transmission chains in longer paths. To avoid this bias, this step uses a geometric mean to normalize the edge qualities in the path, obtaining the geometric mean of the path's edge qualities:
[0113] Based on this, the path is calculated by combining the importance of the destination and the length penalty. Overall score:
[0114] in, The importance weight of the path's endpoint node is used to distinguish the severity of the consequences of different endpoints. For example, paths that ultimately lead to key dimensions such as geopolitical games or algorithmic security should receive higher attention. This is the length penalty coefficient, which is usually less than 1. It causes the score to be multiplied by a decay factor for each additional hop in the path, thereby further adjusting the ranking position of long paths based on the geometric mean and preventing excessively long paths from obtaining unreasonably high scores due to the large number of edges.
[0115] Each final output critical path includes complete traceability information, including the path node sequence, the contribution of each edge to the score, the supporting case identifiers for each edge, the lifecycle stage of the path, and a description of the potential consequences of the endpoint. When two paths have similar scores, the path with more independent evidence and fewer virtual edges is returned first. This is because more independent evidence means that the transmission relationship occurs repeatedly in multiple independent cases, making it statistically more robust; fewer virtual edges indicate that the path relies more on established explicit knowledge, making the reasoning chain more deterministic.
[0116] Through this step, the post-propagation node activations are transformed into a set of ordered, interpretable, and traceable risk transmission paths. The score for each path reflects both the transmission strength and evidence confidence at the edge level, and the overall quality of the path and the severity of the endpoint consequences are demonstrated through geometric mean and endpoint importance. Meanwhile, the length penalty avoids the tendency to simply pursue short paths or ignore long paths. The supporting case identifiers and evidence tracing information in the output enable risk analysts to trace back to specific case evidence from the final conclusion, enhancing the credibility and verifiability of the risk transmission chain deduction results.
[0117] Step 7: Based on the updated activation amount and the critical path, calculate the comprehensive risk value and form a graded early warning system using a fixed quantile benchmark.
[0118] On a historical sample set at the same lifecycle stage, the risk composite score for each sample is calculated using the same method as the current simulation, and the Q50, Q75, and Q90 quantiles of the risk composite score are statistically obtained. These quantiles serve as the fixed benchmark for this early warning judgment and remain unchanged throughout the simulation and subsequent intervention comparisons. Using historical samples from a fixed stage to calculate quantiles ensures the comparability of early warning levels between different enterprises or different intervention plans, avoiding the possibility of unattributable early warning results due to changes in quantile benchmarks with the sample.
[0119] Current overall risk value of enterprises Based on node exposure level Highest path score and the range of affected nodes The three indicators are weighted and combined, and the calculation formula is as follows:
[0120] Among them, node exposure level Highest path score and the range of affected nodes .
[0121] Specifically, the formula for calculating node exposure is as follows:
[0122] That is, first obtain the peak activation amount of each node during the entire propagation cycle, and then calculate the weighted average according to the indicator weights of the current life cycle stage.
[0123] The formula for calculating the highest path score is:
[0124] in This is the set of candidate paths generated in step 6; if no valid path is found, then... .
[0125] The formula for calculating the range of affected nodes is:
[0126] in: It is a set of risk source nodes with an initial activation value greater than 0; It is a fixed effective activation threshold in the map snapshot; This indicates the number of newly affected nodes after propagation; It is the number of nodes that can be affected, excluding the initial source of risk.
[0127] After obtaining the comprehensive risk value, and combining the propagation and path characteristics obtained in steps 5 and 6, a four-level warning system (blue, yellow, orange, and red) is output according to preset rules. The warning level determination considers not only the quantile range of the comprehensive risk value but also rules related to rapid propagation, cross-dimensional diffusion, two-way feedback, and critical consequences. The rapid propagation rule identifies nodes with rapidly increasing activation levels within a few rounds; the cross-dimensional diffusion rule identifies situations where risk propagates between different first-level dimensions; the two-way feedback rule identifies propagation structures with feedback loops that meet feedback thresholds; and the critical consequences rule identifies situations where the path endpoint points to a highly important node. The comprehensive risk value and rule-based determination jointly determine the final warning level, ensuring that the warning results are based on both numerical data and the structural characteristics of risk propagation.
[0128] After a tiered early warning system is established, a comparison is made before and after the intervention. The intervention plan only allows changes to one or more of the specified edge probabilities, node persistence coefficients, or initial activation amounts, without adjusting the graph structure, quantile benchmarks, random seeds, or other propagation parameters.
[0129] Under the same map snapshot, the same random seed, and the same quantile benchmark, the propagation in step 5 and the path sorting in step 6 were re-executed on the parameters after intervention to obtain the node activations, critical paths, and overall risk values after intervention. When evaluating the intervention effect, the decrease in risk score, the reduction in critical paths, and the reduction in high-exposure nodes were used as quantitative indicators. Specifically, the decrease in risk score refers to the difference in the overall risk value before and after intervention; the reduction in critical paths refers to the reduction in the number of critical paths with scores exceeding a preset threshold after intervention; and the reduction in high-exposure nodes refers to the reduction in the number of nodes with activations exceeding a preset high-exposure threshold.
[0130] By employing the aforementioned fixed-benchmark intervention control design, differences in outcomes before and after the intervention can be clearly attributed to changes in parameters specified in the intervention plan, avoiding confounding factors introduced by changes in graph structure, propagation parameters, or early warning benchmarks. Furthermore, the combination of tiered early warning and quantified intervention effects enables risk analysis to not only provide early warning levels but also assess the actual contribution of specific control measures to reducing risk transmission intensity, shortening critical paths, and minimizing high-exposure nodes, providing quantifiable evidence for enterprise risk management decisions.
[0131] In this embodiment, the case chain record at least includes caseId, stage, nodeSequence, evidenceSummary, and sourceRef; the graph edge record at least includes sourceId, targetId, baseProbability, severity, delay, retention, confidence, evidenceCaseIds, and virtualFlag; and the graph snapshot record at least includes snapshotId, graphVersion, weightVersion, thresholdVersion, and createdAt. The input to the propagation interface at least includes snapshotId, stage, initialActivations, and interventions, and the output at least includes nodeExposures, rankedPaths, warningLevel, evidenceTrace, and comparisonMetrics.
[0132] In leave-one-out complete case validation on 149 cases, 234 adjacent edges, and 31 candidate metrics, compared to the baseline that only uses Laplacian smoothing of global transition frequencies, the MRR of the method in this embodiment increased from 0.472724 to 0.491919, a relative improvement of approximately 4.06%; Hit@5 increased from 0.598291 to 0.658120, a relative improvement of approximately 10%.
[0133]
[0134] Example 2 Based on the method described in Embodiment 1, the purpose of this embodiment is to provide a stage-constrained enterprise risk transmission prediction system, including: The basic dictionary and initial activation module are used to build a risk indicator dictionary and a life cycle dictionary, and to acquire multi-source data related to enterprise risks to generate an initial activation vector containing activation volume and evidence identifiers. The Case Analysis and Posterior Statistics module is used to acquire the cases that have occurred and parse them into a sequence of nodes. Using the complete case as the isolation unit, it statistically obtains the hybrid transition posterior of the stage constraints. The virtual candidate edge completion module is used to complete virtual candidate edges based on the hybrid transition posterior, the lifecycle sensitivity of the target node, and the independent case support, and is fixed as a graph snapshot; The bounded synchronous propagation module is used to synchronously propagate explicit edges and virtual edges within the graph snapshot and update the node activation count. The path sorting module is used to sort candidate paths based on the updated activation quantity, edge quality, endpoint importance, and length penalty, starting from the initial activated node, and output the evidenced critical path. The early warning module is used to calculate a comprehensive risk value and form a graded early warning based on the updated activation amount and the critical path, combined with a fixed quantile benchmark.
[0135] Example 3 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.
[0136] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium.
[0137] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.
[0138] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0139] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0140] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A stage-constrained enterprise risk transmission prediction method, characterized in that, include: Establish a risk indicator dictionary and a life cycle dictionary; Acquire multi-source data related to enterprise risk and generate an initial activation vector containing activation levels and evidence identifiers; Obtain the occurrence cases and parse them into a node sequence. Using complete cases as isolation units, statistically obtain the hybrid transition posterior of stage constraints. Virtual candidate edges are completed based on the hybrid transition posterior, the lifecycle sensitivity of the target node, and the independent case support, and then fixed as a graph snapshot; Within the aforementioned graph snapshot, explicit edges and virtual edges are synchronously and boundedly propagated to update node activation values; Starting from the initial activated node, candidate paths are sorted based on the updated activation amount, edge quality, endpoint importance, and length penalty, and the evidenced critical path is output. Based on the updated activation volume and the critical path, a comprehensive risk value is calculated using a fixed quantile benchmark, and a graded early warning system is generated.
2. The enterprise risk transmission prediction method with stage constraints as described in claim 1, characterized in that, The generation of the initial activation vector includes: mapping the enterprise's self-assessment answer, annual report evidence, and manual confirmation result to risk indicator nodes respectively; calculating the self-assessment strength, annual report evidence strength, and confirmation coefficient; synthesizing the initial activation amount of the node according to the preset fusion rules; and writing the evidence identifiers involved in the calculation into the evidence set of the node.
3. The enterprise risk transmission prediction method with stage constraints as described in claim 1, characterized in that, The calculation of the hybrid transfer posterior includes: counting the number of global transfers, the number of stage transfers, the number of global occurrences of the source node, and the number of stage occurrences of the source node, respectively, and obtaining the global transfer posterior and the stage transfer posterior after smoothing; setting the reliability according to the stage sample size, and weighting and mixing the global posterior and the stage posterior.
4. The enterprise risk transmission prediction method with stage constraints as described in claim 1, characterized in that, The completion of virtual candidate edges includes generating virtual candidate edges only when the candidate edge score of the node pair reaches a preset completion threshold and the number of independent evidences reaches a preset lower limit.
5. The enterprise risk transmission prediction method with stage constraints as described in claim 1, characterized in that, The synchronous bounded propagation adopts Noisy-OR merging and adds node persistence terms; in each round, the activation amount of each node is updated synchronously based on the state of the previous round, so that the activation amount is kept in the range of [0,1].
6. The enterprise risk transmission prediction method with stage constraints as described in claim 1, characterized in that, The ranking of candidate paths includes: for each edge in the candidate path, its edge quality is determined by the effective edge coefficient, the node persistence coefficient of the node pointed to by the edge, and the edge confidence; the path score is calculated by combining the geometric mean of the edge quality of each edge in the path, the importance weight of the path endpoint, and the length penalty coefficient; when outputting the path, the path node sequence, support case identifier, life cycle stage, and endpoint consequence information are attached.
7. The enterprise risk transmission prediction method with stage constraints as described in claim 1, characterized in that, The overall risk value is determined by a weighted calculation of the node exposure level, the highest path score, and the range of affected nodes.
8. A stage-constrained enterprise risk transmission prediction system, characterized in that, include: The basic dictionary and initial activation module are used to build a risk indicator dictionary and a life cycle dictionary, and to acquire multi-source data related to enterprise risks to generate an initial activation vector containing activation volume and evidence identifiers. The Case Analysis and Posterior Statistics module is used to acquire the cases that have occurred and parse them into a sequence of nodes. Using the complete case as the isolation unit, it statistically obtains the hybrid transition posterior of the stage constraints. The virtual candidate edge completion module is used to complete virtual candidate edges based on the hybrid transition posterior, the lifecycle sensitivity of the target node, and the independent case support, and is fixed as a graph snapshot; The bounded synchronous propagation module is used to synchronously propagate explicit edges and virtual edges within the graph snapshot and update the node activation count. The path sorting module is used to sort candidate paths based on the updated activation quantity, edge quality, endpoint importance, and length penalty, starting from the initial activated node, and output the evidenced critical path. The early warning module is used to calculate a comprehensive risk value and form a graded early warning based on the updated activation amount and the critical path, combined with a fixed quantile benchmark.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the stage-constrained enterprise risk transmission prediction method according to any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the stage-constrained enterprise risk transmission prediction method as described in any one of claims 1-7.
