A food safety network transaction risk multidimensional monitoring and early warning system and method

Abstract

The application discloses a food safety network transaction risk multidimensional monitoring and early warning system and method, and relates to the technical field of food safety network transaction supervision data processing. The system converts multi-source supervision observation data into risk observation items, maps them into multi-dimensional risk nodes, generates risk association edges and associated risk domains, generates observable risk state fields in the associated risk domains, calculates risk flow conversion balance deviations and generates potential risk state fields, identifies implicit risk transmission edges according to the potential risk state fields, determines risk critical mutation candidate nodes, determines recommended supervision actions and hierarchical early warning results by simulating supervision actions, and corrects related parameters according to disposal feedback. The application can improve the identification accuracy of risk migration, reflux and implicit transmission in network food transactions.

Description

Technical Field

[0001] This invention relates to the field of data processing technology for online food safety transaction supervision, and in particular to a multi-dimensional monitoring and early warning system and method for online food safety transaction risks. Background Technology

[0002] As the scale of online food transactions expands, regulatory information such as food business entities, product information, transaction behavior, complaints and reports, sampling inspection results, license status, and penalty records are scattered across different online trading platforms and regulatory systems. Existing regulatory early warning methods typically rely on risk scoring, risk classification, or rule-based warnings based on a single platform, single business entity, or single product record. While these methods can alert users to exposed risks such as complaints, substandard sampling inspections, abnormal licenses, and historical penalties, they are ill-suited to the complex relationships between entities, the volatile forms of goods, the migratable operating platforms, and the time-bound nature of risk events in online food transactions.

[0003] In actual supervision, food safety risks are often not continuously exposed by the same entity, the same store, or the same platform. For example, after a product or business entity is subject to platform action, inspection notices, or administrative penalties, the related risks may continue to exist by changing stores, changing product names, adjusting sales channels, transferring platforms, using affiliated entities, or changing shipping regions. Such risks may appear as a visible decrease in risk at the original site of action, but as an increase in weak signals at affiliated entities, products, platforms, or regions. If static scoring or single-point warnings are based solely on directly visible data, it is easy to mistake a superficial decrease in risk for completion of action, making it difficult to identify the migration, recurrence, or accumulation of risks after action.

[0004] Therefore, existing technologies still have the following shortcomings in the supervision of online food transactions: insufficient ability to organize and correlate data from multiple sources; lack of unified analysis of continuous changes in risks across different entities, commodities, platforms, regions, and times; and lack of effective judgment on whether risks have been truly reduced or transferred to related parties after handling. These shortcomings can lead to delayed early warning results, unclear risk transmission paths, inaccurate identification of key regulatory targets, and may result in regulatory resources being invested in superficially risky entities while failing to promptly identify related parties that actually bear the risks. Summary of the Invention

[0005] To address the challenge of identifying and generating targeted, tiered early warning results for risks in online food safety transactions based on multi-source regulatory observation data, this invention proposes a multi-dimensional monitoring and early warning system and method for online food safety transactions.

[0006] The present invention achieves the above objectives through the following technical solutions:

[0007] A multi-dimensional monitoring and early warning system for food safety online transaction risks, the system comprising:

[0008] The associated risk domain construction module is used to convert multi-source regulatory observation data of food safety online transactions into risk observation items, map the risk observation items into multi-dimensional risk nodes according to risk dimension combinations, generate risk association edges based on the multi-dimensional risk nodes with correlation relationships, and generate associated risk domains based on the risk association edges; the risk dimension combinations include region, food category, business entity or group of entities, online transaction platform, risk event type, and time window;

[0009] The dual-field deviation analysis module is used to generate an observable risk state field within the associated risk domain, calculate the risk flow balance deviation, and generate a potential risk state field from the risk flow balance deviation.

[0010] The latent transmission identification module is used to identify latent risk transmission edges based on the observable risk state field, the potential risk state field, and the risk flow balance deviation, and to determine candidate nodes for risk critical mutation based on the latent risk flux intensity and the potential risk state intensity.

[0011] The counterfactual warning module for regulatory actions is used to apply simulated regulatory actions to the hidden risk transmission edge or candidate nodes of risk critical mutation, simulate the dual-field response of the observable risk state field and the potential risk state field, and determine the recommended regulatory actions and hierarchical warning results accordingly.

[0012] The early warning memory correction module is used to correct the risk flow balance deviation threshold, risk association edge weight, implicit risk transmission edge weight, critical mutation threshold, and regulatory action response parameters based on the handling feedback.

[0013] As a preferred embodiment of the present invention, the risk observation items include a combination of risk dimensions, risk attributes, risk contribution weights, data credibility, and regulatory observability; the risk contribution weights are determined based on the severity of the risk event, data credibility, time freshness, and regulatory effectiveness; the regulatory observability is used to indicate the degree to which the corresponding risk observation item can be directly observed by complaints and reports, spot check results, platform reports, manual review, certificate verification, or historical penalty records;

[0014] The types of associated risk domains include: associated risk domains of the same region and product category, associated risk domains of the same subject group, associated risk domains of the same product family, associated risk domains of the same platform, and associated risk domains of the same disposal object.

[0015] As a preferred embodiment of the present invention, the associated risk domain construction module includes:

[0016] The risk observation stratification unit is used to perform two-dimensional stratification of risk observation items. Risk observation items with a risk contribution weight not lower than the risk contribution threshold and a regulatory observability lower than the observability threshold are marked as low observability high contribution items, and risk observation items with a regulatory observability not lower than the observability threshold are marked as directly observable items.

[0017] The node mapping unit is used to map the risk observations after two-dimensional hierarchical classification to multi-dimensional risk nodes according to the risk dimension combination, retain the two-dimensional hierarchical label of each risk observation, and generate a node index for recording the risk dimension combination.

[0018] The node attribute aggregation unit is used to determine the node risk characterization value of the multidimensional risk node based on the risk contribution weight, data credibility, and time freshness of the multiple risk observations when multiple risk observations are mapped to the same multidimensional risk node, and to determine the domain priority of the multidimensional risk node in participating in the construction of associated risk domains based on the node risk characterization value.

[0019] The association edge generation unit is used to generate risk association edges between multidimensional risk nodes based on at least two of the following: subject association, product family association, platform migration, regional adjacency, complaint co-occurrence, sampling inspection association, certificate change association, and disposal impact association, and to determine the association type and initial association weight of the risk association edges.

[0020] The association domain unit is used to determine the starting multidimensional risk node according to the domain priority, and to expand from the starting multidimensional risk node along the risk association edge in descending order of the initial association weight, and merge multidimensional risk nodes that meet the corresponding association risk domain type into an association risk domain.

[0021] As a preferred embodiment of the present invention, the dual-field deviation analysis module includes:

[0022] The observable state generation unit is used to perform weighted aggregation of directly observable items mapped to the same multidimensional risk node to obtain observable risk quality, and organize it into an observable risk state field according to the node index and time window.

[0023] The balance deviation calculation unit is used to calculate the risk flow balance deviation of each multi-dimensional risk node based on the observable risk quality, identified risk flux, new risk sources, and regulatory inhibition amount of each multi-dimensional risk node within the same associated risk domain.

[0024] The potential anomaly marking unit is used to mark multidimensional risk nodes whose absolute value of risk flow balance deviation is greater than the risk flow balance deviation threshold as potential anomaly nodes.

[0025] The potential state generation unit is used to generate observation gaps based on the weighted differences between low-observable high-contribution items and directly observable items mapped to the same potential anomaly node, and to generate a potential risk state field by combining the risk flow balance deviation, the association type of risk association edges and the initial association weight; the potential risk state field includes the potential risk quality, potential migration tendency, potential backflow tendency and potential risk state intensity of the corresponding multi-dimensional risk node.

[0026] As a preferred embodiment of the present invention, the formula for calculating the risk transfer balance deviation is:

[0027] ;

[0028] In the formula, For multi-dimensional risk nodes In the time window Risk transfer balance deviation; , Representing multidimensional risk nodes In the time window and time window Observable risk quality; This indicates multi-dimensional risk nodes within the same associated risk domain. The risk is transmitted to the set of neighboring nodes. express Any incoming neighbor node in the process; Indicates within the time window The risk-related edge flows into the neighbor node. Transmitted to multi-dimensional risk nodes Identified risk fluxes; This indicates that multi-dimensional risk nodes are received within the same associated risk domain. The set of outflowing neighbor nodes that transmit the risk. express Any outgoing neighbor node in the list; Indicates within the time window Risk-related edges consist of multi-dimensional risk nodes. Pass to outgoing neighbor nodes Identified risk fluxes; Represents multidimensional risk nodes In the time window New sources of risk; Represents multidimensional risk nodes In the time window The amount of regulatory restraint.

[0029] As a preferred embodiment of the present invention, the implicit transmission identification module includes:

[0030] The deviation node pairing unit is used to identify potential abnormal nodes with a risk flow balance deviation of less than zero as source candidate nodes and potential abnormal nodes with a risk flow balance deviation of greater than zero as target candidate nodes within the same associated risk domain and the same time window, and to generate candidate transmission node pairs when there is a risk association edge between the source candidate node and the target candidate node.

[0031] The candidate flux verification unit is used to determine the candidate latent risk flux of the candidate transmission node pair, and after including the candidate latent risk flux in the outflow item of the source candidate node and the inflow item of the target candidate node, recalculate the corrected risk flow balance deviation between the source candidate node and the target candidate node.

[0032] The implicit transmission confirmation unit is used to determine the candidate implicit risk flux as a confirmed implicit risk flux, determine the directed association between candidate transmission node pairs as implicit risk transmission edges, and determine the implicit risk flux intensity of the corresponding multidimensional risk node based on the confirmed implicit risk flux when the absolute values ​​of the risk flow balance deviations after the source candidate node and the target candidate node are both reduced and the sum of the absolute value reductions is not lower than the deviation improvement threshold.

[0033] The critical mutation identification unit is used to determine the critical mutation degree based on the potential risk state field, the intensity of the latent risk flux, the curvature of the change in the corrected risk flow balance deviation, and the amount of regulatory inhibition, and to identify multidimensional risk nodes with a critical mutation degree greater than the critical mutation threshold as candidate nodes for risk critical mutation.

[0034] As a preferred embodiment of the present invention, the formula for calculating the candidate latent risk flux is as follows:

[0035] ;

[0036] In the formula, Indicates within the time window Internal source candidate nodes Point to the target candidate node Candidate latent risk fluxes; as source candidate nodes Node identifier, For target candidate nodes Node identifier; For the source candidate node set The node identifier of any source candidate node in the dataset; For the target candidate node set The node identifier of any target candidate node in the dataset; , Representing the source candidate nodes respectively Target candidate nodes In the time window Risk transfer balance deviation; Indicates within the time window Internal source candidate nodes Point to the target candidate node Candidate flux allocation weights, Indicates within the time window Internal source candidate nodes The node identifier is The candidate flux allocation weights for the target candidate nodes. Indicates within the time window The internal node is identified as The source candidate node points to the target candidate node Candidate flux allocation weights; This indicates a positive number used to avoid a denominator of zero;

[0037] in, ; Indicates the candidate node for the connection source. With target candidate nodes The initial association weight of the risk-related edges; Indicates within the time window Internally, based on the association type of the risk-related edges, from the source candidate nodes... The directional tendency value selected from the potential migration tendency and potential return tendency; Indicates the target candidate node In the time window Potential risks to quality.

[0038] As a preferred embodiment of the present invention, the critical mutation identification unit determines the curvature of the change of the corrected risk flow balance deviation based on the absolute value of the corrected risk flow balance deviation of the same multidimensional risk node within three consecutive time windows.

[0039] The product of the potential risk state intensity and the latent risk flux intensity is determined as the basic mutation amount;

[0040] The curvature adjustment coefficient is determined based on the portion of the curvature that is greater than zero, and the inhibition adjustment coefficient is determined based on the amount of regulatory inhibition, wherein both the curvature adjustment coefficient and the inhibition adjustment coefficient are not less than 1;

[0041] The critical mutation degree of the multidimensional risk node is obtained by dividing the product of the basic mutation amount and the curvature adjustment coefficient by the inhibition adjustment coefficient.

[0042] As a preferred embodiment of the present invention, the regulatory action counterfactual warning module includes:

[0043] The action generation unit is used to generate simulated regulatory actions for hidden risk transmission edges or candidate nodes of risk critical mutation. The simulated regulatory actions include targeted sampling inspection, platform cooperation investigation, certificate verification, entity relationship penetration, commodity family tracking and cross-regional cooperation investigation.

[0044] The dual-field response simulation unit is used to simulate the observable risk state field and potential risk state field after executing each simulated regulatory action within the corresponding associated risk domain, and to obtain the critical abrupt change reduction amount. The decline in the intensity of potential risk status The decline in the intensity of hidden risk flux Regulatory costs and risk transfer penalty amount ;

[0045] Action revenue determination unit, used to determine simulated regulatory actions according to the following formula. Action benefits :

[0046] ;

[0047] In the formula, to These represent the weights of the corresponding items;

[0048] The action screening and early warning unit is used to identify simulated regulatory actions with the greatest risk transfer penalty amount and the highest action benefit as recommended regulatory actions, and to generate hierarchical early warning results based on candidate nodes of risk critical mutation, implicit risk transmission edges and recommended regulatory actions.

[0049] A method for multi-dimensional monitoring and early warning of food safety risks in online transactions, the method comprising:

[0050] Multi-source regulatory observation data of food safety online transactions are converted into risk observation items. The risk observation items are combined and mapped into multi-dimensional risk nodes according to risk dimensions. Risk correlation edges are generated based on the multi-dimensional risk nodes with correlation relationships, and correlation risk domains are generated based on the risk correlation edges.

[0051] An observable risk state field is generated within the associated risk domain, the risk flow balance deviation is calculated, and a potential risk state field is generated from the risk flow balance deviation.

[0052] The hidden risk transmission edge is identified based on the observable risk state field, the potential risk state field, and the risk flow balance deviation, and the candidate nodes for risk critical mutation are determined based on the hidden risk flux intensity and the potential risk state intensity.

[0053] Simulated regulatory actions are applied to the hidden risk transmission edge or candidate nodes of risk critical mutation, simulating the dual-field response of the observable risk state field and the potential risk state field, and the recommended regulatory actions and tiered early warning results are determined accordingly.

[0054] The risk transfer balance deviation threshold, risk association edge weight, implicit risk transmission edge weight, critical mutation threshold, and regulatory action response parameters are adjusted based on the feedback from the handling.

[0055] The beneficial effects of this invention are: it can unify the organization of scattered online food transaction regulatory information according to multi-dimensional risk relationships, and analyze the changing state of risks among related objects, thus no longer relying solely on a single entity, platform, or event record for risk assessment; by jointly analyzing visible and potential risk changes, it can discover potential risk migration, backflow, or re-aggregation behind apparent risk declines, improving the accuracy of identifying online food transaction risks such as cross-platform operations, entity group avoidance, and product family deformation and recurrence; based on identifying risk transmission relationships and risk mutation tendencies, it can simulate and compare the expected impact of different regulatory actions, thereby outputting more targeted recommended regulatory actions and tiered early warning results; through continuous correction of thresholds, edge weights, and action response parameters through handling feedback, it can continuously revise the judgment criteria according to actual regulatory results, reducing the risk of false alarms and omissions, and improving the long-term early warning stability and regulatory resource utilization efficiency. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of the modular structure of the system according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0057] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0058] This invention provides a multi-dimensional monitoring and early warning system and method for food safety online transaction risks. It is applicable to scenarios such as online food transaction supervision, platform food safety risk monitoring, cross-platform entity risk tracking, risk backflow identification after abnormal sampling, risk migration identification after handling, and tiered early warning of food safety risks. The system can be deployed on market supervision department servers, food safety supervision platforms, platform cooperation systems, or cross-platform data governance platforms, or it can operate as a risk analysis subsystem within these platforms.

[0059] In this embodiment, the food safety online transaction supervision system obtains multi-source supervision and observation data from the supervision platform, online transaction platform, complaint and reporting system, sampling inspection system, license management system, historical penalty system, and disposal feedback system. After desensitization, cleaning, time alignment, subject unification, platform coding, regional coding, and food category coding, it is converted into risk observation items. The risk observation items form multi-dimensional risk nodes according to region, food category, business entity or group of entities, online transaction platform, risk event type, and time window. Risk association edges are generated between multi-dimensional risk nodes based on entity association, product family association, platform migration, regional adjacency, complaint co-occurrence, sampling inspection association, license change association, and disposal impact association. The multi-dimensional risk nodes connected by the risk association edges form risk association edges. Risk nodes form associated risk domains; within these associated risk domains, the system calculates the observable risk state field, risk flow balance deviation, and potential risk state field; then, it uses candidate latent risk flux to explain the deviation transfer between source candidate nodes and target candidate nodes, confirming the latent risk transmission edge; further, it calculates the critical mutation degree based on the latent risk flux intensity, potential risk state intensity, the curvature of the corrected risk flow balance deviation, and the regulatory inhibition amount; finally, it simulates regulatory actions for latent risk transmission edges or risk critical mutation candidate nodes, outputs recommended regulatory actions and tiered early warning results, and corrects the risk flow balance deviation threshold, risk associated edge weight, latent risk transmission edge weight, critical mutation threshold, and regulatory action response parameters based on the handling feedback.

[0060] When this embodiment involves user complaints, user reviews, transaction behavior, platform behavior data, or other personal information, the relevant data processing has been agreed to by the user and complies with legal and regulatory requirements. The system retains only the fields necessary for food safety supervision, and performs anonymization, hashing, generalization, or access control processing on sensitive fields such as user identifiers, contact information, precise addresses, order numbers, and device identifiers. The system does not use attributes such as ethnicity, religion, gender, and age, which are not directly technically related to food safety risks, as the basis for risk judgment, to avoid the formation of discriminatory rules.

[0061] like Figure 1As shown, this is an embodiment of the present invention, which provides a multi-dimensional monitoring and early warning system for food safety online transaction risks, including an associated risk domain construction module, a dual-field deviation analysis module, an implicit transmission identification module, a regulatory action counterfactual early warning module, and an early warning memory correction module.

[0062] The associated risk domain construction module is used to convert multi-source regulatory observation data of food safety online transactions into risk observation items, combine risk observation items according to risk dimensions and map them into multi-dimensional risk nodes, generate risk association edges based on the multi-dimensional risk nodes with correlation relationships, and generate associated risk domains based on the risk association edges.

[0063] Multi-source regulatory observation data includes violation identification results, cross-platform entity association results, complaint and report data, sampling inspection results, platform transaction change data, food category data, license change data, historical penalty data, and handling feedback data. The system first performs field standardization, entity name unification, platform coding, regional coding, food category coding, time window division, and sensitive field anonymization on the above data before converting it into risk observation items.

[0064] Each risk observation item includes a combination of risk dimensions, risk attributes, risk contribution weight, data credibility, and regulatory observability.

[0065] The risk dimension combination includes region, food category, business entity or group of entities, online trading platform, risk event type, and time window. For example, a risk observation item can be represented as: region is a certain district of a certain city, food category is pre-packaged cooked food, business entity or group of entities is store A and its related entities, online trading platform is platform P, risk event type is complaints and reports, and time window is the first week of July 2026.

[0066] Risk attributes are used to record the source of risk and specific content types, including types of violations, types of complaints, types of non-compliance in spot checks, types of abnormal licenses and certificates, types of abnormal platform transactions, handling status, similarity identifiers for product families, and identifiers related to entities.

[0067] Risk contribution weights are used to represent the degree to which risk observations contribute to the calculation of food safety risks, based on the severity of the risk event. Data credibility Freshness of time and regulatory effectiveness Confirmed. In practical implementation, the above four indicators can be normalized to 0 to 1 first, and then calculated according to the following formula: ;in, Indicates risk observation item Risk contribution weight; to This represents the weighting coefficient, with non-negative values, and can be set to sum to 1. The severity of a risk event can be determined based on its type; for example, events involving substandard samples, administrative penalties, and high-risk food categories have higher severity. Data reliability can be determined based on the data source; for example, official sampling inspections, administrative penalties, and official platform handling records have higher reliability. Time freshness can be determined using an exponential decay method, such as... , This indicates the time interval between the generation of the risk observation and the current time window. This is the attenuation coefficient. Regulatory effectiveness is used to indicate whether the risk observation item has been verified and confirmed by regulatory authorities, online trading platforms, or humans.

[0068] Regulatory observability is used to represent the degree to which a corresponding risk observation item can be directly observed through complaints and reports, spot check results, platform reports, manual review, license verification, or historical penalty records. It can be determined based on factors such as data source, availability of direct evidence, whether it has been reviewed, and whether it has been confirmed by the platform or regulatory authorities, and is normalized to 0 to 1. For example, the regulatory observability of unqualified spot checks, administrative penalties, abnormal licenses, and manually verified confirmation records is relatively high; while the regulatory observability of abnormal complaints, abnormal transactions, suspected recurrence of product groups, and suspected association of entity groups is relatively low.

[0069] The associated risk domain construction module includes a risk observation hierarchical unit, a node mapping unit, a node attribute aggregation unit, an associated edge generation unit, and an associated domain formation unit.

[0070] The risk observation stratification unit is used to perform two-dimensional stratification of risk observation items. The two dimensions of the stratification are risk contribution weight and regulatory observability. Risk observation items with a risk contribution weight not lower than the risk contribution threshold and a regulatory observability lower than the observability threshold are marked as low-observability, high-contribution items. Risk observation items with a regulatory observability not lower than the observability threshold are marked as directly observable items. The risk contribution threshold and observability threshold can be determined based on historical data quantiles, expert rules, or training set statistics. For example, the risk contribution threshold can be taken as the 75th quantile of historical risk contribution weights, and the observability threshold can be taken as the 50th quantile of historical regulatory observability. When there are no historical samples, regulatory personnel can set initial experience thresholds, which are then updated by the early warning memory correction module based on the handling feedback after the system is running.

[0071] The node mapping unit is used to map risk observations, after two-dimensional hierarchical processing, to multi-dimensional risk nodes according to risk dimension combinations. It retains the two-dimensional hierarchical labels of each risk observation and generates a node index to record the risk dimension combinations of the corresponding multi-dimensional risk nodes. If multiple risk observations have the same risk dimension combination, or the same combination after subject merging or commodity family merging, they are mapped to the same multi-dimensional risk node. If a risk observation has multiple relationships, risk association edges of the corresponding association type can be generated for each multi-dimensional risk node after the risk observation is mapped to the multi-dimensional risk node.

[0072] The node attribute aggregation unit is used to determine the node risk representation value of a multidimensional risk node based on the risk contribution weights, data reliability, and time freshness of the multiple risk observations when multiple risk observations are mapped to the same multidimensional risk node. Specifically, the following formula can be used: ; Represents multidimensional risk nodes In the time window The node risk characterization value; This represents the set of risk observations mapped to this multidimensional risk node and falling within the time window t. The node risk characterization value is used to determine the domain formation priority of the multidimensional risk node in the construction of associated risk domains; the higher the node risk characterization value, the higher the domain formation priority.

[0073] The association edge generation unit is used to generate risk association edges between multi-dimensional risk nodes based on at least two of the following: subject association, product family association, platform migration, regional adjacency, complaint co-occurrence, sampling inspection association, license change association, and disposal impact association. It also determines the association type and initial association weight of the risk association edges. Risk association edges can be directed or undirected; for associations with a clear temporal order, such as platform migration and disposal impact, directed edges are preferred.

[0074] The initial association weight for entity association can be determined based on the consistency or similarity of unified social credit code, operator identity, contact information hash value, payment account, shipping address, IP address, device fingerprint, and store administrator information. The initial association weight for product family association can be determined based on the similarity of product name, food category, brand, specifications, packaging image, main image similarity, details page keywords, sales entry point, and ingredient information. The initial association weight for platform migration association can be determined based on the time interval between the same entity group or product family appearing on different platforms, product similarity, and the relationship before and after the action. The initial association weight for action-related impact association can be determined based on the action target, action time, recurrence point after action, penalty records, or rectification feedback. All of the above initial association weights can be normalized to 0 to 1.

[0075] The associated domain unit is used to determine the starting multidimensional risk node based on the domain priority, and then expands along the risk association edge from the starting multidimensional risk node in descending order of initial association weight, merging multidimensional risk nodes that meet the corresponding associated risk domain type into an associated risk domain. To avoid excessive expansion of associated risk domains, constraints such as minimum initial association weight, maximum expansion layer, and maximum number of nodes can be set. For example, when the initial association weight is lower than 0.3, expansion along the risk association edge stops; when the expansion layer exceeds 3 layers, expansion stops.

[0076] The types of associated risk domains include those related to the same region and product category, the same entity group, the same product family, the same platform, and the same disposal target. The same multidimensional risk node can enter multiple associated risk domains. For example, if a business entity is disposed of on platform A, and its associated entities sell similar products on platform B, then the corresponding node can simultaneously enter the same entity group, same product family, and same disposal target risk domains.

[0077] The dual-field bias analysis module is used to generate an observable risk state field within the associated risk domain, calculate the risk flow balance bias, and generate a potential risk state field from the risk flow balance bias. The observable risk state field represents the risk state that can be directly observed through complaints, spot checks, platform reports, manual review, certificate verification, or historical penalty records. The potential risk state field represents the potential risk state inferred through low-observable-high-contribution items, observation gaps, risk-related edges, and risk flow balance bias.

[0078] The dual-field deviation analysis module includes an observable state generation unit, an equilibrium deviation calculation unit, a potential anomaly labeling unit, and a potential state generation unit.

[0079] The observable state generation unit is used to perform weighted aggregation of directly observable items mapped to the same multidimensional risk node, obtaining the observable risk quality, and organizing it into an observable risk state field according to the node index and time window. The observable risk quality can be calculated using the following formula: ; Represents multidimensional risk nodes In the time window Observable risk quality; This indicates mapping to multidimensional risk nodes. Furthermore, it belongs to the set of risk observation items that are directly observable. If a multidimensional risk node has no directly observable items within a certain time window, its observable risk quality can be zero or the attenuation value of the observable risk quality in the previous time window. In this embodiment, it is preferable to take zero to avoid low observable risks being incorrectly included in the observable risk state field. The observable risk state field is formed by organizing the observable risk quality of each multidimensional risk node within the associated risk domain according to the node index and time window, and can be represented as a node-time matrix, a tensor, or a graph state with time attributes.

[0080] The balance deviation calculation unit is used to calculate the risk flow balance deviation of each multidimensional risk node based on the observable risk quality, identified risk flux, new risk sources, and regulatory restraint within the same associated risk domain. The formula is as follows:

[0081] ;

[0082] In the formula, For multi-dimensional risk nodes In the time window Risk transfer balance deviation; , Representing multidimensional risk nodes In the time window and time window Observable risk quality; This indicates multi-dimensional risk nodes within the same associated risk domain. The risk is transmitted to the set of neighboring nodes. express Any incoming neighbor node in the process; Indicates within the time window The risk-related edge flows into the neighbor node. Transmitted to multi-dimensional risk nodes Identified risk fluxes; This indicates that multi-dimensional risk nodes are received within the same associated risk domain. The set of outflowing neighbor nodes that transmit the risk. express Any outgoing neighbor node in the list; Indicates within the time window Risk-related edges consist of multi-dimensional risk nodes. Pass to outgoing neighbor nodes Identified risk fluxes; Represents multidimensional risk nodes In the time window New sources of risk; Represents multidimensional risk nodes In the time window The amount of regulatory restraint.

[0083] New risk sources can be weighted by the sum of newly added complaints and reports, abnormal spot checks, illegal identification results, abnormal transaction growth, abnormal licenses, and new platform handling records within the current time window. Regulatory suppression can be weighted by the intensity of regulatory actions such as targeted spot checks, manual review, platform rectification, product delisting, administrative penalties, and license verification. In practice, observable risk quality, identified risk flux, new risk sources, and regulatory suppression are all normalized to the same risk range, such as 0-1 or 0-100. Identified risk flux represents the amount of risk transmitted along risk-related edges, and its dimension is consistent with observable risk quality. New risk sources represent the amount of externally added risk, and its dimension is consistent with observable risk quality. Regulatory suppression represents the reduction in risk amount caused by regulatory actions, and its dimension is consistent with observable risk quality. Therefore, both sides of the formula are normalized risk quantities.

[0084] The identified risk flux can be determined based on the initial association weights of risk-related edges, the association type of risk-related edges, observable changes in risk quality within adjacent time windows, records of the impact of actions taken, platform migration records, commodity family recurrence records, or historical risk transmission samples. As a specific implementation method, the following calculation can be used: ; Indicates the flow into neighboring nodes In the time window Observable risk quality; Indicates flow from neighboring nodes Pointing to risk nodes The initial association weight of the risk-related edges; Indicates the correlation type coefficient; This represents the time decay or transmission reliability coefficient. This formula is an optional implementation method, and in actual implementation, it can also be adjusted in combination with propagation delay, data credibility, or the intensity of the impact of the action.

[0085] when When it approaches zero, it means that the current risk change can be explained by known risk inflows, outflows, new sources, and regulatory restraint; when A value significantly greater than zero indicates an increased risk to the node that cannot be explained by known inflows, potentially indicating unidentified risk inflows, potential exposures, or external risk backflows; when When the value is significantly less than zero, it indicates that there is a risk of decline in the node that cannot be explained by known outflows or regulatory restraints, and there may be unidentified risk outflows, hidden migrations, or transfers outside of regulation.

[0086] The potential anomaly marking unit is used to mark multidimensional risk nodes whose absolute value of risk flow balance deviation is greater than the risk flow balance deviation threshold as potential anomaly nodes. The risk flow balance deviation threshold can be set according to food category, region, platform, or risk event type, or it can be dynamically updated based on historical false alarm rate and confirmation rate.

[0087] The latent state generation unit generates an observation gap based on the weighted difference between low-observable high-contribution items and directly observable items mapped to the same potential anomaly node. Specifically, the observation gap is obtained by subtracting the weighted aggregate value of the directly observable items from the weighted aggregate value of the low-observable high-contribution items; if no directly observable items exist, the weighted aggregate value of the directly observable items is zero; if no low-observable high-contribution items exist, the observation gap is zero or no latent state is generated for that node. Combining the observation gap, risk flow balance deviation, association type of risk-related edges, and initial association weights, a potential risk state field is generated. The potential risk state field includes the potential risk quality, potential migration tendency, potential backflow tendency, and potential risk state intensity corresponding to the multi-dimensional risk node. Among them, the potential risk quality is used to represent the risk quality of potential abnormal nodes that are not fully observed by direct regulatory data but are reflected by observation gaps and risk flow balance deviations, and is determined by weighting the absolute values ​​of observation gaps and risk flow balance deviations; the potential migration tendency is used to represent the tendency of the node's potential risk to migrate to other nodes, and is determined by weighting the initial association weights of platform migration association, regional adjacency association, and cross-platform recurrence association of commodity families; the potential return tendency is used to represent the association strength between the node and the dealt object, subject group recurrence, commodity family recurrence, or reappearance after disposal, and is determined by disposal impact association, subject association, commodity family association, and post-disposal time interval; the potential risk state strength is used to represent the comprehensive strength of the node's potential risk state, which can be obtained by normalizing the potential risk quality, potential migration tendency, and potential return tendency to 0 to 1, and multiplying the potential risk quality by the tendency amplification factor. The tendency amplification factor increases with the increase of potential migration tendency and potential return tendency, and is not less than 1. For example, the tendency amplification factor can be determined based on the positive zero portion of the potential migration tendency and the positive zero portion of the potential return tendency.

[0088] When the association type of the risk-related edge is platform migration or regional adjacency, the observation gap and risk flow balance deviation can be preferentially applied to potential migration tendencies; when the association type is disposal impact association or entity association, it can be preferentially applied to potential return flow tendencies; when the association type is complaint co-occurrence, sampling association, or commodity family association, it can be preferentially applied to potential risk quality. The specific allocation ratio can be determined according to the initial association weight of the corresponding risk-related edge.

[0089] The implicit risk transmission identification module is used to identify implicit risk transmission edges based on the observable risk state field, the potential risk state field, and the risk flow balance deviation, and to determine candidate nodes for critical risk abrupt changes based on the implicit risk flux intensity and the potential risk state intensity. This module is used to identify situations where the apparent risk decreases but the risk transfers to related entities, platforms, commodity families, or regions.

[0090] The implicit transmission identification module includes a deviation node pairing unit, a candidate flux verification unit, an implicit transmission confirmation unit, and a critical mutation identification unit.

[0091] The deviation node pairing unit is used to identify potential abnormal nodes with a risk flow balance deviation of less than zero as source candidate nodes and potential abnormal nodes with a risk flow balance deviation of greater than zero as target candidate nodes within the same associated risk domain and the same time window. It generates candidate transfer node pairs when a risk-related edge exists between the source and target candidate nodes. Only source and target candidate nodes within the same associated risk domain and with a risk-related edge are paired to avoid mismatches between nodes without business or data connections.

[0092] The candidate flux verification unit is used to determine the candidate latent risk fluxes for candidate transit node pairs. After including the candidate latent risk fluxes in the outflow items of the source candidate node and the inflow items of the target candidate node, it recalculates the corrected risk flow balance deviations of the source and target candidate nodes. Including the outflow items of the source candidate node means that the candidate latent risk fluxes are treated as additional outflow risk fluxes in the calculation of the risk flow balance deviation of the source candidate node; including the inflow items of the target candidate node means that the candidate latent risk fluxes are treated as additional inflow risk fluxes in the calculation of the risk flow balance deviation of the target candidate node.

[0093] The formula for calculating candidate latent risk flux is:

[0094] ;

[0095] In the formula, Indicates within the time window Internal source candidate nodes Point to the target candidate node Candidate latent risk fluxes; as source candidate nodes Node identifier, For target candidate nodes Node identifier; For the source candidate node set The node identifier of any source candidate node in the dataset; For the target candidate node set The node identifier of any target candidate node in the dataset; , Representing the source candidate nodes respectively Target candidate nodes In the time window Risk transfer balance deviation; Indicates within the time window Internal source candidate nodes Point to the target candidate node Candidate flux allocation weights, Indicates within the time window Internal source candidate nodes The node identifier is The candidate flux allocation weights for the target candidate nodes. Indicates within the time window The internal node is identified as The source candidate node points to the target candidate node Candidate flux allocation weights; This indicates a positive number used to avoid a denominator of zero, which can be taken as... ;

[0096] in, ; Indicates the candidate node for the connection source. With target candidate nodes The initial association weight of the risk-related edges; Indicates within the time window Internally, based on the association type of the risk-related edges, from the source candidate nodes... The directional tendency value selected from the potential migration tendency and potential return tendency; Indicates the target candidate node In the time window Potential risks to quality.

[0097] If the association type of the risk-related edge is platform migration or regional adjacency, the directional tendency value can be selected from the potential migration tendency of the source candidate node; if the association type is disposal impact association or subject association, the directional tendency value can be selected from the potential return flow tendency of the source candidate node; if the association type meets multiple conditions at the same time, the larger value or weighted value among the corresponding tendency values ​​can be taken.

[0098] The implicit risk transmission confirmation unit is used to determine the candidate implicit risk flux as a confirmed implicit risk flux when the absolute value of the risk flow balance deviation after correction of both the source candidate node and the target candidate node decreases, and the sum of the absolute value reductions is not lower than the deviation improvement threshold. It also determines the directed association between candidate transmission node pairs as implicit risk transmission edges and determines the implicit risk flux intensity of the corresponding multi-dimensional risk node based on the confirmed implicit risk flux. The deviation improvement threshold can be determined based on the statistical quantile of the sum of the absolute value reductions of the deviations of the source candidate node and the target candidate node in historical confirmed implicit risk transmission samples, or it can be set as a fixed threshold by regulators. In the absence of historical samples, an empirical threshold can be used, which is then updated by the early warning memory correction module after the accumulation of feedback from handling actions.

[0099] The intensity of latent risk flux can be determined based on the sum of confirmed latent risk fluxes at the corresponding multidimensional risk nodes for both inflows and outflows. Specifically, the intensity of latent risk flux at a node can be the sum of confirmed latent risk fluxes at the node as a source node, the sum of confirmed latent risk fluxes at the node as a target node, or a weighted sum of the two. The higher the intensity of latent risk flux, the greater the degree to which the node participates in the transmission of latent risks within the current time window.

[0100] The critical mutation identification unit is used to identify potential risk state fields, latent risk flux intensity, the curvature of the corrected risk flow balance deviation, and regulatory inhibition levels. Determine the critical mutation degree, and identify multidimensional risk nodes with a critical mutation degree greater than the critical mutation threshold as candidate nodes for critical mutation.

[0101] In one embodiment, the critical mutation identification unit determines the curvature of the corrected risk flow balance deviation based on the absolute value of the corrected risk flow balance deviation for the same multidimensional risk node within three consecutive time windows. Specifically, the formula for calculating the curvature is: ; Represents multidimensional risk nodes In the time window The curvature of the corrected risk flow balance deviation; , , Representing multidimensional risk nodes The risk flow balance deviation after correction over three consecutive time windows. The curvature of change is used to characterize the degree of bending of the trend of the absolute value of the corrected risk flow balance deviation within consecutive time windows; when... When the value is greater than zero, it indicates that the absolute value of the deviation between two consecutive time windows is increasing; if the absolute value of the deviation in the current time window is greater than the absolute value of the deviation in the previous time window, it indicates that the absolute value of the deviation is accelerating; if the absolute value of the deviation in the current time window is less than the absolute value of the deviation in the previous time window, it indicates that the rate of decrease of the absolute value of the deviation is slowing down.

[0102] The product of the potential risk state intensity and the latent risk flux intensity is determined as the basic mutation quantity. The potential risk state intensity is provided by the potential risk state field, and the latent risk flux intensity is provided by the latent transmission confirmation unit. The curvature adjustment coefficient is determined based on the portion of the curvature change that is greater than zero, and the inhibition adjustment coefficient is determined based on the regulatory inhibition amount, with both the curvature adjustment coefficient and the inhibition adjustment coefficient being no less than 1. The curvature adjustment coefficient is used to characterize the degree of curvature in the trend of the absolute value change of the deviation, and is not used alone to determine whether the absolute value of the deviation is expanding. The candidate nodes for risk critical mutation are still determined jointly based on the potential risk state intensity, the latent risk flux intensity, the curvature change, and the regulatory inhibition amount.

[0103] The formula for calculating the critical mutation degree is: ;in, Represents multidimensional risk nodes In the time window The critical degree of abrupt change; Indicates the intensity of potential risk status; This represents the intensity of latent risk flux; all parameters above are normalized values, therefore the critical mutation degree is a dimensionless risk indicator.

[0104] The critical mutation threshold can be set according to food category, region, platform, or risk event type, or it can be adjusted by the early warning memory correction module based on the handling feedback.

[0105] The counterfactual warning module for regulatory actions is used to apply simulated regulatory actions to hidden risk transmission edges or candidate nodes of risk critical mutations, simulate the dual-field response of the observable risk state field and the potential risk state field, and determine the recommended regulatory actions and hierarchical warning results accordingly.

[0106] The regulatory action counterfactual warning module includes an action generation unit, a dual-field response simulation unit, an action benefit determination unit, and an action screening and warning unit.

[0107] The action generation unit is used to generate simulated regulatory actions for hidden risk transmission edges or candidate nodes of risk critical mutations. Simulated regulatory actions include targeted sampling inspections, platform cooperation investigations, license verification, entity relationship penetration, product family tracking, and cross-regional cooperation investigations. Targeted sampling inspections are used to verify the actual risks of food categories or product families; platform cooperation investigations are used to request online trading platforms to provide information on entities, products, transactions, and disposal; license verification is used to verify the license status of operating entities; entity relationship penetration is used to verify the connections between operating entities or groups of entities; product family tracking is used to track the reappearance of disposed products after changes in name, image, category, or sales entry point; and cross-regional cooperation investigations are used to collaboratively verify risks between operating locations, shipping locations, complaint locations, or sampling locations.

[0108] The dual-field response simulation unit is used to simulate the observable risk state field and potential risk state field after executing each simulated regulatory action within the corresponding associated risk domain, and to obtain the critical abrupt change reduction amount. The decline in the intensity of potential risk status The decline in the intensity of hidden risk flux Regulatory costs and risk transfer penalty amount .

[0109] Different simulated regulatory actions can be applied to both fields using preset response parameters. These response parameters represent the magnitude of an action's impact on observable risk quality, potential risk intensity, latent risk flux intensity, regulatory inhibition, and risk transfer penalties. Different regulatory actions can have different response parameters. For example, the response parameters for targeted sampling primarily affect regulatory observability and regulatory inhibition; the response parameters for platform cooperation actions primarily affect platform migration-related edge weights and latent risk flux intensity; and the response parameters for commodity family tracking actions primarily affect commodity family-related edge weights and potential repatriation tendency.

[0110] Risk transfer penalty amount The risk transfer penalty can be determined based on the increase in the intensity of potential risk status and the increase in the intensity of latent risk flux of non-target nodes within the same associated risk domain after the simulated regulatory action is executed. In practice, the increase in the intensity of potential risk status and the increase in the intensity of latent risk flux of non-target nodes can be normalized and then weighted and summed to obtain the risk transfer penalty. If the simulated regulatory action only reduces the observable risk quality of the targeted node but increases the intensity of potential risk status or latent risk flux of associated nodes, then the corresponding risk transfer penalty will increase.

[0111] The action revenue determination unit is used to determine the simulated regulatory action according to the following formula. Action benefits :

[0112] ;

[0113] Critical abrupt change reduction Determined by the difference in critical abruptness before and after the action; the decrease in the intensity of the potential risk state. Determined by the difference in the intensity of potential risk states before and after the action is executed; the decrease in the intensity of latent risk flux. Determined by the difference in the intensity of implicit risk flux before and after the action is executed. Regulatory costs. This can include sampling costs, manual review costs, platform cooperation costs, cross-regional collaboration costs, and time costs. (Weight) to The optimal regulatory action can be determined based on the allocation of regulatory resources, historical outcomes, expert experience, or training samples. A greater benefit indicates a better simulated regulatory action; the decrease in critical abruptness, the decrease in the intensity of potential risk states, and the decrease in the intensity of latent risk flux are positive benefit items; regulatory costs and the amount of risk transfer penalties are negative penalty items.

[0114] The action screening and early warning unit identifies simulated regulatory actions with the highest risk transfer penalty (below the penalty threshold) as recommended regulatory actions. It then generates tiered early warning results based on candidate nodes for risk critical mutations, implicit risk transmission edges, and recommended regulatory actions. These tiered early warning results can include observable high-risk warnings, potential risk exposure warnings, implicit risk transmission warnings, risk critical mutation warnings, and disposal repatriation warnings. Each warning result can include an early warning number, associated risk domain type, risk dimension combination, candidate nodes for risk critical mutations, implicit risk transmission edges, recommended regulatory action, action benefit, a summary of evidence sources, and a feedback collection entry point.

[0115] After the system outputs tiered early warning results, regulatory personnel or platform staff can take actual action based on the recommended regulatory actions and send feedback back to the system. This feedback includes whether the recommended regulatory actions were executed, whether the risk decreased after the action, whether genuine illegal risks were discovered, whether risk transfer occurred, whether there were any false alarms, the cost of the action, and the time taken to complete the action.

[0116] The early warning memory correction module is used to correct the risk flow balance deviation threshold, risk association edge weight, implicit risk transmission edge weight, critical mutation threshold, and regulatory action response parameters based on the handling feedback. If an implicit risk transmission edge is confirmed to exist through actual handling, the weight of that implicit risk transmission edge is increased; if a risk association edge triggers multiple effective early warnings, the weight of the corresponding risk association edge is increased; if a certain type of risk critical mutation candidate node is repeatedly confirmed as a false alarm, the critical mutation threshold for the corresponding food category, region, platform, or risk event type is appropriately increased; if a simulated regulatory action can reduce the critical mutation degree without causing risk transfer after actual handling, the response parameters of that simulated regulatory action are increased; if risk transfer or backflow occurs after actual handling, the weight corresponding to the risk transfer penalty is increased or the response parameters of the corresponding action are decreased.

[0117] For any parameter to be corrected, an incremental update method can be used. The new parameter value equals the original parameter value plus the product of the learning step size and the feedback bias, and the updated parameter is limited to a preset value range. The feedback bias can be determined by the difference between the actual risk reduction and the simulated risk reduction. If the actual response is better than the simulated result, the corresponding regulatory action response parameter is increased; if the actual response is worse than the simulated result or risk transfer occurs, the corresponding regulatory action response parameter is decreased and the weighting of the risk transfer penalty is increased.

[0118] For example, after a cooked food product on platform A fails a random inspection and is disposed of, the observable risk quality of the corresponding multidimensional risk node on platform A decreases. However, the system finds an increase in low-observable high-contribution items of the same subject group or similar product families on platform B. The system constructs a risk domain associated with the same product family that includes nodes on platform A and platform B. The calculated risk flow balance deviation of node A is less than zero, while that of node B is greater than zero. There is a product family association and a platform migration association between the two, thus generating candidate transmission node pairs. The system calculates the candidate latent risk flux and recalculates the deviation after including it in the outflow items of node A and the inflow items of node B. If the absolute value of the deviation decreases after correction at both ends, and the sum of the decreases exceeds the deviation improvement threshold, then the latent risk transmission edge from platform A to platform B is confirmed. Subsequently, the system calculates the critical abrupt change degree based on the potential risk state strength, latent risk flux strength, and change curvature of node B. If the critical mutation degree exceeds the critical mutation threshold, a risk critical mutation warning is generated. The recommended regulatory action is selected from simulated regulatory actions such as platform investigation, product family tracking, and entity relationship penetration, choosing the action with the highest benefit and a risk transfer penalty amount below the penalty threshold. After the action is completed, if regulatory feedback confirms that Platform B has similar risky products, the system increases the weight of the product family association edge, platform migration association edge, and corresponding implicit risk transmission edge; if the feedback indicates a false alarm, the weight of the corresponding edge is decreased or the corresponding critical mutation threshold is increased.

[0119] like Figure 2 As shown, another embodiment of the present invention provides a multi-dimensional monitoring and early warning method for risks in online food safety transactions, applied to the aforementioned multi-dimensional monitoring and early warning system for risks in online food safety transactions. The method includes:

[0120] S1: Convert multi-source regulatory observation data of food safety online transactions into risk observation items, map the risk observation items into multi-dimensional risk nodes according to risk dimensions, generate risk correlation edges based on the multi-dimensional risk nodes with correlation relationships, and generate correlation risk domains based on the risk correlation edges.

[0121] S2: Generate an observable risk state field within the associated risk domain, calculate the risk flow balance deviation, and generate a potential risk state field from the risk flow balance deviation.

[0122] S3: Based on the sign of the risk flow balance deviation, form source candidate nodes and target candidate nodes, calculate candidate latent risk flux, and confirm the latent risk transmission edge based on the corrected deviation improvement result;

[0123] S4: Calculate the critical mutation degree based on the potential risk state intensity, the latent risk flux intensity, the curvature of the change in the corrected risk flow balance deviation, and the amount of regulatory inhibition, and determine the candidate nodes for critical risk mutation.

[0124] S5: Generate simulated regulatory actions for hidden risk transmission edges or candidate nodes of risk critical mutation, calculate the action benefits, and determine recommended regulatory actions and tiered early warning results;

[0125] S6: Adjust the risk flow balance deviation threshold, risk association edge weight, implicit risk transmission edge weight, critical mutation threshold, and regulatory action response parameters based on the feedback from the handling.

[0126] Through the above implementation methods, this invention can organize multi-source regulatory data of online food safety transactions into associated risk domains, calculate the risk flow balance deviation between observable and potential risks, identify hidden risk transmissions that are insufficiently covered by conventional regulatory data, and provide regulatory action suggestions through counterfactual simulation. Compared with early warning methods that rely solely on a single risk score or ordinary multi-source data fusion, this invention can more effectively identify online food safety risks such as post-disposal risk transfer, platform migration, recurrence after changes in product families, and risk return from related entities.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-dimensional monitoring and early warning system for food safety online transaction risks, characterized in that, The system includes: The associated risk domain construction module is used to convert multi-source regulatory observation data of food safety online transactions into risk observation items, map the risk observation items into multi-dimensional risk nodes according to risk dimension combinations, generate risk association edges based on the multi-dimensional risk nodes with correlation relationships, and generate associated risk domains based on the risk association edges; the risk dimension combinations include region, food category, business entity or group of entities, online transaction platform, risk event type, and time window; The dual-field deviation analysis module is used to generate an observable risk state field within the associated risk domain, calculate the risk flow balance deviation, and generate a potential risk state field from the risk flow balance deviation. The latent transmission identification module is used to identify latent risk transmission edges based on the observable risk state field, the potential risk state field, and the risk flow balance deviation, and to determine candidate nodes for risk critical mutation based on the latent risk flux intensity and the potential risk state intensity. The counterfactual warning module for regulatory actions is used to apply simulated regulatory actions to the hidden risk transmission edge or candidate nodes of risk critical mutation, simulate the dual-field response of the observable risk state field and the potential risk state field, and determine the recommended regulatory actions and hierarchical warning results accordingly. The early warning memory correction module is used to correct the risk flow balance deviation threshold, risk association edge weight, implicit risk transmission edge weight, critical mutation threshold, and regulatory action response parameters based on the handling feedback.

2. The multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 1, characterized in that, The risk observation items include risk dimension combinations, risk attributes, risk contribution weights, data credibility, and regulatory observability; the risk contribution weights are determined based on the severity of the risk event, data credibility, time freshness, and regulatory effectiveness; the regulatory observability is used to indicate the degree to which the corresponding risk observation item can be directly observed by complaints and reports, spot check results, platform reports, manual review, certificate verification, or historical penalty records. The types of associated risk domains include: associated risk domains of the same region and product category, associated risk domains of the same subject group, associated risk domains of the same product family, associated risk domains of the same platform, and associated risk domains of the same disposal object.

3. The multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 2, characterized in that, The associated risk domain construction module includes: The risk observation stratification unit is used to perform two-dimensional stratification of risk observation items. Risk observation items with a risk contribution weight not lower than the risk contribution threshold and a regulatory observability lower than the observability threshold are marked as low observability high contribution items, and risk observation items with a regulatory observability not lower than the observability threshold are marked as directly observable items. The node mapping unit is used to map the risk observations after two-dimensional hierarchical classification to multi-dimensional risk nodes according to the risk dimension combination, retain the two-dimensional hierarchical label of each risk observation, and generate a node index for recording the risk dimension combination. The node attribute aggregation unit is used to determine the node risk characterization value of the multidimensional risk node based on the risk contribution weight, data credibility, and time freshness of the multiple risk observations when multiple risk observations are mapped to the same multidimensional risk node, and to determine the domain priority of the multidimensional risk node in participating in the construction of associated risk domains based on the node risk characterization value. The association edge generation unit is used to generate risk association edges between multidimensional risk nodes based on at least two of the following: subject association, product family association, platform migration, regional adjacency, complaint co-occurrence, sampling inspection association, certificate change association, and disposal impact association, and to determine the association type and initial association weight of the risk association edges. The association domain unit is used to determine the starting multidimensional risk node according to the domain priority, and to expand from the starting multidimensional risk node along the risk association edge in descending order of the initial association weight, and merge multidimensional risk nodes that meet the corresponding association risk domain type into an association risk domain.

4. The multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 3, characterized in that, The dual-field deviation analysis module includes: The observable state generation unit is used to perform weighted aggregation of directly observable items mapped to the same multidimensional risk node to obtain observable risk quality, and organize it into an observable risk state field according to the node index and time window. The balance deviation calculation unit is used to calculate the risk flow balance deviation of each multi-dimensional risk node based on the observable risk quality, identified risk flux, new risk sources, and regulatory inhibition amount of each multi-dimensional risk node within the same associated risk domain. The potential anomaly marking unit is used to mark multidimensional risk nodes whose absolute value of risk flow balance deviation is greater than the risk flow balance deviation threshold as potential anomaly nodes. The potential state generation unit is used to generate observation gaps based on the weighted differences between low-observable high-contribution items and directly observable items mapped to the same potential anomaly node, and to generate a potential risk state field by combining the risk flow balance deviation, the association type of risk association edges and the initial association weight; the potential risk state field includes the potential risk quality, potential migration tendency, potential backflow tendency and potential risk state intensity of the corresponding multi-dimensional risk node.

5. The multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 4, characterized in that, The formula for calculating the risk transfer balance deviation is: ； In the formula, For multi-dimensional risk nodes In the time window Risk transfer balance deviation; , Representing multidimensional risk nodes In the time window and time window Observable risk quality; This indicates multi-dimensional risk nodes within the same associated risk domain. The risk is transmitted to the set of neighboring nodes. express Any incoming neighbor node in the process; Indicates within the time window The risk-related edge flows into the neighbor node. Transmitted to multi-dimensional risk nodes Identified risk fluxes; This indicates that multi-dimensional risk nodes are received within the same associated risk domain. The set of outflowing neighbor nodes that transmit the risk. express Any outgoing neighbor node in the list; Indicates within the time window Risk-related edges consist of multi-dimensional risk nodes. Pass to outgoing neighbor nodes Identified risk fluxes; Represents multidimensional risk nodes In the time window New sources of risk; Represents multidimensional risk nodes In the time window The amount of regulatory restraint.

6. The multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 5, characterized in that, The implicit transmission identification module includes: The deviation node pairing unit is used to identify potential abnormal nodes with a risk flow balance deviation of less than zero as source candidate nodes and potential abnormal nodes with a risk flow balance deviation of greater than zero as target candidate nodes within the same associated risk domain and the same time window, and to generate candidate transmission node pairs when there is a risk association edge between the source candidate node and the target candidate node. The candidate flux verification unit is used to determine the candidate latent risk flux of the candidate transmission node pair, and after including the candidate latent risk flux in the outflow item of the source candidate node and the inflow item of the target candidate node, recalculate the corrected risk flow balance deviation between the source candidate node and the target candidate node. The implicit transmission confirmation unit is used to determine the candidate implicit risk flux as a confirmed implicit risk flux, determine the directed association between candidate transmission node pairs as implicit risk transmission edges, and determine the implicit risk flux intensity of the corresponding multidimensional risk node based on the confirmed implicit risk flux when the absolute values ​​of the risk flow balance deviations after the source candidate node and the target candidate node are both reduced and the sum of the absolute value reductions is not lower than the deviation improvement threshold. The critical mutation identification unit is used to determine the critical mutation degree based on the potential risk state field, the intensity of the latent risk flux, the curvature of the change in the corrected risk flow balance deviation, and the amount of regulatory inhibition, and to identify multidimensional risk nodes with a critical mutation degree greater than the critical mutation threshold as candidate nodes for risk critical mutation.

7. A multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 6, characterized in that, The formula for calculating the candidate latent risk flux is as follows: ； In the formula, Indicates within the time window Internal source candidate nodes Point to the target candidate node Candidate latent risk fluxes; as source candidate nodes Node identifier, For target candidate nodes Node identifier; For the source candidate node set The node identifier of any source candidate node in the dataset; For the target candidate node set The node identifier of any target candidate node in the dataset; , Representing the source candidate nodes respectively Target candidate nodes In the time window Risk transfer balance deviation; Indicates within the time window Internal source candidate nodes Point to the target candidate node Candidate flux allocation weights, Indicates within the time window Internal source candidate nodes The node identifier is The candidate flux allocation weights for the target candidate nodes. Indicates within the time window The internal node is identified as The source candidate node points to the target candidate node Candidate flux allocation weights; This indicates a positive number used to avoid a denominator of zero; in, ; Indicates the candidate node for the connection source. With target candidate nodes The initial association weight of the risk-related edges; Indicates within the time window Internally, based on the association type of the risk-related edges, from the source candidate nodes... The directional tendency value selected from the potential migration tendency and potential return tendency; Indicates the target candidate node In the time window Potential risks to quality.

8. The multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 7, characterized in that, The critical mutation identification unit determines the curvature of the change in the corrected risk flow balance deviation based on the absolute value of the corrected risk flow balance deviation of the same multidimensional risk node within three consecutive time windows. The product of the potential risk state intensity and the latent risk flux intensity is determined as the basic mutation amount; The curvature adjustment coefficient is determined based on the portion of the curvature that is greater than zero, and the inhibition adjustment coefficient is determined based on the amount of regulatory inhibition, wherein both the curvature adjustment coefficient and the inhibition adjustment coefficient are not less than 1; The critical mutation degree of the multidimensional risk node is obtained by dividing the product of the basic mutation amount and the curvature adjustment coefficient by the inhibition adjustment coefficient.

9. A multi-dimensional monitoring and early warning system for food safety online transaction risks according to claim 8, characterized in that, The counterfactual warning module for regulatory actions includes: The action generation unit is used to generate simulated regulatory actions for hidden risk transmission edges or candidate nodes of risk critical mutation. The simulated regulatory actions include targeted sampling inspection, platform cooperation investigation, certificate verification, entity relationship penetration, commodity family tracking and cross-regional cooperation investigation. The dual-field response simulation unit is used to simulate the observable risk state field and potential risk state field after executing each simulated regulatory action within the corresponding associated risk domain, and to obtain the critical abrupt change reduction amount. The amount of decline in the intensity of potential risk status The decline in the intensity of hidden risk flux Regulatory costs and risk transfer penalty amount ; Action revenue determination unit, used to determine simulated regulatory actions according to the following formula. Action benefits : ； In the formula, to These represent the weights of the corresponding items; The action screening and early warning unit is used to identify simulated regulatory actions with the greatest risk transfer penalty amount and the highest action benefit as recommended regulatory actions, and to generate hierarchical early warning results based on candidate nodes of risk critical mutation, implicit risk transmission edges and recommended regulatory actions.

10. A method for multi-dimensional monitoring and early warning of risks in online food safety transactions, characterized in that, The method, applied to a multi-dimensional monitoring and early warning system for food safety online transaction risks as described in any one of claims 1-9, comprises: Multi-source regulatory observation data of food safety online transactions are converted into risk observation items. The risk observation items are combined and mapped into multi-dimensional risk nodes according to risk dimensions. Risk correlation edges are generated based on the multi-dimensional risk nodes with correlation relationships, and correlation risk domains are generated based on the risk correlation edges. An observable risk state field is generated within the associated risk domain, the risk flow balance deviation is calculated, and a potential risk state field is generated from the risk flow balance deviation. The latent risk transmission edge is identified based on the observable risk state field, the potential risk state field, and the risk flow balance deviation, and the candidate nodes for risk critical mutation are determined based on the latent risk flux intensity and the potential risk state intensity. Simulated regulatory actions are applied to the hidden risk transmission edge or candidate nodes of risk critical mutation, simulating the dual-field response of the observable risk state field and the potential risk state field, and the recommended regulatory actions and tiered early warning results are determined accordingly. The risk transfer balance deviation threshold, risk association edge weight, implicit risk transmission edge weight, critical mutation threshold, and regulatory action response parameters are adjusted based on the feedback from the handling.

Images