Market subject management method and system based on multi-source data fusion

Abstract

The application discloses a market subject management method and system based on multi-source data fusion, relates to the technical field of intelligent supervision, and comprises the following steps: based on a plurality of data holders, feature data of a market subject risk state is generated locally; the plurality of data holders interact through a privacy protection cooperation mechanism, participate in the construction process of cross-domain risk cognition consensus, and form a cross-domain risk consensus discriminator locally; the cross-domain risk consensus discriminator is used to filter high-risk market subjects from the feature data of the market subject risk state, and a high-risk market subject list is formed; for each market subject in the high-risk market subject list, an actual behavior track is formed based on the feature data of the market subject risk state. The application realizes local risk cognition model iterative optimization through inspection feedback, and forms a process from risk accurate identification, deep attribution analysis to intelligent supervision intervention and self-sustaining optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent supervision technology, and in particular to a market entity management method and system based on multi-source data fusion. Background Technology

[0002] With the development of information technology, data-driven risk supervision of market entities has become an important trend. Existing technologies mainly utilize the market supervision and management departments' own data to build risk assessment models, such as using machine learning algorithms to score and classify market entities based on risk. More advanced solutions integrate publicly available or authorized data from external departments such as the judiciary and taxation, and improve the model's predictive capabilities through feature engineering. These methods all follow the traditional paradigm of centralized data processing and unified model training, and their effectiveness is highly dependent on the scope and quality of the data collection.

[0003] The current approach is limited by its analytical depth. While existing technologies can identify high-risk entities through multi-source data fusion, they are mainly based on statistical correlation analysis and cannot deeply reveal the specific causes of risks and effective intervention paths. Risk scoring models usually give a comprehensive score that cannot clearly point to the key behaviors or decision-making nodes that lead to risks, nor can they assess the possible effects of different regulatory measures. This results in regulatory actions still relying heavily on experience-based judgments. The chain from risk discovery to precise policy implementation has not been fully established, affecting the optimal allocation of regulatory resources and the further improvement of governance effectiveness. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a market entity management method based on multi-source data fusion to solve the problem in the prior art where risk warning and in-depth attribution and precise intervention decision-making are disconnected, resulting in the inability to achieve intelligent, refined and dynamic allocation of regulatory resources.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a market entity management method based on multi-source data fusion, which includes generating characteristic data of the risk status of market entities locally based on multiple data holders; The multiple data holders interact through a privacy protection collaboration mechanism to participate in the construction of a cross-domain risk awareness consensus and form a cross-domain risk consensus discriminator locally; Using a cross-domain risk consensus discriminator, high-risk market entities are screened out based on the characteristic data of the risk status of market entities, and a list of high-risk market entities is formed. For each market entity in the list of high-risk market entities, based on the actual behavioral trajectory formed by the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator, a virtual development path is constructed in reverse. The actual behavioral trajectory and the virtual development path are used to perform compliance deviation inference, locate the set of key decision nodes that lead to the deviation, and generate an attribution inference report. Based on the attribution and inference report, the list of high-risk market entities is classified and prioritized to generate a regulatory task queue. Assigning tasks from the regulatory task queue to inspection personnel to perform inspections and submit inspection reports triggers a new round of iterative optimization of cross-domain risk awareness consensus.

[0007] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the method involves generating characteristic data of the risk status of market entities locally based on multiple data holders, including the following steps: Based on the data processing of the market supervision and administration department, characteristic data of the market supervision and administration department is generated. Based on data processing of financial institutions, generate characteristic data of financial institutions; Based on data processing of public institutions, generate characteristic data of public institutions; The characteristic data of market entities' risk status are generated by aggregating characteristic data from market supervision and management departments, financial institutions, and public institutions.

[0008] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the multiple data holders interact through a privacy-protected collaborative mechanism to participate in the construction process of cross-domain risk awareness consensus, forming a cross-domain risk consensus discriminator locally, including the following steps: Based on the federated learning framework, multiple data holders interact through encrypted model parameter updates. Under the secure aggregation of the central coordinator, they jointly drive the iterative optimization of local risk perception model parameters and converge to form a collaborative discrimination benchmark with embedded cross-domain risk consensus. Based on the collaborative judgment benchmark with embedded cross-domain risk consensus, the risk cognition of market entities is independently iterated and evolved in the local environment using characteristic data of risk status. This forms a cognitive state evolution direction that reflects the local data pattern. The cognitive state evolution direction is submitted after privacy encryption. The central coordination node then securely integrates and aligns the encrypted multi-party cognitive evolution directions. Through multiple rounds of local cognitive evolution, privacy-encrypted submission, security fusion alignment, and cognitive state synchronization, the local risk cognitive model converges from its initial isolated state to a collaborative cognitive state that reflects cross-domain common risk patterns, thus obtaining a cross-domain risk consensus discriminator.

[0009] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the method involves: using a cross-domain risk consensus discriminator to screen high-risk market entities based on the characteristic data of their risk status, thereby forming a list of high-risk market entities, and includes the following steps: The characteristic data of the market supervision and management department is input into the cross-domain risk consensus discriminator, which outputs a risk attribute label that integrates the cross-domain risk consensus for each market entity. The risk attribute labels that incorporate cross-domain risk consensus are compared with preset risk thresholds. Market entities whose risk attribute labels that incorporate cross-domain risk consensus are higher than the preset risk thresholds are selected and compiled into a list of high-risk market entities.

[0010] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the method involves: for each market entity in the high-risk market entity list, constructing a virtual development path by reverse-engineering the actual behavioral trajectory formed from the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator, including the following steps: Based on the list of high-risk market entities, select one market entity as the current analysis object, extract the characteristic data of the risk status of the market entity corresponding to the current analysis object, and arrange them in time series to form the actual behavior trajectory of the current analysis object. By using a cross-domain risk consensus discriminator to analyze the actual behavioral trajectory of the current analysis object, the feature dimension that contributes the most to the risk attribute label of the current analysis object is identified as the key risk dimension. Based on key risk dimensions, at the starting point of the actual behavior trajectory of the current analysis object, match compliance reference entities with similar initial states in the same dimension in history; By integrating the behavioral patterns of compliance reference entities across key risk dimensions, a virtual development path representing the ideal direction of compliance is synthesized. By comparing the actual behavioral trajectory of the current analysis object with the virtual development path, the decision-making forks that cause the current analysis object to deviate from the compliance direction are revealed, and the virtual development path is constructed in reverse.

[0011] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the following steps are included: performing compliance deviation analysis between actual behavioral trajectories and virtual development paths to identify the set of key decision-making nodes leading to the deviation, and generating an attribution analysis report: Based on key risk dimensions, the actual behavior trajectory of the current analysis object is compared with the virtual development path over time to identify deviation events in which the actual behavior trajectory deviates from the virtual development path; Filter out deviation events that exceed a preset threshold, and for each deviation event, trace the triggering behavior that caused the deviation event. The triggering behavioral points corresponding to multiple deviation events are aggregated to form a set of key decision nodes, and an attribution inference report is generated.

[0012] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the method involves classifying and prioritizing a list of high-risk market entities based on an attribution analysis report to generate a regulatory task queue, including the following steps: Analyze the attribution inference report, extract the key decision node set information from the attribution inference report, conduct risk assessment for each market entity in the list of high-risk market entities, and generate the risk level of each market entity. Based on the aforementioned risk levels, market entities in the list of high-risk market entities are classified and placed into different regulatory categories. Within the regulatory categories, market entities are sorted according to their risk levels to generate a regulatory priority sequence. The regulatory categories and their corresponding regulatory priority sequences are then integrated to form a regulatory task queue.

[0013] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, the following steps are included: assigning tasks from the regulatory task queue to inspection personnel, performing inspections, and submitting inspection reports: Based on the regulatory category and order of tasks in the regulatory task queue, each task in the regulatory task queue is matched with an inspector with corresponding professional qualifications. Tasks in the regulatory task queue are assigned to the matched inspectors, and attribution and inference reports are provided to the inspectors. Based on the guidance of the attribution analysis report, conduct on-site inspections of the market entities corresponding to the assigned tasks and submit structured inspection reports.

[0014] As a preferred embodiment of the market entity management method based on multi-source data fusion described in this invention, triggering a new round of iterative optimization of cross-domain risk awareness consensus includes the following steps: The inspection results in the structured inspection report are linked to the characteristic data of the market entity's risk status. Update the local training sample set of the data holder using the characteristic data of the risk status of market entities with new verification labels; Based on the updated training sample set, the data holder restarts the process of evolving the local risk perception model and exchanging encrypted parameters. Through a new round of interaction and optimization, the parameters of the cross-domain risk consensus discriminator are iteratively updated to optimize the cross-domain risk perception consensus.

[0015] Secondly, the present invention provides a market entity management system based on multi-source data fusion, including a feature data module that generates feature data of the risk status of market entities locally based on multiple data holders. The module is constructed in which the multiple data holders interact through a privacy-preserving collaborative mechanism to participate in the construction process of cross-domain risk awareness consensus and form a cross-domain risk consensus discriminator locally. The screening module uses a cross-domain risk consensus discriminator to screen out high-risk market entities based on the characteristic data of the risk status of market entities, and forms a list of high-risk market entities. The deduction module, for each market entity in the list of high-risk market entities, constructs a virtual development path in reverse based on the actual behavioral trajectory formed by the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator. It then deduces the compliance deviation between the actual behavioral trajectory and the virtual development path, locates the set of key decision nodes that lead to the deviation, and generates an attribution deduction report. The sorting module classifies and prioritizes the list of high-risk market entities based on the attribution inference report, and generates a regulatory task queue. The optimization module assigns tasks from the regulatory task queue to inspection personnel, who then perform inspections and submit inspection reports, triggering a new round of iterative optimization of cross-domain risk awareness consensus.

[0016] The beneficial effects of this invention are as follows: By generating characteristic data of market entity risk status locally on multiple data holders and constructing a cross-domain risk consensus discriminator with the help of a privacy-protected collaborative mechanism, it achieves deep integration of multi-source heterogeneous data under the premise of ensuring data privacy and security, improving the accuracy and comprehensiveness of risk identification. The discriminator is used to screen out high-risk market entities, and for each high-risk entity, compliance deviation is deduced by comparing its actual behavior trajectory with the virtual compliance development path constructed in reverse based on key risk dimensions. This accurately locates the key decision nodes that lead to risks and generates attribution reports, solving the limitation of traditional methods that can only provide risk scores but cannot reveal the specific causes of risks and intervention paths. Based on the attribution reports, high-risk entities are classified, ranked, and a regulatory task queue is generated. The local risk cognition model is iteratively optimized through inspection feedback, forming a process from accurate risk identification and in-depth attribution analysis to intelligent regulatory intervention and self-continuous optimization. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of a market entity management method based on multi-source data fusion.

[0019] Figure 2 This is a schematic diagram of a market entity management system based on multi-source data fusion. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Reference Figures 1-2 This is one embodiment of the present invention, which provides a market entity management method based on multi-source data fusion, including the following steps: S1. Based on multiple data holders, generate characteristic data of the risk status of market entities locally.

[0024] S1.1. Based on the data processing of the market supervision and administration department, generate characteristic data of the market supervision and administration department.

[0025] Furthermore, the market supervision and administration department processes the data locally, identifying and extracting core elements related to the risks of market entities. For example, it extracts penalty types, penalty amounts, and penalty frequencies from administrative penalty records; analyzes complaint themes, resolution status, and recurrence cycles from complaint and report information; and calculates registered capital change rates, years of operation, and equity structure complexity from enterprise annual reports. The extracted elements are quantified into standardized values ​​through predefined conversion rules, such as mapping penalty types to severity weights and converting complaint resolution status into timeliness scores. All quantified values ​​are arranged according to a unified dimension to form characteristic data of the market supervision and administration department.

[0026] S1.2. Based on the data processing of financial institutions, generate characteristic data of financial institutions; Furthermore, financial institutions process their data locally, mining patterns that reflect abnormal financial behavior of market participants from transaction flows, account activities, and credit records. For example, they analyze the concentration of corporate account transaction frequency in specific time periods (such as late at night), the dispersion of fund inflows and outflows to identify suspected idle transactions, and the proportion of small-amount, high-frequency transfers in total transaction volume. These behavioral patterns are converted into numerical features using indicators commonly used in the financial risk control field, such as dispersion index, concentration ratio, and volatility coefficient, forming characteristic data for market supervision and management departments.

[0027] S1.3. Data processing based on public institutions: Data from public institutions is used to generate characteristic data for public institutions.

[0028] Furthermore, public utilities process their data locally, capturing anomalies in the operational status of market entities from continuous monitoring data such as energy consumption and resource usage. For example, analyzing electricity consumption data can identify whether the baseline load during non-working hours is abnormally high, which may indicate hidden production activities; whether the cyclical fluctuations in water consumption deviate significantly from the declared production scale; and whether sudden patterns in network traffic are inconsistent with normal office behavior. Through time series analysis and baseline comparison methods, consumption data is converted into characteristic values ​​such as load deviation, fluctuation stability, and peak anomaly index, forming a characteristic vector for public utilities.

[0029] S1.4. Collect characteristic data from market supervision and management departments, financial institutions, and public institutions to generate characteristic data on the risk status of market entities.

[0030] Furthermore, using a unified market entity identifier as the key, feature vectors from three independent data holders are aligned and associated. Each market entity ultimately corresponds to a multi-dimensional composite feature vector, which contains risk descriptions from three different levels: administrative enforcement, financial transactions, and physical operations. This is a key step in building the foundation for multi-source data fusion. Under the premise of physically dispersed data, through the standardization and logical association of the feature layer, a unified profile covering the multi-dimensional behavior of market entities in terms of supervision, finance, and physical aspects is conceptually constructed, thereby solving the technical challenge of multi-source heterogeneous data fusion at the source.

[0031] S2. Multiple data holders interact through a privacy protection collaboration mechanism to participate in the construction of cross-domain risk awareness consensus and form a cross-domain risk consensus discriminator locally.

[0032] S2.1 Based on the federated learning framework, multiple data holders interact through encrypted model parameter updates. Under the secure aggregation of the central coordinator, they jointly drive the iterative optimization of local risk perception model parameters and converge to form a collaborative discrimination benchmark with embedded cross-domain risk consensus.

[0033] Furthermore, each data holder, namely the market supervision and management department, financial institutions, and public utilities, independently initializes an initial risk perception model with the same structure in their respective computing environments. Each party uses locally generated characteristic data of the risk status of market entities to train its own initialized risk perception model. This training process generates model parameters optimized for local data. Each party calculates the difference between the model parameters before and after this round of training, i.e., the model parameter update amount, and applies a homomorphic encryption algorithm to encrypt the model parameter update amount to obtain the encrypted model parameter update amount. All encrypted model parameter update amounts are sent to the central coordinator.

[0034] Specifically, without decryption, the central coordinator uses a secure aggregation algorithm, such as additive aggregation based on secret sharing, to sum all encrypted model parameter updates, generating an aggregated encrypted global model parameter update. This encrypted global model parameter update is broadcast back to all data holders. Each party decrypts it using its corresponding decryption key to obtain the global model parameter update and applies it to update its local risk perception model parameters. This cycle of local training, encrypted upload, secure aggregation, and decryption update is repeated multiple times, causing the model parameters of each party to gradually converge from a scattered state to a central point. After sufficient iteration, the local risk perception model parameters of each party converge to a highly consistent state. This shared parameter set constitutes the collaborative discrimination benchmark for embedded cross-domain risk consensus.

[0035] It should be noted that by using encrypted model parameter update information as a medium, the direct exchange of original data or complete models is replaced. Mathematically, this is equivalent to joint training on a complete set of multi-source data. This breaks through the dilemma of being unable to effectively integrate knowledge under data privacy barriers, and realizes a privacy-secure collaborative learning paradigm where data remains stationary, the model moves, and knowledge is aggregated. This enables market supervision and management departments to obtain a powerful benchmark model that integrates multi-dimensional external risk knowledge.

[0036] S2.2 Based on the collaborative judgment benchmark of embedded cross-domain risk consensus, the risk cognition of market entities is independently iterated and evolved in the local environment using the characteristic data of the risk status of market entities. This forms a cognitive state evolution direction that reflects the local data pattern. The cognitive state evolution direction is submitted after privacy encryption processing. The central coordination node performs secure fusion and alignment of the encrypted multi-party cognitive evolution directions.

[0037] Furthermore, after obtaining the collaborative discrimination benchmark with embedded cross-domain risk consensus, all parties use the parameters of this benchmark model as the starting point for a new round of training, continuing to conduct additional training rounds using the latest local market entity risk status characteristic data. This process is called fine-tuning. During fine-tuning, the model parameters will be slightly adjusted based on new patterns or unique distributions in the local data. The resulting parameter changes are defined as the cognitive state evolution direction. For example, a financial institution may discover a new cash-out pattern associated with a specific industry that is not fully represented in the global consensus benchmark. Through fine-tuning, the financial institution's local model will learn this pattern, and its cognitive state evolution direction will include this new knowledge. All parties also homomorphically encrypt the cognitive state evolution direction, generating an encrypted cognitive state evolution direction and submitting it to the central coordination node.

[0038] Specifically, after receiving the encrypted cognitive state evolution directions from all parties, the central coordinating node performs secure fusion and alignment operations. Secure fusion also involves weighted averaging within the encrypted state. The alignment operation is achieved by introducing constraints. For example, a norm constraint can be imposed on the encrypted cognitive state evolution directions of all parties before aggregation, or a regularization term can be added to the loss function to penalize updates that deviate too much from the consensus benchmark, ensuring that the merged update direction is a beneficial supplement to the consensus benchmark rather than a destructive change.

[0039] It is important to explain the consensus evolution mechanism. It allows for the safe absorption of dynamically generated, localized, and fine-grained new knowledge from all participants, built upon a solid consensus foundation. The alignment operation of the central coordinating node acts as a knowledge filter and direction adjuster, ensuring that the knowledge absorbed locally is complementary and consistent. This prevents model drift or splitting caused by data heterogeneity or noise, thereby maintaining the robustness of the global model while enhancing its adaptability and agility to new local risks and pattern changes.

[0040] S2.3 Through multiple rounds of local cognitive evolution, privacy-encrypted submission, security fusion alignment, and cognitive state synchronization, the local risk cognitive model converges from the initial isolated state to a collaborative cognitive state that reflects cross-domain common risk patterns, thus obtaining a cross-domain risk consensus discriminator.

[0041] Furthermore, through multiple rounds of encrypted parameter interaction and secure aggregation, the model parameters of all parties converge to form a collaborative discrimination benchmark with embedded cross-domain risk consensus, achieving initial knowledge unification. Using this collaborative discrimination benchmark as a new starting point, a series of sub-loops—local cognitive evolution, privacy-encrypted submission, secure fusion and alignment, and cognitive state synchronization—are repeatedly executed. In each sub-loop, each party fine-tunes the latest collaborative model, generating a direction for cognitive state evolution. After encryption, secure fusion, and alignment, a new round of optimized collaborative model parameters is generated and synchronized to all parties.

[0042] Specifically, this process resembles a co-evolutionary spiral. With each cycle, the local models of each party evolve by absorbing global consensus and new local knowledge. As the iterations deepen, cross-domain risk knowledge, such as administrative penalty patterns from market supervision and management departments, abnormal fund patterns from financial institutions, and abnormal energy consumption patterns from public utilities, is continuously and securely encoded into the same model parameter space. Ultimately, when the changes in model parameters or the performance improvement of the model on independent validation sets become negligible after multiple iterations, the iteration terminates. The local risk perception model of the market supervision and management department has completed the transformation from an isolated model to a cross-domain risk consensus discriminator. The discriminator not only internalizes multi-source static knowledge but also possesses the ability to dynamically update and adaptively optimize through a continuous co-evolutionary mechanism.

[0043] It's important to note that a hierarchical, iterative federated learning framework was constructed. It breaks down the one-time joint training into two closely linked stages: global consensus foundation building and continuous collaborative evolution, with encryption and secure aggregation technologies used throughout to ensure privacy. This makes the final cross-domain risk consensus discriminator a living, evolving model capable of continuously tracking changes in risk patterns and providing persistent and accurate decision support for risk identification.

[0044] S3. Use the cross-domain risk consensus discriminator to screen out high-risk market entities based on the characteristic data of the risk status of market entities, and form a list of high-risk market entities.

[0045] S3.1 Input the characteristic data of the market supervision and management department into the cross-domain risk consensus discriminator and output a risk attribute label that integrates the cross-domain risk consensus for each market entity.

[0046] Furthermore, the market supervision and administration department inputs the characteristic data of the market supervision and administration department from the locally stored characteristic data of the risk status of market entities into the cross-domain risk consensus discriminator that has been trained. For each market entity, the cross-domain risk consensus discriminator uses the aforementioned characteristic data of the market supervision and administration department as input, and after a series of nonlinear transformations and calculations based on weight matrices and activation functions, it finally outputs a scalar value or a multi-dimensional vector. The output is defined as a risk attribute label that integrates the cross-domain risk consensus. The value of the risk attribute label or the direction of the vector comprehensively reflects the overall assessment of the market entity by the joint risk pattern contained in the characteristic data of the market supervision and administration department, the characteristic data of financial institutions, and the characteristic data of public institutions.

[0047] Specifically, for example, for a company that frequently changes its legal representative but whose transaction flow appears normal, the characteristic data from the market supervision and management department alone may indicate a medium risk. However, the cross-domain risk consensus discriminator, by internalizing the risk model of financial institutions regarding the rapid inflow and outflow of funds in shell companies, may output a higher risk attribute label value. Conversely, for a market entity with abnormal electricity consumption patterns but no administrative penalty records, the cross-domain risk consensus discriminator, after integrating the risk perception of abnormal energy consumption by public utilities, may also improve its risk attribute label. The multi-dimensional risk discrimination capability, obtained through privacy-preserving collaborative learning and embedded with knowledge from multiple parties, is specifically applied to the risk assessment practice of individual market entities. The output risk attribute label breaks through the perspective limitations of a single data source and is a more comprehensive and robust risk measurement based on cross-domain risk consensus.

[0048] S3.2. Compare the risk attribute labels that incorporate cross-domain risk consensus with the preset risk threshold, screen out market entities whose risk attribute labels that incorporate cross-domain risk consensus are higher than the preset risk threshold, and compile a list of high-risk market entities.

[0049] Furthermore, after obtaining risk attribute labels from all market entities that incorporate cross-domain risk consensus, the market supervision and management department compares these risk attribute labels with a predefined and set risk threshold. This risk threshold is a baseline determined based on historical regulatory experience, risk tolerance, and acceptance of false warnings. The comparison is conducted on a per-market-entity basis: if a market entity's risk attribute label value exceeds the preset risk threshold, the market entity is identified as a potentially high-risk entity. Subsequently, the identification information of all market entities that meet the condition of risk attribute labels exceeding the preset risk threshold, such as their unified social credit codes, is collected and summarized, and organized in the form of lists or database records, thereby forming a structured list of high-risk market entities.

[0050] Specifically, an absolute and consensus-based risk threshold is introduced as a screening criterion. The risk attribute label output by the cross-domain risk consensus discriminator is itself a comparable relative risk measure that has been calibrated by multiple parties' knowledge. The preset risk threshold provides a unified and clear action trigger point. Combining the two enables the transformation from continuous risk assessment to discrete high-risk decision-making, allowing regulatory resources to be accurately directed to entities whose risks have indeed exceeded the generally accepted warning level from a multi-source fusion perspective. This avoids misjudgments or omissions caused by single-source data bias or incomplete local risk cognition, ensuring the objectivity and consensus basis of regulatory actions.

[0051] S4. For each market entity in the list of high-risk market entities, construct a virtual development path in reverse by combining the actual behavioral trajectory formed by the characteristic data of the market entity's risk status with the key risk dimensions revealed by the cross-domain risk consensus discriminator.

[0052] S4.1. Select a market entity from the list of high-risk market entities as the current analysis object, extract the characteristic data of the risk status of the market entity corresponding to the current analysis object, and arrange them in time series to form the actual behavior trajectory of the current analysis object.

[0053] Furthermore, a market entity identifier is selected sequentially or by priority from the existing list of high-risk market entities as the current analysis object. Based on this identifier, all historical feature records belonging to the current analysis object are retrieved from the database storing feature data on the risk status of market entities. These feature records contain multi-dimensional values ​​from feature data from market supervision and management departments, financial institutions, and public institutions. Each record has a clear timestamp. All feature records are arranged and aligned according to the order of the timestamps to form a multi-dimensional feature matrix organized by time step. This matrix represents the actual behavioral trajectory of the current analysis object. Each row in the actual behavioral trajectory corresponds to a time point, and each column corresponds to a specific feature dimension, depicting the historical evolution path of the current analysis object in the multi-dimensional feature space.

[0054] Specifically, discrete, point-like risk characteristic data are reconstructed into continuous, linear behavioral sequences, thereby providing structured input for subsequent time-series-based comparative analysis. The spatiotemporal anchoring and data organization of the analysis object transform abstract high risk into concrete, analyzable dynamic behavioral sequences, enabling subsequent attribution analysis to unfold along the time axis and trace the dynamic process of risk formation.

[0055] S4.2 Utilize the cross-domain risk consensus discriminator to analyze the actual behavioral trajectory of the current analysis object, identify the feature dimension that contributes the most to the risk attribute label of the current analysis object, and use it as the key risk dimension.

[0056] Furthermore, the actual behavioral trajectory of the current analysis object is input into the cross-domain risk consensus discriminator, either as a whole or in segments. The cross-domain risk consensus discriminator not only outputs the final risk attribute label of the current analysis object, but its internal processing mechanism (such as calculating feature importance through gradient backpropagation or analyzing the attention weight distribution of deep neural networks) can trace and quantify the contribution of each feature dimension in the actual behavioral trajectory to the final high-risk judgment result. By sorting the contribution, the top few feature dimensions with the highest contribution are selected and these dimensions are identified as key risk dimensions.

[0057] Specifically, for example, for a market entity identified as high-risk, the interpretability analysis of the cross-domain risk consensus discriminator may reveal that the abnormal fluctuations in three characteristic dimensions—the frequency of nighttime public-to-private transfers, the monthly change rate of the number of social security contributors, and the proportion of electricity consumption on non-working days—contribute far more to the high-risk score than other characteristics. These three dimensions are identified as key risk dimensions. This deeply explores and utilizes the interpretability potential of the cross-domain risk consensus discriminator, enabling a focus from macro-risk assessment to micro-risk triggers.

[0058] S4.3 Based on key risk dimensions, at the starting point of the actual behavioral trajectory of the current analysis object, match compliance reference entities with similar initial states in the same dimension in history.

[0059] Furthermore, based on the key risk dimension, the actual behavioral trajectory of the current analysis object at the starting time point is extracted, and the specific values ​​of the key risk dimension are formed to constitute a vector describing the initial state. In the historical market entity database, all market entities not marked as high risk are retrieved. The value vectors of these compliant entities at the same historical time point (or similar development stage) on the same key risk dimension are used. The similarity between the initial state vector of the current analysis object and the corresponding vector of each compliant entity is calculated using distance measurement methods (such as Euclidean distance and cosine similarity). A similarity threshold is set, and all compliant market entities with similarity higher than the similarity threshold are selected and identified as compliant reference entities.

[0060] Specifically, for example, if the key risk dimensions are registered capital and first-month transaction volume, then we look for historical entities with similar registered capital and first-month transaction volume at the initial stage of their establishment to the current analysis object. We apply the basic idea of ​​counterfactual causal inference, and by matching control groups with similar initial conditions, we construct a potential reference system for high-risk entities, which is how their development trajectory might have been if they had gone to compliance in the first place. This controls the influence of other initial variables except for subsequent behavioral choices and enhances the internal validity of the attribution analysis.

[0061] S4.4 Integrate the behavioral patterns of compliance reference entities in key risk dimensions to synthesize a virtual development path that represents the ideal compliance trend.

[0062] Furthermore, for each compliance reference entity, complete time-series data on key risk dimensions is extracted from the starting point to the current moment (or a comparable time period), forming multiple compliance behavior trajectories. These compliance reference entities' behavior trajectories on key risk dimensions are then fused. Fusion methods could include obtaining statistical measures of the trajectory's characteristic values ​​at each time point, such as the mean or median, to form a smooth, average trajectory representing the typical behavioral patterns of the compliance group. More complex methods, such as dynamic time normalization followed by alignment and averaging, can also be used to better handle differences in the development pace of different entities. The final synthesized time-series trajectory is the virtual development path, representing a typical and ideal evolutionary pattern followed by a group of market entities successfully maintaining compliance under the same initial conditions on the key risk dimensions.

[0063] Specifically, by summarizing and synthesizing behavioral patterns from multiple real-world compliance cases, an ideal compliance trend benchmark based on historical experience and statistical representativeness is constructed. This avoids the randomness or particularity that may exist with a single reference sample, making the constructed virtual development path more universal and robust, and more reliable as a comparison benchmark.

[0064] S4.5 By comparing the actual behavior trajectory of the current analysis object with the virtual development path, the decision-making fork points that cause the current analysis object to deviate from the compliance direction are revealed, and the virtual development path is constructed in reverse.

[0065] Furthermore, the actual behavioral trajectory of the currently analyzed object is compared with the virtual development path at key risk dimensions on a point-by-point basis. The comparison process includes obtaining the deviation between the actual value and the expected value of the virtual path at each point in time. By setting a deviation threshold or applying a change point detection algorithm, the point(s) at which the actual behavioral trajectory begins to continuously deviate from the virtual development path are identified, and these point(s) are marked as decision fork points. For example, the virtual development path shows that the proportion of nighttime transactions of compliant entities usually remains stable at a low level, while the actual trajectory of the currently analyzed object shows that, starting from a certain month, its proportion of nighttime transactions has been continuously rising and far exceeds the virtual path. This month is then identified as a decision fork point. The core logic of this reverse construction process is: by first matching the reference entity, then synthesizing the virtual path, and finally locating the deviation point, it is essentially deducing from the result (high risk) backward to locate the key behavioral turning point that leads to the difference in results.

[0066] Specifically, the high-risk attribution problem is transformed into a time-series anomaly detection and pattern comparison problem, using a virtual path synthesized based on historical compliance data as a dynamic baseline for normal behavior. This approach clearly reveals that high risk does not arise overnight, but rather stems from decisions or behaviors made at specific points in time that contradict the compliance group. This provides regulators with clear intervention opportunities and specific behavioral correction targets, achieving a leap from risk warning to root cause identification and timely alerts.

[0067] S5. Perform compliance deviation analysis between actual behavior trajectory and virtual development path, identify the set of key decision nodes that lead to deviation, and generate an attribution analysis report.

[0068] S5.1 Based on key risk dimensions, conduct a time-series comparison between the actual behavior trajectory and the virtual development path of the current analysis object to identify deviation events where the actual behavior trajectory deviates from the virtual development path.

[0069] Furthermore, based on the identified key risk dimensions, a point-in-time numerical comparison is made between the actual behavioral trajectory and the virtual development path of the current analysis object. For each key risk dimension, the difference between the observed value of the actual behavioral trajectory and the expected value of the virtual development path at the corresponding time point is calculated, and the trend of this difference over time is continuously tracked. When the actual value of a dimension is consistently higher or lower than the expected value of the virtual development path at multiple consecutive time points, and the accumulation or trend of the difference indicates that it is not a random fluctuation, it is recorded as an independent deviation event. The deviation event includes attributes such as the key risk dimension involved, the time point when the deviation started, the duration of the deviation, and the overall direction of the deviation (such as positive or negative deviation).

[0070] Specifically, for example, in the dimension of the average number of nighttime transactions per month, if the expected value of the virtual development path is stable, but the actual behavior trajectory shows a continuous and rapid increase starting from a certain month, then a continuous positive deviation event is identified from that month. The continuous trajectory difference is discretized into a series of specific and describable abnormal events, providing clear operational objects for in-depth tracing of causes. The complex overall deviation is decomposed into sub-problems in multiple dimensions and different time periods, allowing the attribution analysis to be carried out point by point, avoiding vague and ambiguous judgments.

[0071] S5.2 Filter out deviation events that exceed a preset threshold, and for each deviation event, trace the inducing behavior that caused the deviation event.

[0072] Furthermore, for each deviation event, the degree of deviation is assessed individually, and a uniform deviation magnitude threshold is set. This threshold is determined based on the normal fluctuation range of compliant entity behavior in historical data or expert experience. For each deviation event, the average or maximum deviation magnitude between the actual observed value and the expected value of the virtual path is determined during its duration. Deviation events with deviation magnitudes exceeding the preset threshold are selected as deviation events. For each deviation event, it is necessary to trace its inducing behavior by backtracking and analyzing the data changes of the actual behavior trajectory before and after the starting point of the deviation event. The focus is on examining whether there are other related characteristics or business behaviors that have changed abruptly or abnormally at the same point in time, in addition to the key risk dimensions.

[0073] Specifically, by tracing back, it may be found that before and after the starting point of the deviation, the number of job postings by market entities also decreased sharply, and the outflow of public funds increased significantly. These co-variant related behaviors together constitute the set of induced behaviors that led to the deviation in the number of social security personnel. By combining time correlation and multi-dimensional data covariance analysis, the superficial statistical deviation is linked to the deep behavioral decision-making, so that the attribution analysis can reach the possible specific business actions or strategy choices, thereby enhancing the depth and operability of the attribution.

[0074] S5.3 Aggregate the triggering behavioral points corresponding to multiple deviation events to form a set of key decision nodes and generate an attribution inference report.

[0075] Furthermore, the time points of the triggering behaviors corresponding to each deviation event are summarized and aggregated. Since the triggering behaviors of different deviation events may occur at the same or similar time points, the aggregation process merges or clusters these time points to form a set of key decision-making node time points without redundancy. Each key decision-making node not only contains time information, but also is associated with a series of triggering behaviors that occurred at that node and caused significant deviations in one or more key risk dimensions. For example, it may be found that the triggering behaviors of multiple deviation events (such as abnormal transactions or abnormal employment) are concentrated in the same quarter. This quarter is regarded as a key decision-making node, marking that market participants have made a series of major decisions that led to deviations in the risk path. Based on the set of key decision-making nodes and related information such as deviation events, triggering behaviors, and key risk dimensions, a structured attribution inference report is generated.

[0076] Specifically, this approach elevates the focus from scattered anomalies to concentrated decision-making moments. High risks are often not caused by isolated errors, but rather by a series of related erroneous decisions or behaviors at several key points in time. Through aggregate analysis, attribution analysis reports can clearly identify the critical turning points in the risk evolution process and the combinations of erroneous decisions made during those periods. This provides regulators with highly targeted, in-depth diagnostic reports that focus on when and what behaviors occurred, rather than simple risk descriptions, significantly improving the accuracy and effectiveness of regulatory intervention.

[0077] S6. Based on the attribution analysis report, classify and prioritize the list of high-risk market entities to generate a regulatory task queue.

[0078] S6.1 Analyze the attribution analysis report, extract the key decision-making node set information from the attribution analysis report, conduct risk assessment for each market entity in the list of high-risk market entities, and generate the risk level of each market entity.

[0079] Furthermore, each generated attribution analysis report undergoes structured analysis to extract key decision-making node information recorded in the report. This key decision-making node information includes factors such as the time point of the key decision-making node, the type and number of associated triggering behaviors, and the severity of their impact. A risk assessment function is then constructed, which comprehensively considers multiple dimensions such as the number of key decision-making nodes, the temporal density of the nodes, the potential harm of the triggering behaviors, and the strength of the correlation between the behaviors and risk dimensions. This comprehensive quantitative assessment of each high-risk market entity outputs a quantitative risk score or a discrete risk level, such as high, medium, and low. This transforms the in-depth, qualitative causal relationship analysis in the attribution analysis report into a quantifiable, comparable, and unified macro-risk measure.

[0080] Specifically, risk assessment is no longer based on simple risk scores, but on structured information (a set of key decision-making nodes) derived from attribution analysis that reflects the root causes and evolution of risks. This makes the risk level classification more accurately reflect the complexity and urgency of the causes of risks. For example, even if a market entity does not have the highest absolute risk score, if its key decision-making nodes show a intensive and multi-type tendency to violate regulations in the short term, its risk level may be higher because its risk evolution is more active and complex.

[0081] S6.2 Classify market entities in the list of high-risk market entities based on risk level and divide them into different regulatory categories.

[0082] Furthermore, based on the risk level generated for each market entity in the list of high-risk market entities, regulatory classification rules are set. For example, market entities with a high risk level and whose key decision-making nodes involve financial fraud-related inducements are classified as emergency financial verification entities; market entities with a high risk level but whose inducements mainly involve safety production or environmental protection are classified as key safety and environmental protection inspection entities; market entities with a medium risk level are classified as routine inspection entities; and market entities with a low risk level or whose key decision-making node information is vague and needs to be observed are classified as monitoring and early warning entities.

[0083] Specifically, this achieves a mapping from unified risk measurement to differentiated regulatory strategies. The classification is not only based on the magnitude of the risk, but also incorporates the nature of the risk revealed by the attribution analysis report (reflected by the type of triggering behavior), thereby enabling the allocation of regulatory resources to be precisely diverted according to the handling requirements and professional needs of different risk types, making regulatory actions more professional, targeted, and strategic.

[0084] S6.3 Within each regulatory category, market entities are ranked according to their risk level to generate a regulatory priority sequence. The regulatory categories and their corresponding regulatory priority sequences are then integrated to form a regulatory task queue.

[0085] Furthermore, market entities within each regulatory category are ranked from highest to lowest risk level, generating a regulatory priority sequence within each category. For example, in the emergency financial investigation category, all market entities belonging to the category are arranged in descending order of their risk scores. All regulatory categories and their corresponding internally ranked lists of market entities are integrated to form a structured regulatory task queue, clearly indicating the order of tasks for different categories and the specific execution order within each category.

[0086] Specifically, a two-dimensional, hierarchical task scheduling logic was constructed: the first dimension is category, ensuring that different types of risks are handled by corresponding professional forces; the second dimension is priority, ensuring that, among similar tasks, those with more urgent risks and more active evolution are given priority, enabling limited regulatory resources to be optimally allocated according to the principle of urgency priority, and ultimately transforming the in-depth insights obtained from attribution analysis into clear, executable, and highly efficient regulatory action plans. S7. Assign tasks in the supervisory task queue to inspection personnel, perform inspections, and submit inspection reports.

[0087] S7.1 Based on the regulatory category and order of tasks in the regulatory task queue, match each task in the regulatory task queue with an inspector with the corresponding professional qualifications, assign tasks in the regulatory task queue to the matched inspectors, and provide the inspectors with an attribution analysis report.

[0088] Furthermore, the regulatory task queue is analyzed, and the regulatory category attribute and sort number within the category of each task are read. A database recording all available inspectors and their professional qualifications is maintained. Based on the regulatory category of the task, a set of inspectors with corresponding professional qualifications is selected from the database. Then, combined with the task's sort number and the inspectors' current workload and geographical location constraints, a task allocation algorithm is used to determine the most suitable inspector for each task. After matching, the task details along with the corresponding attribution inference report are assigned to the determined inspectors. The attribution inference report provides in-depth insights into the causes of risk, key decision-making nodes, and suspected problems of the market entity.

[0089] Specifically, it achieves a refined and intelligent matching of regulatory tasks with the professional capabilities of personnel, directly linking the classification of regulatory tasks (based on risk nature attribution) with the professional division of human resources. This ensures that professionals handle professional tasks. For example, a task attributed to suspected financial fraud will be automatically matched to inspectors with financial audit qualifications, while tasks related to safety production hazards will be assigned to personnel with corresponding safety inspection qualifications. Combining sorting and load optimization improves overall scheduling efficiency, enabling in-depth attribution analysis results to directly drive the deployment of professional and efficient on-site inspection forces, thereby enhancing the pertinence and effectiveness of regulatory actions.

[0090] S7.2. In accordance with the guidance of the attribution analysis report, conduct on-site inspections of the market entities corresponding to the assigned tasks and submit a structured inspection report.

[0091] Furthermore, after receiving the assigned tasks and attribution analysis reports, the inspection personnel study the reports. The key decision-making nodes and descriptions of triggering behaviors in the reports provide clear investigative directions and key clues for on-site inspections. The inspection personnel then focus their on-site inspections on verifying the entity's nighttime transaction vouchers, the consistency between contracts and actual business operations, and the rationality of account changes. The inspection personnel go to the market entity's business premises to conduct on-site inspections, verifying the doubts pointed out in the attribution analysis reports by reviewing documents, interviewing personnel, and inspecting the site. After the inspection is completed, the inspection personnel fill in the inspection results in a structured form using a mobile terminal. The form's fields correspond to the key points in the attribution analysis reports, such as verification conclusions of behaviors corresponding to key decision-making nodes, descriptions of problems found, and evidence materials. The submitted structured inspection report not only records the inspection results.

[0092] Specifically, the deep understanding (attribution inference report) is transformed into an actionable on-site inspection guide and structured results feedback, forming a closed loop of analysis-driven inspection and inspection-verification analysis. The attribution inference report is no longer just an analytical conclusion on paper, but a navigation map that precisely guides on-site actions, enabling on-site inspections to transform from traditional general surveys into highly focused and precise verifications.

[0093] S8. Trigger a new round of iterative optimization of cross-domain risk awareness consensus.

[0094] S8.1. Link the inspection results in the structured inspection report to the characteristic data of the market entity's risk status.

[0095] Furthermore, the submitted structured inspection report is read, and the market entity identifier and corresponding final inspection conclusion (such as problems found, no problems found, and some problems requiring rectification) recorded in the report are extracted. Using the market entity identifier as the key index, all historical characteristic data records of the entity within the analysis period are located in the database storing characteristic data of the market entity's risk status. The inspection conclusion is used as a new, on-site verified label data and is associated and bound with the located historical characteristic data.

[0096] Specifically, for example, if a subject's characteristic data within a specific time period is marked as high-risk and an attribution inference report is generated, and an inspection confirms that there is indeed a problem, then the characteristic data for that time period is associated with a verification label indicating verified risk. If no problem is found during the inspection, then a verification label indicating false alarm or low risk is associated with it. The closed-loop result of the regulatory action (inspection report) is fed back to the data source, adding real-world verification labels to historical data. This allows the characteristic data originally used for prediction to obtain valuable, practice-tested supervisory signals, providing high-quality, high-confidence labeled data for the learning and optimization of local risk perception models.

[0097] S8.2. Update the local training sample set of the data holder using the characteristic data of the risk status of market entities with new verification labels.

[0098] Furthermore, data holders (market supervision and management departments, financial institutions, and public utilities) add the characteristic data of the risk status of market entities with newly verified labels to their respective local databases as new training samples to the existing local training sample set. This includes adding newly generated samples and updating or correcting the labels of existing samples. For example, market supervision and management departments add the characteristic data and risk labels of market entities that have been verified to have problems to the training set, and may adjust the labels of characteristic data of market entities that have been falsely reported to low risk or make special markings. The updated training sample set reflects the latest risk distribution and pattern information verified on-site.

[0099] Specifically, a continuous and automated data quality and label optimization mechanism has been established. Each on-site inspection is a calibration of the true values ​​of the model's previous predictions. By feeding back the calibrated samples to the training set, the local knowledge base of the data holder can keep up with the times, continuously correct the cognitive biases of the model, and accumulate experience in new risk patterns or false alarm patterns.

[0100] S8.3 Based on the updated training sample set, the data holder restarts the process of evolving the local risk perception model and exchanging encrypted parameters.

[0101] Furthermore, each data holder restarts a complete new training cycle based on the training sample set of the new verification label samples: each party uses the updated training sample set locally to retrain or incrementally train the current version of the local risk perception model to obtain new model parameter updates. Each party performs encryption operations to generate encrypted model parameter updates and submits them to the central coordinator. The central coordinator executes the secure aggregation algorithm again to generate new encrypted global model parameter updates and distributes them. Each party decrypts and updates its local risk perception model.

[0102] Specifically, the entire privacy-protected collaborative learning mechanism was reactivated, but the starting data and label quality of this learning have been improved due to feedback from previous regulatory practices. The real results generated by regulatory actions (on-site inspections) were transformed into training data, driving the federated learning framework to carry out a new round of knowledge collaborative optimization, enabling the local risk perception model to learn from actual regulatory effects, thereby achieving self-correction and enhancement.

[0103] S8.4. Through a new round of interaction and optimization, drive the parameters of the cross-domain risk consensus discriminator to be updated iteratively to optimize the cross-domain risk cognition consensus.

[0104] Furthermore, the new round of local risk perception model evolution and encrypted parameter interaction process, which was initiated and completed, is reflected in the iterative update of the cross-domain risk consensus discriminator model parameters. After multiple rounds of secure interaction and optimization, the parameters of the local risk perception models of all parties converged again to a new and better stable state. Corresponding to the cross-domain risk consensus discriminator on the market supervision and management department side, its internal parameters were adjusted and optimized, which affected the cross-domain risk perception consensus: the local risk perception model will have a deeper memory of the patterns of verified risks, may reduce the weight of feature combinations that lead to false alarms, and will learn newly discovered risk patterns that were not previously identified.

[0105] Specifically, it has achieved an overall evolution of cross-domain risk perception consensus, making the discriminator more accurate, reliable, and forward-looking in future risk identification and attribution. It seamlessly embeds the regulatory closed loop into the iterative lifecycle of the machine learning model, so that the cross-domain risk consensus discriminator is no longer a static tool that is fixed after deployment, but a living cognitive system that can learn from experience in every regulatory practice and continuously optimize its own judgment capabilities, realizing the deep integration and common growth of regulatory wisdom and artificial intelligence.

[0106] This embodiment also provides a market entity management system based on multi-source data fusion, including: a feature data module, which generates feature data of market entity risk status locally based on multiple data holders; The module allows multiple data holders to interact through a privacy-protected collaborative mechanism, participate in the process of building a cross-domain risk awareness consensus, and form a cross-domain risk consensus discriminator locally. The screening module uses a cross-domain risk consensus discriminator to screen out high-risk market entities based on the characteristic data of the risk status of market entities, and forms a list of high-risk market entities. The deduction module, for each market entity in the list of high-risk market entities, constructs a virtual development path in reverse based on the actual behavioral trajectory formed by the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator. It then deduces the compliance deviation between the actual behavioral trajectory and the virtual development path, locates the set of key decision nodes that lead to the deviation, and generates an attribution deduction report. The sorting module classifies and prioritizes the list of high-risk market entities based on the attribution inference report, and generates a regulatory task queue. The optimization module assigns tasks from the regulatory task queue to inspection personnel, who then perform inspections and submit inspection reports, triggering a new round of iterative optimization of cross-domain risk awareness consensus.

[0107] This embodiment also provides a computer device applicable to the market entity management method based on multi-source data fusion, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the market entity management method based on multi-source data fusion as proposed in the above embodiment.

[0108] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0109] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the market entity management method based on multi-source data fusion as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0110] In summary, this invention generates characteristic data on the risk status of market entities locally on multiple data holders and constructs a cross-domain risk consensus discriminator using a privacy-preserving collaborative mechanism. This achieves deep integration of multi-source heterogeneous data while ensuring data privacy and security, improving the accuracy and comprehensiveness of risk identification. The discriminator filters out high-risk market entities, and for each high-risk entity, it compares its actual behavioral trajectory with a virtual compliance development path constructed in reverse based on key risk dimensions to perform compliance deviation deduction. This accurately locates the key decision-making nodes that lead to risks and generates attribution reports. This overcomes the limitations of traditional methods that can only provide risk scores but cannot reveal specific risk causes and intervention paths. Based on the attribution reports, high-risk entities are classified, ranked, and a regulatory task queue is generated. Through inspection feedback, the local risk cognition model is iteratively optimized, forming a process from accurate risk identification and in-depth attribution analysis to intelligent regulatory intervention and continuous self-optimization.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A market entity management method based on multi-source data fusion, characterized by: This includes generating characteristic data on the risk status of market entities locally based on multiple data holders; The multiple data holders interact through a privacy protection collaboration mechanism to participate in the construction of a cross-domain risk awareness consensus and form a cross-domain risk consensus discriminator locally; Using a cross-domain risk consensus discriminator, high-risk market entities are screened out based on the characteristic data of the risk status of market entities, and a list of high-risk market entities is formed. For each market entity in the list of high-risk market entities, based on the actual behavioral trajectory formed by the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator, a virtual development path is constructed in reverse. The actual behavioral trajectory and the virtual development path are used to perform compliance deviation inference, locate the set of key decision nodes that lead to the deviation, and generate an attribution inference report. Based on the attribution and inference report, the list of high-risk market entities is classified and prioritized to generate a regulatory task queue. Assigning tasks from the regulatory task queue to inspection personnel to perform inspections and submit inspection reports triggers a new round of iterative optimization of cross-domain risk awareness consensus.

2. The market entity management method based on multi-source data fusion as described in claim 1, characterized in that: Based on multiple data holders, characteristic data on the risk status of market participants are generated locally, including the following steps: Based on the data processing of the market supervision and administration department, characteristic data of the market supervision and administration department is generated. Based on data processing of financial institutions, generate characteristic data of financial institutions; Based on data processing of public institutions, generate characteristic data of public institutions; The characteristic data of market entities' risk status are generated by aggregating characteristic data from market supervision and management departments, financial institutions, and public institutions.

3. The market entity management method based on multi-source data fusion as described in claim 2, characterized in that: The multiple data holders interact through a privacy-protection collaboration mechanism to participate in the construction of a cross-domain risk awareness consensus, forming a cross-domain risk consensus discriminator locally, including the following steps: Based on the federated learning framework, multiple data holders interact through encrypted model parameter updates. Under the secure aggregation of the central coordinator, they jointly drive the iterative optimization of local risk perception model parameters and converge to form a collaborative discrimination benchmark with embedded cross-domain risk consensus. Based on the collaborative judgment benchmark with embedded cross-domain risk consensus, the risk cognition of market entities is independently iterated and evolved in the local environment using characteristic data of risk status. This forms a cognitive state evolution direction that reflects the local data pattern. The cognitive state evolution direction is submitted after privacy encryption. The central coordination node then securely integrates and aligns the encrypted multi-party cognitive evolution directions. Through multiple rounds of local cognitive evolution, privacy-encrypted submission, security fusion alignment, and cognitive state synchronization, the local risk cognitive model converges from its initial isolated state to a collaborative cognitive state that reflects cross-domain common risk patterns, thus obtaining a cross-domain risk consensus discriminator.

4. The market entity management method based on multi-source data fusion as described in claim 3, characterized in that: Using a cross-domain risk consensus discriminator to screen high-risk market entities based on their risk status characteristics, a list of high-risk market entities is generated, including the following steps: The characteristic data of the market supervision and management department is input into the cross-domain risk consensus discriminator, which outputs a risk attribute label that integrates the cross-domain risk consensus for each market entity. The risk attribute labels that incorporate cross-domain risk consensus are compared with preset risk thresholds. Market entities whose risk attribute labels that incorporate cross-domain risk consensus are higher than the preset risk thresholds are selected and compiled into a list of high-risk market entities.

5. The market entity management method based on multi-source data fusion as described in claim 4, characterized in that: For each market entity in the high-risk market entity list, a virtual development path is constructed in reverse, based on the actual behavioral trajectory formed by the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator. This includes the following steps: Based on the list of high-risk market entities, select one market entity as the current analysis object, extract the characteristic data of the risk status of the market entity corresponding to the current analysis object, and arrange them in time series to form the actual behavior trajectory of the current analysis object. By using a cross-domain risk consensus discriminator to analyze the actual behavioral trajectory of the current analysis object, the feature dimension that contributes the most to the risk attribute label of the current analysis object is identified as the key risk dimension. Based on key risk dimensions, at the starting point of the actual behavior trajectory of the current analysis object, match compliance reference entities with similar initial states in the same dimension in history; By integrating the behavioral patterns of compliance reference entities across key risk dimensions, a virtual development path representing the ideal direction of compliance is synthesized. By comparing the actual behavioral trajectory of the current analysis object with the virtual development path, the decision-making forks that cause the current analysis object to deviate from the compliance direction are revealed, and the virtual development path is constructed in reverse.

6. The market entity management method based on multi-source data fusion as described in claim 5, characterized in that: By comparing actual behavioral patterns with virtual development paths to identify compliance deviations, a set of key decision-making nodes leading to these deviations is located, and an attribution analysis report is generated. This includes the following steps: Based on key risk dimensions, the actual behavior trajectory of the current analysis object is compared with the virtual development path over time to identify deviation events in which the actual behavior trajectory deviates from the virtual development path; Filter out deviation events that exceed a preset threshold, and for each deviation event, trace the triggering behavior that caused the deviation event. The triggering behavioral points corresponding to multiple deviation events are aggregated to form a set of key decision nodes, and an attribution inference report is generated.

7. The market entity management method based on multi-source data fusion as described in claim 6, characterized in that: Based on the attribution analysis report, the list of high-risk market entities is classified and prioritized to generate a regulatory task queue, including the following steps: Analyze the attribution inference report, extract the key decision node set information from the attribution inference report, conduct risk assessment for each market entity in the list of high-risk market entities, and generate the risk level of each market entity. Based on the aforementioned risk levels, market entities in the list of high-risk market entities are classified and placed into different regulatory categories. Within the regulatory categories, market entities are sorted according to their risk levels to generate a regulatory priority sequence. The regulatory categories and their corresponding regulatory priority sequences are then integrated to form a regulatory task queue.

8. The market entity management method based on multi-source data fusion as described in claim 7, characterized in that: Assigning tasks from the supervisory task queue to inspection personnel, performing inspections, and submitting inspection reports includes the following steps: Based on the regulatory category and order of tasks in the regulatory task queue, each task in the regulatory task queue is matched with an inspector with corresponding professional qualifications. Tasks in the regulatory task queue are assigned to the matched inspectors, and attribution and inference reports are provided to the inspectors. Based on the guidance of the attribution analysis report, conduct on-site inspections of the market entities corresponding to the assigned tasks and submit structured inspection reports.

9. The market entity management method based on multi-source data fusion as described in claim 8, characterized in that: Triggering a new round of iterative optimization of cross-domain risk awareness consensus includes the following steps: The inspection results in the structured inspection report are linked to the characteristic data of the market entity's risk status. Update the local training sample set of the data holder using the characteristic data of the risk status of market entities with new verification labels; Based on the updated training sample set, the data holder restarts the process of evolving the local risk perception model and exchanging encrypted parameters. Through a new round of interaction and optimization, the parameters of the cross-domain risk consensus discriminator are iteratively updated to optimize the cross-domain risk perception consensus.

10. A market entity management system based on multi-source data fusion, based on the market entity management method based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that: This includes a feature data module, which generates feature data on the risk status of market entities locally based on multiple data holders; The module is constructed in which the multiple data holders interact through a privacy-preserving collaborative mechanism to participate in the construction process of cross-domain risk awareness consensus and form a cross-domain risk consensus discriminator locally. The screening module uses a cross-domain risk consensus discriminator to screen out high-risk market entities based on the characteristic data of the risk status of market entities, and forms a list of high-risk market entities. The deduction module, for each market entity in the list of high-risk market entities, constructs a virtual development path in reverse based on the actual behavioral trajectory formed by the characteristic data of the market entity's risk status and the key risk dimensions revealed by the cross-domain risk consensus discriminator. It then deduces the compliance deviation between the actual behavioral trajectory and the virtual development path, locates the set of key decision nodes that lead to the deviation, and generates an attribution deduction report. The sorting module classifies and prioritizes the list of high-risk market entities based on the attribution inference report, and generates a regulatory task queue. The optimization module assigns tasks from the regulatory task queue to inspection personnel, who then perform inspections and submit inspection reports, triggering a new round of iterative optimization of cross-domain risk awareness consensus.

Images