Operation state evaluation method and device, electronic equipment, medium and program product

Abstract

The invention provides an operation state evaluation method and device, electronic equipment, a medium and a program product, and can be applied to the technical field of artificial intelligence and the technical field of big data. The method comprises the following steps: acquiring a multi-source feature set of a target enterprise; obtaining a pre-constructed heterogeneous relation graph, and extracting an associated sub-graph of the target enterprise based on the heterogeneous relation graph; the multi-source feature set and the associated sub-graph are input into a graph attention network for graph calculation, the graph attention network comprises temperature parameters dynamically adjusted based on external environment parameters, and the temperature parameters are used for adjusting the weights of the edges in the graph calculation process; and generating a risk score of the target enterprise based on a result of the graph calculation, and performing operation state evaluation of the target enterprise based on the risk score.

Description

Technical Field

[0001] This application can be applied to the fields of artificial intelligence, big data, and financial technology, and more specifically relates to a method, apparatus, equipment, medium, and program product for assessing business status. Background Technology

[0002] Currently, enterprise operational status assessment systems generally employ rule engines or static scoring models based on structured financial data. These systems rely on historical loan records and financial statements to construct feature spaces, lacking the ability to integrate unstructured information and heterogeneous multi-source data. Furthermore, traditional models often use batch processing or periodic retraining methods, making it difficult to respond in real-time to dynamic changes in enterprise operational status, resulting in delayed risk detection. Graph-based modeling of upstream and downstream operations and holding structures is also inadequate, failing to convey potential risk signals within the graph structure. Summary of the Invention

[0003] In view of the above problems, this application provides methods, apparatus, equipment, media and procedures for assessing operational status.

[0004] According to a first aspect of this application, a method for assessing business status is provided, the method comprising: acquiring a multi-source feature set of a target enterprise; acquiring a pre-constructed heterogeneous relationship graph, and extracting a relational subgraph of the target enterprise based on the heterogeneous relationship graph, wherein the nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes, and multiple industry nodes, and the edges in the heterogeneous relationship graph include relational attributes, the relational attributes including positive cooperation relational attributes and negative competition relational attributes; inputting the multi-source feature set and the relational subgraph into a graph attention network for graph computation, wherein the graph attention network includes a temperature parameter dynamically adjusted based on external environmental parameters, the temperature parameter being used to adjust the weights of the edges during the graph computation process; and generating a risk score for the target enterprise based on the graph computation result, and assessing the business status of the target enterprise based on the risk score.

[0005] According to an embodiment of this application, pre-constructing a heterogeneous relationship graph includes: obtaining multi-source heterogeneous enterprise data from multiple data providers; performing structured processing on the multi-source heterogeneous enterprise data to obtain feature data; updating local parameters on the multiple data providers based on the feature data to obtain multiple update parameters; aggregating the multiple update parameters based on horizontal federated learning to form a joint feature representation; and using the joint feature representation to determine the multiple enterprise nodes, the multiple related party nodes, the multiple industry nodes, and the relationship attributes to construct the heterogeneous relationship graph.

[0006] According to an embodiment of this application, the method further includes at least one of the following: identifying enterprise entities in the feature data based on pseudo-anonymous identifiers generated by the plurality of data providers through a shared key; and during the aggregation to form a joint feature representation, introducing differential privacy noise based on a preset privacy budget to obtain a privacy-processed joint feature representation, wherein the privacy-processed joint feature representation is used to determine the relational attribute.

[0007] According to an embodiment of this application, the step of extracting the associated subgraph of the target enterprise based on the heterogeneous relationship graph includes: taking the target enterprise as the central node, selecting associated nodes within a preset order range in the heterogeneous relationship graph based on the central node; and comparing the weight of the edge between the central node and the associated node with a preset threshold to determine the target node and the target edge, and constructing the associated subgraph based on the target node and the target edge.

[0008] According to an embodiment of this application, the step of generating a risk score for the target enterprise based on the graph computation result and assessing the operating status of the target enterprise based on the risk score includes: vectorizing the graph computation result to obtain a risk feature vector; inputting the risk feature vector into a multilayer perceptron network to calculate the risk score; and comparing the risk score with a target threshold parameter to generate an operating status assessment result for the target enterprise.

[0009] According to an embodiment of this application, the target threshold parameter is updated based on a reinforcement learning algorithm. The state space of the reinforcement learning algorithm includes historical risk score distribution and external economic indicators, and the action space of the reinforcement learning algorithm includes a set of candidate threshold parameters.

[0010] According to an embodiment of this application, the update of the target threshold parameter includes: using the preset threshold parameter as a parameter of a learnable sub-network and jointly modeling it with the parameter of the multilayer perceptron network in the same training process, wherein the target threshold parameter and the parameter of the multilayer perceptron network are jointly updated based on the same loss function.

[0011] According to an embodiment of this application, the method further includes: identifying causal relationships between the multi-source feature set and the node features in the associated subgraph based on causal reasoning rules, and generating a causal relationship graph; and determining a set of influencing factors based on the causal relationship graph, and generating explanatory information based on the set of influencing factors.

[0012] According to an embodiment of this application, during graph computation, a corresponding learnable weight matrix is ​​set for edges with different relational attributes, and the edges with different relational attributes are weighted based on the learnable weight matrix.

[0013] A second aspect of this application provides an operational status assessment device, comprising: a data acquisition module for acquiring a multi-source feature set of a target enterprise; a subgraph extraction module for acquiring a pre-constructed heterogeneous relationship graph and extracting a subgraph of the target enterprise based on the heterogeneous relationship graph, wherein the nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes, and multiple industry nodes, and the edges in the heterogeneous relationship graph include relationship attributes, including positive cooperation relationship attributes and negative competition relationship attributes; a graph calculation module for inputting the multi-source feature set and the subgraph of the target enterprise into a graph attention network for graph calculation, wherein the graph attention network includes a temperature parameter dynamically adjusted based on external environmental parameters, the temperature parameter being used to adjust the weights of the edges during the graph calculation process; and an operational status assessment module for generating a risk score for the target enterprise based on the graph calculation results and assessing the operational status of the target enterprise based on the risk score.

[0014] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0017] According to embodiments of this application, by combining multi-source feature sets with heterogeneous relationship graphs and inputting both into a graph attention network for relationship-aware modeling, a unified representation of the enterprise's own attributes and external relationships, as well as a fine-grained characterization of risk transmission, can be achieved. A temperature parameter, set by external environmental parameters, is introduced to adjust edge weights during graph computation, ensuring the model maintains robustness in connection strength when the macro-environment is stable and highlights high-risk relationships during environmental fluctuations, improving the timeliness of response to risk changes and the accuracy of identification. Through the extraction of associated subgraphs, computation is performed only on nodes and edges strongly related to the target enterprise, reducing redundant graph operations and memory consumption, and improving throughput and latency performance on large-scale graphs. Furthermore, risk scores are generated by combining positive / negative relationship attributes, and operational status assessment results are output, making the assessment conclusions both dynamic and interpretable. Attached Figure Description

[0018] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 The illustrations depict application scenarios of the business status assessment method, apparatus, equipment, medium, and program products according to embodiments of this application.

[0020] Figure 2 A flowchart illustrating an operational status assessment method according to an embodiment of this application is shown schematically.

[0021] Figure 3 The flowchart illustrating a method for pre-constructing a heterogeneous relationship graph according to some exemplary embodiments of this application is shown schematically.

[0022] Figure 4 A flowchart illustrating a method for extracting an association subgraph of a target enterprise according to some exemplary embodiments of this application is shown schematically.

[0023] Figure 5 A schematic diagram illustrating the structure of an operational status assessment device according to an embodiment of this application is shown; and

[0024] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a business status assessment method according to an embodiment of this application. Detailed Implementation

[0025] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0026] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0027] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0028] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0029] The following will explain and clarify the terms used in this article.

[0030] Graph attention networks refer to a graph neural network structure based on an attention mechanism. During graph computation, learnable weights are used to weight and aggregate the features of adjacent nodes to highlight the differences between different relationships.

[0031] The temperature parameter refers to a dynamically adjusted parameter introduced in the weight allocation of a graph attention network. This parameter is set based on external environmental indicators and is used to adjust the distribution of edge weights so that the model can adapt to different macroscopic environments.

[0032] A multilayer perceptron network is a feedforward neural network composed of multiple layers of neurons that receives risk feature vectors as input and outputs risk scoring results.

[0033] Causal reasoning rules refer to reasoning methods that identify causal relationships between input features and node features based on causal relationship theories (such as causal graphs and reasoning rules), and are used to uncover key factors that have a significant impact on risk scores.

[0034] A learnable weight matrix refers to the weight allocation matrix used for edges with different relational attributes during graph computation. This matrix automatically learns the optimal weights through gradient updates during training to reflect the strength of the impact of different relational attributes on risk transmission.

[0035] With the development of big data, artificial intelligence, and fintech, the quantitative assessment of enterprise operating status has gradually become an important research direction for financial institutions, government departments, and third-party credit reporting agencies. Effective assessment of enterprise operating status can not only support credit decisions, risk control, and policy formulation, but also help regulatory authorities gain a macro-level understanding of industry operations. However, existing technologies still have many limitations in practical applications, making it difficult to meet the high timeliness and accuracy requirements of risk assessment in complex and ever-changing enterprise operating scenarios.

[0036] In current practice, enterprise operational status assessment systems are generally built upon structured financial data, typically relying on financial statements, loan records, and partially publicly available credit files. This data, after preprocessing, is input into rule engines or static scoring models to form a feature space, and the results are output in the form of scorecards or classification models. This approach has advantages such as simplicity of implementation and strong interpretability, and is therefore widely used in the risk management systems of traditional financial institutions.

[0037] However, this type of technology has significant shortcomings: First, it lacks a single data dimension. A company's financial statements and loan data can only reflect its cash flow, failing to cover its behavior in areas such as tax declarations, supply chain transactions, contract performance, legal proceedings, and public opinion dynamics. In particular, unstructured information, such as news reports, contract texts, invoice images, and social media comments, often cannot be directly utilized by existing models, thus limiting the input dimensions of the assessment results and making it difficult to comprehensively represent the company's true operating status.

[0038] Second, the modeling approach is static. Existing scoring models mostly rely on batch processing or periodic retraining mechanisms, typically updated monthly or quarterly. When sudden changes occur in the business environment, such as a sharp drop in cash flow, supply chain disruptions, or negative public opinion exposure, the model cannot respond in real time, resulting in a significant lag in risk detection and early warning. This static characteristic severely impacts the timeliness and foresight of the assessment results.

[0039] Third, there is insufficient risk transmission through graph-based risk mapping. Enterprises often have complex transactional and capital relationships with their upstream and downstream suppliers, controlling shareholders, and affiliated companies, and potential risks tend to propagate through multiple levels and paths. However, current technologies for relationship modeling largely remain at the stage of relational database or static graph construction, lacking the application of emerging graph learning methods such as graph neural networks and graph attention mechanisms. This results in risk signals not being effectively transmitted and amplified in the enterprise relationship graph, making it difficult to promptly detect systemic risks caused by related parties.

[0040] Fourth, data sharing and privacy protection are limited. A comprehensive assessment of a company's operational status often requires data fusion across banks, tax authorities, and industry platforms. However, under the existing framework, cross-institutional data flow is strictly limited by compliance and privacy protection requirements. The lack of support from privacy-preserving computation technologies such as federated learning, differential privacy, and secure multi-party computation prevents assessment models from fully utilizing heterogeneous data resources distributed across different institutions. This data barrier results in insufficient model coverage and robustness, making it difficult to form a truly comprehensive corporate profile.

[0041] Furthermore, existing technologies also have shortcomings at the system architecture level. For example, most existing systems use a centralized architecture for data storage and modeling, which poses a single point of failure risk; in high-concurrency scenarios, static scoring models are difficult to support large-scale real-time calls; and in terms of scalability, model updates often rely on manual configuration and lack an automated model evolution mechanism.

[0042] Based on this, embodiments of this application provide a method for assessing business status. The method includes: acquiring a multi-source feature set of a target enterprise; acquiring a pre-constructed heterogeneous relationship graph; extracting a subgraph of associations of the target enterprise based on the heterogeneous relationship graph, wherein the nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes, and multiple industry nodes, and the edges in the heterogeneous relationship graph include relationship attributes, including positive cooperation relationship attributes and negative competition relationship attributes; inputting the multi-source feature set and the subgraph of associations into a graph attention network for graph computation, wherein the graph attention network includes a temperature parameter dynamically adjusted based on external environmental parameters, the temperature parameter being used to adjust the weights of the edges during the graph computation process; and generating a risk score for the target enterprise based on the graph computation result, and assessing the business status of the target enterprise based on the risk score. According to embodiments of this application, a structured interactive interface supports multiple input formats, avoiding semantic ambiguity and logical risks associated with purely natural language descriptions. This allows business personnel to accurately express rule requirements in an intuitive and low-barrier manner, thereby improving user experience. Secondly, initial prompt words are generated based on structured requirement data, and the automatic comparison between the large model inference results and expected labels accurately identifies discrepancies and generates prompt word variants, effectively reducing manual debugging and trial-and-error steps and enabling rapid iterative optimization of prompt words. Furthermore, an iterative verification mechanism is used to evaluate the performance of prompt word variants, and the final prompt word is output when preset performance conditions are met. This not only ensures the accuracy and stability of the generated results but also avoids invalid calculations by introducing performance threshold control, thereby reducing computational resource consumption.

[0043] It should be noted that the business status assessment methods, apparatus, equipment, media, and program products defined in this application can be used in the fields of artificial intelligence technology, big data technology, and fintech technology, and can also be used in a variety of fields other than artificial intelligence technology, big data technology, and fintech technology. The application fields of the business status assessment methods, apparatus, equipment, media, and program products provided in the embodiments of this application are not limited.

[0044] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0045] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0046] Figure 1 The illustration shows an application scenario of the business status assessment method, apparatus, equipment, medium, and program product according to embodiments of this application.

[0047] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0048] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0049] In the embodiments of this application, the first terminal device 101 can be an example of the first device, and the second terminal device 102 and / or the third terminal device 103 can be an example of at least one second device. The first device and the second device can communicate collaboratively through an internal client mechanism to implement the data distribution and rendering logic described in the business status assessment method.

[0050] In some embodiments, the first device and at least one second device may be different display modules, windows or screens on the same computing terminal (such as a host), or multiple physical devices that work together through a network, such as different client instances deployed on a desktop computer, tablet terminal or mobile device respectively.

[0051] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smart mobile terminals, tablet computers, laptop computers, and desktop computers.

[0052] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0053] It should be noted that the business status assessment method provided in this application embodiment can generally be executed by server 105. Correspondingly, the business status assessment device provided in this application embodiment can generally be located in server 105. The business status assessment method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the business status assessment device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0054] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0055] The following will be based on Figure 1 The described scene, through Figures 2-4The operational status assessment method of the disclosed embodiments is described in detail.

[0056] Figure 2 A flowchart illustrating an operational status assessment method according to an embodiment of this application is shown schematically.

[0057] like Figure 2 As shown, the business status assessment method 200 of this embodiment includes operations S210 to S240.

[0058] In operation S210, a multi-source feature set of the target enterprise is obtained.

[0059] In the embodiments of this application, original information related to the target enterprise can be extracted from multiple heterogeneous data sources, and the information can be structured, standardized and vectorized to form a unified feature representation that can be used as input for subsequent graph computing models.

[0060] In some embodiments, a unified data dictionary and identification system can be established for the target enterprise, using the enterprise's unified social credit code as the primary key. Data from sources such as the target enterprise's financial statements, loan ledgers, mortgage and pledge registrations, tax declaration records, invoice circulation records, business registration and change records, and publicly available public opinion texts can be integrated into the same feature space. For data from different sources, alignment and conversion can be performed according to timestamp and caliber differences to unify them to the set accounting period and unit of measurement.

[0061] In some embodiments, historical data can be extracted and verified in batches; near real-time data can be stored on an hourly or daily basis via a message queue, and metadata tags such as collection time, source, and processing serial number can be attached to each record. For multiple reports of the same business fact, version management can be carried out according to timestamp and source credibility, only the latest valid version is imported into the feature set, while historical versions are retained for subsequent time window features and backtracking audits.

[0062] In some embodiments, for public opinion and announcement texts, word segmentation, stop word filtering, synonym merging and domain dictionary expansion can be used to construct a keyword set for the theme of enterprise operation and compliance, and calculate word frequency and polarity score; for long texts (news, annual reports, announcements), sentence vectors or segment vectors can be introduced to form a time series of public opinion intensity in chronological order, which is used to measure the lagging / leading correlation with tax and cash flow series, thereby obtaining cross-source linkage features.

[0063] In some embodiments, outlier handling and data quality control processes can also be established. For example, quantile truncation or robust scaling can be used for continuous features, and rare values ​​can be merged for discrete features; another example is to use a weighted fusion method for multi-source observations of the same indicator and record the source weights; to retrospectively revise historical data and record the revision time and differences between the previous and subsequent values ​​to ensure feature temporal consistency; and to set logical verification rules for important features, and generate conflict markers and alternative features when conflicts are found to avoid the propagation of anomalies.

[0064] In some embodiments, different feature templates and observation windows can be configured for micro and small enterprises and large enterprises respectively. For example, the feature template for micro and small enterprises focuses more on cash flow, tax declaration stability, and invoice credibility, while the template for large enterprises adds features related to mergers and acquisitions, overseas entities, cross-regional operations, and complex equity structures. Regarding observation windows, micro and small enterprises use shorter periods to capture rapid changes, while large enterprises use multiple periods in parallel to take into account long-term trends. Different templates are mapped to a unified feature naming and data type specification when entered into the database to ensure direct consumption in subsequent graph calculation and scoring stages.

[0065] In operation S220, a pre-constructed heterogeneous relationship graph is obtained, and the association subgraph of the target enterprise is extracted based on the heterogeneous relationship graph. The nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes, and multiple industry nodes. The edges in the heterogeneous relationship graph include relationship attributes, which include positive cooperation relationship attributes and negative competition relationship attributes.

[0066] In the embodiments of this application, the construction of the heterogeneous relationship graph can be carried out on the entire enterprise group. Specifically, firstly, enterprise entity information is extracted from multiple data sources such as business registration, taxation, finance, and industry statistics, and different enterprises are uniformly mapped as enterprise nodes; then, association information is extracted from third-party entities such as shareholders, holding companies, supply chain partners, and investors to establish corresponding related party nodes; and finally, industry nodes are established from industry classification standards and industrial chain distribution information. Various types of nodes are connected through cross-source data, and relationship attribute labels are assigned to edges according to the nature of the relationship.

[0067] For example, the positive cooperative relationship attribute of an edge can be determined based on supply chain transaction data between enterprises. When there are continuous and stable transactions between two enterprises, and the transaction amount exceeds a preset threshold, an edge with a positive cooperative relationship attribute can be established in the heterogeneous relationship graph. Specifically, if enterprise A and enterprise B have a procurement contract relationship of more than three years and the contract fulfillment rate exceeds 90%, then the edge can be assigned a positive cooperative relationship attribute, with the contract amount or transaction frequency as the weight parameter.

[0068] For example, the negative competitive relationship attribute of an edge can be determined based on market monitoring and industry statistics. When two companies are in the same sub-sector and their main product market shares overlap, an edge with a negative competitive relationship attribute is created in the heterogeneous relationship graph. For instance, if company C and company D have a market share overlap of more than 60% in the smartphone market, the edge is labeled as a negative competitive relationship, and different weights can be set according to the degree of market overlap.

[0069] In some embodiments, the positive cooperative relationship attribute can also encompass investment and financing relationships and equity relationships. When one company invests in another company with an equity stake exceeding a certain threshold, an edge with a positive cooperative attribute is established between the two. Such edges can reflect controlling or minority stake relationships and serve as important channels for risk transmission. For example, if company E holds 30% of the equity in company F, the edge between them is labeled as having a positive cooperative attribute, and this edge is assigned a higher weight during graph calculation.

[0070] In some embodiments, negative competition attributes can also be identified through judicial and administrative penalty information. For example, when two companies are involved in a business competition lawsuit or are simultaneously penalized due to an antitrust investigation in the same industry, an edge with a negative competition attribute can be established between them. For instance, if company G and company H are involved in intellectual property litigation that has lasted for more than two years, the edge is marked as having a negative competition attribute, and the amount of the litigation is used as an additional weighting feature.

[0071] In some embodiments, the relational attributes of edges can also be dynamically adjusted based on public opinion data. For example, when online public opinion shows that two companies have issued cooperation announcements or reached strategic agreements, a positive cooperation edge can be temporarily added; when public opinion shows that two companies are widely reported on due to price wars or vicious competition, a negative competition edge can be established. The weight of such edges can decay over time, gradually weakening and eventually disappearing after the public opinion heat subsides.

[0072] In some embodiments, the construction of edges relies not only on static business registration information but also on dynamic behavioral data. For example, transaction links between enterprises and suppliers / customers can be identified from transaction flows and invoice circulation, and assigned a positive cooperative relationship attribute; enterprises with high overlap in sales areas can be identified from market monitoring data, and assigned a negative competitive relationship attribute. These edges from different sources are superimposed on the graph, enabling the heterogeneous relationship graph to simultaneously represent static structure and dynamic behavior.

[0073] To ensure the stability and scalability of the global heterogeneous relationship graph, multi-dimensional labels can be introduced for nodes and edges. For example, enterprise nodes can be labeled with their size, type, registered location, and credit rating; affiliate nodes can be labeled with the strength of their connections with multiple enterprises; and industry nodes can be labeled with industry prosperity indicators. Edge labels can include cooperation strength, contract amount, market share overlap, etc.

[0074] In the embodiments of this application, the constructed heterogeneous relationship graph can be stored as a global graph. When a risk assessment of a target enterprise is required, the associated subgraph of the target enterprise can be extracted from this global heterogeneous relationship graph. By setting extraction rules, such as taking the target enterprise as the central node and expanding outwards to first- or second-order neighbors, and filtering positive cooperative relationships or negative competitive relationships based on the relationship attributes of the edges, the associated subgraph of the target enterprise can be obtained. This approach ensures both the local specificity of the subgraph and preserves the integrity and reusability of the global graph.

[0075] Furthermore, the extraction of relational subgraphs can be dynamically adjusted by incorporating time windows. For example, when analyzing short-term risks for a company, a subgraph can be constructed by selecting only suppliers, customers, and competitive events from the past year; when conducting long-term trend analysis, the time window can be expanded to include relational nodes within a three- or five-year timeframe. This dynamic windowing subgraph extraction method can better balance computational overhead and analytical accuracy, while improving the timeliness and stability of risk assessment results.

[0076] To avoid computational complexity issues caused by excessively large subgraphs, edge weight thresholds can be set during the extraction process, retaining only edges with weights greater than the threshold and their corresponding nodes. For example, in positive partnerships, only partners with annual transaction volumes exceeding a certain amount are retained; in negative competitive relationships, only competitors with market share overlap exceeding a certain percentage are retained. This filtering method ensures that the associated subgraph covers the main risk relationships while reducing computational resource consumption and improving overall computational efficiency.

[0077] In operation S230, the multi-source feature set and the associated subgraph are input into a graph attention network for graph computation. The graph attention network includes a temperature parameter that is dynamically adjusted based on external environmental parameters. The temperature parameter is used to adjust the weight of the edges during the graph computation process.

[0078] In some embodiments, graph computation can employ an attention-based graph neural network structure, taking the multi-source feature set of the target enterprise and the node features and edge attributes of its associated subgraphs as input. In this process, the weight of each edge depends not only on the feature similarity between adjacent nodes but also on a temperature parameter. The temperature parameter can be dynamically adjusted based on external environmental parameters to alter the steepness of the attention distribution, thereby highlighting high-risk relationships during periods of significant economic fluctuation and maintaining the smoothness of overall relationships during periods of economic stability. This ensures the model can adapt to environmental changes and optimize its sensitivity to risk identification.

[0079] In some embodiments, graph computation can also employ a graph convolutional network-based structure. Specifically, when processing the associated subgraph, the features of the target enterprise and its neighboring nodes can be aggregated through convolution operations, with the convolution kernel weights determined jointly by the adjacency matrix and the feature matrix. By stacking multiple layers of convolutional operations, multi-level relationships between the target enterprise and distant nodes can be gradually captured. In this implementation, the temperature parameter can be applied to the normalization coefficient of the convolutional layer to dynamically adjust the contribution of different adjacent edges in the aggregation process, thereby maintaining coordination with external environmental parameters.

[0080] In some embodiments, graph computation can employ graph isomorphic networks, which can distinguish different subgraph structures through stronger expressive power, avoiding the loss of discriminative power in high-order neighbor aggregation of traditional convolution. In this approach, the temperature parameter can be used as a scaling factor before nonlinear activation to control the signal strength of different edges during aggregation, enabling the model to more finely distinguish potential risk paths when facing complex competitive or cooperative relationships.

[0081] In some embodiments, graph computation can also be implemented using graph temporal neural networks. This approach not only models the static relationships between nodes and edges but also incorporates information such as edge occurrence time and event frequency into the calculation. In this scenario, the temperature parameter can be combined with a time decay factor to assign higher weights to recent events and gradually decay weights to older events, making the risk transmission more consistent with the actual dynamic patterns over time.

[0082] In the embodiments of this application, the temperature parameter can be set not only based on a single macroeconomic indicator, but also by combining multiple external environmental factors. For example, the producer price index, interest rate level, and industry capacity utilization rate can be used as inputs to obtain a dynamic temperature parameter through linear or nonlinear function mapping. Changes in this temperature parameter can affect the calculation of attention weights or convolution factors, causing the weights of the edges to shrink or expand as a whole with changes in the external environment, thus achieving environmental adaptability without retraining.

[0083] In some embodiments, the adjustment of temperature parameters can also be time-sensitive. The system samples and smooths external environmental parameters within different time windows, allowing the temperature parameters to be updated gradually over time. In environments with short-term, drastic fluctuations, the model can respond quickly and increase its focus on risk relationships; while in long-term stable environments, the model can reduce the impact of short-term anomalies and maintain the stability of risk assessment.

[0084] Furthermore, the temperature parameter can be used not only to adjust edge weights across the entire graph, but also to set differentiated strategies for different relationship attributes. For example, in positive cooperative relationships, the temperature parameter can be kept low to enhance cooperative stability; in negative competitive relationships, the temperature parameter can be moderately increased to amplify the risk impact of competitive events.

[0085] In some embodiments, dynamic adjustment of the temperature parameter can also be achieved through a learnable mapping function. This function takes external environmental parameters as input, generates temperature values ​​via a multilayer perceptron or other nonlinear network, and is trained together with a graph neural network.

[0086] In operation S240, a risk score for the target company is generated based on the results of graph calculation, and the operating status of the target company is assessed based on the risk score.

[0087] In some embodiments, after completing the computation of the graph attention network, the system can transform the obtained target enterprise embedding vector into a risk feature representation, and perform nonlinear mapping through a fully connected layer or a multilayer perceptron to output a continuous risk score. The numerical range of the risk score can be set to a decimal between 0 and 1 to characterize the risk level of the target enterprise's operating status. Subsequently, the system compares the risk score with a preset threshold range to classify the target enterprise into multiple levels such as low risk, controllable risk, and high risk, and uses the result as the final output of the operating status assessment.

[0088] In some embodiments, the generation of risk scores is not limited to a single numerical form, but can be a multi-dimensional risk vector. For example, the system can output sub-scores such as financial stability risk, tax compliance risk, supply chain dependence risk, and market competition risk, and then generate a comprehensive risk score by weighted fusion or hierarchical aggregation of these sub-scores.

[0089] In some embodiments, the risk score generation process can also be dynamically adjusted by incorporating historical data. Specifically, while generating a real-time risk score, the system can compare the score result with the target company's risk score series over several past periods and use moving averages or weighted smoothing methods to correct for abnormal fluctuations.

[0090] In some embodiments, risk scores are used not only as internal quantitative indicators but also for comparison with external benchmark models or industry standards. For example, a target company's risk score can be compared with the risk distribution of its peers in the industry to generate a relative position indicator, such as being in the top 25%, below the median, or at the bottom of the industry.

[0091] In some embodiments, the output of risk scoring can also be presented in a visual form, such as displaying the trend of risk score changes of the target enterprise as a time series curve, displaying the sub-score distribution of different risk dimensions as a radar chart, or displaying the risk transmission path between the enterprise and its affiliates in a graph network as a heat map.

[0092] According to embodiments of this application, by acquiring a multi-source feature set of the target enterprise, a complete data input can be formed based on multi-dimensional information such as finance, taxation, business registration, and public opinion, avoiding assessment bias caused by relying solely on a single financial statement. By constructing a heterogeneous relationship graph containing enterprise nodes, related party nodes, and industry nodes, and extracting related subgraphs closely related to the target enterprise, the system can characterize the enterprise's position and role in a multi-level network, revealing potential cooperative and competitive relationships. Furthermore, a temperature parameter dynamically adjusted based on external environmental parameters is introduced in the graph calculation stage, enabling the edge weights to adaptively adjust with changes in the macroeconomic and industry environment, thereby improving the flexibility and sensitivity of risk transmission modeling. Finally, while generating a risk score, the system maps the results to an operational status assessment conclusion, providing a quantitative representation of the enterprise's risk level and supporting dynamic comparisons based on time and relationships, enhancing the accuracy, interpretability, and computational efficiency of the overall assessment.

[0093] The operational status assessment method of this application will be specifically described below by way of preferred embodiments.

[0094] In the embodiments of this application, a method for pre-constructing heterogeneous relationship graphs is proposed to support graph computation tasks for subsequent enterprise operation status assessment. This enables joint modeling of multi-source heterogeneous data while ensuring data security and privacy, and comprehensively depicts the relationships between enterprises, related parties, and industries through graph structures.

[0095] Figure 3 The flowchart illustrating a method for pre-constructing a heterogeneous relationship graph according to some exemplary embodiments of this application is shown schematically.

[0096] like Figure 3 As shown, the method for pre-constructing heterogeneous relationship graphs includes operations S310 to S340.

[0097] In operation S310, multi-source heterogeneous data of an enterprise is obtained from multiple data providers, and the multi-source heterogeneous data of the enterprise is processed in a structured manner to obtain feature data.

[0098] In the embodiments of this application, data providers may include financial institutions, tax authorities, business registration departments, industry associations, and third-party public opinion platforms. Because the sources of enterprise multi-source heterogeneous data vary significantly, and the data may be in the form of tables, text, images, or logs, it is necessary to perform structured processing through steps such as cleaning, deduplication, standardization, and feature extraction. For example, financial data can be used to extract asset-liability ratios and cash flow indicators; tax data can be used to extract filing frequency and tax payment stability; business registration data can be used to extract equity structure and changes; and public opinion data can be used to identify positive or negative corporate events through natural language processing technology. Through structured processing, data of different formats and dimensions can be uniformly mapped into a comparable and computable feature space.

[0099] In operation S320, based on the feature data, local parameter updates are performed on the multiple data providers to obtain multiple updated parameters.

[0100] Specifically, each data provider can independently run the model update process based on its locally held feature data, generating update parameters corresponding to that data provider. For example, banks can update risk parameters using lending behavior, repayment cycles, and collateral information; industrial and commercial banks can update operational stability parameters using equity change records and pledge frequency; and public opinion monitoring can update external risk parameters based on litigation and penalty events identified by text mining models. In this way, each participating party only needs to transmit update parameters without sharing the original data, thereby avoiding privacy leaks and reducing the security risks of cross-institutional data transmission.

[0101] In operation S330, the multiple update parameters are aggregated based on horizontal federated learning to form a joint feature representation.

[0102] Horizontal federated learning can be accomplished by a central coordinator or through protocols based on secure multi-party computation. The aggregation process can employ weighted averaging, gradient fusion, or other secure aggregation strategies. To ensure that data providers contribute effective information despite differences in feature size and sample distribution, a dynamic weight allocation mechanism based on data quality or sample size can be introduced during aggregation. The resulting joint feature representation does not contain any original information from a single provider but rather integrates statistical features from multiple dimensions, including financial, tax, business, and public opinion data, comprehensively reflecting the target company's behavioral patterns across different dimensions.

[0103] In operation S340, the joint feature representation is used to determine the multiple enterprise nodes, the multiple related party nodes, the multiple industry nodes, and the relationship attributes to construct the heterogeneous relationship graph.

[0104] Specifically, enterprise nodes represent individual enterprise entities, affiliate nodes represent third-party entities with direct or indirect economic ties to the enterprises, and industry nodes reflect the industry environment in which the enterprises operate. During edge attribute determination, if long-term stable transactions, equity investments, or cooperative relationships exist in the joint feature representation, the corresponding edge is marked as a positive cooperative relationship attribute; if overlapping market share, patent litigation, or competitive behavior are reflected in the joint feature representation, the corresponding edge is marked as a negative competitive relationship attribute. The constructed heterogeneous relationship graph not only presents the direct connections between enterprises but also reveals multi-level relationships across industries and affiliates, providing an intuitive network structure for subsequent risk transmission modeling.

[0105] It should be noted that, to balance the needs of data privacy protection and cross-institutional relationship modeling, the construction of edge relationships can adopt a combination of distributed edge extraction and centralized fusion. Specifically, each data provider identifies the inter-enterprise edge relationships it can confirm based only on locally observable data, and generates a local edge candidate set for these edges, including edge type, weight estimate, and confidence index. This local edge candidate set will be submitted to the central coordinator or shared through a cryptographic federation protocol without containing the original data. The central coordinator will then uniformly fuse and resolve conflicts among the edge candidates submitted by multiple parties.

[0106] Through the above process, the joint feature representation formed by horizontal federated learning can effectively integrate information from different dimensions, thereby obtaining a more comprehensive expression of enterprise characteristics than a single data source. The heterogeneous relationship graph constructed on this basis covers enterprise nodes, related party nodes, and industry nodes, and, combined with edge attributes of positive cooperative relationships and negative competitive relationships, can reveal the position and role of enterprises in complex economic networks from a global perspective.

[0107] In some embodiments, to further enhance privacy protection in cross-organizational data collaboration, the system can introduce pseudo-anonymous identifiers and differential privacy mechanisms during the construction of heterogeneous relationship graphs. Specifically, multiple data providers can generate pseudo-anonymous identifiers through a shared key to replace the original enterprise identifiers. This ensures that even during cross-organizational data interaction or parameter aggregation, external observers cannot directly identify the true identity of the enterprise, effectively preventing the leakage of sensitive information. In this approach, each enterprise entity is labeled with a unified pseudo-anonymous identifier during joint modeling, ensuring data consistency and traceability while concealing the true identity characteristics of the enterprise.

[0108] In some embodiments, the system can also introduce differential privacy noise based on a preset privacy budget during the aggregation of update parameters. Specifically, before the update parameters from multiple data providers are processed centrally, noise can be injected into their parameter gradients or statistical features respectively, making it impossible to accurately infer the local information of a single data provider. Through the differential privacy mechanism, the resulting joint feature representation still retains the distribution characteristics of the enterprise group in an overall statistical sense, but the original data cannot be reversed at the level of a single enterprise. Furthermore, the privacy-processed joint feature representation is used to determine relationship attributes, such as determining whether there are stable cooperative or competitive relationships between different enterprises, thereby ensuring both the accuracy and security of the relationship graph construction.

[0109] By introducing pseudo-anonymous identifiers and differential privacy mechanisms, the embodiments of this application not only achieve privacy protection in cross-institutional data joint modeling, but also enhance the applicability of the system in the face of regulatory compliance requirements and real-world application environments, enabling the constructed heterogeneous relationship graph to still have high-quality relationship expression capabilities while ensuring data security.

[0110] In the embodiments of this application, local structures closely related to the target enterprise can also be extracted from the global heterogeneous relationship graph to provide targeted input for subsequent graph calculation and business status assessment. This can effectively reduce the redundant overhead of the full graph calculation, while improving the calculation efficiency and the accuracy of the assessment while retaining the key risk transmission path.

[0111] Figure 4 The flowchart illustrates a method for extracting an association subgraph of a target enterprise according to some exemplary embodiments of this application.

[0112] like Figure 4 As shown, the method for extracting the association subgraph of the target enterprise includes operations S410 to S420.

[0113] In operation S410, with the target enterprise as the central node, related nodes within a preset order range are selected in the heterogeneous relationship graph based on the central node.

[0114] The preset order can be first-order, second-order, or higher, depending on the application scenario. For example, a first-order subgraph only includes nodes directly related to the target company, such as direct suppliers, customers, or competitors; a second-order subgraph further expands to include the related objects of these nodes, such as the upstream raw material suppliers of suppliers or the shareholder companies of competitors. By controlling the range of orders, the system can achieve a flexible balance between the scope of associations and computational complexity.

[0115] In operation S420, the weight of the edge between the central node and the associated node is compared with a preset threshold to determine the target node and the target edge, and the associated subgraph is constructed based on the target node and the target edge.

[0116] Specifically, when the weight of a certain side is higher than a preset threshold, it indicates that the relationship has a significant influence on risk transmission, and therefore the edge and its connected nodes are included in the association subgraph. If the weight is lower than the preset threshold, the edge may only reflect a weak connection or a noisy relationship and can be discarded. Finally, the target node and the target edge together constitute the association subgraph, which maintains the closeness between the target enterprise and the core relationship network while avoiding the computational redundancy brought about by large-scale full graph modeling.

[0117] In some embodiments, the preset threshold can be dynamically set according to different types of relationship attributes. For example, a higher threshold can be used for positive cooperative relationships to ensure that only long-term stable cooperative relationships are retained; a relatively lower threshold can be used for negative competitive relationships to capture potential competitive risk signals. Through this differentiated threshold setting, the association subgraph can better reflect the actual impact of different relationship attributes on risk transmission.

[0118] Using the methods described above, the embodiments of this application can quickly extract subgraphs closely related to the target enterprise from large-scale heterogeneous relationship graphs, effectively reducing unnecessary graph computations and improving overall computational performance and memory utilization efficiency. Because the extraction process incorporates dual constraints of order range and edge weight thresholds, the associated subgraphs not only contain the most important cooperation and competition relationships but also avoid interference from weak connections and noisy edges, thereby improving the accuracy and robustness of risk assessment. Furthermore, the relatively small size of the associated subgraphs allows subsequent graph attention networks to achieve more efficient risk transmission modeling with limited computational resources, providing a more targeted and interpretable input basis for enterprise operational status assessment.

[0119] In the embodiments of this application, after the graph attention network computation is completed, the results can be further transformed into a quantitative representation for risk assessment. Specifically, the output of the graph computation is first vectorized to form a risk feature vector. The risk feature vector integrates the multi-source feature information of the target enterprise and its network structure features in the associated subgraph, and can simultaneously reflect the enterprise's own operating attributes and the transmission effect of external relationships.

[0120] Risk feature vectors can be input into a multilayer perceptron network (MLB), which generates a risk score through multilayer nonlinear transformations. The MLB network can capture potential high-order nonlinear relationships within the risk feature vectors and comprehensively model features from different dimensions, thereby improving the accuracy of the risk score. The resulting risk score can be a continuous numerical value, used to characterize the risk level of a target company over a certain period.

[0121] Furthermore, the risk score can be compared with the target threshold parameter, and an assessment conclusion on the target company's operational status can be generated based on the comparison results. The target threshold parameter can be preset to multiple range values, such as low risk, controllable risk, and high risk levels, or it can be dynamically adjusted according to the external environment or training process.

[0122] Using the above method, complex graph calculation results are transformed into numerical risk feature vectors, and efficient nonlinear modeling is achieved through a multilayer perceptron. This not only improves the discriminative power of risk scores but also enhances the stability and interpretability of operational status assessments. Furthermore, based on the comparison between risk scores and target threshold parameters, the system can output clear assessment conclusions, supporting both quantitative analysis and intuitive application in practice, thus contributing to improved efficiency in risk identification and reliability in decision-making.

[0123] In the embodiments of this application, the target threshold parameter can be dynamically updated through reinforcement learning algorithms to adapt to changes in the external environment and enterprise risk characteristics.

[0124] Specifically, the state space of a reinforcement learning algorithm includes the distribution of historical risk scores and information on external economic indicators. The distribution of historical risk scores comprehensively reflects the changing trends of a company's risk level across different time windows, such as whether the risk score has increased over a continuous period, or whether there are cyclical fluctuations or sudden anomalies. Meanwhile, external economic indicators, as another type of input state, can encompass producer price indices, interest rates, industry prosperity indices, raw material price fluctuations, and macroeconomic policies. These indicators comprehensively characterize the overall economic environment's prosperity and volatility, and have a direct or indirect impact on the company's operational stability. By incorporating historical risk sequences and macroeconomic factors into the state space, the reinforcement learning agent can learn and optimize threshold strategies within a more comprehensive context.

[0125] In this implementation, the action space of the reinforcement learning algorithm can be composed of a set of candidate threshold parameters. Each candidate threshold parameter corresponds to a possible scheme in risk level classification, such as a boundary value distinguishing between low and high risk. During system execution, the reinforcement learning agent selects candidate threshold parameters based on the current state and applies them to the risk score determination. Subsequently, the system generates reward signals based on prediction accuracy, timeliness of warnings, or other performance indicators, and feeds them back to the reinforcement learning agent, thereby driving it to continuously optimize the threshold parameter selection strategy.

[0126] During reinforcement learning, the system evaluates the effectiveness of threshold selection through a reward function. The reward signal can consist of multiple performance metrics, including prediction accuracy, timeliness of risk warnings, balance between false positive and false negative rates, and the degree of deviation from the actual business operation status. After receiving reward feedback, the system gradually optimizes the threshold update strategy, enabling the target threshold parameter to gradually converge to the optimal solution over long-term operation.

[0127] In this way, the target threshold parameter can be gradually adjusted according to changes in historical risk distribution and the external environment, achieving adaptive dynamic updates. Compared with traditional methods that rely on manual setting or fixed interval division, this mechanism not only improves the flexibility of risk level classification, but also maintains the validity and robustness of the assessment results when economic fluctuations or significant changes occur in enterprise risk characteristics.

[0128] In other embodiments, the target threshold parameter can also be updated during the same training process through joint modeling. Specifically, a preset threshold parameter can be incorporated as the parameter of a learnable sub-network and modeled together with the parameters of the multilayer perceptron network used to calculate the risk score. During the training phase, while the risk feature vector is input into the multilayer perceptron network and the risk score result is output, the threshold parameter sub-network also participates in the computation in the same computation graph, so that the threshold update process and the risk score optimization process are synchronized.

[0129] Furthermore, the system can jointly update the parameters of the multilayer perceptron network and the target threshold parameters based on a unified loss function. This loss function not only considers the difference between the predicted results and the true labels, but also incorporates metrics such as the accuracy of risk level classification and early warning performance. Through this end-to-end joint training approach, the threshold parameters can be automatically adjusted during gradient descent optimization to maintain an optimal match with the risk score output.

[0130] Compared with the traditional separate training and parameter tuning methods, the embodiments of this application achieve the synergistic optimization of threshold parameters and risk prediction model parameters through joint modeling. This makes the setting of the threshold no longer static or dependent on human experience, but can automatically converge to the optimal range according to the data distribution and model performance during the training process, thereby improving the accuracy and adaptability of business status assessment.

[0131] In the embodiments of this application, to improve the interpretability of risk assessment results, the system introduces a causal inference process after graph computation. Specifically, the system identifies causal relationships between node features in a multi-source feature set and related subgraphs based on causal inference rules, thereby generating a causal relationship graph. This causal relationship graph not only represents the correlation between different features but also identifies potential causal dependencies through a causal inference mechanism. For example, the system can identify the impact path of tax filing frequency on cash flow stability or the transmission mechanism of supply chain concentration on market competition risk.

[0132] Furthermore, the set of influencing factors can be determined based on the generated causal relationship diagram. This set of influencing factors indicates key characteristics that significantly impact changes in risk scores, such as macroeconomic fluctuations, tax policy adjustments, abnormal corporate debt ratios, or increased industry concentration. Subsequently, the system generates explanatory information based on the set of influencing factors. This explanatory information may include the contribution of key variables, visualizations of risk transmission paths, or explanatory conclusions comparing the company's historical behavior.

[0133] By introducing causal reasoning, the embodiments of this application can effectively avoid misjudgments that may arise from relying solely on correlation analysis, making the risk assessment results more interpretable and credible. In practical applications, this mechanism can assist financial institutions and regulatory authorities in decision-making, and also provide enterprises with a reference for improving their business strategies.

[0134] In the embodiments of this application, to enhance the ability to distinguish different relationship attributes during graph computation, the system sets corresponding learnable weight matrices for edges with different relationship attributes during modeling. Specifically, the edges in the heterogeneous relationship graph not only reflect the connections between enterprises and related parties or industry nodes, but also have different relationship attributes, such as positive cooperative relationships and negative competitive relationships. During the computation of graph attention networks or other graph neural networks, the system defines a separate set of weight matrices for each relationship attribute and learns and updates these weight matrices during training.

[0135] When performing graph computation, the system can use corresponding weight matrices to weight and aggregate the node features of edges of a given type. For example, if an edge represents a positive cooperative relationship, the weight matrix corresponding to that relationship is used to map and combine the node features; if it represents a negative competitive relationship, a different weight matrix is ​​used for computation. Through this differentiated weighting strategy, edges of different relationship types can play different roles in information transmission and feature aggregation, thus avoiding the information loss problem caused by treating different relationships the same in traditional methods.

[0136] By adopting the above mechanism, the system can explicitly model the influence of different relationship attributes during graph computation, ensuring that risk transmission paths are not only constrained by structural connections but also reflect the differentiated contributions of relationship types to the results. This design enhances the model's expressive power and the accuracy of risk identification, while also improving the interpretability of the assessment results. It helps to differentiate the impacts of cooperation-enhancing factors and competition-intensifying factors on the company's operating status in subsequent analyses.

[0137] According to embodiments of this application, by acquiring a multi-source feature set of enterprises and combining it with multi-dimensional data such as financial, tax, business registration, and public opinion data, comprehensive coverage of information sources is achieved, overcoming the limitations of traditional methods that rely on single financial statements. By performing parameter updates locally on multiple data providers and forming joint feature representations based on horizontal federated learning, data privacy and security are effectively guaranteed, and the feasibility of cross-institutional collaboration is improved. The heterogeneous relationship graph constructed on this basis covers enterprise nodes, related party nodes, and industry nodes, and by distinguishing between positive cooperative relationships and negative competitive relationships, it can comprehensively depict the position and role of enterprises in complex economic networks. Furthermore, order constraints and edge weight threshold screening are introduced in the subgraph extraction stage, so that the associated subgraphs can cover key risk transmission paths while avoiding redundant information, thereby improving computational efficiency and memory utilization. In the graph calculation process, a temperature parameter dynamically adjusted based on external environmental parameters is introduced, so that the edge weights can adaptively adjust with changes in the macro environment. At the same time, combined with learnable weight matrices of different relationship attributes, differentiated modeling of relationship types is achieved. By vectorizing risk feature vectors and modeling with a multilayer perceptron, the system can generate accurate risk scores and dynamically update thresholds based on reinforcement learning, or simultaneously optimize threshold parameters and prediction model parameters through joint modeling, thereby improving the flexibility and stability of risk level classification. Furthermore, embodiments of this application introduce a causal reasoning mechanism, identifying a set of key influencing factors and outputting explanatory information simultaneously with the generation of evaluation results, making risk conclusions not only quantifiable but also interpretable.

[0138] Corresponding to the above-described business status assessment method, embodiments of this application also provide a business status assessment device.

[0139] Figure 5 A schematic block diagram of an operational status assessment device according to an embodiment of this application is shown.

[0140] like Figure 5 As shown, the business status assessment device 500 of this embodiment includes a data acquisition module 510, a related subgraph extraction module 520, a graph calculation module 530, and a business status assessment module 540.

[0141] The data acquisition module 510 can be used to acquire a multi-source feature set of the target enterprise. In one embodiment, the data acquisition module 510 can be used to perform the operation S210 described above, which will not be repeated here.

[0142] The subgraph extraction module 520 can be used to obtain a pre-constructed heterogeneous relationship graph, and extract the subgraph of the target enterprise based on the heterogeneous relationship graph. The nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes, and multiple industry nodes. The edges in the heterogeneous relationship graph include relationship attributes, including positive cooperation relationship attributes and negative competition relationship attributes. In one embodiment, the subgraph extraction module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0143] The graph computation module 530 can be used to input the multi-source feature set and the associated subgraph into a graph attention network for graph computation. The graph attention network includes a temperature parameter that is dynamically adjusted based on external environmental parameters. This temperature parameter is used to adjust the weights of the edges during the graph computation process. In one embodiment, the graph computation module 530 can be used to perform the operation S230 described above, which will not be repeated here.

[0144] The operational status assessment module 540 can be used to generate a risk score for the target enterprise based on the results of graph calculations, and to assess the operational status of the target enterprise based on the risk score. In one embodiment, the operational status assessment module 540 can be used to perform the operation S240 described above, which will not be repeated here.

[0145] According to an embodiment of this application, the business status assessment device 500 may further include a graph construction module. The graph construction module can be used to obtain multi-source heterogeneous enterprise data from multiple data providers, perform structured processing on the multi-source heterogeneous enterprise data to obtain feature data; based on the feature data, perform local parameter updates on the multiple data providers to obtain multiple update parameters; aggregate the multiple update parameters based on horizontal federated learning to form a joint feature representation; and use the joint feature representation to determine the multiple enterprise nodes, the multiple related party nodes, the multiple industry nodes, and the relationship attributes to construct the heterogeneous relationship graph.

[0146] According to embodiments of this application, the graph construction module can also be used to identify enterprise entities in the feature data based on pseudo-anonymous identifiers generated by the multiple data providers through a shared key; and / or, during the process of aggregating to form a joint feature representation, to introduce differential privacy noise based on a preset privacy budget to obtain a privacy-processed joint feature representation, wherein the privacy-processed joint feature representation is used to determine the relation attributes.

[0147] According to an embodiment of this application, the association subgraph extraction module 520 can also be used to select association nodes within a preset order range in the heterogeneous relationship graph based on the target enterprise as the central node; and to compare the weight of the edge between the central node and the association node with a preset threshold to determine the target node and the target edge, and to construct the association subgraph based on the target node and the target edge.

[0148] According to an embodiment of this application, the business status assessment module 540 can also be used to vectorize the results of graph calculation to obtain a risk feature vector; and input the risk feature vector into a multilayer perceptron network to calculate the risk score, and compare the risk score with a target threshold parameter to generate a business status assessment result for the target enterprise.

[0149] In some embodiments, the target threshold parameter is updated based on a reinforcement learning algorithm, wherein the state space of the reinforcement learning algorithm includes historical risk score distribution and external economic indicators, and the action space of the reinforcement learning algorithm includes a set of candidate threshold parameters.

[0150] In some embodiments, updating the target threshold parameter includes: using the preset threshold parameter as a parameter of a learnable sub-network and jointly modeling it with the parameter of the multilayer perceptron network during the same training process, wherein the target threshold parameter and the parameter of the multilayer perceptron network are jointly updated based on the same loss function.

[0151] According to an embodiment of this application, the business status assessment device 500 may further include a causal reasoning module. The causal reasoning module can be used to identify causal relationships between the multi-source feature set and the node features in the correlation subgraph based on causal reasoning rules, generating a causal relationship graph; and to determine a set of influencing factors based on the causal relationship graph, generating explanatory information based on the set of influencing factors.

[0152] According to an embodiment of this application, the graph computation module 530 can also be used to set corresponding learnable weight matrices for edges with different relational attributes during the graph computation process, and to weight the edges with different relational attributes based on the learnable weight matrices.

[0153] According to embodiments of this application, any multiple modules among the data acquisition module 510, the association subgraph extraction module 520, the graph calculation module 530, and the business status assessment module 540 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 510, the association subgraph extraction module 520, the graph calculation module 530, and the business status assessment module 540 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the data acquisition module 510, the associated subgraph extraction module 520, the graph calculation module 530, and the business status assessment module 540 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0154] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a business status assessment method according to an embodiment of this application.

[0155] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0156] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0157] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0158] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0159] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0160] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the business status assessment method provided in the embodiments of this application.

[0161] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0162] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0163] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0164] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0165] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0166] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0167] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A method for assessing operational status, characterized in that, The method includes: Obtain the multi-source feature set of the target enterprise; Obtain a pre-constructed heterogeneous relationship graph, and extract the associated subgraph of the target enterprise based on the heterogeneous relationship graph. The nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes, and multiple industry nodes. The edges in the heterogeneous relationship graph include relationship attributes, including positive cooperation relationship attributes and negative competition relationship attributes. The multi-source feature set and the associated subgraph are input into a graph attention network for graph computation. The graph attention network includes a temperature parameter dynamically adjusted based on external environmental parameters, which is used to adjust the edge weights during graph computation. A risk score for the target company is generated based on the results of graph computation, and the operational status of the target company is assessed based on the risk score.

2. The method according to claim 1, characterized in that, Pre-construct a heterogeneous relationship graph, including: Obtain multi-source heterogeneous data from multiple data providers, perform structured processing on the multi-source heterogeneous data, and obtain feature data; Based on the feature data, local parameter updates are performed at the multiple data providers to obtain multiple updated parameters; The multiple update parameters are aggregated based on lateral federated learning to form a joint feature representation; and The joint feature representation is used to determine the multiple enterprise nodes, the multiple related party nodes, the multiple industry nodes, and the relationship attributes to construct the heterogeneous relationship graph.

3. The method according to claim 2, characterized in that, The method further includes at least one of the following: Based on the pseudo-anonymous identifiers generated by the multiple data providers through a shared key, the enterprise entities in the feature data are identified; During the aggregation process to form the joint feature representation, differential privacy noise is introduced based on a preset privacy budget to obtain a privacy-processed joint feature representation, which is used to determine the relational attribute.

4. The method according to claim 1, characterized in that, The step of extracting the association subgraph of the target enterprise based on the heterogeneous relationship graph includes: Using the target enterprise as the central node, related nodes within a preset order range are selected in the heterogeneous relationship graph based on the central node; and The weights of the edges between the central node and the associated nodes are compared with preset thresholds to determine the target node and the target edge, and the associated subgraph is constructed based on the target node and the target edge.

5. The method according to claim 1, characterized in that, The graph-based computation results generate a risk score for the target company, and the operational status of the target company is assessed based on the risk score, including: The results of the graph computation are vectorized to obtain the risk feature vector; and The risk feature vector is input into a multilayer perceptron network to calculate the risk score. Based on the risk score and the target threshold parameter, an assessment result of the target enterprise's operating status is generated.

6. The method according to claim 5, characterized in that, The target threshold parameter is updated based on a reinforcement learning algorithm. The state space of the reinforcement learning algorithm includes historical risk score distribution and external economic indicators, and the action space of the reinforcement learning algorithm includes a set of candidate threshold parameters.

7. The method according to claim 5, characterized in that, The update of the target threshold parameter includes: The preset threshold parameter is used as the parameter of the learnable sub-network and is jointly modeled with the parameter of the multilayer perceptron network in the same training process. The target threshold parameter and the parameter of the multilayer perceptron network are jointly updated based on the same loss function.

8. The method according to claim 1, characterized in that, The method further includes: Based on causal inference rules, causal relationships are identified among the node features in the multi-source feature set and the associated subgraph, generating a causal relationship graph; and The set of influencing factors is determined based on the causal relationship diagram, and explanatory information is generated based on the set of influencing factors.

9. The method according to claim 1, characterized in that, During graph computation, a corresponding learnable weight matrix is ​​set for edges with different relational attributes, and the edges with different relational attributes are weighted based on the learnable weight matrix.

10. An operational status assessment device, characterized in that, The device includes: The data acquisition module is used to: acquire a multi-source feature set of the target enterprise; The associated subgraph extraction module is used to: obtain a pre-constructed heterogeneous relationship graph, and extract the associated subgraph of the target enterprise based on the heterogeneous relationship graph, wherein the nodes in the heterogeneous relationship graph include multiple enterprise nodes, multiple related party nodes and multiple industry nodes, and the edges in the heterogeneous relationship graph include relationship attributes, the relationship attributes including positive cooperation relationship attributes and negative competition relationship attributes; The graph computation module is configured to: input the multi-source feature set and the associated subgraph into a graph attention network for graph computation, wherein the graph attention network includes a temperature parameter dynamically adjusted based on external environmental parameters, the temperature parameter being used to adjust the weights of the edges during the graph computation process; and The operational status assessment module is used to: generate a risk score for the target company based on the results of graph calculations, and assess the operational status of the target company based on the risk score.

11. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

Images