Risk prediction method and device, electronic equipment, medium and program product

Abstract

The invention provides a risk prediction method and device, electronic equipment, a medium and a program product, and can be applied to the technical field of big data and the technical field of artificial intelligence. The method comprises the steps of obtaining target enterprise data, wherein the target enterprise data at least comprises public information source data, supply chain data and enterprise financial data; processing the public information source data by using a sequential relation graph to generate target public information features; performing embedded representation on the supply chain data by using a graph neural network to generate a supply chain embedded representation, and generating enterprise risk features based on the supply chain embedded representation and the enterprise financial data; and at least inputting the target public information features and the enterprise risk features into a multi-modal fusion prediction model, and outputting a risk prediction result.

Description

Technical Field

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

[0002] Existing enterprise risk prediction methods often rely on regression modeling of enterprise financial indicators, expert scoring systems, or industry experience rule bases, which have the following technical bottlenecks: First, based solely on enterprise financial data, it is difficult to capture the dynamic impact of changes in the external environment on enterprise risk levels, especially when facing sudden and rapidly evolving industry trends or public information events, resulting in a delayed response; Second, supply chain risks often have multi-hop transmission and upstream and downstream linkages, and traditional methods are unable to model the structural information of inter-enterprise transaction networks and their indirect impact on risks; In addition, it is difficult to integrate different types of data into a unified risk assessment framework, resulting in fragmented feature sets and limited predictive capabilities. Summary of the Invention

[0003] In view of the above problems, this application provides risk prediction methods, apparatus, equipment, media and program products.

[0004] According to a first aspect of this application, a risk prediction method is provided, the method comprising: acquiring target enterprise data, the target enterprise data including at least publicly available information source data, supply chain data, and enterprise financial data; processing the publicly available information source data using a time-series graph to generate target publicly available information features; embedding the supply chain data using a graph neural network to generate a supply chain embedded representation; generating enterprise risk features based on the supply chain embedded representation and the enterprise financial data; and inputting at least the target publicly available information features and the enterprise risk features into a multimodal fusion prediction model to output a risk prediction result.

[0005] According to an embodiment of this application, the target enterprise data further includes rule information data, and the method further includes: obtaining structured target rule features based on the rule information data by using rule extraction and field normalization; and inputting the target rule features, the target public information features, and the enterprise risk features into a multimodal fusion prediction model to output risk prediction results.

[0006] According to an embodiment of this application, the step of processing the publicly available information source data using a temporal relationship graph to generate target publicly available information features includes: extracting event information units from the publicly available information source data; identifying temporal, semantic, and / or causal relationships between the event information units to form a relationship set, wherein the event information unit includes an event subject, event behavior, event object, event time information, and source labeling; constructing a temporal relationship graph with the event information units as information event nodes and the relationship set as information relationship edges, wherein the information relationship edges have relationship type information; and based on the temporal relationship graph, using the relationship graph neural network to perform multi-round message propagation and adjacent node aggregation to generate the target publicly available information features.

[0007] According to an embodiment of this application, the multi-round message propagation and adjacent node aggregation using the relational graph neural network includes: generating an initial semantic embedding vector for each information event node in the temporal relational graph, encoding the event time information into a time feature vector, and concatenating the initial semantic embedding vector and the time feature vector to form a time-enhanced representation; updating the time-enhanced representation of adjacent nodes based on the relation type information in each round of message propagation; and fusing the updated time-enhanced representations after multiple rounds of message propagation to obtain the target public information feature.

[0008] According to an embodiment of this application, the step of using a graph neural network to embed the supply chain data and generate a supply chain embedded representation includes: constructing a supply chain graph based on the supply chain data, wherein the supply chain graph has enterprises as nodes and transaction relationships between enterprises as edges, the nodes have node attributes for representing enterprise operating characteristics, and the edges have edge attributes for representing sensitivity to changes in conditions; and based on the supply chain graph, aggregating the node attributes and the edge attributes using a graph neural network model to generate the supply chain embedded representation.

[0009] According to an embodiment of this application, generating enterprise risk characteristics based on the supply chain embedded representation and the enterprise financial data includes: calculating a cash flow pressure index based on the enterprise financial data; fusing the cash flow pressure index and the supply chain embedded representation to form an enterprise joint feature vector; and inputting the enterprise joint feature vector into a risk prediction model to generate the enterprise risk characteristics.

[0010] According to an embodiment of this application, the step of inputting at least the target publicly available information features and the enterprise risk features into a multimodal fusion prediction model and outputting a risk prediction result includes: concatenating the target publicly available information features and the enterprise risk features to form a multimodal input vector; and inputting the multimodal input vector into a multimodal fusion prediction model based on a gradient boosting tree algorithm and outputting the risk prediction result, wherein the risk prediction result includes at least the default probability change trend of the target enterprise, the operational pressure index, and support strategy recommendations.

[0011] According to an embodiment of this application, the method further includes: generating a source traceability identifier and a version marker for the rule information data; and monitoring correction or withdrawal events based on the source traceability identifier, and updating the rule information data and the target rule feature in response to detecting a correction or withdrawal event.

[0012] A second aspect of this application provides a risk prediction device, comprising: a data acquisition module for acquiring target enterprise data, the target enterprise data including at least publicly available information source data, supply chain data, and enterprise financial data; a publicly available information processing module for processing the publicly available information source data using a time-series graph to generate target publicly available information features; an enterprise data processing module for embedding the supply chain data using a graph neural network to generate a supply chain embedded representation, and generating enterprise risk features based on the supply chain embedded representation and the enterprise financial data; and a risk prediction module for inputting at least the target publicly available information features and the enterprise risk features into a multimodal fusion prediction model and outputting a risk prediction result.

[0013] 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.

[0014] 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.

[0015] 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.

[0016] According to embodiments of this application, by introducing a temporal relationship graph to structure and incrementally update publicly available information source data, the evolution of events and their temporal sequence can be captured in a timely manner when sudden changes occur in the external environment, thus compensating for the problem of delayed response to external shocks when relying solely on financial regression models. By constructing a graph on supply chain data and using graph neural networks to generate supply chain embedded representations, multi-hop transmission between upstream and downstream and structural relationships between enterprises can be explicitly modeled, addressing the shortcomings of traditional methods in characterizing indirect network effects. By unifying the modeling of target publicly available information features and enterprise risk features in a multimodal fusion prediction model, textual events, network structures, and financial indicators can be incorporated into the same evaluation framework, eliminating feature fragmentation and improving prediction robustness and foresight. Attached Figure Description

[0017] 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:

[0018] Figure 1 The illustrations depict application scenarios of risk prediction methods, apparatus, devices, media, and program products according to embodiments of this application.

[0019] Figure 2 A flowchart illustrating a risk prediction method according to an embodiment of this application is shown schematically.

[0020] Figure 3 This schematically illustrates a flowchart of a method for modeling publicly available information source data according to some exemplary embodiments of this application;

[0021] Figure 4 A flowchart illustrating a method for generating a supply chain embedded representation according to some exemplary embodiments of this application is shown schematically;

[0022] Figure 5 A schematic diagram illustrating the structure of a risk prediction device according to an embodiment of this application is shown; and

[0023] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a risk prediction method according to an embodiment of this application. Detailed Implementation

[0024] 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.

[0025] 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.

[0026] 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.

[0027] 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.).

[0028] The following will explain and clarify the terms used in this article.

[0029] Existing enterprise risk prediction methods mostly rely on regression modeling based on financial statements, expert scoring systems, or empirical rule bases. Their technical bottlenecks are mainly reflected in the following aspects.

[0030] First, the data sources are singular and mainly consist of low-frequency structured indicators (such as quarterly / annual financial items), making it difficult to depict the real-time impact and short-cycle fluctuations of the external environment on a company's status. When macroeconomic regulations change, industry trends, or publicly available information events evolve rapidly, models often can only absorb the effects that have already taken place after the fact, resulting in delayed signal acquisition, difficulty in time-series alignment, and insufficient sensitivity of prediction results to sudden scenarios. Second, corporate risks exhibit significant network externalities and multi-hop transmission characteristics: factors such as upstream supply disturbances, downstream demand contraction, and the substitutability of key components can amplify or mitigate risks through complex transaction relationship chains. Traditional linear models that assume independent and identically distributed samples are unable to express the structural dependencies and indirect impacts between companies, cannot distinguish between node attribute risks and network location risks, and are even less able to explicitly quantify the differentiated contributions of link strength and path length.

[0031] Secondly, multi-source heterogeneous data is difficult to integrate into a unified evaluation framework. Real-world scenarios simultaneously contain tabular financial indicators, semi-structured list data, and unstructured text from announcements, media, and research reports. These different sources vary significantly in terms of scope, time granularity, missing data patterns, and confidence levels. The lack of unified mechanisms for entity alignment, unit and currency normalization, time window alignment, and source credibility weighting easily leads to feature fragmentation and semantic conflicts, resulting in inconsistent model inputs, training / inference distribution drift, and decreased reproducibility of evaluation results. Furthermore, traditional linear or semi-parametric models have limited ability to express high-dimensional nonlinearity and temporal feedback, making it difficult to capture the coupling relationship between external changes, steady-state business operations, and market expectations. They also struggle to handle robust feature extraction and threshold alerts under high-frequency noise, leading to high false positive / false negative rates at critical points.

[0032] Finally, shortcomings remain at the engineering and governance levels: The lack of unified identification and traceability across systems makes data reconciliation difficult, and version changes and recalculations are costly; the model lacks explicit modeling and explanation of source uncertainties, feature contributions are untraceable, and it struggles to support auditing and compliance; the separation of offline training and online inference links restricts incremental updates, resulting in significant end-to-end latency, high resource consumption, and poor user readability and interactive experience. In summary, current methods are deficient in real-time performance, networked representation, heterogeneous data fusion, nonlinear time series characterization, engineering reproducibility, and interpretability, making it difficult to provide high-confidence, low-latency, and auditable risk prediction results in complex external environments.

[0033] Based on this, embodiments of this application provide a risk prediction method, the method comprising: acquiring target enterprise data, the target enterprise data including at least publicly available information source data, supply chain data, and enterprise financial data; processing the publicly available information source data using a time-series relationship graph to generate target publicly available information features; embedding the supply chain data using a graph neural network to generate a supply chain embedded representation; generating enterprise risk features based on the supply chain embedded representation and the enterprise financial data; and inputting at least the target publicly available information features and the enterprise risk features into a multimodal fusion prediction model to output a risk prediction result. According to embodiments of this application, by introducing a temporal relationship graph to structure and incrementally update publicly available information source data, the evolution of events and their temporal sequence can be captured in a timely manner when sudden changes occur in the external environment, thus compensating for the problem of delayed response to external shocks when relying solely on financial regression models. By constructing a graph on supply chain data and using graph neural networks to generate supply chain embedded representations, multi-hop transmission between upstream and downstream and structural relationships between enterprises can be explicitly modeled, addressing the shortcomings of traditional methods in characterizing indirect network effects. By unifying the modeling of target publicly available information features and enterprise risk features in a multimodal fusion prediction model, textual events, network structures, and financial indicators can be incorporated into the same evaluation framework, eliminating feature fragmentation and improving prediction robustness and foresight.

[0034] It should be noted that the risk prediction methods, apparatus, devices, media, and program products defined in this application can be used in the fields of big data technology, artificial intelligence technology, and fintech technology, and can also be used in various other fields besides big data technology, artificial intelligence technology, and fintech technology. The application fields of the risk prediction methods, apparatus, devices, media, and program products provided in the embodiments of this application are not limited.

[0035] 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.

[0036] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all offer users corresponding operation entry points, allowing them to choose to agree to or reject the automated decision results; if the user chooses to reject, the process proceeds to expert decision-making. Here, "expert decision-making" refers to the decision-making activities of personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0037] Figure 1 The illustrations depict application scenarios of risk prediction methods, apparatus, devices, media, and program products according to embodiments of this application.

[0038] 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.

[0039] 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).

[0040] 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 risk prediction method.

[0041] 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.

[0042] 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.

[0043] 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.

[0044] It should be noted that the risk prediction method provided in this application embodiment can generally be executed by server 105. Correspondingly, the risk prediction device provided in this application embodiment can generally be located in server 105. The risk prediction 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 risk prediction 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.

[0045] 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.

[0046] The following will be based on Figure 1 The described scene, through Figures 2-4 The risk prediction method of the disclosed embodiments is described in detail.

[0047] Figure 2 A flowchart illustrating a risk prediction method according to an embodiment of this application is shown.

[0048] like Figure 2 As shown, the risk prediction method 200 of this embodiment includes operations S210 to S240.

[0049] In operation S210, target enterprise data is obtained, which includes at least publicly available information source data, supply chain data, and enterprise financial data.

[0050] Regarding publicly available information data, embodiments of this application support data collection from multiple types of legitimate information channels to achieve a comprehensive understanding of the target company's external environment. Data sources may include publicly released information, analysis reports from research institutions or industry associations, industry dynamics disclosed by mainstream media, operational information disclosed by the company itself, professional databases, and authorized data service interfaces. The system employs a structure-aware parsing method to automatically segment and identify different formats of collected text, including main text, tables, and numerical paragraphs. It also extracts key time points from the text using a time information recognition model, such as the information generation time, initial disclosure time, and data entry time, thereby forming a traceable time-series record.

[0051] In some embodiments, semantic similarity matching and unique identifier encoding can be used to clean and merge repetitive or highly similar information, ensuring data uniqueness and accuracy. To improve data reliability, the system can also assess the credibility of information based on source authority, publication time, and cross-validation results, generating structured data fragments. If necessary, a manual review mechanism can be set up to re-examine fields with low credibility, thereby ensuring that the data input to the model has high timeliness and consistency.

[0052] Regarding supply chain data, embodiments of this application can focus on cooperation and transaction information between enterprises. Data can originate from an enterprise's internal supplier management system, procurement and contract system, warehousing and logistics system, customs declaration and transportation records, and can also be supplemented by compliant and authorized industry data platforms. The system extracts key characteristic indicators including transaction cycles, order fulfillment status, material flow, inventory turnover, and changes in customer demand, and performs field standardization processing. Through entity recognition technology, suppliers, customers, and logistics nodes from different sources can be mapped to a unified enterprise identifier, ensuring the relevance of various data in subsequent modeling.

[0053] In some embodiments, to support dynamic modeling and historical traceability, snapshots and change records of supply chain nodes, edge relationships, and time slices can be maintained. Specifically, a sliding time window mechanism can be used to capture and synchronize data updates, while a compensation merging strategy is employed for delayed data to ensure the continuity of the supply chain structure at the temporal level. This mechanism enables the system to perform incremental calculations without reloading all data, effectively improving computing performance and real-time response capabilities.

[0054] Regarding corporate financial data, in addition to basic financial statement indicators, dynamic indicators related to operating activities and cash flow can be introduced, such as accounts receivable and accounts payable turnover, inventory change rate, expense ratio, capital expenditure plans, credit rating changes, and financing records. Simultaneously, relevant external market indicators can be collected as needed, such as the company's performance in the capital market or publicly available rating results. For missing or outlier data, imputation and boundary correction methods based on historical data series can be provided to ensure data integrity and statistical stability.

[0055] In some embodiments, to integrate data from different sources, a multi-source data integration platform can be established to realize data extraction, transformation, loading, and standardized management. The multi-source data integration platform has master data management and entity association functions, and records the source, version, and processing path of the data.

[0056] In some embodiments, data can be updated in two modes: real-time and periodic. Public information data adopts a real-time stream processing architecture, which can update the data warehouse in a short time; supply chain and financial data are batch-ingested on an hourly or daily basis. Small-batch incremental update channels can be configured for important indicators to reduce latency while ensuring consistency in statistical methods. The system uses time window aggregation and latency control mechanisms to ensure that data from different sources are statistically consistent and time-series consistent during aggregation.

[0057] It should be noted that the acquisition of publicly available information source data, supply chain data, and corporate financial data involved in this application is all conducted through legal and compliant means. Furthermore, the data collection, transmission, and storage processes comply with relevant laws, regulations, and industry standards. Where necessary, data sharing agreements are signed or data anonymization and other technical means are used to ensure that data processing activities do not infringe upon the legitimate rights and interests of third parties, and to guarantee the compliance and security of the entire data acquisition process.

[0058] In operation S220, the publicly available information source data is processed using a time-series relationship diagram to generate target publicly available information features.

[0059] In the embodiments of this application, a relational graph neural network is used to perform temporal and structural modeling of publicly available information source data. Specifically, the relational graph neural network can first parse the publicly available information text into an event structure containing event trigger words, time, subject, and object, and then construct a graph structure based on these event nodes according to temporal order and semantic association. Through the iterative propagation mechanism of the graph neural network, it can capture the causal chains and contextual dependencies between publicly available information events, thereby generating feature representations that reflect the dynamic evolution trend of publicly available information. For example, when a company is widely reported by the media due to supply chain disruption, the relational graph neural network can not only capture the event feature of "supply chain disruption," but also identify its correlation with subsequent events such as "stock price decline" and "customer churn," thereby generating publicly available information features that reflect the overall risk situation.

[0060] In another embodiment, the temporal relational graph neural network can also jointly model multi-source heterogeneous public information. For the same event, data from news reports, research institution reports, industry announcements, or corporate announcements may contain different information dimensions. By abstracting this information from different sources into heterogeneous nodes (such as text nodes, image nodes, numerical nodes, etc.) and establishing corresponding semantic or temporal connections in the graph, the system can achieve complementary modeling of multi-source information. Taking the event of "facility shutdown" as an example, textual materials describe the reasons for the shutdown, statistical data reflects the trend of capacity changes, and image information shows the on-site status. The system achieves cross-modal semantic alignment through learning the aggregation weights of relation edges, thereby obtaining a more complete event representation.

[0061] In another embodiment, the temporal relationship graph neural network can incorporate sentiment or situational parameters into the node attributes by combining the semantic polarity analysis results of events. The system uses natural language processing technology to identify the emotional tendency in the reports and classify event nodes into positive, neutral, or negative categories. For example, when two reports about the same company describe "expanding production capacity" and "experiencing operational difficulties" respectively, the system will assign differentiated weights to event nodes with different emotional polarities, giving negative events a higher influence in the overall features.

[0062] In another embodiment, the temporal relationship graph neural network can be combined with external knowledge systems for contextual completion and semantic expansion. The system can introduce industry knowledge graphs or macroeconomic knowledge bases to expand information such as industry categories, related enterprises, regions, and supply chain relationships involved in event nodes. For example, when a record of "a certain region adjusting export quotas" appears in public information, the system can use the knowledge graph to associate it with the major export enterprises in that region and the upstream and downstream relationships in the industrial chain, thereby obtaining graph structure features with contextual understanding capabilities. This expansion mechanism helps the model better identify indirect risk propagation paths and potential associated impacts during prediction.

[0063] Furthermore, to reflect the breadth and intensity of the dissemination of public information, embodiments of this application can introduce propagation relationship edges into the time-series graph to describe the scale and speed of information diffusion. The system calculates the propagation weight based on the frequency of occurrence, propagation path, and interaction depth of the same event across different channels, and quantifies the diffusion effect through the aggregation of propagation relationships between nodes. When a certain type of information appears repeatedly on multiple platforms for a long period of time, its propagation weight increases accordingly, and the model considers it a high-attention event, so that the generated public information features not only reflect the importance of the event itself, but also its dissemination scope and lasting influence.

[0064] In operation S230, the supply chain data is embedded using a graph neural network to generate a supply chain embedded representation. Based on the supply chain embedded representation and the enterprise financial data, enterprise risk characteristics are generated.

[0065] In the embodiments of this application, graph neural networks can be used to perform structured modeling of supply chain data. Enterprises, suppliers, customers, and logistics links in the supply chain are abstracted as nodes, and a graph structure is constructed through the transaction, cooperation, and logistics relationships between nodes. Through the iterative propagation mechanism of the graph neural network, local information of a node can be fused with the features of its neighboring nodes, thereby generating an embedded representation reflecting the entire supply chain structure and dynamics. Based on this, the enterprise's financial data (such as balance sheets, cash flow statements, and profit and loss statements) is input and fused with the supply chain embedded representation to generate risk features that reflect the enterprise's operational stability and supply chain dependence. For example, when an upstream supplier's financial situation becomes abnormal, the graph neural network can transmit its risk impact to downstream enterprises through edge propagation relationships, thus revealing potential disruption risks in the risk features.

[0066] In another embodiment, graph neural networks are not limited to modeling direct transaction relationships but can also be extended to multi-layered supply chain networks. For example, a company's supply chain includes not only direct suppliers but also second- and third-tier suppliers. By introducing propagation computation of multi-hop neighbor nodes, graph neural networks can capture deep risk transmission paths within the supply chain. For instance, when a second-tier supplier of a key raw material experiences capacity shortages, the graph neural network can reflect its indirect impact on the target company in the embedded representation, making the generated company risk characteristics more systematic and forward-looking.

[0067] In another embodiment, the generation of the supply chain embedding representation can also incorporate time-series information, forming a time-series dynamic graph neural network modeling approach. Specifically, supply chain relationships are not static; some suppliers may change in different quarters, and transaction amounts and frequencies fluctuate over time. By embedding the time dimension into the input of the graph neural network, the supply chain embedding representation can be dynamically updated, thereby reflecting the evolution trajectory of the supply chain structure. Combined with dynamic changes in corporate financial data, such as fluctuations in quarterly cash flow and accounts receivable, more realistic corporate risk characteristics can be generated.

[0068] In another embodiment, the generation of supply chain embedding representations can also incorporate enhanced modeling based on external macro-environmental factors. For example, when a natural disaster or geopolitical conflict occurs in a region, the supply chain nodes in that region will be significantly impacted. Graph neural networks can use such environmental factors as additional features of nodes or edges during embedding generation, thus enabling enterprise risk characteristics to not only include the state of the enterprise and its supply chain itself, but also reflect the potential impact of the external environment on supply chain stability.

[0069] Furthermore, graph neural networks can be combined with knowledge graphs to expand the semantic relationships of supply chain data. For example, if a knowledge graph already records a supplier's main business, financing status, and partnerships, the graph neural network can use this information to enhance node modeling. Combined with corporate financial data, more comprehensive risk features can be generated. For instance, when identifying the node event of "a supplier's financing failure," the graph neural network can not only reflect its impact on direct customers but also extend to potential industry chain reactions through the knowledge graph, making the generation of risk features more accurate and multi-dimensional.

[0070] In operation S240, at least the target public information features and the enterprise risk features are input into the multimodal fusion prediction model, and the risk prediction results are output.

[0071] In the embodiments of this application, the multimodal fusion prediction model can adopt a multi-input structure based on deep neural networks, processing target public information features and enterprise risk features as independent input channels. During the fusion phase, the model uses an attention mechanism to weight different modal features, highlighting the modal information that is more critical in the current prediction task.

[0072] In another embodiment, the multimodal fusion prediction model can be based on a framework combining graph convolution and sequence modeling. Specifically, enterprise risk characteristics may inherently possess supply chain network relationships, while publicly available information features exhibit temporal dynamics. By employing a joint modeling approach using graph convolutional networks and long short-term memory networks during the fusion phase, the model can simultaneously capture the spatial correlations and temporal evolution between different modalities. For example, when supply chain disruptions and negative publicly available information occur simultaneously, the model can identify their cumulative effect and provide a higher risk warning score.

[0073] In some embodiments, the multimodal fusion prediction model may employ a cross-modal alignment mechanism based on a transformer structure. This mechanism establishes a shared semantic representation space within the model, mapping publicly available information features generated from the time-series relationship graph to supply chain risk features and corporate financial features into a unified vector space, and utilizes a cross-modal attention mechanism for information association and weighting. This enables semantic correspondence and complementarity between different types of data, avoiding bias or information loss caused by single-modal features. During training, the model aligns the semantic centers of different modalities, allowing the potential causal relationships between publicly available information signals, supply chain fluctuations, and financial indicators to be perceived and expressed in a unified space. For example, when changes in certain publicly available information cause a slowdown in the operation of upstream and downstream enterprises, accompanied by changes in cash flow, the model can capture this cross-modal time-series correlation, generating risk prediction results with higher credibility and more timely response.

[0074] In another embodiment, the multimodal fusion prediction model can further incorporate external knowledge systems to enhance its interpretability through knowledge distillation or rule constraints. The system can embed industry patterns, experiential judgments, or regulatory rules preset by domain experts as prior information into the model training process, guiding the update direction of model weights through soft constraints. For example, adding rules such as "the risk assessment weight needs to be increased when export restriction information is superimposed on supply chain node fluctuations" to the distillation layer allows the model to not only rely on data distribution characteristics during the learning process but also refer to expert experience to correct the results. This ensures that the prediction results maintain the adaptive capabilities of the deep model while possessing interpretability and verifiability, meeting the technical requirements of actual business supervision and compliance review.

[0075] In another embodiment, the multimodal fusion prediction model can maintain its timeliness and adaptability through online learning and dynamic update mechanisms. The system continuously monitors newly incoming data streams, and when it detects new trends in publicly available market information or anomalies in the supply chain, it can trigger incremental training of local parameters to achieve adaptive updates of model weights.

[0076] Furthermore, to achieve model stability and continuous optimization during long-term operation, embodiments of this application may employ a hierarchical update mechanism. The base layer maintains general features that remain in a long-term steady state, such as industry attributes and enterprise structural characteristics; the time series layer updates short-cycle changes rapidly, such as market fluctuations and enterprise transaction status; and the fusion layer dynamically adjusts the weight allocation of different modal features, enabling the model to maintain prediction accuracy and response speed even under external disturbances.

[0077] According to embodiments of this application, by constructing a complete technical process encompassing data acquisition, feature extraction, and multimodal fusion prediction, dynamic, structured, and interpretable modeling of enterprise risk is achieved. First, at the data level, the system can collect relevant enterprise information from multiple channels, forming a comprehensive and timely basic data system, overcoming the limitations of traditional methods that rely on a single data source and have long update cycles. Second, at the feature extraction level, regulatory information is transformed into structured features through semantic parsing and field normalization, allowing complex clauses to be quantitatively expressed; public information is modeled through time-series relationship graphs, depicting the temporal and logical relationships between events, and combining sentiment and dissemination attributes to generate time-varying features reflecting market dynamics; supply chain and financial data are embedded through graph neural networks, reflecting the stability of the supply chain and the characteristics of capital operation, forming a multi-dimensional enterprise risk profile. Finally, at the prediction level, the multimodal fusion model achieves unified representation and interactive modeling of cross-modal features, not only adaptively adjusting the importance weights of different features but also identifying potential nonlinear dependencies, generating interpretable, real-time, and forward-looking risk prediction results, thereby improving the overall model's computational efficiency and user experience.

[0078] The risk prediction method of this application will be specifically described below by way of preferred embodiments.

[0079] In the embodiments of this application, the rule information data contained in the target enterprise data can be structured through rule extraction and field normalization methods to generate target rule features that can be directly input into the model. This embodiment can transform unstructured rule text into formalized, computable feature expressions, thereby enhancing the interpretability and timeliness of the risk prediction model in regulatory change scenarios.

[0080] Specifically, rule information data can originate from management regulations, technical standards, industry guidelines, and revised notices. Such texts are often lengthy and complex, containing multiple levels of entries, constraints, and applicable explanations; directly modeling the entire text as a whole can easily lead to semantic confusion. Therefore, the embodiments of this application can first perform clause-level decomposition and logical structure identification of the rule text, independently segmenting the clause title, applicable objects, main clauses, and supplementary explanations to achieve fine-grained modeling. This process can effectively distinguish the scope of application and direction of action of different clauses.

[0081] In the embodiments of this application, the system can further combine semantic rule templates and entity recognition algorithms to automatically extract key elements from the rule text. By identifying and extracting the main entities (such as enterprise categories, industry types, and transaction objects), constraints (such as threshold parameters and scope of restrictions), enforcement measures (such as regulatory requirements and compliance standards), and exceptions (such as exemption conditions and temporary clauses) appearing in the clauses, the complex rule logic can be transformed into structured fields. Subsequently, the system normalizes these elements according to preset field standards, mapping the text content to a standardized tag set and a unified data structure to form target rule features.

[0082] In some embodiments, to enhance time series consistency, the system can attach time-related attributes, such as effective date, expiration date, and applicable period, to each rule element. By introducing time parameters, the model can identify the effectiveness and lasting impact of rules within a specific time period, thereby achieving dynamic modeling in the time dimension during the risk prediction phase. For example, when an industry standard takes effect in a specific quarter with a transition period, the system will encode this time window into the feature vector, making the prediction results more accurately reflect the phased characteristics of the rule's effect.

[0083] Furthermore, during rule extraction and field normalization, contextual semantic relationships can be incorporated to identify logical dependencies and hierarchical relationships between clauses. For clauses with references or revisions (such as "This clause is revised from clause × of the previous version of the standard"), the system can establish corresponding association identifiers to maintain the internal consistency of the regulatory system. This hierarchical modeling approach enables rule features to have higher interpretability and logical integrity when input into the multimodal fusion model.

[0084] According to embodiments of this application, in the input stage of the multimodal fusion prediction model, the generated target rule features can be input into the fusion prediction model along with multimodal features such as public information features and enterprise risk features. The model can dynamically adjust the feature contribution based on the correlation and weight between different features, thereby achieving joint modeling of changes in the external regulatory environment, dynamics of public information, and enterprise operating conditions.

[0085] This embodiment achieves the standardization and structured transformation of rule information data through clause-level segmentation, rule element extraction, field normalization, and time attribute annotation. Compared with traditional methods that rely on manual screening or keyword statistics, this technical solution can automatically identify and quantify the correlation and temporal effectiveness of complex regulatory information, enabling the target rule features to have higher expression accuracy and model adaptability.

[0086] In embodiments of this application, it may further include: generating source traceability identifiers and version markers for rule information data, and realizing dynamic monitoring and updating of rule data based on the source traceability identifiers.

[0087] Specifically, the sources of rule information data may include announcements from industry regulatory authorities, standard revision notices, technical specifications, or other authorized regulatory materials. To ensure the reliability and traceability of model input data, the system generates a unique source traceability identifier for each rule information entry upon data entry. This identifier may include fields such as document number, issuing organization, publication date, and collection date, used to record the data source and generation path. Simultaneously, the system assigns a version tag to each rule information entry or revision, managing the rule content, field structure, and feature vectors in a versioned manner, thereby ensuring that different versions of regulatory data can be clearly distinguished in subsequent modeling or backtracking analysis.

[0088] In some embodiments, the system can continuously monitor rule information correction, revision, or withdrawal events based on source traceability identifiers. When a rule file is detected to be updated, supplemented, or abolished, an update process can be automatically triggered to synchronously adjust the corresponding rule information data and its corresponding target rule features. For example, the update process may include: comparing the differences between the old and new versions, recording the changed fields, updating the timestamp, and regenerating the corresponding structured features to ensure that the feature set of the input model is always up-to-date.

[0089] In some embodiments, the system can also maintain a version history chain of rule information to record the content and time of each version change. Through this mechanism, the model can not only call upon the rule features of the current version but also trace back to historical versions when necessary, thereby enabling time-series-based feature evolution analysis. For example, when the model needs to assess the risk trends of an enterprise under different regulatory cycles, it can use the version history chain to trace back the regulatory change process to quantify the impact intensity at different stages.

[0090] Furthermore, the system can monitor and statistically analyze the reliability of the source and update frequency of rule data to form data quality assessment indicators. For data with frequently revised sources or low reliability, its influence in feature input can be reduced through confidence weighting, thereby improving the overall model's stability and prediction reliability.

[0091] The embodiments of this application enable end-to-end traceability, dynamic updating, and version controllable management of rule information data. This mechanism allows the system to maintain data consistency and timeliness even as the regulatory environment continues to change, thereby ensuring that the risk prediction model has continuous adaptability, version interpretability, and result verifiability in long-term operation, providing a high-quality, compliant, and traceable data foundation for subsequent enterprise risk analysis and decision support.

[0092] In the embodiments of this application, unstructured public information data can be transformed into a structured time-series relationship diagram representation, thereby extracting the inherent logical relationships and propagation patterns between public information events.

[0093] Figure 3 The flowchart illustrating a method for modeling publicly available information source data according to some exemplary embodiments of this application is shown schematically.

[0094] like Figure 3 As shown, the method for modeling publicly available information source data includes operations S310 to S330.

[0095] In operation S310, event information units are extracted from the publicly available information source data, and the temporal, semantic and / or causal relationships between the event information units are identified to form a relationship set. The event information unit includes event subject, event behavior, event object, event time information and source labeling.

[0096] In the embodiments of this application, event-level semantic parsing can be performed first on the collected public information text. Specifically, a pre-trained semantic recognition model or an extraction algorithm based on dependency parsing can be used to identify event information units containing complete semantic meaning from unstructured text. Each event information unit may include event subject, event behavior, event object, event time information, and source annotation.

[0097] Among them, the event subject can be used to indicate the party that issued or is responsible for the action, which can be an enterprise, institution, industry organization or other entity; the event action is used to describe the specific actions or changes performed by the subject, such as production delays, contract terminations, technology releases, etc.; the event object is used to represent the affected or related entities, which may include products, projects, upstream and downstream partners, etc.; the event time information is used to identify the specific time point or time range in which the event occurred; the source label is used to record the source of the event, such as the information collection channel, release time or issuing organization, to ensure the traceability and credibility of the event data.

[0098] After identifying multiple event information units, the relationship types between these units can be further determined through temporal feature analysis, semantic similarity calculation, and logical dependency modeling. Temporal relationships reflect the sequence or overlap of events over time; semantic relationships identify events that are similar, complementary, or contrasting in content; and causal relationships infer the chain of influence between events through contextual dependency and trigger word analysis. For example, if public information includes two events—limited raw material supply and delayed production plans—the system can identify a temporal sequence and a logical causal relationship between them.

[0099] In operation S320, a temporal relationship graph is constructed with the event information unit as the information event node and the relationship set as the information relationship edge, wherein the information relationship edge has relationship type information.

[0100] In the embodiments of this application, after the system completes the extraction of event information units and relationship identification, it can map each event unit as a node object in a graph structure, and use the temporal, semantic, and causal relationships between the nodes as connecting edges, thereby forming a temporal relationship graph with time attributes and semantic constraints. Specifically, each information event node contains corresponding multi-dimensional attributes such as event subject, event behavior, event object, event time, and source labeling, to comprehensively characterize the basic features and semantic content of the event.

[0101] To reflect the logical connections and evolutionary paths between events, the system constructs information relationship edges based on the relationship set. Each information relationship edge represents a specific type of connection between two event nodes and includes a relationship type information field to identify the association category corresponding to the edge, such as chronological order, semantic similarity, semantic complementarity, causal influence, etc. By introducing the relationship type field, the system can distinguish different association patterns in the subsequent modeling stage, avoiding misjudging unrelated events as strongly related, thereby improving the semantic accuracy of the graph structure.

[0102] In some embodiments, the information relationship edge may further carry weight information to characterize the importance or confidence level of different relationship types. The weight value can be automatically generated based on the similarity score, causal inference strength, or time interval calculation results in the relationship identification stage. For example, when two events occur consecutively in a short period of time and semantically present a causal triggering relationship, the system can assign a higher weight to the edge; conversely, for event pairs with a large time span and weak semantic association, the weight is relatively low.

[0103] In operation S330, based on the time-series relationship graph, the relationship graph neural network is used to perform multiple rounds of message propagation and neighbor node aggregation to generate the target public information features.

[0104] Specifically, relational graph neural networks embed the semantic information and relationship types of neighboring nodes into the target node by propagating messages multiple times within the graph structure. This ensures that the representation of each event node not only includes the event's own features but also incorporates its associated contextual information. Through multiple iterations, the model can gradually capture event dependencies over longer distances, thereby generating target public information features that reflect the overall structure of public information. For example, in the context of financial public information, the distant event of "external environment adjustment" may also affect "corporate financing difficulties" through multi-hop causal paths. Relational graph neural networks can identify and reflect this cross-level impact during the aggregation process.

[0105] According to embodiments of this application, dispersed, heterogeneous, and unstructured public information data can be transformed into a unified temporal relationship graph structure representation, which not only retains the core information of the event itself, but also highlights the temporal, causal, and semantic relationships between events. This allows for a deeper capture of the dynamic evolution and propagation patterns of public information events, thereby generating more comprehensive and predictive target public information features.

[0106] In the embodiments of this application, the event information unit includes at least event time information. When modeling publicly available information source data, event time information not only reflects the time of occurrence of the event, but also reflects the chronological order and duration of the event in the entire evolution chain of publicly available information. Therefore, introducing time information as a key feature in the process of modeling using a relational graph neural network helps to more accurately capture the dynamic relationships between publicly available information events.

[0107] Specifically, in the relational graph neural network, an initial semantic embedding vector is first generated for each event node in the temporal relational graph. This vector reflects the semantic features of the event content itself. Simultaneously, the event's temporal information is encoded into a temporal feature vector to characterize the event's temporal attributes. Subsequently, the initial semantic embedding vector and the temporal feature vector are concatenated to form a temporally enhanced representation, thus enabling each event node to contain not only semantic information but also temporal dimension features.

[0108] In each round of message propagation, the relational graph neural network can update the temporally enhanced representations of adjacent nodes based on relation type information. As described in the embodiments above, relation type information can include temporal relations, semantic relations, and causal relations. Different relation types assign different weights and update methods to adjacent nodes during propagation. For example, causal relations may have a stronger influence, while temporal relations emphasize the order of events. By dynamically adjusting the contribution of adjacent nodes during message propagation, the model can gradually aggregate and strengthen the multidimensional connections between events.

[0109] After multiple rounds of message propagation, the relational graph neural network fuses the updated time-enhanced representations to obtain the target public information features. These features not only encompass the semantic content of individual events but also integrate the event's evolutionary trajectory over time and its relationships within the event network, achieving a deep fusion of semantics and temporal sequence. This enhances the expressive power and predictive value of the extracted public information features.

[0110] Figure 4 The flowchart illustrating a method for generating a supply chain embedded representation according to some exemplary embodiments of this application is shown schematically.

[0111] like Figure 4As shown, the method for generating the embedded representation of the supply chain includes operations S410 to S420.

[0112] In operation S410, a supply chain graph is constructed based on the supply chain data, wherein the supply chain graph uses enterprises as nodes and the transaction relationships between enterprises as edges.

[0113] Specifically, the system first structures and identifies entities in the collected supply chain data to determine the role of each enterprise within the supply chain system, including upstream suppliers, midstream manufacturers, and downstream distributors. Each enterprise is mapped as a node object in the graph, accompanied by a set of node attributes characterizing its operational features. For example, node attributes may include indicators such as the enterprise's main business type, revenue scale, asset structure, and production capacity, reflecting the enterprise's basic operational status and robustness within the supply network.

[0114] Once the nodes are established, the connections between enterprises can be determined based on multi-dimensional data such as transaction records, contract data, order history, logistics tracking information, and payment settlement relationships, thereby constructing the network structure of the supply chain graph. Each edge represents an actual or potential business transaction between enterprises, reflecting the path of material flow, capital flow, or information flow in the supply chain system.

[0115] To improve the interpretability and modeling accuracy of the graph, the system introduces edge attributes into the edge structure to describe the sensitivity of the transaction relationship to changes in external conditions. Sensitivity to changes in conditions refers to the degree to which the transaction relationship is affected in terms of performance stability, delivery cycle, or cost structure when external conditions such as the macro environment, industry regulations, raw material prices, or exchange rates change. The system quantifies the sensitivity of each edge by statistically analyzing historical transaction data, including fluctuation characteristics, order fulfillment time differences, and raw material price transmission coefficients. For example, companies that rely on a single raw material supplier will have a higher sensitivity to price or supply changes in their transaction relationships; while diversified procurement companies will have relatively lower sensitivity.

[0116] In operation S420, based on the supply chain graph, the node attributes and edge attributes are aggregated using a graph neural network model to generate the supply chain embedded representation.

[0117] Specifically, graph neural networks fuse the attributes of each node with those of its neighboring nodes through a multi-round message passing mechanism, while also considering the transaction sensitivity represented by the edge attributes. During the iterative propagation process, the embedded representation of enterprise nodes is continuously updated, ensuring that their final features not only contain their own operational information but also fully integrate their relative position and external influences within the entire supply chain network.

[0118] By constructing a supply chain graph and introducing a graph neural network modeling mechanism, not only is a structured expression path for multi-dimensional transaction relationships between enterprises established, but explicit modeling of external shock factors is also achieved. Compared with traditional risk assessment methods based on the characteristics of the enterprises themselves, the embodiments of this application can incorporate the chain transmission effects in the supply chain network into the scope of risk identification, thereby improving the systematicness and accuracy of risk prediction.

[0119] In embodiments of this application, the process of generating enterprise risk characteristics based on supply chain embedded representation and enterprise financial data may include the following: First, calculating a cash flow stress index based on enterprise financial data. This cash flow stress index can comprehensively reflect the enterprise's cash inflows and outflows over a certain period, including indicators such as operating cash flow, investing cash flow, and financing cash flow. By quantifying their differences and fluctuations, it can intuitively characterize the enterprise's financial health and the potential debt repayment and operational pressures it may face.

[0120] Building upon this, the cash flow stress index can be integrated with the supply chain embedding representation to form a joint feature vector for the enterprise. The supply chain embedding representation reflects the enterprise's position in the supply chain network and the stability of its upstream and downstream relationships, while the cash flow stress index reflects the robustness of the enterprise's internal capital operations. Combining the two can simultaneously cover the enterprise's external supply chain risks and internal financial health in the feature space, thereby obtaining a more comprehensive and discriminative joint feature.

[0121] Subsequently, the joint feature vector of the enterprises can be input into a risk prediction model, such as a tree model, for modeling and learning. Tree models, such as Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), or Light Gradient Boosting Machine (LightGBM), can perform non-linear segmentation and hierarchical aggregation of input features, thereby generating more detailed risk profiles. By fitting historical samples during training, tree models can automatically learn the complex interaction patterns between cash flow pressure and supply chain relationships, and map them into quantifiable enterprise risk characteristics.

[0122] For example, the joint feature vector of the enterprise can be input into a pre-trained XGBoost model for inference. During the training phase, this model has already learned the mapping relationship between enterprise risk characteristics and historical risk events based on a large number of labeled samples. During the inference phase, XGBoost uses its multi-layered decision tree structure to split and aggregate the input features, automatically identifying nonlinear interactions between key variables, thereby generating high-confidence enterprise risk feature outputs. This output can serve as an important input for subsequent multimodal fusion risk prediction models, or it can be directly used in scenarios such as enterprise rating, risk warning, and credit assessment.

[0123] By calculating the cash flow stress index, integrating supply chain embedding representation, and combining tree modeling, we can not only capture the dual risk factors of internal financial pressure and external network dependence, but also generate more discriminative enterprise risk characteristics through efficient nonlinear model learning, thereby providing accurate data support for subsequent risk prediction and decision-making.

[0124] In the embodiments of this application, the multimodal fusion prediction model constructs a unified multimodal input vector by splicing the target's publicly available information features and the enterprise's risk features. This vector is then input into a GBDT-based fusion model for risk prediction, enabling dynamic assessment of the target enterprise's default trends, operational pressures, and coping strategies in a complex macroeconomic environment.

[0125] For example, the model can respond quickly to signals of changes in the external environment. For instance, when the input information is "a certain region adjusted its export product management measures at a specific time," the system can automatically identify and extract elements such as the corresponding product classification code, scope of application, and adjustment range. Combined with the effective date and execution cycle, it generates standardized structured features. Subsequently, through an entity alignment mechanism, the system can identify potentially affected entities from enterprise business data, such as manufacturing enterprises primarily engaged in export business and assembly enterprises focused on component procurement. By combining the enterprise's financial indicators with its position in the supply chain network, the system forms a multimodal fusion input, providing multi-dimensional data support for subsequent model inference.

[0126] During the model inference phase, the fusion prediction model utilizes hierarchical feature segmentation capabilities to perform nonlinear combination and weighted discrimination of data features from different sources, outputting fine-grained risk prediction results. For example, the model can predict that the operational stability of a manufacturing company will decrease by approximately 12% under specific environmental changes, based on factors such as rising raw material costs, high market concentration, and insufficient price transmission capabilities. For integrated enterprises, the model may generate multiple strategic recommendations, such as appropriately increasing interest rates and introducing risk mitigation measures like inventory pledging in supply chain finance; or, under the constraints of macroeconomic guidance and industrial planning goals, providing differentiated credit support recommendations to help maintain the coordinated stability and long-term resilience of the industrial chain.

[0127] Through the embodiments of this application, the multimodal fusion prediction model not only has the ability to integrate heterogeneous features, but also combines upstream and downstream structures, comprehensively analyzes changes in external conditions and financial stability to achieve accurate risk assessment at the enterprise level, and outputs customized strategy suggestions, which are applicable to various application scenarios such as bank risk control, regulatory early warning, and supply chain finance decision-making.

[0128] According to embodiments of this application, by introducing publicly available information time-series relationship graph modeling, supply chain graph embedding, and entity alignment mechanisms, the ability to model the transmission path of external shocks and the vulnerability structure of enterprises is improved. The multimodal fusion prediction model can not only output risk scores but also provide strategy suggestions, possessing high interpretability and business operability, significantly enhancing the real-time, forward-looking, and accurate nature of risk assessment.

[0129] Corresponding to the above-described risk prediction method, embodiments of this application also provide a risk prediction device.

[0130] Figure 5 A schematic block diagram of a risk prediction device according to an embodiment of this application is shown.

[0131] like Figure 5 As shown, the risk prediction device 500 of this embodiment includes a data acquisition module 510, a public information processing module 520, an enterprise data processing module 530, and a risk prediction module 540.

[0132] The data acquisition module 510 can be used to acquire target enterprise data, which includes at least publicly available information source data, supply chain data, and enterprise financial data. In one embodiment, the data acquisition module 510 can be used to perform the operation S210 described above, which will not be repeated here.

[0133] The public information processing module 520 can be used to process the public information source data using a time-series relationship diagram to generate target public information features. In one embodiment, the public information processing module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0134] The enterprise data processing module 530 can be used to embed the supply chain data into a graph neural network to generate a supply chain embedded representation, and based on the supply chain embedded representation and the enterprise financial data, generate enterprise risk characteristics. In one embodiment, the enterprise data processing module 530 can be used to perform the operation S230 described above, which will not be repeated here.

[0135] The risk prediction module 540 can be used to input at least the target public information features and the enterprise risk features into a multimodal fusion prediction model and output a risk prediction result. In one embodiment, the risk prediction module 540 can be used to perform the operation S240 described above, which will not be repeated here.

[0136] According to embodiments of this application, the risk prediction device 500 may further include a rule information data processing module. The rule information data processing module can be used to obtain structured target rule features based on the rule information data, utilizing rule extraction and field normalization; and to input the target rule features, the target public information features, and the enterprise risk features into a multimodal fusion prediction model, outputting risk prediction results.

[0137] According to an embodiment of this application, the rule information data processing module can also be used to generate a source traceability identifier and a version mark for the rule information data; and to monitor correction or withdrawal events based on the source traceability identifier, and in response to the detection of a correction or withdrawal event, to update the version of the rule information data and the target rule feature.

[0138] According to an embodiment of this application, the public information processing module 520 can also be used to extract event information units from the public information source data, identify the temporal relationship, semantic relationship and / or causal relationship between the event information units, and form a relationship set, wherein the event information unit includes event subject, event behavior, event object, event time information and source label; construct a temporal relationship graph with the event information units as information event nodes and the relationship set as information relationship edges, wherein the information relationship edges have relationship type information; and based on the temporal relationship graph, use the relationship graph neural network to perform multi-round message propagation and adjacent node aggregation to generate the target public information feature.

[0139] According to an embodiment of this application, the public information processing module 520 can also be used to generate an initial semantic embedding vector for each information event node in the time sequence graph, encode the event time information into a time feature vector, and concatenate the initial semantic embedding vector and the time feature vector to form a time-enhanced representation; in each round of message propagation, update the time-enhanced representation of adjacent nodes based on the relationship type information; and after multiple rounds of message propagation, fuse the updated time-enhanced representations to obtain the target public information feature.

[0140] According to an embodiment of this application, the enterprise data processing module 530 can also be used to construct a supply chain graph based on the supply chain data, wherein the supply chain graph has enterprises as nodes and transaction relationships between enterprises as edges, the nodes have node attributes for representing enterprise operating characteristics, and the edges have edge attributes for representing sensitivity to changes in conditions; and based on the supply chain graph, the node attributes and the edge attributes are aggregated using a graph neural network model to generate the supply chain embedded representation.

[0141] According to an embodiment of this application, the enterprise data processing module 530 can also be used to calculate a cash flow pressure index based on the enterprise financial data; fuse the cash flow pressure index and the supply chain embedded representation to form an enterprise joint feature vector; and input the enterprise joint feature vector into a risk prediction model to generate the enterprise risk characteristics.

[0142] According to an embodiment of this application, the risk prediction module 540 can also be used to concatenate the target public information features and the enterprise risk features to form a multimodal input vector; and input the multimodal input vector into a multimodal fusion prediction model based on the gradient boosting tree algorithm to output the risk prediction result, wherein the risk prediction result includes at least the default probability change trend of the target enterprise, the operational pressure index, and support strategy suggestions.

[0143] According to embodiments of this application, any multiple modules among the data acquisition module 510, public information processing module 520, enterprise data processing module 530, and risk prediction module 540 can be combined 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, public information processing module 520, enterprise data processing module 530, and risk prediction 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 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 public information processing module 520, the enterprise data processing module 530, and the risk prediction module 540 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0144] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a risk prediction method according to an embodiment of this application.

[0145] 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.

[0146] 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.

[0147] 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.

[0148] 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.

[0149] 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.

[0150] 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 risk prediction method provided in the embodiments of this application.

[0151] 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.

[0152] 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.

[0153] 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.

[0154] 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).

[0155] 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.

[0156] 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.

[0157] 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 risk prediction method, characterized in that, The method includes: Acquire target company data, which includes at least publicly available information source data, supply chain data, and corporate financial data; The publicly available information source data is processed using a time-series relationship diagram to generate target publicly available information features; The supply chain data is embedded using a graph neural network to generate a supply chain embedded representation. Based on the supply chain embedded representation and the enterprise financial data, enterprise risk characteristics are generated. The target public information features and the enterprise risk features are input into the multimodal fusion prediction model to output the risk prediction results.

2. The method according to claim 1, characterized in that, The target enterprise data also includes rule information data, and the method further includes: Based on the aforementioned rule information data, structured target rule features are obtained through rule extraction and field normalization; and The target rule features, the target public information features, and the enterprise risk features are input into the multimodal fusion prediction model, and the risk prediction results are output.

3. The method according to claim 1, characterized in that, The process of processing the publicly available information source data using a time-series relationship diagram to generate target publicly available information features includes: Event information units are extracted from the publicly available information source data, and the temporal, semantic and / or causal relationships between the event information units are identified to form a relationship set. The event information unit includes event subject, event behavior, event object, event time information and source labeling. Construct a temporal relationship graph with the event information units as information event nodes and the relationship sets as information relationship edges, wherein the information relationship edges have relationship type information; and Based on the temporal relationship graph, the relationship graph neural network is used to perform multiple rounds of message propagation and neighbor node aggregation to generate the target public information features.

4. The method according to claim 3, characterized in that, The process of using the relational graph neural network for multi-round message propagation and neighbor node aggregation includes: An initial semantic embedding vector is generated for each information event node in the time sequence graph. The event time information is encoded into a time feature vector. The initial semantic embedding vector and the time feature vector are concatenated to form a time-enhanced representation. In each round of message propagation, the time-enhanced representation of adjacent nodes is updated based on the relationship type information; and After multiple rounds of message propagation, the updated time-enhanced representations are fused to obtain the target public information features.

5. The method according to claim 1, characterized in that, The step of using a graph neural network to embed the supply chain data and generate a supply chain embedded representation includes: A supply chain graph is constructed based on the aforementioned supply chain data. This graph uses enterprises as nodes and the transaction relationships between enterprises as edges. Nodes have attributes representing enterprise operational characteristics, and edges have attributes representing sensitivity to changes in conditions. Based on the supply chain graph, the node attributes and edge attributes are aggregated using a graph neural network model to generate the supply chain embedded representation.

6. The method according to claim 1 or 5, characterized in that, The generation of enterprise risk characteristics based on the embedded representation of the supply chain and the enterprise financial data includes: Calculate the cash flow pressure index based on the aforementioned company's financial data; The cash flow stress index and the supply chain embedding representation are fused to form a joint enterprise feature vector; and The enterprise's joint feature vector is input into the risk prediction model to generate the enterprise's risk features.

7. The method according to claim 1, characterized in that, The step of inputting at least the target publicly available information features and the enterprise risk features into a multimodal fusion prediction model and outputting risk prediction results includes: The publicly available information features of the target and the risk features of the enterprise are concatenated to form a multimodal input vector; and The multimodal input vector is input into a multimodal fusion prediction model based on the gradient boosting tree algorithm, and the risk prediction result is output. The risk prediction result includes at least the default probability trend of the target enterprise, the operational pressure index, and support strategy recommendations.

8. The method according to claim 2, characterized in that, The method further includes: Generate source traceability identifiers and version tags for the rule information data; and Based on the source traceability identifier, the system monitors correction or withdrawal events. In response to the detection of a correction or withdrawal event, the system updates the rule information data and the target rule features.

9. A risk prediction device, characterized in that, The device includes: The data acquisition module is used to: acquire target enterprise data, wherein the target enterprise data includes at least publicly available information source data, supply chain data, and enterprise financial data; The public information processing module is used to: process the public information source data using a time-series relationship diagram to generate target public information features; The enterprise data processing module is used to: embed the supply chain data into a graph neural network to generate a supply chain embedding representation; and generate enterprise risk characteristics based on the supply chain embedding representation and the enterprise financial data; and The risk prediction module is used to: input at least the publicly available information features of the target and the risk features of the enterprise into the multimodal fusion prediction model, and output the risk prediction results.

10. 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 8.

11. 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 8.

12. 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 8.

Images