An industry chain multi-relationship modeling method based on a dynamic relationship graph

Abstract

The application relates to the technical field of knowledge graphs, in particular to an industry chain multi-relation modeling method based on a dynamic relation graph, which comprises the following steps: receiving industry chain data comprising fact data and operation data; constructing a fact graph framework by using the fact data; quantifying entity nodes and event relations in the graph framework by using the operation data, and giving dynamic quantitative attributes; reasoning through a fusion analysis model, identifying a risk transmission path based on the event graph framework, calculating a risk index of each path in combination with the dynamic quantitative attributes, and outputting an intermediate result; converting the intermediate result into resource scheduling decision support suggestions through a strategy optimization model, taking the intermediate result as a constraint condition and an incentive function of multi-agent reinforcement learning, and generating resource scheduling suggestions based on heterogeneous agent roles and structured constraints of each entity in the industry chain. The application realizes industry chain risk analysis and resource optimization scheduling by constructing a dynamic relation graph.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically to a method for modeling multiple relationships in an industrial chain based on dynamic relationship graphs. Background Technology

[0002] Modern industrial chains are becoming increasingly complex, exhibiting networked and dynamic characteristics. Fluctuations in the operations of any entity within the chain can be transmitted through various pathways, including supply chains, guarantee chains, and funding chains, potentially triggering systemic risks. Therefore, effective analysis and risk management of industrial chains are crucial.

[0003] Existing supply chain analysis methods generally have limitations. On the one hand, traditional methods rely heavily on static financial statements or isolated operating data, making it difficult to depict the complex and intertwined dynamic relationships between entities, resulting in insufficient predictive power regarding risk transmission paths and probabilities. On the other hand, while existing technologies use knowledge graphs to represent relationships between entities, they are mostly limited to static relationship displays and fail to effectively integrate time-varying operating data. Furthermore, some attempts to use artificial intelligence models for risk prediction have also suffered from a lack of understanding of the inherent logic of the supply chain, leading to weak interpretability of the analysis results and making it difficult to guide actual collaborative decision-making.

[0004] To address this, a multi-relationship modeling method for the industrial chain based on dynamic relationship graphs is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-relationship modeling method for the industrial chain based on dynamic relationship graphs. By constructing dynamic relationship graphs, risk analysis and resource optimization scheduling of the industrial chain are realized. The method involves receiving industrial chain data including event-related data and operational data; constructing an event-related graph framework using the event-related data; quantifying the relationships between entity nodes and events in the graph framework using operational data and assigning dynamic quantitative attributes; performing reasoning through a fusion analysis model, identifying risk transmission paths based on the event graph framework, calculating the risk index of each path based on the dynamic quantitative attributes, and outputting intermediate results; and transforming the intermediate results into resource scheduling decision support suggestions through a strategy optimization model. These intermediate results are then used as constraints and incentive functions for multi-agent reinforcement learning, and resource scheduling suggestions are generated based on the heterogeneous agent roles and structured constraints of each entity in the industrial chain.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for multi-relationship modeling of industrial chains based on dynamic relationship graphs includes:

[0008] Receive industry chain data, which includes theoretical data and operational data;

[0009] Using the aforementioned event data, a event graph framework representing the topological structure and logical sequence of events is constructed; using the aforementioned operational data, the relationships between entity nodes and events in the event graph framework are quantified, and dynamic quantitative attributes of operational health and transaction activity are assigned.

[0010] By using a fusion analysis model for reasoning, the event graph framework based on the graph identifies the risk transmission path, and the risk index of the risk transmission path is calculated by combining dynamic quantitative attributes. The entire simulation process is ensured to comply with the preset business rules, and the intermediate results that include risk assessment and compliance judgment are output and are mutually verified.

[0011] The intermediate results are transformed into resource scheduling decision support suggestions through a strategy optimization model. These intermediate results are then used as constraints and incentive functions for multi-agent reinforcement learning. Based on the heterogeneous agent roles and structured constraints of each entity in the industry chain, resource scheduling suggestions are generated.

[0012] Preferably, the event data includes enterprise registration information, judicial litigation information, and contract performance information; the operational data includes enterprise financial statement data, tax declaration data, and supply chain transaction data.

[0013] Preferably, the process of constructing a context graph framework representing the topological structure and logical sequence of events using the context data includes: processing the context data using a pre-trained industry chain domain language model, simultaneously performing event trigger word recognition and entity recognition; extracting event arguments based on the event trigger words to obtain structured event information; analyzing the structured event information using an event association evaluation model to evaluate the probabilistic dependencies between events and construct an event chain representing the potential risk transmission logic; and constructing the context graph framework by using the identified entities as entity nodes, the structured event information as event nodes, and the evaluated event chain as directed edges connecting the event nodes.

[0014] Preferably, the process of quantifying the entity nodes and event relationships in the graph framework using the operational data includes: establishing a node indicator system representing the operational status of entities and an edge indicator system representing the interaction strength of relationships; calculating the operational data based on the node indicator system and the edge indicator system to generate node attribute values ​​and edge attribute values; and periodically attaching the generated node attribute values ​​and edge attribute values ​​to the entity nodes and event relationships in the event graph framework to generate the dynamic relationship graph.

[0015] Preferably, the fusion analysis model includes:

[0016] The graph representation layer is used to receive the dynamic relationship graph and convert it into a vectorized representation that the model can process;

[0017] A path identification layer is used to perform graph traversal computation on the vectorized representation, search for and output all potential risk propagation paths;

[0018] The risk assessment layer is used to conduct a comprehensive risk assessment for each risk transmission path, combining the dynamic quantitative attributes of each node on the path, and output a risk index.

[0019] The compliance verification layer is used to load a preset rule base, verify each risk transmission path, and output a compliance judgment result.

[0020] Preferably, the strategy optimization model includes:

[0021] The environmental perception layer is used to receive the intermediate results and parse them into observable state information, action constraints and reward signals for the agent.

[0022] The value network layer is used to evaluate the long-term expected return of performing different actions based on the current state information and output value evaluation data.

[0023] The policy network layer is used to receive the state information and the value assessment data, calculate and output the probability distribution of the optimized action to be performed in the current state;

[0024] The action output layer is used to sample based on the optimized action probability distribution, determine and output the final decision action of the current agent.

[0025] Preferably, the process based on the heterogeneous intelligent agent roles and structured constraints of various entities in the industry chain includes:

[0026] The entity nodes in the dynamic relationship graph are mapped to heterogeneous intelligent agent roles based on their business functions and industry positioning. These heterogeneous intelligent agent roles include core manufacturing enterprises, upstream suppliers, downstream distributors, and logistics service providers.

[0027] Each type of intelligent agent role is configured with a dedicated action space and observation space. The action space defines the resource scheduling operations that the heterogeneous intelligent agent role can perform, and the observation space defines the decision information that the heterogeneous intelligent agent role can obtain.

[0028] The topological connections in the graph are transformed into interaction constraints in the multi-agent reinforcement learning environment, which limit the interaction mode of information flow and material flow between agents.

[0029] Preferably, the process of using the intermediate results as constraints and incentive functions for multi-agent reinforcement learning includes: setting the risk transmission path and compliance judgment results in the intermediate results as constraints for policy evaluation in a training environment containing historical cases; constructing the incentive function for the multi-agent reinforcement learning environment by combining the aggregated risk assessment values ​​in the intermediate results with the actual final results in historical cases; and using agents representing different roles in the industry chain to perform offline learning and policy mining on historical data in the static training environment, and outputting the set of policies with the highest correlation to historical successful cases after learning convergence as resource scheduling suggestions.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. By integrating factual data (business registration, litigation, contract information) and operational data (financial, tax, transaction data), a dynamic relationship graph is constructed to clearly depict the relationships between entities and the logic of events. Compared with traditional isolated data analysis, this significantly enhances the comprehensiveness and transparency of the analysis.

[0032] 2. Enhance interpretability through risk path identification. The integrated analysis model accurately identifies risk transmission paths and quantifies their risk levels through graph representation, path identification, probabilistic reasoning, and compliance verification. This improves the interpretability problem caused by black-box prediction in traditional methods, making risk analysis more intuitive and traceable.

[0033] 3. By mapping entities to heterogeneous intelligent agent roles and configuring interaction constraints, a decision-making environment that truly reflects the logic of the industrial chain is constructed, which improves the problem of general AI model decisions violating business common sense, makes the suggestions more in line with actual operation, and enhances credibility. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of a multi-relationship modeling method for the industrial chain based on dynamic relationship graphs, provided in an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the fusion analysis model provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the strategy optimization model provided in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1:

[0039] Please see Figures 1 to 3 This invention provides a method for multi-relationship modeling of industrial chains based on dynamic relationship graphs, the technical solution of which is as follows:

[0040] A method for multi-relationship modeling of industrial chains based on dynamic relationship graphs, specifically as follows: Figure 1 As shown, it includes:

[0041] Receive industry chain data, which includes theoretical data and operational data;

[0042] Using the aforementioned event data, a event graph framework representing the topological structure and logical sequence of events is constructed; using the aforementioned operational data, the relationships between entity nodes and events in the event graph framework are quantified, and dynamic quantitative attributes of operational health and transaction activity are assigned.

[0043] The system uses a fusion analysis model to reason, identifies risk transmission paths based on an event graph framework, calculates risk indices for these paths by combining dynamic quantitative attributes, and ensures that the entire simulation process complies with pre-defined business rules. It outputs mutually verified intermediate results that include risk assessment and compliance judgment.

[0044] The intermediate results are transformed into resource scheduling decision support suggestions through a strategy optimization model. These intermediate results are then used as constraints and incentive functions for multi-agent reinforcement learning. Based on the heterogeneous agent roles and structured constraints of each entity in the industry chain, resource scheduling suggestions are generated.

[0045] Furthermore, the operational data includes enterprise registration information, legal proceedings information, and contract performance information. The legal proceedings information was obtained by connecting to a judicial document disclosure website, capturing a recent lawsuit against a supplier, showing that the supplier was sued for patent infringement due to a core technology control issue by a third-party company. The contract performance information, after authorization, retrieved the core enterprise's annual purchase contracts with suppliers and machine tool sales contracts with downstream customers from its contract management system, and extracted key content such as delivery terms and breach of contract liabilities in a structured manner. The operational data includes enterprise financial statement data, tax declaration data, and supply chain transaction data. The enterprise financial statement data was obtained by accessing the National Enterprise Credit Information Publicity System, acquiring data on the business registration, shareholder structure, historical change records, and branches of the core enterprise, supplier, and downstream customer. The quarterly and annual financial reports of the core enterprise (a listed company) and downstream customer were also obtained, with a focus on indicators such as debt-to-asset ratio, net cash flow, and inventory turnover. The tax declaration data, with the supplier's authorization, accessed a tax interface to obtain its VAT declaration data for the past six months. Data shows that its sales revenue in the most recent quarter declined significantly compared to the previous quarter; the supply chain transaction data is connected with the core enterprise's enterprise resource planning (ERP) system, and this method obtains in real time the purchase order records issued by the enterprise to suppliers, the batches of goods delivered by suppliers and the acceptance rate, as well as the real-time inventory level of this material.

[0046] By clearly defining the sources of factual and operational data, the comprehensiveness and interpretability of supply chain analysis are enhanced. By integrating multi-source data, the legal, contractual, and operational status of entities is comprehensively depicted, providing a more accurate basis for risk assessment. This data integration enhances the transparency and traceability of analytical results, providing reliable and comprehensive data support for collaborative decision-making.

[0047] Furthermore, the process of constructing a contextual graph framework representing the topological structure and logical sequence of events using the contextual data includes: processing the contextual data using a pre-trained industry chain domain language model, simultaneously performing event trigger word recognition and entity recognition; extracting event arguments based on the event trigger words to obtain structured event information; the pre-trained industry chain domain language model is specifically an enhanced model with a dual-channel fusion architecture. This dual-channel fusion architecture includes a text semantic channel responsible for processing unstructured text content, and a structured information channel specifically responsible for extracting key numerical and categorical data. Before extracting event arguments, a fusion layer fuses the feature representations output from these two channels to generate a fused feature representation. By fusing text semantics and numerical logic, the accuracy and comprehensiveness of event feature extraction are improved, laying a more solid and reliable data foundation for subsequent risk assessment and effectively avoiding misjudgments from a single model. Using an event association assessment model, the structured event information is analyzed to evaluate the probabilistic dependencies between events and construct an event chain representing the potential risk transmission logic. Identified entities are used as entity nodes, structured event information as event nodes, and the evaluated event chain as directed edges connecting these event nodes, thus constructing the event graph framework. A pre-trained industry chain domain language model is used to process the collected event data (unstructured data such as legal documents and contract texts), simultaneously performing event trigger word recognition and entity recognition. This aims to automatically discover words describing key business events in the text (such as filing a lawsuit, signing a contract, or breach of contract) and the entities involved in the events. Then, based on the identified event trigger words, an event argument extraction module is activated to extract various elements related to the event from the text context (such as time, location, and participants), and integrate them into a structured event information, for example: {Event ID: [unique identifier], Event Type: [e.g., legal risk, cooperative relationship], Event Subject: [entity name], ...}.

[0048] Then, the structured event information is analyzed using an event correlation assessment model. This model, combined with a built-in business logic and risk knowledge base, constructs potential event transmission paths by evaluating the strength of correlation and logical order between different events. For example, the model might determine that a contract default event is a high-risk precursor to a supply chain disruption event, thus establishing a logical connection between the two. The business logic and risk knowledge base is specifically an event relationship library built on ontology. This library predefines event categories such as legal proceedings, contract signing, and supply chain disruptions, as well as logical relationships such as precursor, potential trigger, and belonging. For example, the library contains the following knowledge entry: {Subject: Legal proceedings, Relationship: may_cause, Object: Business deterioration, Condition: Litigation amount > 1 million}.

[0049] Finally, the graph framework is constructed. It creates all identified entities as entity nodes and all structured event information as event nodes. Based on supply and sales, guarantee, and other relationships parsed from contracts and other data, directed edges are established between entity nodes. Simultaneously, the evaluated event chains are used as directed edges connecting different event nodes, thus constructing a complete event graph framework. The pre-trained industry chain domain language model is specifically a model based on the BERT architecture. After basic pre-training using general corpora, this model is further pre-trained using publicly available research reports, news information, and legal documents related to the industry chain to enhance its understanding of specific business terms and event descriptions. The process of calculating the association strength index by the event association evaluation model includes: for two structured events A and B, searching the knowledge base for a possible path from A to B; if it exists, then based on historical data, calculating the conditional probability of event B occurring within a time window T (e.g., 90 days) after event A occurs. Combining the attributes of the event itself (such as the amount in dispute or the amount of breach of contract), a preset sigmoid function is used to map it into an influence factor. Finally, the event index is obtained by multiplying the conditional probability of event B by the influence factor. When the event index is greater than a preset threshold (such as 0.75), an event chain is established.

[0050] By using automated event correlation analysis, scattered textual information is transformed into a logically clear event transmission network. It constructs logical connections between events and uncovers potential transmission paths, rather than simply presenting isolated events themselves. It transforms the black-box relationships that are difficult to explain in traditional models into white-box event transmission paths that can be reviewed and traced. This makes each identified risk path have clear logic and narrative, improving the transparency of the analysis process and the credibility of the conclusions.

[0051] Furthermore, the process of quantifying the entity nodes and event relationships in the graph framework using the operational data includes: establishing a node indicator system representing the operational status of entities and an edge indicator system representing the interaction strength of relationships; calculating the operational data based on the indicator system to generate node attribute values ​​and edge attribute values; and periodically attaching the generated node attribute values ​​and edge attribute values ​​to the entity nodes and event relationships in the event graph framework to generate the dynamic relationship graph. When establishing the node indicator system representing the operational status of entities, a data-driven backtesting mechanism is introduced. Specifically, the system pre-maintains a candidate indicator pool containing multiple industry-standard financial and operational indicators. When an evaluation model needs to be built for a specific industry (e.g., the automotive manufacturing industry), the system initiates an automated backtesting process. This process iterates through each candidate indicator in the indicator pool, using historical data from the industry to independently test whether the indicator has significant predictive power for real-world risk events (such as corporate defaults, major operational crises, etc.) over the past few years. For example, it tests whether a decline in inventory turnover is highly correlated with subsequent historical cases of operational crises. After testing, the system ranks the indicators based on their predictive power and automatically selects the top-ranked indicators (e.g., the top five). These are then combined into an optimal indicator system specifically for the industry, validated by historical data. Subsequent business health scores are calculated based on this optimal indicator system. By replacing subjective expert experience with data-driven backtesting, the system automatically selects the most predictive indicator combinations for different industries, ensuring the objectivity and accuracy of risk assessment and grounding final decisions in more reliable data.

[0052] Specifically, the operational health score of the node is derived through a weighted comprehensive evaluation of multiple key financial indicators. Specifically, this method selects core indicators such as the debt-to-equity ratio, net cash flow, and inventory turnover. First, the value of each original financial indicator is converted into a standardized percentage score using a preset normalization function. This normalization function is a piecewise linear mapping function, which sets mapping rules based on the median value of the industry to which the indicator belongs. Specifically: if the indicator value equals the industry median value, it is mapped to a standard score of 60; if the indicator value is better than the industry median value (e.g., lower for the debt-to-equity ratio; higher for net cash flow), it is linearly mapped between 60 and 100 points, with indicators reaching the top 10% of the industry receiving a score of 90; if the indicator value is worse than the industry median value, it is linearly mapped between 0 and 60 points, with indicators reaching the bottom 10% of the industry receiving a score of 30. Then, a preset weight is assigned to the standardized score of each indicator. The preset weights aim to reflect the relative importance of different indicators in specific links of the industrial chain. Their determination is as follows: First, an expert knowledge base containing weight configurations for multiple industries is loaded. This knowledge base is pre-built by industry analysts and sets benchmark weights for different indicators in different industries. Second, principal component analysis is used to calculate historical data for specific industrial chains. By analyzing the contribution of each financial indicator to the fluctuation of overall operating conditions, a set of data-driven weight suggestions is objectively generated. Finally, the benchmark weights of the expert knowledge base are integrated with the data-driven weight suggestions, and a weighted average method is used to obtain the preset weights for final application. For example, in the weight allocation of the manufacturing industry, the final determination is to allocate higher weights to the debt-to-equity ratio and net cash flow, each accounting for 40%, while allocating a lower weight to the inventory turnover rate, accounting for 20%. Finally, the weighted scores of each indicator are summed to obtain the final comprehensive score for operating health.

[0053] The transaction activity score is calculated by comprehensively considering transaction frequency and transaction amount. This method statistically analyzes the total number of transactions and total transaction amount within a specific period (e.g., the last 90 days). To smooth out the impact of extreme data, these two values ​​are first logarithmically processed, then multiplied by an adjustable coefficient to balance their importance. Finally, the two processed results are added together to obtain the final transaction activity score.

[0054] Finally, the generated node attribute values ​​and edge attribute values ​​are periodically (e.g., daily or weekly) appended to the entity nodes and relationship edges in the event graph framework. For example, {Operating Health: 85, Credit Risk: Low} can be appended as an attribute to a specific entity node. This process is continuously performed through automated data update tasks, thereby transforming the static event graph framework into a dynamic relationship graph that can dynamically reflect the latest status of the industry chain.

[0055] By quantifying the relationships between entity nodes and events within a contextual graph framework using operational data, the interpretability of the analysis results and its ability to guide collaborative decision-making are significantly improved. By constructing node and edge indicator systems, financial, tax, and supply chain transaction data are transformed into dynamic attribute values, which are periodically updated to the graph. This quantification method makes entity health and transaction activity clearly visible, enhances the traceability of risk transmission paths, and provides an intuitive, data-driven basis for collaborative decision-making across the industrial chain, thereby effectively guiding practical operations.

[0056] Furthermore, the specific flowchart of the fusion analysis model is as follows: Figure 2 As shown, it includes:

[0057] The graph representation layer is used to receive the dynamic relationship graph and convert it into a vectorized representation that the model can process;

[0058] A path identification layer is used to perform graph traversal computation on the vectorized representation, search for and output all potential risk propagation paths;

[0059] The risk assessment layer is used to conduct a comprehensive risk assessment for each risk transmission path, combining the dynamic quantitative attributes of each node on the path, and output a risk index.

[0060] The compliance verification layer is used to load a preset rule base, verify each risk transmission path, and output a compliance judgment result.

[0061] Specifically, firstly, the graph representation layer receives the generated dynamic relation graph. This layer uses graph convolutional network technology to uniformly transform the topology, node attributes, and edge attributes of the entire graph into a high-dimensional vectorized representation that the model can process.

[0062] Next, the path identification layer performs calculations on the generated vectorized representation. This layer uses a depth-first search algorithm, starting from the identified risk event nodes, to search along entity and event relationship edges in the graph, thereby identifying and outputting all potential risk transmission paths. Then, the risk assessment layer performs a quantitative assessment of each risk transmission path output by the path identification layer. This layer extracts dynamic quantitative attributes such as the operational health score of each node and the transaction dependence of each edge along the specified risk transmission path. These attribute values ​​are then used as input evidence and fed into a Bayesian network model for comprehensive reasoning. The model ultimately calculates and outputs a normalized risk index that characterizes the overall risk level of the path. The structure of this Bayesian network is dynamically generated based on the relationship edges of the event graph, and its conditional probability table (CPT) is learned through maximum likelihood estimation based on quantitative attributes such as the operational health of nodes and the transaction activity of edges. The specific steps for calculating the risk index in the risk assessment layer are as follows: First, the continuous scores, such as the operational health score of each node, are converted into discrete states such as high risk, medium risk, and low risk according to a preset interval. These discrete states are then used as observation evidence in the Bayesian network model to update and calculate the posterior probability of a specific risk occurring at each node on the path. To aggregate the risk of the entire path, the model first calculates the safety probability of each node, i.e., subtracts the risk index of that node from 1. Then, by multiplying the safety probabilities of all nodes on the path, a total path safety probability is obtained. Finally, this total path safety probability is subtracted from 1 to obtain the final index representing the risk level of the entire path. This index is output after being normalized to a percentage score.

[0063] Finally, the compliance verification layer constrains and verifies the above process. This layer pre-loads a rule base related to the business scenario (e.g., specific clauses in financial institutions' credit policies regarding the limit of credit concentration to a single customer not exceeding 20%). It compares each identified risk transmission path with the clauses in the rule base, outputting a compliance judgment result on whether the path violates a specific rule. Through the collaborative work of these four layers, the model ultimately outputs a mutually verified intermediate result containing path explanation, risk assessment, and compliance judgment.

[0064] The graph representation layer transforms dynamic relationship maps into processable data; the path identification layer clearly identifies risk transmission paths; the risk assessment layer accurately calculates risk indices based on dynamic attributes; and the compliance verification layer ensures that the results comply with business rules. This structured analysis not only makes risk paths and probabilities transparent and traceable, but also enhances the reliability of the results through compliance verification, providing accurate and credible guidance for collaborative decision-making across the industry chain.

[0065] Furthermore, the specific flowchart of the strategy optimization model is as follows: Figure 3 As shown, it includes:

[0066] The environmental perception layer is used to receive the intermediate results and parse them into observable state information, action constraints and reward signals for the agent.

[0067] The value network layer is used to evaluate the long-term expected return of performing different actions based on the current state information and output value evaluation data.

[0068] The policy network layer is used to receive the state information and the value assessment data, calculate and output the probability distribution of the optimized action to be performed in the current state;

[0069] The action output layer is used to sample based on the optimized action probability distribution, determine and output the final decision action of the current agent.

[0070] First, the environment perception layer serves as the model's input interface. Its core responsibility is to receive intermediate results generated by the fusion analysis model and, according to the parsing rules defined in this invention, transform the input dynamic relationship graph into a standardized state information feature vector. During offline training, this layer is responsible for batch reading and parsing historical states, actions, and actual outcomes from the historical case database. In the online application phase, it is responsible for analyzing the current real-time risk scenario, providing decision-making support for subsequent network layers.

[0071] The value network layer acts as a commentator, and its core function is to evaluate the merits of decision-making actions. During offline training, this layer learns to predict the long-term rewards that can be obtained by performing specific historical actions under specific historical states. To address the problem of overestimating the value of out-of-distribution actions, the algorithmic idea of ​​conservative Q-learning is introduced. This method adds an extra regularization term to the standard value function learning. This regularization term is used to penalize and reduce the evaluation value of actions that have not appeared in the historical dataset, thereby ensuring that the value network layer can make robust evaluations and effectively avoid unrealistic evaluations.

[0072] The policy network layer acts as the actor, receiving state information from the environment perception layer as input. During the offline training phase, to ensure that the generated policy does not deviate from the reliable action range verified in historical data, the learning process of this layer is also constrained. It must not only learn to maximize the conservative evaluation value given by the value network layer, but also its output action distribution must be constrained to a range similar to the behavioral policies in historical data. In this way, the policy learned by this layer pursues high returns without producing unrealistic and radical decision actions.

[0073] Finally, the action output layer is the model's final output, receiving the optimized action probability distribution for the current state calculated by the policy network layer. This layer then determines and selects the action with the optimal probability based on this distribution, designating it as the agent's final decision action and transforming it into a concrete, understandable resource scheduling suggestion for output. Simultaneously, the current state and suggested action can be input into the pre-trained value network layer to output a confidence score for the current decision suggestion, providing additional reference for the decision-maker. Further, when the output decision confidence score is higher than a preset high threshold, the action output layer classifies the decision suggestion as highly reliable and automatically routes it to the downstream automated processing system, enabling rapid response to high-certainty opportunities. When the confidence score is lower than a preset low threshold, the decision is automatically suspended, triggering a human expert review process. During human review, the system not only alerts to the risks but also submits the highest-probability candidate actions output by the policy network, along with their respective confidence scores, to experts for review. Decision suggestions falling between the two thresholds enter the standard human review process. Efficiency is improved by automating highly certain suggestions, while uncertainties and risks are reported to experts for review to ensure safety. In this embodiment, the preset high threshold is 0.95, and the preset low threshold is 0.4.

[0074] The process involves an environmental perception layer that analyzes intermediate results to provide the agent with usable state, constraint, and reward information; a value network layer that evaluates long-term returns; a policy network layer that generates probabilities for optimized actions; and an action output layer that outputs the final decision action. This hierarchical approach transforms risk analysis into concrete action recommendations, ensuring that decisions align with the dynamic logic of the industry chain, enhancing the operability and interpretability of the results, and providing precise guidance for resource scheduling.

[0075] Furthermore, the process based on the heterogeneous intelligent agent roles and structured constraints of various entities in the industry chain includes:

[0076] The entity nodes in the dynamic relationship graph are mapped to heterogeneous intelligent agent roles based on their business functions and industry positioning. These heterogeneous intelligent agent roles include core manufacturing enterprises, upstream suppliers, downstream distributors, and logistics service providers.

[0077] Each type of intelligent agent role is configured with a dedicated action space and observation space. The action space defines the resource scheduling operations that the heterogeneous intelligent agent role can perform, and the observation space defines the decision information that the heterogeneous intelligent agent role can obtain.

[0078] The topological connections in the graph are transformed into interaction constraints in the multi-agent reinforcement learning environment, which limit the interaction mode of information flow and material flow between agents.

[0079] The process of mapping entities to intelligent agent roles involves traversing each entity node in a dynamic relationship graph and matching it with a heterogeneous intelligent agent role based on its business functions and industry positioning information using a pre-defined rule table. For example, if an entity node belongs to the manufacturing industry and occupies a core position in the industrial chain, it is mapped to a core manufacturing enterprise role; if its industry belongs to raw material supply, it is mapped to an upstream supplier role; if it belongs to product sales, it is mapped to a downstream distributor role; and if it belongs to the transportation industry, it is mapped to a logistics service provider role. Each mapped intelligent agent role is then configured with its own dedicated action space and observation space. Based on the role type, corresponding configuration information is loaded from a predefined configuration library. The action space is defined as a discrete set containing all resource scheduling operations that the role can execute. For example, the action space loaded for the upstream supplier role is {deliver on time, delay delivery, adjust production plan}. The observation space is defined as a fixed-dimensional vector, where each dimension corresponds to a decision information that the role can obtain when making decisions. For example, the observation space vector loaded for the core manufacturing enterprise role includes its own inventory level, order status from upstream, and order quantity from downstream.

[0080] Finally, the topological connections in the graph are transformed into interaction rules in a multi-agent reinforcement learning environment. This process traverses every directed edge in the graph. For an edge pointing from entity A to entity B, a one-way communication channel is established in the reinforcement learning environment from agent A to agent B. This channel is configured to allow agent A to send state updates of information and material flows to agent B, while prohibiting agent B from initiating communication with agent A through this channel. After this process is completed, a role-based multi-agent decision-making environment reflecting the real topological constraints of the industry chain is constructed.

[0081] By generating resource scheduling suggestions based on heterogeneous intelligent agent roles and structured constraints, the realism and effectiveness of collaborative decision-making in the industrial chain are significantly improved. This is achieved by mapping entities in the dynamic relationship graph to roles such as core manufacturing enterprises, suppliers, and distributors, configuring dedicated actions and observation spaces to ensure that decisions align with role functions. Furthermore, by transforming graph topological relationships into interaction constraints, a decision-making environment that realistically reflects the logic of the industrial chain is constructed. This approach makes scheduling suggestions highly consistent with actual operations, enhancing credibility and operability.

[0082] Furthermore, the process of using the intermediate results as constraints and incentive functions for multi-agent reinforcement learning includes: setting the risk transmission path and compliance judgment results in the intermediate results as constraints for policy evaluation in a training environment containing historical cases; constructing the incentive function of the multi-agent reinforcement learning environment by combining the aggregated risk assessment values ​​in the intermediate results with the actual final results in historical cases; and using agents representing different roles in the industry chain to perform offline learning and policy mining on historical data in the static training environment, and outputting the set of policies that have converged after learning and have the highest correlation with historical successful cases as resource scheduling suggestions.

[0083] The first step is to build and label historical training data. For each case in the historical case database, the model will first run a fusion analysis model to calculate the intermediate results of that case at that historical point in time, namely the risk transmission path, risk index, and compliance judgment.

[0084] Secondly, the incentive function is constructed. The model's incentive function is not calculated in real time, but directly derived from the actual final results of historical cases. For example, if a company in a historical case develops healthily within a year of making a decision, the reward signal for that case is set to a high positive value; if the company's operations deteriorate or it defaults, it is set to a negative value. When constructing the incentive function, the aggregated value of risk assessment is used as a moderating factor. Specifically, the final incentive value is composed of three parts: the first part is the basic reward based on the actual outcome, obtained by multiplying the quantified outcome by the first weight; the second part is the risk-related additional incentive, obtained by multiplying the quantified outcome, the normalized risk value, and the second weight; the third part is the execution cost of the action. The final incentive value is obtained by adding the first two parts and subtracting the third part.

[0085] Then, offline learning and policy mining are initiated. In this stage, agents representing different roles in the industry chain learn in a static training environment composed of massive historical data. The model uses the labeled intermediate results as constraints for policy learning and uses the incentive function constructed from historical real-world outcomes as the optimization objective. Through offline reinforcement learning algorithms, it continuously iterates and optimizes its internal policy network, aiming to find a policy that maps the current state to the decision action with the highest correlation to historical success. Specifically, the model uses the labeled intermediate results as constraints for policy learning in the following way: at each time step in the reinforcement learning environment, if the agent's chosen action triggers a risk transmission path warning or violates compliance judgment, a preset fixed penalty value (100 in this embodiment) is subtracted from the task reward when calculating its total reward.

[0086] Finally, resource scheduling suggestions are generated and output. Once the offline training process of the entire model meets the preset termination conditions (such as reaching the maximum number of training steps or model performance convergence), the stable policy sets learned by each agent will be permanently saved. In practical applications, this pre-trained model can be directly invoked to quickly output the final resource scheduling suggestions based on the current risk scenario.

[0087] By transforming risk analysis into resource scheduling suggestions through multi-agent reinforcement learning, the practicality of collaborative decision-making in the industrial chain is significantly improved. By generating globally optimal scheduling suggestions, the operability, interpretability, and adaptability to industrial chain dynamics of decision-making are enhanced.

[0088] By constructing a multi-relationship modeling method for the industrial chain based on dynamic relationship graphs, and integrating event-related data (business registration, litigation, and contract information) and operational data (financial, tax, and transaction data), a dynamic graph is built to clearly present entity relationships and event logic, enhancing analytical transparency. The fusion analysis model, through graph representation, path identification, quantitative assessment, and compliance verification, clearly constructs risk transmission paths and evaluates their risk levels. The strategy optimization model utilizes multi-agent reinforcement learning to transform risk analysis into resource scheduling suggestions. By simulating the real logic of the industrial chain through heterogeneous agent roles and interaction constraints, it generates reference solutions for collaborative decision-making. This method makes risk paths more traceable and interpretable, and enhances the commercial applicability of suggestions through a simulated decision-making environment, thereby optimizing the practicality and operability of industrial chain risk management and collaborative decision-making.

[0089] Example 2:

[0090] Company B applies for credit from a financial institution. To comprehensively assess the company's creditworthiness, operational stability, and potential risks within the industry chain, the financial institution adopts a multi-relationship modeling method for the industry chain based on dynamic relationship graphs, as provided in this invention. The specific implementation process is as follows:

[0091] This method first initiates data collection. For this credit approval scenario, two types of data related to the applicant company are collected. The first type is factual data, which specifically includes the company's business registration change records for the past two years, a recent judicial judgment related to product quality from its core downstream customer, and information on several technology patents held by the applicant company itself. The second type is operational data, which, after authorization, includes the details of VAT invoices issued by the applicant company to its core downstream customers in the past 12 months, its own tax declaration data, and utility bill payment records.

[0092] After data collection, the next stage is graph construction and quantification. First, a graph framework is constructed using event-based data: through natural language processing, the product quality lawsuit event is extracted from judicial judgments and treated as an event node, linked to the core downstream customer entity node. Simultaneously, based on the buyer and seller information clearly stated in the invoice data, a directed edge is established from the applicant company to the core downstream customer, representing the supply relationship between the two, thus forming the basic topology and event logic. Next, the graph is quantified using operational data: based on tax and water / electricity consumption data, a dynamic operational health score of 85 is calculated and assigned to the applicant company node; furthermore, based on invoice data showing that 80% of the applicant company's sales come from this core customer, the directed edge connecting the two is assigned a dynamic transaction dependency attribute of 0.8.

[0093] After the graph is constructed, this method initiates a fusion analysis model for deep reasoning. The model's path identification layer first performs traversal calculations on the graph data, quickly identifying a risk transmission path: starting with a product quality litigation event at the core downstream customer node, it propagates along transaction relationship edges with a high dependency attribute of 0.8, ultimately reaching the applicant company node. Next, the model's risk assessment layer intervenes, combining the credit rating downgrade of the core downstream customer due to the lawsuit loss, and the high transaction dependency attribute of 0.8, calculating a risk index of 75 points for the applicant company's revenue to decline by more than 20% within the next six months. Simultaneously, the compliance verification layer loads the credit approval rule base of financial institutions and determines that a risk index of 75 points exceeds the acceptable risk threshold for this type of company. Finally, an intermediate result that includes the aforementioned risk transmission path, risk index assessment, and compliance judgment is output, mutually verified.

[0094] Upon receiving the intermediate results, this method initiates a strategy optimization model to transform the analytical conclusions into an actionable solution. First, the model receives the applicant company's risk scenario, including its high risk index of 75 and a transaction dependence of 0.8 on a single customer. Then, the model performs a high-speed search in its internal historical case database to find historical credit cases with similar risk profiles. The model successfully matches several historical cases with similar industry backgrounds and high customer concentration risk (dependency higher than 70%, risk score in the 70-80 range).

[0095] Next, the model analyzes these historical cases, focusing on the different credit decisions made at the time and the actual final results of these decisions when the cases were archived. The analysis shows that for companies with such high dependence risk, a high proportion of historical cases where standard loans were directly approved later experienced repayment difficulties; while in the few historical cases where structured credit schemes were adopted that were linked to the diversification of the applicant company's customers, the final outcome was that the company successfully optimized its customer structure and achieved stable operation.

[0096] The policy network of this model has learned this pattern through offline training on all historical data. Therefore, when the current risk scenario of the applicant company is input, the model will design a structured credit granting scheme that is linked to customer diversification, which is the best historical practice and is identified as the decision action with the highest expected return.

[0097] Ultimately, the model translates this strategy into specific resource allocation recommendations: suggesting that financial institutions approve a structured loan, with 60% as unconditional credit and the remaining 40% linked to the applicant company successfully establishing business relationships with new potential clients representing at least 15% of its total sales within the next year. This approach proactively guides and optimizes identified supply chain risks through the leverage of credit resources, and its recommendations are based on in-depth analysis of real-world successful experiences.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multi-relationship modeling of industrial chains based on dynamic relationship graphs, characterized in that, include: Receive industry chain data, which includes theoretical data and operational data; Using the aforementioned event data, a event graph framework representing the topological structure between entities and the logical sequence of events is constructed; The operational data is used to quantify the relationships between entity nodes and events in the event graph framework, and to assign dynamic quantitative attributes to operational health and transaction activity. The system uses a fusion analysis model to reason, identifies risk transmission paths based on a reasoning graph framework, calculates risk indices for these paths by combining dynamic quantitative attributes, and ensures that the entire deduction process complies with preset business rules. It outputs mutually verified intermediate results that include risk assessment and compliance judgment. The intermediate results are transformed into resource scheduling decision support suggestions through a strategy optimization model. These intermediate results are then used as constraints and incentive functions for multi-agent reinforcement learning. Based on the heterogeneous agent roles and structured constraints of various entities in the industry chain, resource scheduling suggestions are generated. The process based on the heterogeneous agent roles and structured constraints of various entities in the industry chain includes: mapping entity nodes in the dynamic relationship graph to heterogeneous agent roles according to their business functions and industry positioning. These heterogeneous agent roles include core manufacturing enterprises, upstream suppliers, downstream distributors, and logistics service providers. Each agent role is configured with a dedicated action space and observation space. The action space defines the resource scheduling operations that the heterogeneous agent role can execute, and the observation space defines the decision information that the heterogeneous agent role can obtain. The topological connections in the graph are transformed into interaction constraints in the multi-agent reinforcement learning environment, which limit the interaction mode of information flow and material flow between agents. This process traverses each directed edge in the graph. For an edge from entity A to entity B, a one-way communication channel from agent A to agent B is established in the reinforcement learning environment. This channel is set to allow agent A to send state updates of information flow and material flow to agent B, while prohibiting agent B from initiating communication with agent A through this channel.

2. The method for multi-relationship modeling of an industrial chain based on dynamic relationship graphs according to claim 1, characterized in that, The aforementioned data includes enterprise registration information, legal litigation information, and contract performance information; the aforementioned operational data includes enterprise financial statement data, tax declaration data, and supply chain transaction data.

3. The method for multi-relationship modeling of the industrial chain based on dynamic relationship graphs according to claim 1, characterized in that, The process of constructing a contextual graph framework representing the topological structure and logical sequence of events using the contextual data includes: processing the contextual data using a pre-trained industry chain domain language model, simultaneously performing event trigger word recognition and entity recognition; extracting event arguments based on the event trigger words to obtain structured event information; analyzing the structured event information using an event association evaluation model to evaluate the probabilistic dependencies between events and construct an event chain representing the potential risk transmission logic; and constructing the contextual graph framework by using the identified entities as entity nodes, the structured event information as event nodes, and the evaluated event chain as directed edges connecting the event nodes.

4. The method for multi-relationship modeling of the industrial chain based on dynamic relationship graphs according to claim 1, characterized in that, The process of quantifying the relationships between entity nodes and events in the graph framework using the operational data includes: establishing a node indicator system representing the operational status of entities and an edge indicator system representing the interaction strength of relationships; calculating the operational data based on the node indicator system and the edge indicator system to generate node attribute values ​​and edge attribute values; and periodically attaching the generated node attribute values ​​and edge attribute values ​​to the entity nodes and events in the event graph framework to generate the dynamic relationship graph.

5. The method for multi-relationship modeling of an industrial chain based on a dynamic relationship graph according to claim 4, characterized in that, The fusion analysis model includes: The graph representation layer is used to receive the dynamic relationship graph and convert it into a vectorized representation that the model can process; A path identification layer is used to perform graph traversal computation on the vectorized representation, search for and output all potential risk propagation paths; The risk assessment layer is used to conduct a comprehensive risk assessment for each risk transmission path, combining the dynamic quantitative attributes of each node on the path, and output a risk index. The compliance verification layer is used to load a preset rule base, verify each risk transmission path, and output a compliance judgment result.

6. The method for multi-relationship modeling of an industrial chain based on dynamic relationship graphs according to claim 1, characterized in that, The strategy optimization model includes: The environmental perception layer is used to receive the intermediate results and parse them into observable state information, action constraints and reward signals for the agent. The value network layer is used to evaluate the long-term expected return of performing different actions based on the current state information and output value evaluation data. The policy network layer is used to receive the state information and the value assessment data, calculate and output the probability distribution of the optimized action to be performed in the current state; The action output layer is used to sample based on the optimized action probability distribution, determine and output the final decision action of the current agent.

7. The method for multi-relationship modeling of an industrial chain based on dynamic relationship graphs according to claim 1, characterized in that, The process of using the intermediate results as constraints and incentive functions for multi-agent reinforcement learning includes: setting the risk transmission path and compliance judgment results in the intermediate results as constraints for policy evaluation in a training environment containing historical cases; constructing the incentive function of the multi-agent reinforcement learning environment by combining the aggregated risk assessment values ​​in the intermediate results with the actual final results in historical cases; and using agents representing different roles in the industry chain to perform offline learning and policy mining on historical data in the static training environment, and outputting the set of policies that have converged after learning and have the highest correlation with historical successful cases as resource scheduling suggestions.

Images