Industrial chain multi-relation modeling method based on dynamic relation graph

Abstract

The invention relates to the technical field of knowledge graphs, in particular to an industrial chain multi-relation modeling method based on a dynamic relation graph, which comprises the following steps: receiving industrial chain data including affair data and operation data; utilizing the affair data to construct an affair graph framework; quantifying entity nodes and event relationships in the atlas framework by using the operation data, and endowing dynamic quantitative attributes; reasoning is carried out through a fusion analysis model, risk conduction paths are identified based on an event atlas framework, risk indexes of all the paths are calculated in combination with dynamic quantitative attributes, and an intermediate result is output; and through a strategy optimization model, converting the intermediate result into a resource scheduling decision support suggestion, taking the intermediate result as a constraint condition and an excitation function of multi-agent reinforcement learning, and generating a resource scheduling suggestion based on the heterogeneous agent role and the structured constraint of each entity in the industry chain. According to the method, industrial chain risk analysis and resource optimization scheduling are realized by constructing the dynamic relation graph.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph, in particular to an industry chain multi-relation modeling method based on dynamic relation graph. BACKGROUND

[0002] Modern industry chain structure is increasingly complex, showing networked and dynamic characteristics. The fluctuation of the operation of any entity on the chain may be transmitted through various relationship paths such as supply chain, guarantee chain and fund chain, and may cause systemic risk. Therefore, it is crucial to effectively analyze and manage the risk of the industry chain.

[0003] The existing industry chain analysis method usually has limitations. On the one hand, traditional methods mostly rely on static financial statements or isolated operating data, which are difficult to depict the complex and dynamic relationship between entities, resulting in insufficient prediction ability of risk transmission path and probability. On the other hand, although the existing technology uses knowledge graph to represent the relationship between entities, it is mostly limited to static relationship display and fails to effectively integrate time-varying operating data. In addition, some attempts to use artificial intelligence models for risk prediction also lack understanding of the internal logic of the industry chain, resulting in weak explainability of the analysis results and difficulty in guiding actual collaborative decision-making.

[0004] Therefore, an industry chain multi-relation modeling method based on dynamic relation graph is proposed. SUMMARY

[0005] The purpose of the present application is to provide an industry chain multi-relation modeling method based on dynamic relation graph, which realizes industry chain risk analysis and resource optimization scheduling by constructing a dynamic relation graph. Receive industry chain data including transaction data and operating data; use transaction data to construct a transaction graph framework; use operating data to quantify entity nodes and event relationships in the graph framework and give dynamic quantitative attributes; infer through a fusion analysis model, identify risk transmission paths based on the event graph framework, calculate risk indexes of each path based on dynamic quantitative attributes, and output intermediate results; through a strategy optimization model, convert the intermediate results into resource scheduling decision support suggestions, use the intermediate results as constraint conditions and incentive functions of multi-agent reinforcement learning, and generate resource scheduling suggestions based on the heterogeneous agent roles and structured constraints of each entity in the industry chain.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] An industry chain multi-relation modeling method based on dynamic relation graph, comprising:

[0008] Receive industry chain data, the industry chain data including transaction data and operating data;

[0009] The transactional data is used to construct a transactional graph framework representing the topological structure and logical sequence of events between entities; the operational data is used to quantify the entity nodes and event relationships in the transactional graph framework, and dynamic quantitative attributes of operational health and transactional activity are given;

[0010] Through reasoning by the fusion analysis model, a risk transmission path is identified based on the graph-based event graph framework, a risk index of the risk transmission path is calculated in combination with the dynamic quantitative attributes, and it is ensured that the entire deduction process conforms to preset business rules, and intermediate results including risk assessment and compliance judgment are output, which are mutually verified;

[0011] Through the strategy optimization model, the intermediate results are converted into resource scheduling decision support suggestions, and the intermediate results are used as constraint conditions and incentive functions of multi-agent reinforcement learning, and resource scheduling suggestions are generated based on the heterogeneous agent roles and structured constraints of entities in the industrial chain.

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

[0013] Preferably, the process of constructing a transactional graph framework representing the topological structure and logical sequence of events between entities using the transactional data includes: using a pre-trained industrial chain field language model to process the transactional data, and simultaneously performing event trigger word recognition and entity recognition; based on the event trigger word, event argument extraction is performed to obtain structured event information; using an event correlation evaluation model, the structured event information is analyzed to evaluate the probabilistic dependence relationship between events, and an event chain representing the potential risk transmission logic is constructed; the identified entities are used as entity nodes, the structured event information is used as event nodes, and the evaluated event chain is used as a directed edge connecting the event nodes, and the transactional graph framework is constructed.

[0014] Preferably, the process of quantifying the entity nodes and event relationships in the graph framework using the operational data includes: establishing a node index system representing the operational state of the entity and an edge index system representing the interaction strength of the relationship; according to the node index system and the edge index system, the operational data is calculated to generate node attribute values and edge attribute values; the generated node attribute values and edge attribute values are periodically attached to the entity nodes and event relationships in the transactional graph framework, respectively, to generate the dynamic relationship graph.

[0015] Preferably, the fusion analysis model includes:

[0016] a graph representation layer for receiving the dynamic relationship graph and converting it into a vectorized representation processable by the model;

[0017] a path recognition layer configured to perform graph traversal calculation on the vectorized representation, search and output all potential risk transmission paths;

[0018] a risk assessment layer configured to perform comprehensive risk assessment on each risk transmission path in combination with dynamic quantitative attributes of nodes on the path, and output a risk index;

[0019] a compliance verification layer configured to load a preset rule library, perform item-by-item verification on the risk transmission paths, and output compliance judgment results.

[0020] Preferably, the strategy optimization model comprises:

[0021] an environment perception layer configured to receive the intermediate results and parse them into state information, action constraints and reward signals observable by the agent;

[0022] a value network layer configured to evaluate long-term expected returns of performing different actions according to the current state information, and output value evaluation data;

[0023] a strategy network layer configured to receive the state information and the value evaluation data, calculate and output an optimized action probability distribution to be performed in the current state;

[0024] an action output layer configured to sample according to the optimized action probability distribution, determine and output a final decision action of the current agent.

[0025] Preferably, the process based on the heterogeneous agent roles and structured constraints of entities in the industrial chain comprises:

[0026] mapping the entity nodes in the dynamic relationship graph into heterogeneous agent roles according to their business functions and industry positioning, the heterogeneous agent roles including core manufacturing enterprises, upstream suppliers, downstream distributors and logistics service providers;

[0027] configuring exclusive action spaces and observation spaces for each agent role, the action spaces defining resource scheduling operations executable by the heterogeneous agent roles, and the observation spaces defining decision information obtainable by the heterogeneous agent roles;

[0028] converting the topological connection relationships in the graph into interaction constraints in the multi-agent reinforcement learning environment, and limiting the interaction modes of information flow and material flow between agents.

[0029] Preferably, the process of taking the intermediate result as a constraint condition and incentive function of multi-agent reinforcement learning comprises: setting the risk transmission path in the intermediate result and the compliance judgment result as a constraint condition for policy evaluation in a training environment containing historical cases; combining the risk evaluation aggregation value in the intermediate result with the true final result in the historical cases to build an incentive function of the multi-agent reinforcement learning environment; and using agents representing different roles in an industrial chain to perform offline learning and policy mining on historical data in a static training environment, and outputting a strategy set with the highest correlation to historical successful cases after learning convergence as a resource scheduling suggestion

[0030] Compared with the prior art, the present application has the following beneficial effects:

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

[0032] 2. The explainability is enhanced by risk path identification, and the fusion analysis model performs graph representation, path identification, probability reasoning and compliance verification to accurately identify the risk transmission path and quantitatively evaluate the risk level, thereby improving the explainability problem caused by black box prediction in traditional methods and making the risk analysis more intuitive and traceable.

[0033] 3. By mapping entities to heterogeneous 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 that general AI models make decisions that violate business common sense, makes the suggestions more in line with actual operation, and enhances the credibility. BRIEF DESCRIPTION OF DRAWINGS

[0034] Fig. 1 A multi-relationship modeling method flowchart based on a dynamic relationship graph is provided for the embodiments of the present application.

[0035] Fig. 2 A fusion analysis model flowchart is provided for the embodiments of the present application.

[0036] Fig. 3 A strategy optimization model flowchart is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0037] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0038] Embodiment one:

[0039] Please refer to Figs. 1 to 3 The present application provides a kind of based on dynamic relationship graph industry chain multi-relationship modeling method, technical scheme is as follows:

[0040] A kind of based on dynamic relationship graph industry chain multi-relationship modeling method, specifically as Fig. 1 As shown, comprising:

[0041] Receive industry chain data, the industry chain data includes event logic data and operating data;

[0042] Utilize the event logic data to build out the topological structure and event logic sequence of entity representation event graph framework;Utilize operating data to quantify the entity node and event relationship in event graph framework, give dynamic quantitative attribute of operating health degree and transaction activity degree;

[0043] Through fusion analysis model reasoning, risk transmission path is identified based on event graph framework, risk index of risk transmission path is calculated in combination with dynamic quantitative attribute, and it is ensured that the entire deduction process meets the preset business rules, and the intermediate results containing risk assessment and compliance judgment, mutual checking are output;

[0044] Through strategy optimization model, the intermediate results are converted into resource scheduling decision support suggestions, and the intermediate results are used as constraint condition and incentive function of multi-agent reinforcement learning, and based on the heterogeneous agent role and structured constraint of each entity in industry chain, resource scheduling suggestion is generated.

[0045] Further, the transaction data includes enterprise business information, judicial litigation information, and contract performance information; the judicial litigation information collects the latest litigation information about the supplier by connecting with the judicial public document website, showing that the supplier is involved in a patent infringement lawsuit by a third-party company for controlling core technology; the contract performance information, after obtaining authorization, retrieves the annual procurement contract signed with the supplier and the machine tool sales contract signed with the downstream customer from the contract management of the core enterprise, and extracts the structured key contents such as delivery terms and breach of contract responsibilities in the contract. The operating data includes enterprise financial statement data, tax reporting data, and supply chain transaction data. The enterprise financial statement data obtains the business registration information, shareholder structure, historical change record, and branch information of the core enterprise, supplier, and downstream customer by accessing the national enterprise credit information public, and obtains the quarterly and annual financial reports of the core enterprise and downstream customer as listed companies, focusing on the asset-liability ratio, net cash flow, and inventory turnover rate. The tax reporting data, with the authorization of the supplier, accesses the tax interface to obtain its value-added tax reporting data for the past six months. The data shows that the sales revenue of the supplier has declined significantly in the last quarter. The supply chain transaction data is obtained by connecting with the enterprise resource planning (ERP) of the core enterprise, which provides real-time procurement order records, supplier delivery batch and acceptance qualification rate, and real-time inventory level of the precision-controlled material.

[0046] By clearly defining the sources of transaction data and operating data, the comprehensiveness and interpretability of the industry chain analysis are improved. By integrating multi-source data, the legal, performance, and operating status between entities are fully described, providing a more accurate basis for risk assessment. This data integration enhances the transparency and traceability of the analysis results, providing reliable and comprehensive data support for collaborative decision-making.

[0047] Further, the process of constructing the fact graph framework representing the topological structure between entities and the logical sequence of events using the fact data comprises: processing the fact data using a pre-trained industrial chain field language model, synchronously performing event trigger word recognition and entity recognition; based on the event trigger word, performing event argument extraction to obtain structured event information; the pre-trained industrial chain field language model is specifically an enhanced model with a dual-channel fusion architecture, the dual-channel fusion architecture of the model includes a text semantic channel responsible for processing unstructured text content, and a structured information channel specially responsible for extracting key numerical and category data, before event argument extraction, a fusion layer fuses the feature representations output by the 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, effectively avoiding analysis mistakes of a single model. Using an event correlation evaluation model, the structured event information is analyzed to evaluate the probability dependence relationship between events, and an event chain representing the potential risk transmission logic is constructed; the identified entities are taken as entity nodes, the structured event information is taken as event nodes, and the evaluated event chain is taken as a directed edge connecting the event nodes, and the fact graph framework is constructed; the pre-trained industrial chain field language model is used to process the collected fact data (judicial documents, contract texts and other unstructured data), and event trigger word recognition and entity recognition are synchronously performed, aiming to automatically find the words describing key business events (such as filing a lawsuit, signing a contract, breach of contract) and the entities participating in the events in the text. Then, based on the identified event trigger words, an event argument extraction module is started to extract various elements related to the event (such as time, place, participants, etc.) from the text context and integrate them into a structured event information, for example: {Event ID: [unique identifier], event type: [such as legal risk, cooperation relationship], event subject: [entity name],...}.

[0048] Then, the structured event information is analyzed using an event correlation evaluation model. The model combines built-in business logic and risk knowledge base to construct potential event transmission paths by evaluating the correlation strength and logical sequence between different events. For example, the model will determine that the contract breach event is a high-risk precursor event of the supply chain disruption event, thereby establishing a logical connection between the two. The business logic and risk knowledge base are specifically an event relationship library constructed based on ontology, which pre-defines event categories such as legal litigation, contract signing, and supply chain disruption, as well as logical relationships such as precursor, possible cause, and belonging. For example, the library contains a piece of knowledge: {subject: legal litigation, relationship: may_cause, object: business deterioration, condition: litigation subject amount > 1 million}.

[0049] Finally, the construction of the graph framework is performed. It creates all identified entities as entity nodes and all structured event information as event nodes. According to the supply and sale, guarantee and other relationships parsed from data such as contracts, directed edges are established between entity nodes; at the same time, the evaluated event chain is taken as a directed edge connecting different event nodes, thereby constructing a complete graph framework of the reason, and the pre-trained industrial chain field language model is specifically a model based on the BERT architecture. After the basic pre-training using general corpus, the model is further pre-trained in the field using public research reports, news information and legal documents related to the industrial chain to enhance the understanding ability of specific business terms and event descriptions. The process of the event correlation evaluation model calculating the correlation strength index includes: for two structured events A and B, whether there is a may_cause path from A to B in the knowledge base is retrieved, if there is, the conditional probability of the occurrence of event B within a time window T (such as 90 days) after the occurrence of event A is calculated according to historical data. Combining the attributes of the event itself (such as the amount of litigation subject matter, the amount of default), through a pre-set sigmoid function, it is mapped into an influence factor, and finally the conditional probability of the occurrence of event B is multiplied by the influence factor to obtain an event index. When the event index is greater than a preset threshold (such as 0.75), an event chain is established.

[0050] Through automatic event correlation analysis, scattered text information is converted into a logically clear event transmission network. The logical relationship between events is constructed, and potential transmission paths are mined, rather than just presenting isolated events themselves. The black box correlation in traditional models is converted into a white box event transmission path that can be reviewed and traced, so that each identified risk path has clear logic and narrative, improving the transparency of the analysis process and the credibility of the conclusion.

[0051] Further, the process of quantifying the entity nodes and event relationships in the graph framework using the operation data includes: establishing a node index system representing the operation state of the entity and an edge index system representing the interaction strength of the relationship; calculating the operation data according to the index system to generate node attribute values and edge attribute values; periodically attaching the generated node attribute values and edge attribute values to the entity nodes and event relationships in the event graph framework respectively to generate the dynamic relationship graph. When establishing the node index system representing the operation state of the entity, a data-driven backtracking test mechanism is introduced. Specifically, the system pre-maintains a candidate index pool containing a plurality of industry standard financial and operation indexes. When an evaluation model needs to be constructed for a specific industry (for example, the automobile manufacturing industry), the system will start an automated backtracking test process. This process will traverse each candidate index in the index pool and independently verify whether the index has significant predictive ability for real risk events (such as enterprise defaults, major operation crises, etc.) in the past few years using the historical data of the industry. For example, it will be verified whether the decrease of inventory turnover rate has a high correlation with the historical cases of subsequent operation crises. After the test is completed, the system will sort the indexes according to their predictive ability performance and automatically select the top-ranked indexes (for example, the top five) to combine them into an optimal index system specially designed for the industry and verified by historical data. The subsequent operation health score calculation is based on the optimal index system. Through data backtracking test instead of subjective expert experience, the most predictive index combination is automatically selected for different industries, ensuring the objectivity and accuracy of risk assessment, and making the final decision based on more reliable data.

[0052] Specifically, the operating health score of the node is obtained by weighted comprehensive evaluation of multiple key financial indicators. Specifically, the method selects core indicators such as asset-liability ratio, net cash flow, and inventory turnover rate; first, the numerical value of each original financial indicator is converted into a standardized percentage score by a pre-set normalization function, and the normalization function is a piecewise linear mapping function, which sets the mapping rule according to the median value of the industry to which the indicator belongs. Specifically: if the indicator value is equal to the industry median, it is mapped to a standard score of 60 points; if the indicator value is better than the industry median (for example, for asset-liability ratio, the value is lower; for net cash flow, the value is higher), it is linearly mapped between 60 points and 100 points, and the indicator value reaching the top 10% of the industry is mapped to 90 points; if the indicator value is worse than the industry median, it is linearly mapped between 0 points and 60 points, and the indicator value reaching the bottom 10% of the industry is mapped to 30 points; then, a pre-set weight is assigned to each standardized score. The pre-set weight is intended to reflect the relative importance of different indicators in a specific industry chain link, and is determined as follows: first, load an expert knowledge base containing weight configurations for multiple industries, which is pre-built by industry analysts and sets baseline weights for different indicators in different industries; second, use principal component analysis to calculate historical data for a specific industry chain, and objectively generate a set of data-driven weight suggestions by analyzing the contribution of each financial indicator to the overall operating condition fluctuations; finally, fuse the baseline weights of the expert knowledge base with the data-driven weight suggestions, and use weighted average method to obtain the final pre-set weight; for example, in the weight configuration of the manufacturing industry, the final determination is to allocate higher weights to asset-liability ratio and net cash flow, accounting for 40% and 20% respectively, and allocate a lower weight to inventory turnover rate, accounting for 20%; finally, sum the weighted scores of each indicator to obtain the final operating health comprehensive score.

[0053] The transaction activity score of the edge is calculated by comprehensively considering the transaction frequency and transaction amount. The method counts the total number of transactions and the total amount of transactions in a specific period (such as the last 90 days). In order to smooth the influence of extreme data, the two values are first logarithmically processed, then multiplied by an adjustable coefficient to balance their importance, and finally the processed results are added to obtain the final transaction activity score.

[0054] Finally, the generated node attribute values and edge attribute values are periodically (such as daily or weekly) attached to entity nodes and relationship edges in the ontology graph framework, for example, {operating health: 85, credit risk: low} is attached to an entity node as an attribute. This process is continuously carried out through automated data updating tasks, thereby converting the static ontology graph framework into a dynamic relationship graph that can dynamically reflect the latest status of the industry chain.

[0055] By quantifying the entity nodes and event relationships in the business data-based rational map framework, the interpretability of the analysis results and the ability to guide collaborative decision-making are significantly improved. By constructing a node index system and an edge index system, financial, tax, and supply chain transaction data are converted into dynamic attribute values and periodically updated to the map. This quantification method makes the entity health and transaction activity clear and visible, enhances the traceability of the risk transmission path, and provides an intuitive, data-driven basis for industry chain collaborative decision-making, thereby effectively guiding actual operations.

[0056] Further, the specific process diagram of the fusion analysis model is as shown in Fig. 2 , which includes:

[0057] a graph representation layer, configured to receive the dynamic relationship map and convert it into a vectorized representation that can be processed by the model;

[0058] a path recognition layer, configured to perform graph traversal calculation on the vectorized representation, search for and output all potential risk transmission paths;

[0059] a risk assessment layer, configured to, for each risk transmission path, combine the dynamic quantitative attributes of the nodes on the path to perform comprehensive risk assessment and output a risk index;

[0060] a compliance verification layer, configured to load a preset rule library, verify the risk transmission paths one by one, and output compliance judgment results.

[0061] Specifically, first, the graph representation layer receives the generated dynamic relationship map. This layer uses graph convolution network technology to uniformly convert the topological structure of the entire map, node attributes, and edge attributes into a high-dimensional vectorized representation that can be processed by the model.

[0062] Next, the path recognition layer performs calculation on the above generated vectorized representation. This layer utilizes a depth-first search algorithm, taking the identified risk event node as the starting point, searching along the entity relationship edges and event relationship edges in the graph, thereby identifying and outputting all potential risk transmission paths. Then, the risk assessment layer performs quantitative evaluation on each risk transmission path output by the path recognition layer. This layer extracts the operating health score of each node and the transaction dependence of each edge along the specified risk transmission path, and then these attribute values are sent as input evidence into a Bayesian network model for comprehensive reasoning. The structure of the Bayesian network is dynamically generated according to the relationship edges of the graph, and the conditional probability table (CPT) is learned by maximum likelihood estimation according to the quantitative attributes such as the operating health of the node and the transaction activity of the edge. The specific steps of the risk assessment layer to calculate the risk index are as follows: first, the continuous scores of the operating health scores of the nodes are converted into discrete states such as high risk, medium risk, and low risk according to the preset interval, and these discrete states are then used as observed evidence in the Bayesian network model to update and calculate the posterior probability of each node occurring a specific risk. To aggregate the risk of the entire path, the model first calculates the safety probability of each node by subtracting the risk index of the node from 1, and then multiplies the safety probabilities of all nodes on the path to obtain a total path safety probability. Finally, subtracting the total path safety probability from 1 gives the final index representing the risk level of the entire path, which is output after normalization to a percentage score.

[0063] Finally, the compliance verification layer constrains and verifies the above process. This layer preloads a rule library related to the business scenario (e.g., containing specific provisions in the credit policy of financial institutions that the concentration of single customer credit should not exceed 20%). It compares each identified risk transmission path with the provisions in the rule library one by one, and outputs the compliance judgment result of whether the path touches specific rules. Through the cooperative work of the above four levels, the model finally outputs the intermediate results containing path explanation, risk assessment, and compliance judgment, which are mutually verified.

[0064] The dynamic relationship graph is converted into processable data by the graph representation layer, the risk transmission path is clearly identified by the path recognition layer, the risk index is accurately calculated by the risk assessment layer combined with dynamic attributes, and the compliance verification layer ensures that the results meet the business rules. This structured analysis not only makes the risk path and probability transparent and traceable, but also enhances the reliability of the results through compliance verification, providing precise and reliable guidance for industry chain collaborative decision-making.

[0065] Further, the specific flowchart of the strategy optimization model is as followsFig. 3 as shown, comprising:

[0066] an environment perception layer, configured to receive the intermediate result and parse it into state information observable by the agent, action constraints and reward signals;

[0067] a value network layer, configured to evaluate long-term expected returns of performing different actions according to the current state information, and output value evaluation data;

[0068] a policy network layer, configured to receive the state information and the value evaluation data, calculate and output an optimized action probability distribution to be performed in the current state;

[0069] an action output layer, configured to sample according to the optimized action probability distribution, determine and output a final decision action of the current agent.

[0070] First, the environment perception layer is the input interface of the model, and the core responsibility is to receive the intermediate result generated by the fusion analysis model, and according to the parsing rules defined in the application, the input dynamic relationship graph is converted into a standardized state information feature vector. In the offline training phase, this layer is responsible for batch reading and parsing historical state, historical action and real outcome data from the historical case database. In the online application stage, it is responsible for parsing the current real-time risk scenario to provide decision basis for the subsequent network layer.

[0071] The value network layer plays the role of a critic, and its core function is to evaluate the pros and cons of decision actions. In the offline training process, this layer learns to predict the long-term return that can be brought by performing a specific historical action in a specific historical state. To solve the problem of overestimation of the value of out-of-distribution actions, the algorithm idea of conservative Q-learning is introduced. This method adds an additional regularization term to the standard value function learning, which is used to punish and reduce the evaluation value of actions that have not appeared in the historical data set, so as to ensure that the value network layer can make a stable evaluation and effectively avoid unrealistic evaluation.

[0072] The policy network layer plays the role of an actor, which receives the state information transmitted by the environment perception layer as input. In the offline training stage, to ensure that the generated strategy does not deviate from the reliable action range verified in the historical data, the learning process of this layer is also constrained. It not only needs to learn to maximize the conservative evaluation value given by the value network layer, but also the output action distribution must be constrained within a range similar to the behavior strategy of the historical data. In this way, the strategy learned by this layer not only pursues high returns, but also does not produce unrealistic and aggressive decision actions.

[0073] Finally, the action output layer is the final outlet of the model, which receives the optimized action probability distribution calculated by the policy network layer for the current state. Then, the layer determines and selects the action with the highest probability according to the probability distribution, sets it as the final decision action of the current agent, and converts it into a specific and understandable resource scheduling suggestion for output. At the same time, the current state and the suggested action can be input into the pre-trained value network layer to output a confidence score for the decision suggestion, providing an additional reference for the decision maker. Further, when the decision confidence score output by the action output layer is higher than the preset high threshold, the decision suggestion is determined to be highly reliable, and it is automatically routed to the downstream automated processing system to achieve fast response to high certainty opportunities. When the confidence score is lower than the preset low threshold, the decision is automatically suspended, and the artificial expert review process is triggered. When the artificial review is triggered, the system not only prompts the risk, but also submits the candidate actions with the highest probability output by the policy network along with their respective confidence scores to the expert for review. Decision suggestions between the two thresholds enter the standard artificial review process. Through automatic execution of high certainty suggestions to improve efficiency, and reporting of uncertain risks 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 intermediate results analyzed by the environment perception layer are state, constraint and reward information available to the agent, the value network layer evaluates long-term returns, the policy network layer generates optimized action probabilities, and the action output layer outputs the final decision action. This hierarchical method converts risk analysis into specific action suggestions, ensures that the decision conforms to the dynamic logic of the industry chain, and enhances the operability and explainability of the results, providing precise guidance for resource scheduling.

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

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

[0077] Each type of agent role is configured with a dedicated action space and an observation space, the action space defines the resource scheduling operations that can be performed by the heterogeneous agent role, and the observation space defines the decision information that can be obtained by the heterogeneous agent role.

[0078] The topological connection relationship in the graph is converted into an interaction constraint in the multi-agent reinforcement learning environment, limiting the interaction mode of information flow and material flow between agents.

[0079] A mapping process of the execution entity to the intelligent agent role is performed, which traverses each entity node in the dynamic relationship graph, and according to the business function and industry positioning information of the entity node, a preset rule table is matched to assign a heterogeneous intelligent agent role to each entity node. The rule representation example is: if the industry classification of an entity node belongs to manufacturing industry and is in a core position in the industry chain, it is mapped to a core manufacturing enterprise role; if it 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; if it belongs to the transportation industry, it is mapped to a logistics service provider role; the intelligent agent role generated for each mapping is configured with its exclusive action space and observation space. According to the type of the role, the 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 can be executed by the role, for example, the action space loaded for the upstream supplier role is {on-time delivery, delayed delivery, adjust production plan}, and the observation space is defined as a fixed-dimensional vector, each dimension of which corresponds to a decision information that the role can obtain when making a decision, 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, etc.

[0080] Finally, the topological connection relationship in the graph is converted into the interaction rule in the multi-agent reinforcement learning environment. This process traverses each directed edge in the graph, and for an edge from entity A pointing to entity B, a one-way communication channel from intelligent agent A to intelligent agent B is established in the reinforcement learning environment, which is set to allow intelligent agent A to send information flow and material flow state updates to intelligent agent B, while prohibiting intelligent agent B from initiating communication to intelligent agent A through this channel. After this process is completed, a multi-agent decision environment based on roles reflecting the real topological constraints of the industry chain is constructed.

[0081] Through the generation of resource scheduling suggestions based on heterogeneous intelligent agent roles and structured constraints, the authenticity and effectiveness of industry chain collaborative decision-making are significantly improved. By mapping entities in the dynamic relationship graph to core manufacturing enterprises, suppliers, distributors, etc., and configuring exclusive action and observation spaces, it is ensured that the decision-making conforms to the role functions; by converting the graph topology relationship into interaction constraints, a decision-making environment is constructed that truly reflects the logic of the industry chain. This way makes the scheduling suggestions highly consistent with the actual operation, enhancing the credibility and operability.

[0082] Further, the process of taking the intermediate result as a constraint condition and incentive function of multi-agent reinforcement learning includes: setting the risk transmission path and compliance judgment result in the intermediate result as a constraint condition for policy evaluation in a training environment containing historical cases; combining the risk evaluation aggregation value in the intermediate result with the real final result in the historical cases to build an incentive function of the multi-agent reinforcement learning environment; and using agents representing different roles in an industrial chain to perform offline learning and policy mining on historical data in a static training environment, and outputting a policy set with the highest correlation to historical successful cases after learning convergence as a resource scheduling suggestion.

[0083] Firstly, historical training data is constructed and labeled. For each case in the historical case database, the model will first run the fusion analysis model to calculate the intermediate result of the case at the historical node, i.e., the risk transmission path, the risk index and the compliance judgment.

[0084] Secondly, the incentive function is built. The incentive function of the model is not calculated in real time, but directly derived from the real final result in the historical cases. For example, if an enterprise in a historical case develops healthily within a year after decision-making, the reward signal of the case is set to a high positive value; if the enterprise deteriorates or defaults, it is set to a negative value. When building the incentive function, the risk evaluation aggregation value is used as an adjustment factor. The specific construction logic is that the final incentive value is composed of three parts: the first part is the basic reward based on the real outcome, which is obtained by multiplying the quantified outcome by the first weight; the second part is the additional incentive of risk correlation, which is 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 then subtracting the third part.

[0085] Then, offline learning and policy mining are started. In this stage, agents representing different roles in an industrial chain learn in a static training environment composed of massive historical data. The model uses the above labeled intermediate results as constraints for policy learning, and uses the incentive function built from the historical real outcome as the optimization target. Through offline reinforcement learning algorithm, the model iterates and optimizes its internal policy network, aiming to find a policy that can map the current state to the decision action with the highest correlation to historical successful experience. The model uses the above labeled intermediate results as constraints for policy learning, and the specific implementation is as follows: at each time step of the reinforcement learning environment, if the action selected by the agent triggers a pre-warning risk transmission path or violates the compliance judgment, a preset fixed penalty value (100 in this embodiment) will be subtracted from the task reward when calculating the total reward.

[0086] Finally, the resource scheduling suggestion is generated and output. When the offline training process of the entire model meets the preset termination condition (such as reaching the maximum training step or the model performance converges), the stable strategy set learned by each agent will be fixed and saved. In actual application, this pre-trained model can be directly called to quickly output the final resource scheduling suggestion according to the input current risk scenario.

[0087] By converting risk analysis into resource scheduling suggestion through multi-agent reinforcement learning, the practicability of industry chain collaborative decision-making is significantly improved. By generating a globally optimal scheduling suggestion, the operability, explainability and adaptability to industry chain dynamics of the decision are enhanced.

[0088] By constructing an industry chain multi-relation modeling method based on a dynamic relationship graph, by fusing fact data (business, litigation, contract information) and operation data (financial, tax, transaction data), a dynamic graph is constructed to clearly present entity relationships and event logic and improve analysis transparency. The fusion analysis model clearly constructs risk transmission paths and evaluates their risk levels through graph representation, path identification, quantitative evaluation and compliance verification. The strategy optimization model uses multi-agent reinforcement learning to convert risk analysis into resource scheduling suggestion. By simulating the real logic of the industry chain through heterogeneous agent roles and interaction constraints, a reference scheme for collaborative decision-making is generated. This method makes the risk path more traceable and explainable, and improves the commercial applicability of the suggestion by simulating a realistic decision-making environment, thereby optimizing the practicability and operability of industry chain risk management and collaborative decision-making.

[0089] Embodiment two:

[0090] B company applies for credit from a financial institution. To comprehensively assess the credit quality, operating stability and potential risks of the company in the industry chain, the financial institution introduces an industry chain multi-relation modeling method based on a dynamic relationship graph provided by the present application. The specific implementation process is as follows:

[0091] First, the data collection is started. For this credit approval scene, two types of data related to the applicant company are collected. The first type is fact data, which specifically collects the company's business change records in the past two years, a recent product quality-related judicial litigation judgment of its core downstream customers, and several technical patent information held by the applicant company. The second type is operation data. After authorization, the applicant company's value-added tax invoice details issued to its core downstream customers in the past 12 months, its tax return data and water and electricity payment records are collected.

[0092] After the data collection, the method enters the graph construction and quantification stage. First, the graph framework is constructed using the case data: through natural language processing technology, the product quality litigation event is extracted from the judicial judgment book and associated with the entity node of the core downstream customer. At the same time, according to the clear information of the buyer and seller in the invoice data, a directed edge from the applicant enterprise to the core downstream customer is established, representing the supply relationship between the two, thus forming the basic topological structure and event logic. Subsequently, the graph is quantified using the business data: according to the tax data and water and electricity consumption data, a dynamic business health score of 85 is calculated for the applicant enterprise node; at the same time, according to the invoice data showing that 80% of the sales of the applicant enterprise come from the core customer, a transaction dependence dynamic attribute with a value of 0.8 is assigned to the directed edge connecting the two.

[0093] After the graph construction is completed, the method starts the fusion analysis model for deep reasoning. The path recognition layer of the model first performs traversal calculation on the graph data and quickly identifies a risk transmission path: starting from the product quality litigation event of the core downstream customer node, along the transaction relationship edge with a high dependence attribute of 0.8, and finally transmitting to the applicant enterprise node. Then, the risk assessment layer of the model intervenes, which combines the core downstream customer's credit rating downgraded due to the lawsuit and the high transaction dependence of 0.8 to calculate the risk index of the applicant enterprise's revenue decline of more than 20% in the next six months as 75. At the same time, the compliance verification layer loads the credit approval rule library of the financial institution to judge that the risk index of 75 has exceeded the acceptable risk threshold for this type of enterprise. Finally, an intermediate result containing the above risk transmission path, risk index evaluation and compliance judgment is output, which are mutually verified.

[0094] After receiving the intermediate result, the method starts the strategy optimization model to convert the analysis conclusion into an executable plan. First, the model receives the risk scenario of the applicant enterprise, including its risk index of up to 75 and transaction dependence on a single customer of 0.8. Subsequently, the model performs high-speed retrieval in its internal historical case database to find similar historical credit cases with similar risk profiles. The model successfully matches several historical cases with similar industry backgrounds and high customer concentration risks (dependence higher than 70% and risk score in the 70-80 interval).

[0095] Next, the model analyzes these historical cases, focusing on the different credit decisions taken at the time, and the real end results of these decisions when the cases were archived. The analysis shows that for such high-dependence risk enterprises, a higher proportion of historical cases that directly approved a standard loan ended up with repayment difficulties later on; while a few historical cases that took a structured credit solution linked to the diversification process of the applicant enterprise client ended up with the enterprise successfully optimizing its client structure and achieving sound operation.

[0096] The strategy network of this model has learned this rule through offline training on all historical data, so when the risk scenario of the current applicant enterprise is input, the model will identify the historical best practice of designing a structured credit solution linked to client diversification as the decision action with the highest expected return.

[0097] Finally, the model translates this strategy into a specific resource allocation recommendation output: it recommends that the financial institution approve a structured loan, 60% of which is unconditional credit, and the other 40% of which is linked to the applicant enterprise successfully establishing a business relationship with a new candidate client that is not less than 15% of total sales within the next year. This solution actively guides and optimizes the identified industry chain risk through the leverage of credit resources, and its recommendation is based on the deep mining of real historical success experience.

[0098] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for multi-relationship modeling of an industrial chain based on a dynamic relationship graph, characterized in that, The method comprises the following steps: receiving industry chain data, wherein the industry chain data comprises transaction data and operation data; constructing a transaction graph framework representing the topological structure and logical sequence of events between entities by using the transaction data; quantifying the entity nodes and event relationships in the transaction graph framework by using the operation data, and assigning dynamic quantitative attributes of operation health and transaction activity to the entity nodes and event relationships; performing reasoning by using a fusion analysis model, identifying risk transmission paths based on the event graph framework, calculating risk indexes of the risk transmission paths in combination with the dynamic quantitative attributes, and ensuring that the entire reasoning process complies with preset business rules, and outputting intermediate results including risk assessment and compliance judgment, which are cross-verified; transforming the intermediate results into resource scheduling decision support suggestions by using a strategy optimization model, and using the intermediate results as constraint conditions and incentive functions of multi-agent reinforcement learning, and generating resource scheduling suggestions based on the heterogeneous agent roles and structured constraints of entities in the industry chain. 2.The industry chain multi-relationship modeling method based on dynamic relationship graph according to claim 1, characterized in that, The transaction data comprises enterprise business information, judicial litigation information, and contract performance information; the operation data comprises enterprise financial statement data, tax declaration data, and supply chain transaction data. 3.The industry chain multi-relationship modeling method based on dynamic relationship graph according to claim 1, characterized in that, The process of constructing a transaction graph framework representing the topological structure and logical sequence of events between entities by using the transaction data comprises the following steps: processing the transaction data by using a pre-trained industry chain field language model, and simultaneously performing event trigger word recognition and entity recognition; performing event argument extraction based on the event trigger words to obtain structured event information; analyzing the structured event information by using an event correlation evaluation model to evaluate the probabilistic dependency relationship between events, and constructing an event chain representing the potential risk transmission logic; identifying entities as entity nodes, structured event information as event nodes, and the evaluated event chain as directed edges connecting the event nodes, and constructing the transaction graph framework. 4.The industry chain multi-relationship modeling method based on dynamic relationship graph according to claim 1, characterized in that, The process of quantifying the entity nodes and event relationships in the graph framework by using the operation data comprises the following steps: establishing a node index system representing the operation state of entities and an edge index system representing the interaction intensity of relationships; calculating the node attribute values and edge attribute values according to the node index system and the edge index system; and periodically attaching the generated node attribute values and edge attribute values to the entity nodes and event relationships in the transaction graph framework, respectively, to generate the dynamic relationship graph.

5. The method of claim 1, wherein, The fusion analysis model comprises: a graph representation layer for receiving the dynamic relationship graph and transforming it into a vectorized representation processable by the model; a path identification layer for performing graph traversal calculation on the vectorized representation, searching for and outputting all potential risk transmission paths; a risk assessment layer for performing comprehensive risk assessment on each risk transmission path in combination with the dynamic quantitative attributes of the nodes on the path, and outputting a risk index; a compliance verification layer for loading a preset rule library to verify each risk transmission path, and outputting a compliance judgment result.

6. The method of claim 1, wherein, The strategy optimization model comprises: An environment perception layer is configured to receive the intermediate results and parse them into state information, action constraints and reward signals observable by the agent. A value network layer is configured to evaluate long-term expected returns of performing different actions based on the current state information and output value evaluation data. A policy network layer is configured to receive the state information and the value evaluation data, calculate and output an optimized action probability distribution to be performed in the current state. An action output layer is configured to sample the optimized action probability distribution, determine and output a final decision action of the current agent.

7. The method of claim 1, wherein, The process based on the heterogeneous agent roles and structured constraints of entities in the industrial chain includes: The entity nodes in the dynamic relationship graph are mapped into heterogeneous agent roles according to their business functions and industry positioning, and the heterogeneous agent roles include core manufacturing enterprises, upstream suppliers, downstream distributors and logistics service providers. Each type of agent role is configured with a dedicated action space and observation space, the action space defines the resource scheduling operations that can be performed by the heterogeneous agent role, and the observation space defines the decision information that can be obtained by the heterogeneous agent role. The topological connection relationship in the graph is converted into interaction constraints in the multi-agent reinforcement learning environment to limit the interaction modes of information flow and material flow between agents. 8.The industry chain multi-relationship modeling method based on dynamic relationship graph according to claim 1, characterized in that, The process of using the intermediate results as constraint conditions and incentive functions of multi-agent reinforcement learning includes: setting the risk transmission path and compliance judgment result in the intermediate results as constraint conditions for policy evaluation in a training environment containing historical cases; combining the risk evaluation aggregation value in the intermediate results with the real final results in the historical cases to build an incentive function for the multi-agent reinforcement learning environment; using agents representing different roles in the industrial chain to perform offline learning and policy mining on historical data in a static training environment, and outputting a strategy set with the highest correlation to successful historical cases after learning convergence as resource scheduling suggestions.

Images