Management decision method and system based on knowledge base construction technology

Abstract

The application discloses a kind of management decision-making method and system based on knowledge base construction technology.Therein, the method includes: in the time sequence relationship extraction of multi-source financial data, obtain the time sequence relationship set related to query content;Event-entity association matrix corresponding to time sequence relationship set is constructed;According to time sequence relationship set and event-entity association matrix, dynamic knowledge graph is constructed;Adopt double machine learning model, determine the causal effect parameter in causal graph structure;According to causal graph structure and causal effect parameter, construct structural causal model;Generate counterfactual prediction result using structural causal model, and generate the decision scheme corresponding to query content based on counterfactual prediction result.The application solves the technical problem that the decision information including accurate causal basis and forward-looking simulation information cannot be generated due to the difficulty in determining the causal relationship between financial data and the inability to dynamically deduce intervention effect in the related management decision-making method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a management decision method and system based on knowledge base construction technology. BACKGROUND

[0002] The related financial decision support method is mainly based on statistical analysis and correlation mining of historical data, which is difficult to effectively distinguish the causal relationship and correlation relationship between variables. Due to the lack of deep modeling of the causal mechanism in the financial dynamic system, the related method cannot perform counterfactual reasoning and effect estimation on various intervention policies, resulting in the generated decision scheme lacking precise causal basis and forward-looking simulation support.

[0003] At present, no effective solution has been proposed for the above problems. SUMMARY

[0004] The present application provides a management decision method and system based on knowledge base construction technology, to at least solve the technical problem that it is difficult to determine the causal relationship between financial data and unable to dynamically deduce the intervention effect in the related management decision method, resulting in the inability to generate decision information including precise causal basis and forward-looking simulation information.

[0005] According to one aspect of the present application, a management decision method based on knowledge base construction technology is provided, comprising: receiving query content, performing time sequence relationship extraction on multi-source financial data according to the query content to obtain a time sequence relationship set related to the query content, wherein the time sequence relationship set is a structured data set for dynamically representing the association relationship between financial entities; constructing an event-entity association matrix corresponding to the time sequence relationship set, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities; constructing a dynamic knowledge graph according to the time sequence relationship set and the event-entity association matrix, and identifying a causal graph structure corresponding to the financial variables in the query content in the dynamic knowledge graph by using a hybrid causal discovery method, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and transmission path between financial variables; determining a causal effect parameter in the causal graph structure by using a double machine learning model, wherein the causal effect parameter is used to quantitatively evaluate the causal influence degree of the intervention measure on the target variable for the query content; constructing a structural causal model according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes a structural equation for defining the causal mechanism between financial variables; generating a counterfactual prediction result by using the structural causal model, and generating a decision scheme corresponding to the query content based on the counterfactual prediction result.

[0006] Optionally, according to the set of time sequence relations and the event-entity association matrix, a dynamic knowledge graph is constructed, including: based on the subject, object, relation and timestamp information in the set of time sequence relations, an initial graph structure carrying time annotation is constructed, wherein the timestamp information is used to identify the starting time or effective time period of the relation; based on the similarity of entity attributes, the nodes in the initial graph structure are aligned to obtain a reference graph; according to the timestamp information in the set of time sequence relations, a time interval attribute is added to each relation edge in the reference graph, wherein the time interval attribute is defined by the starting timestamp and the ending timestamp of the relation; and according to the pre-defined confidence or the statistical features based on the co-occurrence frequency in the set of time sequence relations, the relation strength weight of each relation edge is calculated; the reference graph carrying the relation strength weight and the time interval attribute is determined as the main graph structure, the event feature vector of the entity node in the main graph structure is generated by using the event-entity association matrix, and the event feature vector is used as the initial feature of the node, and the topological connection, relation strength and time interval in the main graph structure are learned simultaneously to obtain the graph embedding representation; the reference graph and the graph embedding representation obtained by the time sequence graph embedding algorithm are stored in the graph database to obtain the dynamic knowledge graph, wherein the graph embedding representation is used as the vectorized attribute of the graph node and is stored in association with the topological structure of the reference graph.

[0007] Optionally, a hybrid causal discovery method is used to identify the causal graph structure corresponding to the financial variable in the query content in the dynamic knowledge graph, including: extracting a set of core financial variables related to the query intent of the query content in the dynamic knowledge graph, and obtaining the graph embedding representation corresponding to the core financial variables; using a constraint-based causal discovery algorithm, taking the set of core financial variables as nodes, and taking the time priority constraint included in the set of time sequence relations in the dynamic knowledge graph as a necessary condition for determining the causal direction, an initial causal skeleton graph is constructed, wherein the initial causal skeleton graph is used to represent whether there is a causal association between variables; based on the initial causal skeleton graph, a score-based causal discovery algorithm is used, the similarity between the graph embedding representations is integrated into the graph structure scoring function, and the key event association edges in the event-entity association matrix that are higher than a preset threshold are used as directional constraints for intensive search, and a candidate causal graph structure is obtained by optimization; based on the timestamp information in the set of time sequence relations, the causal edges in the candidate causal graph structure that do not conform to the preset time logic are removed, and the confidence score of the retained causal edges is calculated using the historical data in the dynamic knowledge graph; according to the causal edges that pass the time sequence verification and the confidence score of the causal edges that pass the time sequence verification, a causal graph structure is generated, wherein the causal graph structure includes: time sequence lag attribute and confidence weight.

[0008] Optionally, the causal effect parameter in the causal graph structure is determined by using a double machine learning model, including: determining an input feature set of causal inference based on a causal path in the causal graph structure, wherein the input feature set at least includes a treatment variable, an outcome variable, and a set of confounding variables, the treatment variable is used to represent an intervention to be implemented, the outcome variable is used to represent a target indicator to be evaluated, and the set of confounding variables is used to control background factors that have a common influence on the treatment variable and the outcome variable; using a first machine learning model to predict the treatment variable with the set of confounding variables as input, obtaining a first predicted value of the treatment variable, and calculating a first residual between an actual value of the treatment variable and the first predicted value; using a second machine learning model to predict the outcome variable with the set of confounding variables as input, obtaining a second predicted value of the outcome variable, and calculating a second residual between an actual value of the outcome variable and the second predicted value, wherein the second machine learning model and the first machine learning model are independent of each other in model structure or model parameter; calculating a correlation between the first residual and the second residual by an orthogonalization technique to obtain an average treatment effect parameter, wherein the average treatment effect parameter is used to represent a global average causal influence of the treatment variable on the outcome variable; based on a heterogeneity analysis model, taking the set of confounding variables as a conditional variable to calculate a conditional average treatment effect parameter in different subgroups, wherein the heterogeneity analysis model is a model based on a decision tree or a meta-learner, and the different subgroups are sample sets with similar characteristics divided according to the confounding variable feature space; and the average treatment effect parameter and the conditional average treatment effect parameter are jointly determined as the causal effect parameter.

[0009] Optionally, a structural causal model is constructed according to the causal graph structure and the causal effect parameter, including: constructing a structural equation framework based on nodes and edges in the causal graph structure, wherein each structural equation in the structural equation framework adopts a linear function form, and a current variable is represented as a linear combination of all parent variables of the current variable in the causal graph structure; based on the causal effect parameter, an initial estimated value of each path coefficient in the structural equation framework is back-calculated by path coefficient calculation based on a structural equation model to obtain an initialized structural equation; based on historical data in the dynamic knowledge graph, a maximum likelihood estimation method is used to perform joint parameter estimation on the initialized structural equation to obtain an optimized structural equation parameter set, wherein the optimized structural equation parameter set includes a point estimate value of the path coefficient and a confidence interval of the path coefficient; and the optimized structural equation parameter set is configured to a corresponding edge of the causal graph structure to obtain the structural causal model.

[0010] Optionally, the counterfactual prediction result is generated by using the structural causal model, including: determining a set of intervention operations based on the query content, wherein each intervention operation in the set of intervention operations is represented by a preset operator, and the preset operator is used to simulate intervention on a preset treatment variable in the causal graph structure; for each intervention operation in the set of intervention operations, calculating a potential result of the intervention operation on the result variable based on an optimized structural equation in the structural causal model, to obtain a basic counterfactual prediction; quantifying the uncertainty of the basic counterfactual prediction based on a confidence interval of a path coefficient in the set of optimized structural equation parameters, to obtain a counterfactual prediction distribution including the confidence interval; and integrating the counterfactual prediction distributions corresponding to all intervention operations in the set of intervention operations, to obtain the counterfactual prediction result.

[0011] Optionally, based on the counterfactual prediction result, a decision scheme corresponding to the query content is generated, including: determining an optimization target and a set of constraint conditions based on a business target of the query content, wherein the optimization target at least includes: maximizing revenue and minimizing risk, and the set of constraint conditions is derived from business rules and risk limits in the dynamic knowledge graph; establishing a multi-objective optimization function based on an expected effect value in the counterfactual prediction result as a revenue estimation basis and a confidence interval in the counterfactual prediction result as a risk assessment basis; constructing a risk constraint based on a risk transmission path identified in the causal graph structure, and adding the risk constraint to the set of constraint conditions; solving the multi-objective optimization function by using a multi-objective optimization algorithm to obtain a set of optimal solutions, wherein each solution in the set of optimal solutions corresponds to a candidate decision strategy and an expected performance of the candidate decision strategy on each optimization target; calculating a robustness score of each strategy based on a causal path length and a node influence range of the candidate decision strategy in the causal graph structure; and selecting an optimal decision strategy from the set of optimal solutions according to the robustness score, and converting the optimal decision strategy into an executable decision scheme.

[0012] According to still another aspect of the present application, a management decision system based on knowledge base construction technology is also provided, comprising: a receiving module configured to receive query content, perform time series relationship extraction on multi-source financial data according to the query content, and obtain a time series relationship set related to the query content, wherein the time series relationship set is a structured data set used to dynamically represent the association relationship between financial entities; a first construction module configured to construct an event-entity association matrix corresponding to the time series relationship set, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities; a second construction module configured to construct a dynamic knowledge graph according to the time series relationship set and the event-entity association matrix, and identify a causal graph structure corresponding to a financial variable in the query content in the dynamic knowledge graph by using a hybrid causal discovery method, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and transmission path between the financial variables; a determination module configured to determine a causal effect parameter in the causal graph structure by using a double machine learning model, wherein the causal effect parameter is used to quantitatively evaluate the causal influence degree of an intervention measure on a target variable for the query content; and a third construction module configured to construct a structural causal model according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes a structural equation used to define the causal mechanism between the financial variables; and a generation module configured to generate counterfactual prediction results by using the structural causal model, and generate a decision scheme corresponding to the query content based on the counterfactual prediction results.

[0013] According to still another aspect of the present application, a non-volatile storage medium is also provided, which comprises a stored program, wherein the program controls a device where the storage medium is located to perform the above management decision method based on knowledge base construction technology when the program is running.

[0014] According to still another aspect of the present application, an electronic device is also provided, which comprises a memory and a processor, and the processor is configured to run a program stored in the memory, wherein the program performs the above management decision method based on knowledge base construction technology when the program is running.

[0015] According to still another aspect of the present application, a computer program is also provided, wherein the computer program is executed by a processor to implement the above management decision method based on knowledge base construction technology.

[0016] According to still another aspect of the present application, a computer program product is also provided, which comprises a non-volatile computer readable storage medium, wherein the non-volatile computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the above management decision method based on knowledge base construction technology.

[0017] In the present application, the query content is received, the time sequence relationship extraction is performed on the multi-source financial data according to the query content, and a time sequence relationship set related to the query content is obtained, wherein the time sequence relationship set is a structured data set for dynamically representing the association relationship between financial entities; an event-entity association matrix corresponding to the time sequence relationship set is constructed, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities; a dynamic knowledge graph is constructed according to the time sequence relationship set and the event-entity association matrix, and a hybrid causal discovery method is used to identify a causal graph structure corresponding to a financial variable in the query content in the dynamic knowledge graph, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and the transmission path between the financial variables; a double machine learning model is used to determine a causal effect parameter in the causal graph structure, wherein the causal effect parameter is used to quantitatively evaluate the causal influence degree of the intervention measure on the target variable for the query content; a structural causal model is constructed according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes a structural equation for defining the causal mechanism between the financial variables; an anti-fact prediction result is generated by using the structural causal model, and a decision scheme corresponding to the query content is generated based on the anti-fact prediction result, which achieves the purpose of determining the causal relationship between the financial data and dynamically deducing the intervention effect, thereby realizing the technical effect of generating decision information including precise causal basis and forward-looking simulation information, and further solving the technical problem that it is difficult to determine the causal relationship between the financial data and unable to dynamically deduce the intervention effect in the related management decision method, resulting in the inability to generate decision information including precise causal basis and forward-looking simulation information. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings, which are included to provide a further understanding of the present application, constitute a part of the present application and illustrate the illustrative embodiments of the present application and their description serve to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0019] Figure 1 is a flowchart of a management decision method based on a knowledge base construction technology according to an embodiment of the present application;

[0020] Figure 2 is a structural diagram of a management decision system based on a knowledge base construction technology according to an embodiment of the present application;

[0021] Figure 3 is a hardware structure block diagram of a computer terminal of a management decision method based on a knowledge base construction technology according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application, so that those skilled in the art can better understand the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present application.

[0023] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] According to the embodiments of the present application, a method embodiment of a management decision method based on knowledge base construction technology is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0025] Figure 1 is a flowchart of a management decision method based on knowledge base construction technology according to the embodiments of the present application, as Figure 1 shown, the method comprises the following steps:

[0026] Step S101, receiving query content, according to the query content, performing time sequence relationship extraction on multi-source financial data to obtain a time sequence relationship set related to the query content, wherein the time sequence relationship set is a structured data set for dynamically representing the association relationship between financial entities.

[0027] In step S101, a query request input by a user in a natural language form is received, a query intention is parsed through a semantic understanding technology, and then a time sequence relationship related to the query content is extracted from multi-source financial data. Specifically, a relationship extraction method based on a pre-trained language model can be used to identify financial entities and their associations from unstructured texts such as financial news, company announcements, and research reports, and to label a timestamp and a confidence for each relationship, thereby forming a time sequence relationship set composed of elements such as a subject, a relationship, an object, and a timestamp, to dynamically record the evolution process of the association between financial entities.

[0028] In step S102, an event-entity association matrix corresponding to the time sequence relationship set is constructed, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities.

[0029] In step S102, a standardized financial event classification system is first established, including types such as monetary policy adjustment, financial report release, and merger and reorganization, and then the association strength between events and entities is comprehensively evaluated by calculating indicators such as co-occurrence frequency, semantic similarity, and influence degree, and finally an association matrix with events as rows and entities as columns is formed, wherein each matrix element represents the quantitative association degree between a specific event and an entity.

[0030] In step S103, a dynamic knowledge graph is constructed according to the time sequence relationship set and the event-entity association matrix, and a hybrid causal discovery method is used to identify a causal graph structure corresponding to a financial variable in the query content in the dynamic knowledge graph, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and the transmission path between financial variables.

[0031] In step S103, the time sequence relationship set and the event-entity association matrix are fused to construct a dynamic knowledge graph supporting time sequence queries. Specifically, steps such as entity alignment, relationship disambiguation, and graph embedding learning are included to form a comprehensive knowledge representation containing topological structure, time dimension, and semantic information. Based on this dynamic knowledge graph, a hybrid causal discovery method is used to identify the causal graph structure between financial variables: first, a constraint-based causal discovery algorithm is used to construct an initial causal skeleton to determine the basic association between variables; then, a score-based causal discovery algorithm is used for structure optimization, and a causal graph with clear causal direction and transmission path is generated by optimizing the graph structure scoring function and introducing a time priority constraint.

[0032] In step S104, a dual machine learning model is used to determine a causal effect parameter in the causal graph structure, wherein the causal effect parameter is used to quantitatively evaluate the causal influence of an intervention measure on a target variable for the query content.

[0033] In step S104, the causal effect parameter is estimated by using the dual machine learning model. The dual machine learning model ensures the accuracy of the estimation through two-stage processing: the first stage uses a machine learning model to predict the treatment variable and the outcome variable respectively, and the second stage calculates the causal effect based on the residual. This method can effectively control the influence of confounding variables and accurately estimate the average treatment effect and the conditional average treatment effect.

[0034] In step S105, a structural causal model is constructed according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes structural equations for defining the causal mechanism between financial variables.

[0035] In step S105, based on the obtained causal graph structure and causal effect parameter, a structural causal model is constructed to formalize the causal mechanism between financial variables. The structural causal model consists of a set of structural equations, each equation describing the functional relationship between a variable and its direct cause variable, and the equation parameters are initialized by the causal effect parameter and optimized by the maximum likelihood estimation method to form a complete mathematical representation of the causal system.

[0036] In step S106, counterfactual prediction results are generated using the structural causal model, and based on the counterfactual prediction results, a decision scheme corresponding to the query content is generated.

[0037] In step S106, counterfactual reasoning is performed using the structural causal model to simulate potential results under different decision scenarios through intervention operations. Specifically, do-algorithm and Monte Carlo simulation methods can be used to generate counterfactual prediction distribution including confidence interval. Based on the counterfactual prediction distribution, a multi-objective optimization model is established, considering multiple factors such as revenue, risk, and constraints, and the Pareto optimal solution set is solved by optimization algorithm, and the robustness of each scheme is evaluated, and finally the optimal decision scheme with causal basis and forward-looking simulation support is generated, completing the intelligent transformation from data to decision.

[0038] Optionally, the decision scheme can be to control different agents to perform different operations. For example, in generating the decision scheme, different intervention operations are mapped to corresponding business agent execution units, such as assigning interest rate adjustment instructions to the business agent of the financial institution, issuing credit policy to the agent of the financial institution, and assigning investment strategy to the agent of the asset management institution. Each agent receives instructions that are customized based on counterfactual prediction results, and by analyzing the transmission path and influence range of different intervention measures in the causal network, the optimal operation parameters and execution time of each agent are determined.

[0039] Specifically, in the decision-making scheme generation process, first, a multi-agent collaborative optimization model is established based on counterfactual prediction results. The multi-agent collaborative optimization model comprehensively considers the interaction between different agent operations, analyzes the collaborative effect and conflict risk between operations through causal graph structure, and ensures that the operations of multiple agents can form a combined force rather than cancel each other out. For example, when it is predicted that the interest rate raising operation of the monetary policy agent conflicts with the tax reduction measure of the fiscal policy agent, the operation parameters and execution sequence are automatically adjusted to optimize the overall effect.

[0040] And the specific operation scheme generated for each agent contains complete execution elements: clear operation instructions, expected effect indicators, execution time window, necessary risk control measures and coordination requirements with other agents. These schemes are delivered to the corresponding business system through a standardized interface, and the actual effect data generated by the agent in the execution process is fed back to the business system to verify the accuracy of the causal model and optimize subsequent decisions. This multi-agent decision-making mechanism based on causal reasoning enables each component in the financial system to operate collaboratively based on unified causal cognition, significantly improving decision-making efficiency and execution effect in complex financial environments.

[0041] Through the above steps S101 to S106, the purpose of determining the causal relationship between financial data and dynamically deducing the intervention effect is achieved, thereby realizing the technical effect of generating decision-making information including accurate causal basis and forward-looking simulation information.

[0042] The steps shown in FIG. 10 are exemplarily illustrated and explained as follows. Figure 1

[0043] ​According to some optional embodiments of the present application, the dynamic knowledge graph is constructed according to the set of time sequence relationships and the event-entity association matrix, which can be realized by the following method: based on the subject, object, relationship and timestamp information in the set of time sequence relationships, an initial graph structure carrying time annotation is constructed, wherein the timestamp information is used to identify the starting time or effective time period of the relationship; based on the similarity of entity attributes, the nodes in the initial graph structure are aligned to obtain a reference graph; according to the timestamp information in the set of time sequence relationships, a time interval attribute is added to each relationship edge in the reference graph, wherein the time interval attribute is defined by the starting timestamp and the ending timestamp of the relationship; and according to the predefined confidence or the statistical characteristics based on the co-occurrence frequency in the set of time sequence relationships, the relationship strength weight of each relationship edge is calculated; the reference graph carrying the relationship strength weight and the time interval attribute is determined as the main graph structure, the event feature vector is generated for the entity nodes in the main graph structure by using the event-entity association matrix, and the time sequence graph embedding algorithm is used to learn the topology connection, relationship strength and time interval in the main graph structure by taking the event feature vector as the initial feature of the node, so as to obtain the graph embedding representation; the reference graph and the graph embedding representation obtained by the time sequence graph embedding algorithm are stored in the graph database together to obtain the dynamic knowledge graph, wherein the graph embedding representation is used as the vectorized attribute of the graph node and is stored in association with the topology structure of the reference graph.

[0044] In the present embodiment, first, an initial graph structure carrying time annotation is constructed based on the subject, object, relationship and timestamp information in the set of time sequence relationships, wherein the timestamp information is used to accurately identify the starting time and effective time period corresponding to each relationship, so as to ensure that the graph can reflect the time-sensitive characteristics of the financial entity association relationship. Through the calculation method based on the similarity of entity attributes, the entity alignment processing is performed on the nodes in the initial graph, the ambiguity of the nodes pointing to the same entity in different data sources is eliminated, and the reference graph with unified entity identification is formed.

[0045] Secondly, according to the timestamp information in the set of time sequence relationships, a time interval attribute defined by the starting timestamp and the ending timestamp is added to each relationship edge in the reference graph, and at the same time, according to the predefined confidence or the statistical characteristics based on the co-occurrence frequency, the quantitative relationship strength weight of each relationship edge is calculated, so that the graph can accurately reflect the effective period of the relationship and also reflect the reliability of the relationship.

[0046] The reference graph carrying the relationship strength weight and the time interval attribute is determined as the main graph structure again, and an event feature vector is generated for an entity node in the main graph structure by using an event-entity association matrix. The event feature vector can capture the association mode between the entity and various financial events. A time sequence graph embedding algorithm is used to take the event feature vector as the initial feature of the node, and simultaneously learn the topological connection characteristics, relationship strength weight and time interval information in the main graph structure. High-quality knowledge graph embedding representation is generated by comprehensively considering the network structure, relationship attribute and time sequence feature.

[0047] Finally, the reference graph and the graph embedding representation obtained by the time sequence graph embedding algorithm are stored in the graph database together to construct a complete dynamic knowledge graph. In the dynamic knowledge graph, the graph embedding representation is stored in association with the topological structure of the reference graph as the vectorized attribute of the graph node, which not only retains the explicit semantic relationship of the graph but also has the implicit representation ability of deep learning.

[0048] The above steps break through the static limitation of the traditional knowledge graph by converting the timestamp information into explicit time interval attributes, not only record the existence of the relationship between the financial entities, but also accurately depict the life cycle and effective period of these relationships. The time sequence labeling mechanism enables the graph to dynamically reflect the relationship network that changes with policy adjustment and market cycle in the financial market, providing a data basis for analyzing the time-varying association of economic variables.

[0049] The data heterogeneity problem is solved by entity alignment, the confidence evaluation is integrated by relationship strength weight, and the external event context is introduced by event feature vector. Finally, the topological structure, relationship strength, time dimension and event feature are unified in the same vector space through graph embedding learning. Multi-level fusion effectively overcomes the limitations of single data source, forms a stereoscopic depiction of financial entities and their relationships, and can significantly improve the completeness and accuracy of knowledge representation.

[0050] The dual storage strategy of "reference graph + embedding representation" is innovatively adopted. The reference graph retains the semantic transparency and interpretability of the original relationship, which is convenient for manual understanding and verification. The graph embedding representation extracts the potential semantic features by deep learning, which supports efficient similarity calculation and relationship reasoning. This design not only meets the strict requirements of the financial field for decision interpretability, but also provides computational feasibility for large-scale complex analysis.

[0051] According to some optional embodiments of the present application, the mixed causal discovery method is adopted to identify the causal graph structure corresponding to the financial variable in the query content in the dynamic knowledge graph, which can be implemented by the following method: a core financial variable set related to the query intent of the query content is extracted in the dynamic knowledge graph, and a graph embedding representation corresponding to the core financial variable is obtained; a constraint-based causal discovery algorithm is adopted, the core financial variable set is taken as a node, and the time priority constraint included in the time sequence relationship set in the dynamic knowledge graph is taken as a necessary condition for determining the causal direction, to construct an initial causal skeleton graph, wherein the initial causal skeleton graph is used to represent whether there is a causal correlation between variables; on the basis of the initial causal skeleton graph, a score-based causal discovery algorithm is adopted, the similarity between the graph embedding representations is integrated into the graph structure scoring function, and the key event association edges in the event-entity association matrix that are higher than a preset threshold are taken as directional constraints for reinforced search, to obtain a candidate causal graph structure through optimization; based on the timestamp information in the time sequence relationship set, causal edges in the candidate causal graph structure that do not conform to the preset time logic are removed, and the confidence score of the retained causal edges is calculated using the historical data in the dynamic knowledge graph; a causal graph structure is generated according to the causal edges that pass the time sequence verification and the confidence score of the causal edges that pass the time sequence verification, wherein the causal graph structure includes a time lag attribute and a confidence weight.

[0052] In the present embodiment, a core financial variable set closely related to the query intent is first extracted in the dynamic knowledge graph, the core financial variable can be selected from the graph nodes through semantic matching and correlation analysis, and the graph embedding representation corresponding to the variable is obtained, wherein the embedding vector captures the structural features and semantic information of the variable in the graph. A constraint-based causal discovery algorithm is adopted, the core financial variable set is taken as an analysis node, and the time priority constraint implied in the time sequence relationship set is introduced as a necessary condition for determining the causal direction, i.e. the cause variable must occur before the result variable in time, and an initial causal skeleton graph is constructed through conditional independence test and causal edge orientation, which can explicitly represent whether there is a potential causal correlation between variables.

[0053] Further, a score-based causal discovery algorithm is adopted for structure optimization. The score-based causal discovery algorithm integrates the similarity between the graph embedding representations into the graph structure scoring function, so that variables with similar semantics in the vector space are more likely to establish causal connections, and the key event association edges in the event-entity association matrix that are higher than a preset threshold are taken as directional constraints for reinforced search, guiding the algorithm to preferentially consider the causal direction with strong event support, and obtaining a more accurate candidate causal graph structure through continuous optimization of the overall graph structure score.

[0054] Then based on the detailed timestamp information in the set of temporal relations, the candidate causal graph structure is strictly verified by temporal logic, and edges that do not conform to the cause-effect law are removed, such as invalid connections where the result appears before the cause. At the same time, using the rich historical data in the dynamic knowledge graph, the confidence score of the retained causal edges is calculated through statistical testing and hypothesis testing, ensuring that each causal edge is fully supported by data.

[0055] Finally, all the causal edges that pass the temporal verification and their corresponding confidence scores are integrated to generate a complete causal graph structure that includes temporal lag attributes and confidence weights. The complete causal graph structure not only clearly shows the causal relationship network between financial variables, but also provides a strength measure and time characteristics for each causal relationship.

[0056] The above steps combine constraint-based causal discovery algorithms and score-based causal discovery algorithms to fully utilize statistical rules in the data and incorporate temporal logic and event prior knowledge in the financial field. This hybrid method can effectively overcome false causal relationships that may arise from relying solely on data-driven methods, and can avoid subjective bias based purely on domain knowledge, thereby forming a more reliable and comprehensive causal discovery framework.

[0057] Innovatively, the similarity between graph embedding representations is incorporated into the graph structure scoring function, making variables with similar semantics more likely to establish causal connections. At the same time, using key event association edges in the event-entity association matrix as directional constraints can significantly improve the accuracy of causal direction determination.

[0058] By strictly verifying the time logic to remove false causes that violate the time sequence, and combining historical data to calculate the confidence score, the final causal graph structure not only contains the causal relationship between variables, but also provides the time lag characteristics and reliability measure of each causal relationship. The rich meta-information makes the causal graph not just a simple connection graph, but a complex causal network containing strength, time sequence, credibility, and other multi-dimensional attributes.

[0059] In some optional embodiments of the present application, the determination of the causal effect parameter in the causal graph structure can be achieved by the following method: based on the causal path in the causal graph structure, determining an input feature set for causal inference, wherein the input feature set at least includes: a treatment variable, a result variable, and a set of confounding variables, the treatment variable is used to represent the intervention to be implemented, the result variable is used to represent the target index to be evaluated, and the set of confounding variables is used to control the background factors that have a common influence on the treatment variable and the result variable; using a first machine learning model to predict the treatment variable with the set of confounding variables as input, obtaining a first predicted value of the treatment variable and calculating a first residual between the actual value and the first predicted value of the treatment variable; using a second machine learning model to predict the result variable with the set of confounding variables as input, obtaining a second predicted value of the result variable and calculating a second residual between the actual value and the second predicted value of the result variable, wherein the second machine learning model and the first machine learning model are independent of each other in model structure or model parameter; calculating the correlation between the first residual and the second residual by an orthogonalization technique to obtain an average treatment effect parameter, wherein the average treatment effect parameter is used to represent the global average causal influence of the treatment variable on the result variable; based on a heterogeneity analysis model, taking the set of confounding variables as conditional variables to calculate conditional average treatment effect parameters in different subgroups, wherein the heterogeneity analysis model is a model based on a decision tree or a meta-learner, and the different subgroups are sample sets with similar characteristics divided according to the confounding variable feature space; and the average treatment effect parameter and the conditional average treatment effect parameter are jointly determined as the causal effect parameter.

[0060] In the present embodiment, first, based on the causal path identified in the causal graph structure, an input feature set required for causal inference is determined, wherein the input feature set includes: a treatment variable representing an intervention to be implemented, such as interest rate adjustment or credit policy change; a result variable corresponding to a target index to be evaluated, such as enterprise financing cost or market volatility; and a set of confounding variables covering background factors that have an influence on the treatment variable and the result variable at the same time, such as macroeconomic environment and industry cycle, and by controlling these confounding factors, the accuracy of causal inference can be ensured.

[0061] Then two independent machine learning models are established for parameter estimation. The first machine learning model is used to predict the processing variable with the set of confounding variables as input features, to obtain the first predicted value, and then to calculate the first residual between the actual value and the predicted value of the processing variable, which represents the net change of the processing variable after removing the confounding factors. At the same time, the second machine learning model is used to predict the result variable with the set of confounding variables as input, to obtain the second predicted value, and to calculate the second residual between the actual value and the predicted value of the result variable, which reflects the independent variation of the result variable after removing the confounding factors. It should be noted that the two machine learning models are independent in model structure or model parameters, which can effectively avoid overfitting problem and ensure the robustness of the estimation.

[0062] Secondly, the internal relationship between the first residual and the second residual is analyzed by orthogonalization technique, and the average processing effect parameter is calculated. The average processing effect parameter is used to represent the global average causal effect of the processing variable on the result variable, and its calculation process effectively eliminates the bias caused by confounding variables, which can provide a reliable estimate of the overall effect of the intervention measure. In order to further reveal the heterogeneity characteristics of causal effect, based on the heterogeneity analysis model such as decision tree or meta-learner, the set of confounding variables is taken as the conditional variable, and the conditional average processing effect parameter under different feature subgroups is calculated. Among them, the subgroups are sample sets with similar features divided by clustering analysis according to the feature space of confounding variables, so that the causal effect estimation can reflect the differential influence under different market environment and different enterprise characteristics.

[0063] Finally, the average processing effect parameter and the conditional average processing effect parameter are determined as the causal effect parameter, and the average processing effect parameter and the conditional average processing effect parameter respectively from the overall average and the group heterogeneity two dimensions completely depict the causal influence degree of the processing variable on the result variable. The above double machine learning framework combined with the method of heterogeneity analysis effectively overcomes the sensitivity of model setting in the row and column causal inference method, and can provide more accurate and detailed causal effect evaluation in complex financial scenarios.

[0064] The above steps effectively solve the estimation bias problem caused by confounding variables in traditional causal inference by using double machine learning model. Two independent machine learning models are used to predict the processing variable and the result variable respectively, and then the influence of confounding factors is eliminated by residual calculation, which can obtain the pure causal effect of the processing variable on the result variable without relying on strong model assumptions, significantly improving the accuracy and reliability of the estimation result.

[0065] In addition, not only the global impact assessment of the intervention is provided by the average treatment effect parameter, but also the heterogeneous performance of the causal effect in different subgroups is revealed by the conditional average treatment effect parameter. This double quantification mechanism enables decision-makers to grasp both the overall trend and identify differentiated impacts, providing fine-grained data support for precise policy-making.

[0066] As some optional embodiments of the present application, according to the causal graph structure and the causal effect parameter, the structural causal model is constructed, which can be realized by the following method: based on the nodes and edges in the causal graph structure, a structural equation framework is constructed, wherein each structural equation in the structural equation framework adopts a linear function form, representing the current variable as a linear combination of all parent variables of the current variable in the causal graph structure; based on the causal effect parameter, the initial estimated value of each path coefficient in the structural equation framework is back-calculated through path coefficient calculation based on the structural equation model, to obtain an initialized structural equation; based on the historical data in the dynamic knowledge graph, a maximum likelihood estimation method is used to perform joint parameter estimation on the initialized structural equation, to obtain an optimized structural equation parameter set, wherein the optimized structural equation parameter set includes: point estimate value of the path coefficient and confidence interval of the path coefficient; and the optimized structural equation parameter set is configured to the corresponding edge of the causal graph structure, to obtain the structural causal model.

[0067] In the present embodiment, based on the determined nodes and directed edge relationships in the causal graph structure, a structural equation framework is constructed, and each structural equation in the structural equation framework adopts a linear function form, representing the current variable as a linear combination of all parent variables of the current variable in the causal graph structure, thereby establishing a mathematical basis for the causal dependence relationship between variables. Based on the estimated causal effect parameter, the initial estimated value of each path coefficient in the structural equation framework is back-calculated through the path coefficient calculation method in the structural equation model, to complete the establishment of the initialized structural equation.

[0068] The maximum likelihood estimation method is used to perform joint parameter estimation on the initialized structural equation, based on the rich historical data in the dynamic knowledge graph. The optimization process can adjust each path coefficient by maximizing the likelihood function of the observed data, fully considering the collaborative change pattern between variables and the statistical characteristics of the data, to obtain an optimized structural equation parameter set. The structural equation parameter set not only contains the accurate point estimate value of each path coefficient, but also provides the corresponding confidence interval, thereby completely characterizing the accuracy and reliability degree of the parameter estimation.

[0069] The optimized structural equation parameter set is configured to the corresponding edge of the causal graph structure, to complete the construction of the structural causal model. The above configuration process enables the original causal graph structure to be transformed into a structural causal model with explicit mathematical relationships and quantitative parameters, each causal edge is assigned with a specific path coefficient and confidence interval, forming a complete causal representation that is both interpretable and computable.

[0070] The above steps successfully realize the conversion from qualitative causal relationship to quantitative mathematical expression by converting the causal diagram structure into a computable structural equation framework. Each structural equation corresponds to a causal path in the causal diagram, so that the abstract causal relationship has a specific function form. A phased parameter estimation strategy is adopted: first, the initial value of the path coefficient is back calculated based on the previous causal effect parameter, and then the historical data is used to optimize the joint optimization by maximum likelihood estimation. This "prior guidance + data driven" hybrid method makes full use of the results of previous causal discovery, and verifies and corrects the parameter estimation through historical data, which can significantly improve the accuracy and robustness of parameter estimation, and effectively avoid the local optimal problem caused by improper parameter initialization in traditional methods.

[0071] In some optional embodiments of the present application, the counterfactual prediction result generated by using the structural causal model can be realized by the following method: determining a set of intervention operations based on the query content, wherein each intervention operation in the set of intervention operations is represented by a preset operator, and the preset operator is used to simulate intervention on a preset treatment variable in the causal graph structure; for each intervention operation in the set of intervention operations, calculating the potential result of the intervention operation on the result variable based on the optimized structural equation in the structural causal model to obtain a basic counterfactual prediction; quantifying the uncertainty of the basic counterfactual prediction based on the confidence interval of the path coefficient in the optimized structural equation parameter set to obtain a counterfactual prediction distribution including the confidence interval; and integrating the counterfactual prediction distributions corresponding to all intervention operations in the set of intervention operations to obtain the counterfactual prediction result.

[0072] In the embodiment, firstly, the intervention operation set to be simulated is determined based on the business demand explicitly in the query content, wherein each intervention operation is formalized by a preset do-operator, which can be used to simulate the operation semantics of applying external intervention to a specific treatment variable in the causal graph structure, ensuring the logical rigor of the intervention. For each specific intervention in the intervention operation set, the optimized structural equation in the structural causal model is calculated, the intervention variable is set to a specific value by substitution method, and the influence is propagated along the causal path to accurately calculate the potential result of the intervention operation on the target result variable, thereby obtaining the basic counterfactual prediction value. Further, the uncertainty quantification analysis is performed based on the confidence interval of each path coefficient in the optimized structural equation parameter set. Specifically, the Monte Carlo simulation method can be used to simulate the transmission of parameter uncertainty to the prediction result by repeatedly sampling from the posterior distribution of the parameters, and the counterfactual prediction distribution containing the confidence interval is obtained, wherein the counterfactual prediction distribution not only provides the central tendency of the prediction, but more importantly, depicts the uncertainty range of the prediction. Finally, the counterfactual prediction distributions corresponding to all intervention operations in the intervention operation set are integrated to form a complete counterfactual prediction result, which presents the effect comparison under different intervention scenarios in a standardized data structure, including the expected effect value, confidence interval and distribution characteristics of each intervention.

[0073] As some optional embodiments of the present application, based on the counterfactual prediction result, a decision scheme corresponding to the query content can be generated by the following method: based on the business target of the query content, an optimization target and a constraint condition set are determined, wherein the optimization target at least includes: maximizing the revenue and minimizing the risk, and the constraint condition set is automatically determined based on the business rules and risk limits in the knowledge graph; the expected effect value in the counterfactual prediction result is taken as the revenue estimation basis, and the confidence interval in the counterfactual prediction result is taken as the risk evaluation basis, and a multi-objective optimization function is established; based on the risk transmission path identified in the causal graph structure, a risk constraint is constructed, and the risk constraint is added to the constraint condition set; a multi-objective optimization algorithm is used to solve the multi-objective optimization function to obtain an optimal solution set, wherein each solution in the optimal solution set corresponds to a candidate decision strategy and the expected performance of the candidate decision strategy on each optimization target; based on the causal path length and node influence range of the candidate decision strategy in the causal graph structure, the robustness score of each strategy is calculated; according to the robustness score, the optimal decision strategy is selected from the optimal solution set, and the optimal decision strategy is converted into an executable decision scheme.

[0074] In this embodiment, first, the objective function and constraint conditions of the optimization problem are determined based on the explicit business target in the query content, wherein the optimization target needs to cover at least two core dimensions of maximum revenue and minimum risk, and the constraint condition set is extracted from the dynamic knowledge graph related business rules and risk limit requirements to ensure that the decision scheme meets the business specifications and risk control standards. The expected effect value in the counterfactual prediction result is used as the quantitative basis of the revenue target, and the confidence interval of the prediction result is used to construct a risk assessment index. By integrating these two elements into a mathematical programming model to establish a multi-objective optimization function, a comprehensive optimization framework that takes into account expected revenue and uncertainty can be formed.

[0075] To further improve the risk control mechanism, special risk constraint conditions are constructed based on the risk transmission paths identified in the causal graph structure. By analyzing the propagation path and influence range of risk factors in the causal network, the risk consideration is converted into a mathematical constraint expression, and the risk constraint is added to the original constraint condition set, so that the optimization model can effectively avoid adverse decisions that may trigger a chain of risks. A multi-objective optimization algorithm is used to solve the established optimization function, and the optimal solution set is obtained through Pareto optimal frontier analysis. Each solution in the optimal solution set corresponds to a candidate decision strategy and its expected performance in each optimization target, providing a diverse decision-making space.

[0076] Further, the robustness score of each strategy is calculated based on the characteristics of the candidate decision strategy in the causal graph structure. By analyzing the length of the causal path corresponding to the strategy and the influence range of the nodes, the sensitivity of the strategy execution effect to the model assumptions and parameter estimation is evaluated. Strategies with shorter path lengths and concentrated influence ranges usually have higher robustness. Finally, the best decision strategy with the best overall performance is selected from the optimal solution set based on the calculated robustness score. The best decision strategy is converted into an executable decision scheme that includes specific operation steps, expected effect evaluation, and risk control measures, ensuring that the final output decision meets the optimization target requirements and has good implementation stability.

[0077] The above steps successfully build a complete link from causal reasoning to decision generation by directly using counterfactual prediction results as inputs to the optimization model. Unlike traditional optimization methods that rely on historical statistical laws, this scheme predicts the potential effects of intervention measures based on causal models, enabling the optimization process to accurately estimate the chain reaction of each decision in a complex causal network, thereby generating more forward-looking decision schemes.

[0078] Figure 2 is a structural diagram of a management decision system based on knowledge base construction technology according to an embodiment of the present application, as shown in Figure 2 The system includes:

[0079] The receiving module 21 is configured to receive query content, perform time sequence relationship extraction on multi-source financial data according to the query content, and obtain a time sequence relationship set related to the query content, wherein the time sequence relationship set is a structured data set used to dynamically represent the association relationship between financial entities.

[0080] The first construction module 22 is configured to construct an event-entity association matrix corresponding to the time sequence relationship set, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities.

[0081] The second construction module 23 is configured to construct a dynamic knowledge graph according to the time sequence relationship set and the event-entity association matrix, and identify a causal graph structure corresponding to a financial variable in the query content in the dynamic knowledge graph by using a hybrid causal discovery method, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and the transmission path between the financial variables.

[0082] The determining module 24 is configured to determine a causal effect parameter in the causal graph structure by using a double machine learning model, wherein the causal effect parameter is used to quantitatively evaluate the causal influence degree of an intervention measure on a target variable for the query content.

[0083] The third construction module 25 is configured to construct a structural causal model according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes a structural equation used to define the causal mechanism between the financial variables.

[0084] The generating module 26 is configured to generate counterfactual prediction results by using the structural causal model, and generate a decision scheme corresponding to the query content based on the counterfactual prediction results.

[0085] Optionally, according to the set of time sequence relations and the event-entity association matrix, a dynamic knowledge graph is constructed, including: based on the subject, object, relation and timestamp information in the set of time sequence relations, an initial graph structure carrying time annotation is constructed, wherein the timestamp information is used to identify the starting time or effective time period of the relation; based on the similarity of entity attributes, the nodes in the initial graph structure are aligned to obtain a reference graph; according to the timestamp information in the set of time sequence relations, a time interval attribute is added to each relation edge in the reference graph, wherein the time interval attribute is defined by the starting timestamp and the ending timestamp of the relation; and according to the pre-defined confidence or the statistical features based on the co-occurrence frequency in the set of time sequence relations, the relation strength weight of each relation edge is calculated; the reference graph carrying the relation strength weight and the time interval attribute is determined as the main graph structure, the event feature vector of the entity node in the main graph structure is generated by using the event-entity association matrix, and the event feature vector is used as the initial feature of the node, and the topological connection, relation strength and time interval in the main graph structure are learned simultaneously to obtain the graph embedding representation; the reference graph and the graph embedding representation obtained by the time sequence graph embedding algorithm are stored in the graph database to obtain the dynamic knowledge graph, wherein the graph embedding representation is used as the vectorized attribute of the graph node and is stored in association with the topological structure of the reference graph.

[0086] Optionally, a hybrid causal discovery method is used to identify the causal graph structure corresponding to the financial variable in the query content in the dynamic knowledge graph, including: extracting a set of core financial variables related to the query intent of the query content in the dynamic knowledge graph, and obtaining the graph embedding representation corresponding to the core financial variables; using a constraint-based causal discovery algorithm, taking the set of core financial variables as nodes, and taking the time priority constraint included in the set of time sequence relations in the dynamic knowledge graph as a necessary condition for determining the causal direction, an initial causal skeleton graph is constructed, wherein the initial causal skeleton graph is used to represent whether there is a causal association between variables; based on the initial causal skeleton graph, a score-based causal discovery algorithm is used, the similarity between the graph embedding representations is integrated into the graph structure scoring function, and the key event association edges in the event-entity association matrix that are higher than a preset threshold are used as directional constraints for intensive search, and a candidate causal graph structure is obtained by optimization; based on the timestamp information in the set of time sequence relations, the causal edges in the candidate causal graph structure that do not conform to the preset time logic are removed, and the confidence score of the retained causal edges is calculated using the historical data in the dynamic knowledge graph; according to the causal edges that pass the time sequence verification and the confidence score of the causal edges that pass the time sequence verification, a causal graph structure is generated, wherein the causal graph structure includes: time sequence lag attribute and confidence weight.

[0087] Optionally, the causal effect parameter in the causal graph structure is determined by using a double machine learning model, including: determining an input feature set of causal inference based on a causal path in the causal graph structure, wherein the input feature set at least includes a treatment variable, an outcome variable, and a set of confounding variables, the treatment variable is used to represent an intervention to be implemented, the outcome variable is used to represent a target indicator to be evaluated, and the set of confounding variables is used to control background factors that have a common influence on the treatment variable and the outcome variable; using a first machine learning model to predict the treatment variable with the set of confounding variables as input, obtaining a first predicted value of the treatment variable, and calculating a first residual between an actual value of the treatment variable and the first predicted value; using a second machine learning model to predict the outcome variable with the set of confounding variables as input, obtaining a second predicted value of the outcome variable, and calculating a second residual between an actual value of the outcome variable and the second predicted value, wherein the second machine learning model and the first machine learning model are independent of each other in model structure or model parameter; calculating a correlation between the first residual and the second residual by an orthogonalization technique to obtain an average treatment effect parameter, wherein the average treatment effect parameter is used to represent a global average causal influence of the treatment variable on the outcome variable; based on a heterogeneity analysis model, taking the set of confounding variables as a conditional variable to calculate a conditional average treatment effect parameter in different subgroups, wherein the heterogeneity analysis model is a model based on a decision tree or a meta-learner, and the different subgroups are sample sets with similar characteristics divided according to the confounding variable feature space; and the average treatment effect parameter and the conditional average treatment effect parameter are jointly determined as the causal effect parameter.

[0088] Optionally, a structural causal model is constructed according to the causal graph structure and the causal effect parameter, including: constructing a structural equation framework based on nodes and edges in the causal graph structure, wherein each structural equation in the structural equation framework adopts a linear function form, and a current variable is represented as a linear combination of all parent variables of the current variable in the causal graph structure; based on the causal effect parameter, an initial estimated value of each path coefficient in the structural equation framework is back-calculated by path coefficient calculation based on a structural equation model to obtain an initialized structural equation; based on historical data in the dynamic knowledge graph, a maximum likelihood estimation method is used to perform joint parameter estimation on the initialized structural equation to obtain an optimized structural equation parameter set, wherein the optimized structural equation parameter set includes a point estimate value of the path coefficient and a confidence interval of the path coefficient; and the optimized structural equation parameter set is configured to a corresponding edge of the causal graph structure to obtain the structural causal model.

[0089] Optionally, the counterfactual prediction result is generated by using the structural causal model, including: determining a set of intervention operations based on the query content, wherein each intervention operation in the set of intervention operations is represented by a preset operator, and the preset operator is used to simulate intervention on a preset treatment variable in a causal graph structure; for each intervention operation in the set of intervention operations, calculating a potential result of the intervention operation on a result variable based on an optimized structural equation in the structural causal model to obtain a basic counterfactual prediction; quantifying the uncertainty of the basic counterfactual prediction based on a confidence interval of a path coefficient in the set of optimized structural equation parameters to obtain a counterfactual prediction distribution including the confidence interval; and integrating the counterfactual prediction distributions corresponding to all intervention operations in the set of intervention operations to obtain the counterfactual prediction result.

[0090] Optionally, based on the counterfactual prediction result, a decision scheme corresponding to the query content is generated, including: determining an optimization target and a set of constraint conditions based on a business target of the query content, wherein the optimization target at least includes: maximizing revenue and minimizing risk, and the set of constraint conditions is based on business rules and risk limits in the knowledge graph; taking an expected effect value in the counterfactual prediction result as a revenue estimation basis and taking a confidence interval in the counterfactual prediction result as a risk evaluation basis to establish a multi-objective optimization function; constructing a risk constraint based on a risk transmission path identified in the causal graph structure and adding the risk constraint to the set of constraint conditions; solving the multi-objective optimization function by using a multi-objective optimization algorithm to obtain a set of optimal solutions, wherein each solution in the set of optimal solutions corresponds to a candidate decision strategy and an expected performance of the candidate decision strategy on each optimization target; calculating a robustness score of each strategy based on a causal path length and a node influence range of the candidate decision strategy in the causal graph structure; and selecting an optimal decision strategy from the set of optimal solutions according to the robustness score and converting the optimal decision strategy into an executable decision scheme.

[0091] It should be noted that the above Figure 2 Each module can be a program module (for example, a program instruction set implementing a certain specific function) or a hardware module. For the latter, it can be in the form of, but not limited to: a processor, or the functions of the above-mentioned modules are implemented by a processor.

[0092] It should be noted that Figure 2 The preferred embodiments of the embodiments shown in the drawings can be referred to the related description of the embodiments shown in the drawings, which will not be repeated here. Figure 1 The preferred embodiments of the embodiments shown in the drawings can be referred to the related description of the embodiments shown in the drawings, which will not be repeated here.

[0093] Figure 3 A hardware structure block diagram of a computer terminal for implementing a management decision-making method based on knowledge base construction technology is shown. As Figure 3As shown, the computer terminal 30 may include one or more processors 302 (shown as 302a, 302b, ..., 302n in the figure) 302 (processor 302 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 304 for storing data, and a transmission module 306 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 3 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 30 may also include... Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown.

[0094] It should be noted that the aforementioned one or more processors 302 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 30. As involved in the embodiments of this application, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0095] The memory 304 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the management decision-making method based on knowledge base construction technology in this embodiment of the application. The processor 302 executes various functional applications and data processing by running the software programs and modules stored in the memory 304, thereby realizing the aforementioned management decision-making method based on knowledge base construction technology. The memory 304 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 304 may further include memory remotely located relative to the processor 302, and these remote memories can be connected to the computer terminal 30 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0096] The transmission module 306 is configured to receive or send data via a network. The network can include a wireless network provided by a communication provider of the computer terminal 30. In one example, the transmission module 306 includes a network interface controller (NIC) that can be connected to other network devices through a base station to communicate with the Internet. In one example, the transmission module 306 can be a radio frequency (RF) module configured to communicate with the Internet via wireless communication.

[0097] The display can be a liquid crystal display (LCD) that is touch screen, for example, which can enable a user to interact with the user interface of the computer terminal 30.

[0098] It should be noted that in some alternative embodiments, the above Figure 3 The computer terminal shown can include hardware elements (including circuitry), software elements (including computer code stored on a computer readable medium), or a combination of both hardware and software elements. It should be noted that in some embodiments, the functions of the computer terminal described above can be provided by one or more of the hardware elements and / or software elements. Figure 3 is merely one example of a particular implementation and is intended to illustrate the types of components that can be present in the computer terminal described above.

[0099] It should be noted that Figure 3 The computer terminal shown is configured to perform the Figure 1 The management decision-making method based on the knowledge base construction technology shown is thus applicable to the electronic device, and the relevant explanations in the method of executing the above commands are also applicable to the electronic device, which will not be described herein.

[0100] The embodiments of the present application also provide a non-volatile storage medium, which includes a stored program, wherein the program controls a device in which the storage medium is located to execute the above management decision-making method based on the knowledge base construction technology when the program is running.

[0101] The non-volatile storage medium executes a program for receiving query content, performing time series relationship extraction on multi-source financial data according to the query content to obtain a time series relationship set related to the query content, wherein the time series relationship set is a structured data set for dynamically representing the association relationship between financial entities; constructing an event-entity association matrix corresponding to the time series relationship set, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities; constructing a dynamic knowledge graph according to the time series relationship set and the event-entity association matrix, and identifying a causal graph structure corresponding to a financial variable in the query content in the dynamic knowledge graph by using a hybrid causal discovery method, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and the transmission path between the financial variables; determining a causal effect parameter in the causal graph structure by using a double machine learning model, wherein the causal effect parameter is used to quantitatively evaluate the causal influence degree of an intervention measure on a target variable for the query content; constructing a structural causal model according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes a structural equation for defining the causal mechanism between the financial variables; generating counterfactual prediction results by using the structural causal model, and generating a decision scheme corresponding to the query content based on the counterfactual prediction results.

[0102] The embodiments of the present application also provide an electronic device, comprising a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program performs the above knowledge base construction technology-based management decision method.

[0103] The processor is configured to run a program to perform the following functions: receiving query content, performing time sequence relation extraction on multi-source financial data according to the query content to obtain a time sequence relation set related to the query content, wherein the time sequence relation set is a structured data set for dynamically representing the association relationship between financial entities; constructing an event-entity association matrix corresponding to the time sequence relation set, wherein the event-entity association matrix is used to quantitatively represent the association strength between financial events and financial entities; constructing a dynamic knowledge graph according to the time sequence relation set and the event-entity association matrix, and identifying a causal graph structure corresponding to a financial variable in the query content in the dynamic knowledge graph by using a hybrid causal discovery method, wherein the hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm, and the causal graph structure is a graphical representation of the causal relationship and the transmission path between the financial variables; determining a causal effect parameter in the causal graph structure by using a double machine learning model, wherein the causal effect parameter is used to quantitatively evaluate the causal influence degree of an intervention measure on a target variable for the query content; constructing a structural causal model according to the causal graph structure and the causal effect parameter, wherein the structural causal model includes a structural equation for defining the causal mechanism between the financial variables; generating counterfactual prediction results by using the structural causal model, and generating a decision scheme corresponding to the query content based on the counterfactual prediction results.

[0104] The above sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0105] In the above embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0106] In the above embodiments of the present application, the collected information is information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary protection measures are taken, do not violate public order and good customs, and provide corresponding operation portal for user to choose authorization or refusal.

[0107] In the several embodiments provided by the present application, it should be understood that the disclosed technology can be implemented in other ways. Of course, the unit described as the division is only a logic function division, and there can be other division ways in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0108] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e. may be located in one place, or may be distributed to multiple units. Part or all of the units may be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0109] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0110] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part that contributes to the related art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various program code storage media.

[0111] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A management decision-making method based on knowledge base construction technology, characterized in that, include: Receive query content, and extract time-series relationships from multi-source financial data based on the query content to obtain a time-series relationship set related to the query content. The time-series relationship set is a structured data set used to dynamically represent the relationship between financial entities. Construct an event-entity association matrix corresponding to the time-series relationship set, wherein the event-entity association matrix is ​​used to quantitatively represent the association strength between financial events and financial entities; Based on the time-series relationship set and the event-entity association matrix, a dynamic knowledge graph is constructed, and a hybrid causal discovery method is used to identify the causal graph structure corresponding to the financial variables in the query content in the dynamic knowledge graph. The hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm. The causal graph structure is a graphical representation of the causal relationships and transmission paths between the financial variables. A dual machine learning model is used to determine the causal effect parameters in the causal graph structure, wherein the causal effect parameters are used to quantitatively evaluate the degree of causal influence of intervention measures on the query content on the target variable; Based on the causal graph structure and the causal effect parameters, a structural causal model is constructed, wherein the structural causal model includes structural equations for defining the causal mechanism between the financial variables; The structural causal model is used to generate counterfactual prediction results, and based on the counterfactual prediction results, a decision scheme corresponding to the query content is generated. A dual machine learning model is used to determine the causal effect parameters in the causal graph structure, including: determining the input feature set for causal inference based on the causal paths in the causal graph structure, wherein the input feature set includes at least: a treatment variable, an outcome variable, and a set of confounding variables, wherein the treatment variable represents the intervention to be implemented, the outcome variable represents the target indicator to be evaluated, and the set of confounding variables controls for background factors that jointly affect the treatment variable and the outcome variable; using a first machine learning model, with the set of confounding variables as input, predicting the treatment variable to obtain a first predicted value of the treatment variable and calculating a first residual between the actual value of the treatment variable and the first predicted value; using a second machine learning model, with the set of confounding variables as input, predicting the outcome variable to obtain a second residual of the outcome variable. The system predicts a value and calculates a second residual between the actual value of the outcome variable and the second predicted value, wherein the second machine learning model and the first machine learning model are independent of each other in terms of model structure or model parameters; the system calculates the correlation between the first residual and the second residual using orthogonalization techniques to obtain the average treatment effect parameter, wherein the average treatment effect parameter is used to characterize the global average causal influence of the treatment variable on the outcome variable; based on a heterogeneity analysis model, using the set of confounding variables as conditional variables, the system calculates the conditional average treatment effect parameter for different subgroups, wherein the heterogeneity analysis model is a model based on a decision tree or meta-learner, and the different subgroups are sample sets with similar characteristics divided according to the feature space of the confounding variables; the average treatment effect parameter and the conditional average treatment effect parameter are jointly determined as the causal effect parameter.

2. The method according to claim 1, characterized in that, Based on the aforementioned temporal relationship set and event-entity association matrix, a dynamic knowledge graph is constructed, including: Based on the subject, object, relationship, and timestamp information in the time-series relationship set, an initial graph structure carrying time stamps is constructed, wherein the timestamp information is used to identify the start time or effective time period of the relationship; Based on the similarity of entity attributes, the nodes in the initial graph structure are aligned to obtain a baseline graph. Based on the timestamp information in the time-series relation set, a time interval attribute is added to each relation edge in the baseline graph, wherein the time interval attribute is defined by the start and end timestamps of the relation; and a relation strength weight is calculated for each relation edge based on the predefined confidence level or statistical features based on co-occurrence frequency in the time-series relation set. The baseline graph carrying the relation strength weights and the time interval attributes is determined as the main graph structure. The event-entity association matrix is ​​used to generate event feature vectors for entity nodes in the main graph structure. The time-series graph embedding algorithm is used to take the event feature vectors as the initial features of the nodes. At the same time, the topological connections, relation strengths and time intervals in the main graph structure are learned to obtain the graph embedding representation. The baseline graph and the graph embedding representation obtained by the temporal graph embedding algorithm are jointly stored in the graph database to obtain the dynamic knowledge graph. The graph embedding representation is stored as a vectorized attribute of the graph node and is associated with the topological structure of the baseline graph.

3. The method according to claim 1, characterized in that, A hybrid causal discovery method is used to identify the causal graph structure corresponding to the financial variables in the query content within the dynamic knowledge graph, including: Extract a set of core financial variables related to the query intent of the query content from the dynamic knowledge graph, and obtain the graph embedding representation corresponding to the core financial variables; A constraint-based causal discovery algorithm is adopted, with the core financial variable set as nodes and the time priority constraints included in the time sequence relationship set in the dynamic knowledge graph as necessary conditions for determining the causal direction, to construct an initial causal skeleton graph, wherein the initial causal skeleton graph is used to characterize whether there is a causal relationship between variables; Based on the initial causal skeleton graph, a score-based causal discovery algorithm is adopted. By incorporating the similarity between the graph embedding representations into the graph structure scoring function, and using the key event association edges in the event-entity association matrix that are higher than a preset threshold as directional constraints to strengthen the search, the candidate causal graph structure is optimized. Based on the timestamp information in the time sequence relation set, causal edges that do not conform to the preset time logic in the candidate causal graph structure are removed, and the confidence score of the retained causal edges is calculated using historical data in the dynamic knowledge graph. The causal graph structure is generated based on the time-tested causal edges and their confidence scores, wherein the causal graph structure includes: time-lag attributes and confidence weights.

4. The method according to claim 1, characterized in that, Based on the causal graph structure and the causal effect parameters, a structural causal model is constructed, including: Based on the nodes and edges in the causal graph structure, a structural equation framework is constructed, wherein each structural equation in the structural equation framework adopts a linear function form, and the current variable is represented as a linear combination of all parent variables of the current variable in the causal graph structure. Based on the causal effect parameters, the initial estimates of each path coefficient in the structural equation model are inferred by calculating the path coefficients based on the structural equation model, and the initial structural equation is obtained. Based on historical data in the dynamic knowledge graph, the maximum likelihood estimation method is used to perform joint parameter estimation on the initial structural equation to obtain an optimized structural equation parameter set, wherein the optimized structural equation parameter set includes: point estimates of path coefficients and confidence intervals of the path coefficients; The optimized set of structural equation parameters is configured onto the corresponding edges of the causal graph structure to obtain the structural causal model.

5. The method according to claim 4, characterized in that, Generating counterfactual prediction results using the aforementioned structural causal model includes: An intervention operation set is determined based on the query content, wherein each intervention operation in the intervention operation set is represented by a preset operator, and the preset operator is used to simulate the intervention on a preset processing variable in the causal graph structure; For each intervention in the set of interventions, based on the optimized structural equation in the structural causal model, the potential consequences of the intervention on the outcome variable are calculated to obtain the basic counterfactual prediction; Based on the confidence intervals of the path coefficients in the optimized structural equation parameter set, the uncertainty of the basic counterfactual prediction is quantified to obtain the counterfactual prediction distribution including the confidence intervals. The counterfactual prediction distributions corresponding to all intervention operations in the intervention operation set are integrated to obtain the counterfactual prediction results.

6. The method according to claim 1, characterized in that, Based on the counterfactual prediction results, a decision scheme corresponding to the query content is generated, including: Based on the business objectives of the query content, an optimization objective and a set of constraints are determined, wherein the optimization objective includes at least: maximizing revenue and minimizing risk, and the set of constraints comes from the business rules and risk limits in the dynamic knowledge graph; Based on the expected effect value in the counterfactual prediction results, and based on the confidence interval in the counterfactual prediction results, a multi-objective optimization function is established. Based on the risk transmission path identified in the causal graph structure, risk constraints are constructed and added to the constraint set. The multi-objective optimization function is solved using a multi-objective optimization algorithm to obtain an optimal solution set, wherein each solution in the optimal solution set corresponds to a candidate decision strategy and the expected performance of the candidate decision strategy on each optimization objective; Based on the causal path length and node influence range of the candidate decision-making strategies in the causal graph structure, a robustness score is calculated for each strategy. Based on the robustness score, an optimal decision strategy is selected from the optimal solution set, and the optimal decision strategy is transformed into an executable decision scheme.

7. A management decision-making system based on knowledge base construction technology, characterized in that, include: The receiving module is used to receive query content and extract time-series relationships from multi-source financial data based on the query content to obtain a time-series relationship set related to the query content. The time-series relationship set is a structured data set used to dynamically represent the relationship between financial entities. The first construction module is used to construct the event-entity association matrix corresponding to the time-series relationship set, wherein the event-entity association matrix is ​​used to quantitatively represent the association strength between financial events and financial entities; The second construction module is used to construct a dynamic knowledge graph based on the time-series relationship set and the event-entity association matrix, and to identify the causal graph structure corresponding to the financial variables in the query content in the dynamic knowledge graph using a hybrid causal discovery method. The hybrid causal discovery method includes a constraint-based causal discovery algorithm and a score-based causal discovery algorithm. The causal graph structure is a graphical representation of the causal relationships and transmission paths between the financial variables. The determination module is used to determine the causal effect parameters in the causal graph structure using a dual machine learning model, wherein the causal effect parameters are used to quantitatively evaluate the degree of causal influence of intervention measures targeting the query content on the target variable; The third construction module is used to construct a structural causal model based on the causal graph structure and the causal effect parameters, wherein the structural causal model includes structural equations for defining the causal mechanism between the financial variables; The generation module is used to generate counterfactual prediction results using the structural causal model, and to generate a decision scheme corresponding to the query content based on the counterfactual prediction results. The determining module is further configured to perform the following steps: based on the causal paths in the causal graph structure, determine the input feature set for causal inference, wherein the input feature set includes at least: a processing variable, an outcome variable, and a set of confounding variables, wherein the processing variable represents the intervention to be implemented, the outcome variable represents the target indicator to be evaluated, and the set of confounding variables controls for background factors that jointly influence the processing variable and the outcome variable; using a first machine learning model, with the set of confounding variables as input, predict the processing variable to obtain a first predicted value of the processing variable and calculate a first residual between the actual value of the processing variable and the first predicted value; using a second machine learning model, with the set of confounding variables as input, predict the outcome variable to obtain a second predicted value of the outcome variable and calculate a second residual between the actual value of the processing variable and the first predicted value. The second residual between the actual value of the outcome variable and the second predicted value is calculated, wherein the second machine learning model and the first machine learning model are independent of each other in terms of model structure or model parameters; the correlation between the first residual and the second residual is calculated using orthogonalization techniques to obtain the average treatment effect parameter, wherein the average treatment effect parameter is used to characterize the global average causal influence of the treatment variable on the outcome variable; based on a heterogeneity analysis model, using the set of confounding variables as conditional variables, the conditional average treatment effect parameter under different subgroups is calculated, wherein the heterogeneity analysis model is a model based on a decision tree or meta-learner, and the different subgroups are sample sets with similar characteristics divided according to the feature space of the confounding variables; the average treatment effect parameter and the conditional average treatment effect parameter are jointly determined as the causal effect parameter.

8. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device where the non-volatile storage medium is located to execute the management decision-making method based on knowledge base construction technology as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the management decision-making method based on knowledge base construction technology as described in any one of claims 1 to 6.

Images