Multi-scale cross-domain index transaction attribution method and system based on knowledge enhancement

Abstract

The invention discloses a multi-scale cross-domain index transaction attribution method and system based on knowledge enhancement, and relates to the technical field of index transaction attribution, and the method comprises the steps: carrying out the two-stage training of a large language model through employing a breeding industry production field knowledge base and a data set with a causal relationship, constructing an index causal knowledge graph by using a large language model, carrying out graph fusion on the index causal knowledge graph and the expert knowledge graph, constructing a fusion knowledge enhanced Bayesian network and a regression model of a depth network based on prior constraints, selecting multi-scale data from the fusion knowledge graph as a training set, identifying and extracting a causal relationship pair, and carrying out model training to obtain an expert knowledge graph; and obtaining target index transaction from the financial indexes and the production indexes, importing the target index transaction into the trained regression model, and obtaining attribution data corresponding to the target index transaction. According to the method, the problem that abnormal index positioning and analysis basically depend on human experience due to various data and complex dimensions is solved, and enterprise efficient operation and data-driven decision making of enterprises are facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of indicator change attribution, and more specifically, to a multi-scale cross-domain indicator change attribution method and system based on knowledge enhancement. Background Art

[0002] In agricultural production, there has always been a data gap between finance and production. How to build an effective knowledge bridge to achieve cross-domain connection and integration is a major problem currently faced by various agricultural production companies. Among them, in view of the intricate relationship between financial indicators and production indicators in the breeding industry, how to comprehensively and clearly capture the relationship between various indicators is crucial for business problem location, decision optimization and business adjustment.

[0003] However, in the traditional attribution method of indicator changes, either it completely relies on experts for manual analysis, which takes a lot of time and has high professional requirements for analysts; or it completely relies on a clear calculation formula given in a simple operating environment. Once it is expanded to a complex operating environment that cannot provide a clear relationship, it will be immediately helpless. For example, in the scenario facing the present invention, the changes between financial indicators and production indicators, if forcibly solved by traditional methods, there is a key problem that the relevant influencing indicators cannot be judged in combination with domain business knowledge, resulting in the inability to locate risk indicators. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a multi-scale cross-domain indicator change attribution method and system based on knowledge enhancement. Through the high integration of knowledge graph and indicator attribution tasks, it helps enterprises to monitor abnormal fluctuations of various business indicators in real time from finance to production, quickly locate the root cause of the problem, and provide in-depth analysis and intelligent diagnosis to support decision optimization and business adjustment.

[0005] The first aspect of the present invention provides a multi-scale cross-domain indicator change attribution method based on knowledge enhancement, comprising the following steps:

[0006] Use the knowledge base of the aquaculture production field and the data set with causal relationships to conduct two-stage training of the large language model, build an indicator causal knowledge graph through the trained large language model, and perform relationship verification on the indicator causal knowledge graph;

[0007] Obtain an expert knowledge graph, and fuse the indicator causal knowledge graph with the expert knowledge graph to generate a fused knowledge graph;

[0008] Construct a regression model, select multi-scale data from the fused knowledge graph as a training set, identify and extract causal relationship pairs, perform univariate Bayesian regression on the causal indicators in the causal relationship pairs, and read the coefficient constraint range;

[0009] Based on the coefficient constraint range, a linear analysis of causal indicators is constructed, dimension additional weights are introduced, and deep learning training is performed using a neural network to obtain a regression model that integrates a knowledge-enhanced Bayesian network and a deep network based on prior constraints;

[0010] Obtain target indicator changes from financial indicators and production indicators and import them into the regression model to obtain the impact coefficient of target indicator changes on corresponding causal indicators, and integrate them to generate final attribution data.

[0011] In this solution, the knowledge base of the aquaculture production field and the dataset with causal relationships are used to train the large language model in two stages, specifically:

[0012] Initialize the large language model, obtain a first-stage training sample based on the professional knowledge base and corpus in the field of aquaculture production, use the first-stage training sample to train the large language model, process the first-stage training sample into a number of question-answer pairs through the prompt template of the large language model, and enable the large language model to understand the professional terms, complex semantic environment and internal logical relationships in the field of aquaculture through iterative training;

[0013] Based on the data set with causal relationship annotations, two-stage training samples are obtained, and the large language model is trained again using the two-stage training samples. Multi-task training is adopted in the two-stage training process, and two task heads are set to be responsible for domain knowledge fine-tuning and causal relationship extraction respectively. Weight parameters are set to control the proportion of the two tasks. Through iterative training, the large language model has the ability of recognition and attribution analysis, and can capture the potential causal relationship between indicators.

[0014] In this solution, the indicator causal knowledge graph is constructed through the trained large language model, and the relationship of the indicator causal knowledge graph is verified, specifically:

[0015] Obtain a large language model that has mastered the knowledge in the field of aquaculture production after training, use the large language model to extract entities and causal relationships from documents related to the pre-selected projects, and use the entities and causal relationships as nodes and edge structures of the knowledge graph to construct an indicator causal knowledge graph;

[0016] The actual data of the pre-selected items are obtained as the basis for verification, and the causal relationship existing in the indicator causal knowledge graph is verified using the Granger causality test. When the relationship verification meets the preset standards, the expression of the corresponding causal relationship in the indicator causal knowledge graph is strengthened; when the relationship verification does not meet the preset standards, the corresponding causal relationship is removed from the indicator causal knowledge graph.

[0017] In this solution, historical indicator change attribution instances are obtained, expert-defined indicator relationship data involved in the historical indicator change attribution instances are extracted, the extracted indicator relationship data are preprocessed to generate expert knowledge, and an expert knowledge graph is obtained based on the expert knowledge;

[0018] The trained large language model is used to calculate the semantic similarity of entities. The indicator causal knowledge graph is aligned with the entities in the expert knowledge graph according to the semantic similarity. The relationships in the two knowledge graphs are aligned based on the expert knowledge graph, and a fused knowledge graph is generated through graph fusion.

[0019] In this solution, a regression model is constructed, multi-scale data is selected from the fusion knowledge graph as a training set, and causal relationship pairs are identified and extracted. Specifically:

[0020] The regression model is constructed by integrating the knowledge-enhanced Bayesian network and the deep network two-stage regression method based on prior constraints. Multi-scale data is constructed according to different time dimensions and spatial dimensions. The cause and effect indicators involved in the multi-scale data are extracted based on the fusion knowledge graph recognition.

[0021] Data partitioning is introduced into the regression model, and the result indicator is segmented according to the fluctuation value. Different segments are modeled separately to establish a data-specific regression model. Data from adjacent time periods are selected as training sets to evaluate the impact of cause indicators on result indicators.

[0022] In this solution, the causal indicators in the causal relationship pair are subjected to univariate Bayesian regression respectively, and the coefficient constraint range is read, specifically:

[0023] For the automatically extracted causal relationship pairs, the cause indicator and the effect indicator are subjected to separate Bayesian regression, expressed as: y = α n ·x n , where y is the fluctuation of the fruit index, x n Due to the fluctuation of the indicator, α n is the influence coefficient of the nth factor indicator;

[0024] Obtaining the influence coefficient α through Bayesian regression n The mean and variance of the regression model are used to obtain the coefficient constraint range of the subsequent regression network [μ-3*σ2 ,μ+3*σ 2 ].

[0025] In this solution, a linear analysis of causal indicators is constructed based on the coefficient constraint range, dimension additional weights are introduced, and a neural network is used for deep learning training, specifically:

[0026] After obtaining the coefficient constraint range, the cause indicator and the effect indicator are combined to construct a linear regression, which is expressed as: Among them, y is the fluctuation of the fruit index, x n Due to the fluctuation of the indicator, α n is the influence coefficient of the nth factor, and the influence coefficient is controlled within the coefficient constraint range each time the back propagation gradient descends;

[0027] In addition, the multiple dimensions of the training samples are processed comprehensively, and the dimensions are encoded and represented using one-hot encoding to obtain the embedding vectors of the dimensions. The embedding vectors of multiple dimensions are imported into the fully connected layer output to generate dimension additional weights to characterize the influence differences of different dimensions.

[0028] The dimension additional weight is combined with the linear regression to fit into the neural network, which is expressed as: α n ij Add weights to the i-th dimension and the j-th dimension. After iterative deep training, according to α n ij α n Characterizes the complete influence coefficient of the nth factor indicator.

[0029] The second aspect of the present invention provides a multi-scale cross-domain indicator change attribution system based on knowledge enhancement, the system comprising a knowledge graph construction unit, an indicator change attribution unit and an attribution result output unit;

[0030] The knowledge graph construction unit uses a large language model to construct an indicator causal knowledge graph, fuses the indicator causal knowledge graph with an expert knowledge graph based on business expert knowledge to generate a fused knowledge graph, and extracts causal association information between any indicator and other indicators based on the fused knowledge graph;

[0031] The indicator change attribution unit integrates the knowledge-enhanced Bayesian network and the deep network two-stage regression method based on prior constraints to construct a regression model, selects corresponding multi-scale data as a training set, identifies and extracts causal relationship pairs, performs univariate Bayesian regression on the causal indicators in the causal relationship pairs, and reads the coefficient constraint range; constructs a linear analysis of the causal indicators based on the coefficient constraint range, introduces dimensional additional weights, uses neural networks for deep learning training, and obtains the weights of each causal indicator; inputs the target indicator change to obtain the impact coefficient of the target indicator change on the corresponding causal indicator, and integrates to generate the final attribution data.

[0032] The attribution result output unit visualizes the attribution data of the target indicator movement to obtain the mutual influence of the financial indicator and the production indicator changes.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention proposes and implements knowledge modeling of causal relationships of aquaculture indicators based on data-driven, as well as an automated construction method of indicator causal relationship knowledge graph based on domain knowledge. Combined with the basic knowledge graph provided by business experts, a multi-scale cross-domain indicator change attribution algorithm for aquaculture enterprises is proposed.

[0035] In the layer-by-layer penetration analysis of abnormal indicators, the present invention integrates knowledge graphs and indicator attribution tasks. The system helps enterprises monitor abnormal fluctuations of various business indicators in real time, quickly locate the root causes of problems, and provide in-depth analysis and intelligent diagnosis to support decision optimization and business adjustments, from macro trends to micro details (multi-scale), from finance to production (cross-domain). BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0037] Figure 1 A flowchart of a multi-scale cross-domain indicator change attribution method based on knowledge enhancement is shown;

[0038] Figure 2 A schematic diagram of the process of constructing a fusion knowledge graph in an embodiment is shown;

[0039] Figure 3 A flowchart of multi-scale indicator attribution modeling in an embodiment is shown;

[0040] Figure 4The block diagram of the multi-scale cross-domain indicator change attribution system based on knowledge enhancement is shown. DETAILED DESCRIPTION

[0041] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0042] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0043] like Figure 1 As shown, in the first embodiment of the present invention, a multi-scale cross-domain indicator change attribution method based on knowledge enhancement is provided, including:

[0044] S102, using the knowledge base of the aquaculture production field and the data set with causal relationships to perform two-stage training on the large language model, constructing an indicator causal knowledge graph through the trained large language model, and performing relationship verification on the indicator causal knowledge graph;

[0045] S104, obtaining an expert knowledge graph, fusing the indicator causal knowledge graph with the expert knowledge graph to generate a fused knowledge graph;

[0046] S106, constructing a regression model, selecting multi-scale data from the fused knowledge graph as a training set, identifying and extracting causal relationship pairs, performing univariate Bayesian regression on the causal indicators in the causal relationship pairs, and reading the coefficient constraint range;

[0047] S108, constructing a linear analysis of causal indicators based on the coefficient constraint range, introducing dimension additional weights, using a neural network for deep learning training, and obtaining a regression model of a Bayesian network enhanced by fusion knowledge and a deep network based on prior constraints;

[0048] S110, obtaining target indicator changes from financial indicators and production indicators and importing them into the regression model, obtaining the influence coefficient of the target indicator changes on the corresponding cause indicators, and integrating them to generate the final attribution data.

[0049] By using advanced large preview model technology, domain knowledge and causal relationship extraction capabilities are mastered through two-stage learning, and the relationship between indicators that are not covered by expert knowledge can be extracted from the text materials commonly used by enterprises. In the first stage, the large language model systematically learns the professional knowledge of a specific industry from the domain-based unstructured corpus to ensure that the model can understand and identify the core concepts and terms related to the enterprise field. In the second stage, through causal relationship extraction, the model further has the ability of identification and attribution analysis, enabling it to efficiently and accurately capture the potential causal relationship between indicators in practical applications. The large language model can transcend the limitations of expert knowledge, automatically explore and reveal complex and hidden indicator associations in the field, and form a more comprehensive knowledge system. This process not only reduces the cost of manual screening and annotation, but also can dynamically adapt to data changes and continuously optimize the structure and content of the knowledge graph. Compared with traditional methods, this method has higher flexibility and scalability, providing enterprise managers with more accurate and in-depth decision support, thereby improving the overall intelligence level of data analysis and management.

[0050] Figure 2 A schematic diagram of the process of constructing a fused knowledge graph according to an embodiment is shown.

[0051] Initialize the large language model, obtain the first-stage training samples based on the professional knowledge base and corpus in the field of aquaculture production, including monographs in the field, scientific research papers, professional forum content, etc., use the first-stage training samples to train the large language model, process the first-stage training samples into several question-answer pairs through the prompt template of the large language model, and make the large language model understand the professional terminology, complex semantic environment and internal logical relationship in the field of aquaculture through iterative training; the first-stage training not only includes the adjustment of the parameters of the large language model and the improvement of the algorithm, but more importantly, the in-depth learning and mastery of domain knowledge. The first-stage training samples need to cover a large number of text resources related to aquaculture production, such as the production process of pigs, the whole process information from production to sales, and private documents within the enterprise. Input the first-stage training samples as the training set into the large language model, and through multiple rounds of iterative training, help the large language model gradually master the professional terminology, conceptual framework and its internal logical relationship in this field, and build a deep understanding of aquaculture production to support more accurate causal reasoning and application needs.

[0052] In order to extract the quantitative relationship logic that is not determined between indicators, two-stage training samples are obtained based on data sets with causal relationship annotations, such as the Causal-TimeBank data set, and the large language model is trained again using the two-stage training samples. In order to ensure the perfect combination of domain knowledge fine-tuning and causal relationship extraction, multi-task training is adopted in the two-stage training process. Two task heads are set to be responsible for domain knowledge fine-tuning and causal relationship extraction respectively. Weight parameters are set to control the proportion of the two tasks to form a multi-task loss function. In each batch, data from the two tasks are randomly sampled to ensure that the learning data of the two tasks is balanced. Through iterative training, the large language model has the ability to identify and attribute analysis, and can capture the potential causal relationship between indicators.

[0053] Knowledge graph is a method for structured representation of knowledge, which can display different entities, concepts and their relationships in the form of a graph. When there is no clear quantitative relationship logic between indicators, the large language model (LLM) can be used to train the knowledge graph to build a model in order to accurately mine the potential relationship between indicators.

[0054] After training, a large language model with knowledge of the aquaculture production field is obtained. The large language model is used to extract entities and causal relationships from documents related to the pre-selected projects. The documents related to the pre-selected projects include document data in the financial and production fields within the enterprise. The entities and causal relationships are used as nodes and edge structures of the knowledge graph, and an indicator causal knowledge graph is constructed based on the basic framework of the knowledge graph. In the construction of the indicator causal knowledge graph, the domain knowledge and causal relationship extraction capabilities of the large language model play a key role, verifying the necessity of the first two stages of training. The model can identify professional terms and contextual logical relationships in the data, thereby helping to build a more accurate and comprehensive knowledge graph.

[0055] The actual data of the pre-selected items are obtained as the basis for verification. The actual data can reflect the complex operation process and key details in detail. In order to ensure the accuracy and dynamic adaptability of the knowledge graph, the Granger causality test is used to verify the causal relationship existing in the indicator causal knowledge graph. When the relationship verification meets the preset standards, the expression of the corresponding causal relationship in the indicator causal knowledge graph is strengthened. When the relationship verification does not meet the preset standards, the corresponding causal relationship is removed from the indicator causal knowledge graph.

[0056] The core idea of ​​using the Granger causality test is that if a change in a variable X can help predict the future change of another variable Y, then X is the Granger cause of Y. Assume that there is a causal relationship between X and Y in the knowledge graph, that is, X is the cause of Y. Set the financial time series X t and the production time series Y t, check X t Does Granger cause Y t The following two models are used for regression:

[0057] No-hysteresis model: Y t =α+β1Y t-1 +∈ t

[0058] Model with lag: Y t =α+β1Y t-1 +γ1X t-1 +∈ t

[0059] Among them, α, β1, γ1 are linear parameters, ∈ t The key to the Granger test is to compare the goodness of fit of the two models. Granger causality test is usually two-way. t Does Granger cause Y t , we also need to test Y t Does Granger cause X t Through the two-way test, we can more fully understand the financial time series X t and the production time series Y t If the model is significantly improved after the introduction of the lag value (i.e., γ1 is significantly not zero), then it can be considered that X t Granger leads to Y t , retain and strengthen the expression of the relationship in the graph. However, if γ1 is 0, consider removing it from the knowledge graph to avoid invalid or weak related information interfering with the structure of the graph. By calculating the Granger causal relationship between nodes, the reliability and accuracy of the knowledge graph are further improved. Through data-driven verification and optimization, it is ensured that the relationships and nodes in the graph not only meet the expectations of domain experts, but also truly reflect the actual situation, thereby providing solid and reliable support for more in-depth analysis and application.

[0060] Obtain historical indicator change attribution instances, extract expert-defined indicator relationship data involved in the historical indicator change attribution instances, preprocess the extracted indicator relationship data, generate expert knowledge, and obtain an expert knowledge graph based on the expert knowledge; use the trained large language model to calculate entity semantic similarity, align the indicator causal knowledge graph with the entities in the expert knowledge graph according to the semantic similarity, and align the relationships in the two knowledge graphs based on the expert knowledge graph. After alignment, if there are entities or relationships in the indicator causal knowledge graph that are not covered by expert knowledge, add this knowledge to the expert knowledge graph to enhance its completeness, and generate a fused knowledge graph through graph fusion.

[0061] By introducing the expert knowledge graph, the system can integrate the insights and analysis results accumulated by experts over a long period of time, and combine them with the self-trained graph of the large model to obtain a more comprehensive and rich knowledge graph. The fused knowledge graph not only has the potential causal relationship information of the indicators learned by the large language model, but also retains the expert's deep understanding of the relationship between key indicators, and can better reflect the business logic and actual operation. This fusion method has obvious advantages in formulating a stable indicator relationship attribution model, which can provide a more reliable causal analysis basis, help improve the accuracy and robustness of the attribution model, and meet actual business needs. The introduction of expert knowledge not only improves the scientificity and accuracy of the graph construction, but also ensures the rationality of the relationships and nodes in the graph, thus providing a solid foundation for subsequent data analysis, logical reasoning and decision support.

[0062] Figure 3 A flowchart of multi-scale value indicator attribution modeling in an embodiment is shown.

[0063] The regression model is constructed by integrating the knowledge-enhanced Bayesian network and the deep network two-stage regression method based on prior constraints. Multi-scale data is constructed according to different time dimensions and spatial dimensions. It supports data of different time scales such as monthly, quarterly, semi-annual, and annual in terms of time, and supports the selection of various organizational levels in terms of space. For example, when performing the monthly attribution task of the five-level unit, the monthly data of the five-level unit should be selected as the training set to ensure that the data for model training is consistent with the target scale of the task, thereby improving the accuracy and effectiveness of the attribution analysis. Based on the fusion of knowledge graphs, the cause and effect indicators involved in the extraction of multi-scale data are identified, and the impact of each cause indicator on the effect indicator is comprehensively evaluated.

[0064] In order to ensure the stability and adaptability of the model, data partitioning is introduced into the regression model to solve the problem of high volatility and large data changes in training data. The data is segmented according to the fluctuation value of the result indicator, and different segments are modeled separately to establish a data-specific regression model. In addition, data from adjacent time periods are selected as training sets to evaluate the impact of cause indicators on result indicators. The mechanism of dynamic update of data partitioning can better reflect the timeliness of data and ensure that the model remains effective and accurate in a constantly changing real environment.

[0065] For the automatically extracted causal relationship pairs, the cause indicator and the effect indicator are subjected to separate Bayesian regression, expressed as: y = α n ·x n , where y is the fluctuation of the fruit index, x n Due to the fluctuation of the indicator, α n is the influence coefficient of the nth factor indicator; the influence coefficient α is obtained through Bayesian regression nThe mean and variance of the regression model are used to obtain the coefficient constraint range of the subsequent regression network [μ-3*σ 2 ,μ+3*σ 2 ], which can solve the problem of data sparsity and limit the influence coefficient to prevent overfitting. For the problem of more sparse data caused by partitioning, Bayesian is used to solve the prior distribution modeling of data sparsity, and then the network is constrained based on the prior distribution to achieve accurate modeling under the condition of a small number of samples.

[0066] After obtaining the coefficient constraint range, the cause indicator and the effect indicator are combined to construct a linear regression, which is expressed as: Among them, y is the fluctuation of the fruit index, x n Due to the fluctuation of the indicator, α n is the influence coefficient of the nth factor indicator. During each back-propagation gradient descent, the influence coefficient is controlled within the coefficient constraint range. During the regression model training process, the training samples are used to learn the influence weights of each factor indicator to quantify its effect size.

[0067] In the modeling process, although the causal relationship between indicators has been determined through the knowledge graph, it is still necessary to overcome the challenges of too many dimensions of enterprise data and fine data granularity. Multiple dimensions bring the problem of insufficient training samples, which affects the adequacy of model training, and may also lead to redundant model parameters and increased complexity. The multiple dimensions of the training samples are processed comprehensively, and the dimensions are encoded and represented using one-hot encoding to obtain the embedding vectors of the dimensions. The embedding vectors of multiple dimensions are imported into the fully connected layer output to generate dimension additional weights to characterize the difference in the influence of different dimensions; so that the features of different dimensions can affect the effect of "cause" indicators on "effect" indicators in the form of adaptive weights, thereby effectively simplifying the dimension problem. For scenarios with many dimensions, the problem of small number of samples and difficulty in training due to fine granularity can be solved by vectorizing these dimensions, which not only effectively solves the problem of dimensional diversity, but also maintains simplicity.

[0068] The dimension additional weight is combined with the linear regression to fit into the neural network, which is expressed as: α n ij Add weights to the i-th dimension and the j-th dimension. After iterative deep training, according to α n ij α n Characterizes the complete influence coefficient of the nth factor indicator. The model compresses information from multiple dimensions into a dynamic and adaptive weight structure, effectively reducing redundant parameters, improving the training efficiency and generalization ability of the model, and ensuring that indicator relationship modeling can still be performed accurately and stably in multi-dimensional and complex enterprise data.

[0069] Obtain the target indicator change in the multi-scale data corresponding to the financial indicator and the production indicator, import the regression model, obtain the influence coefficient of the target indicator change on the corresponding cause indicator, and integrate to generate the final attribution data. Further, the regression model construction proposed in the embodiment is mainly applicable to single-jump causal relationships, that is, the result indicator is directly affected by the cause indicator, for example, A directly causes B. In this case, the impact of A's change on B's change can be obtained through the regression model. For the multi-jump causal chain problem, a preferred embodiment of the present invention uses a chain model calculation to gradually derive the influence coefficient in the multi-level causal relationship. Specifically, the influence coefficient of B's ​​change on C's change can be obtained through the regression model between B and C, and the influence coefficient of A's change on B can be obtained through the regression model between A and B. Since the change of B is the same in the two regression models, the change of B can be aligned with the influence coefficient of B on C, so as to ensure the continuity of the causal chain. Next, by adjusting the influence coefficient of A on B in proportion, it is aligned with the influence coefficient of B on C to ensure that the proportional relationship is maintained during the transmission process. This multi-hop derivation method allows the flexible calculation of the transmission effect between multi-level indicators without modifying the original model, thereby achieving effective modeling of the complex relationship of the causal chain. Through this solution, the system can more accurately capture the indirect impact coefficient of A on C, providing a simple and efficient solution for the attribution analysis of complex causal chains, which is suitable for multi-scale and multi-dimensional business applications.

[0070] like Figure 4 As shown, the second embodiment of the present invention provides a multi-scale cross-domain indicator change attribution system based on knowledge enhancement, the system includes a knowledge graph construction unit 401, an indicator change attribution unit 402 and an attribution result output unit 403;

[0071] The knowledge graph construction unit uses a large language model to construct an indicator causal knowledge graph, fuses the indicator causal knowledge graph with an expert knowledge graph based on business expert knowledge to generate a fused knowledge graph, and extracts causal association information between any indicator and other indicators based on the fused knowledge graph;

[0072] In this unit, the large language model is first fully trained based on the professional knowledge base and rich corpus in the aquaculture field, so that it can accurately understand the professional terms, complex semantic environment and internal logical relationships in the aquaculture field; then, the large language model further learns the common causal relationship extraction corpus, which significantly improves the accuracy and wide applicability of its causal relationship extraction. Combined with actual data for verification, it can automatically judge the authenticity and validity of each connection in the knowledge graph to ensure the timeliness and reliability of the graph content. At the same time, by integrating the knowledge graph built based on expert knowledge, the reliability and credibility of the generated graph are further enhanced, providing scientific and systematic support for the indicator correlation analysis of the aquaculture industry.

[0073] The indicator change attribution unit integrates the knowledge-enhanced Bayesian network and the deep network two-stage regression method based on prior constraints to construct a regression model, selects corresponding multi-scale data as a training set, identifies and extracts causal relationship pairs, performs univariate Bayesian regression on the causal indicators in the causal relationship pairs, and reads the coefficient constraint range; constructs a linear analysis of the causal indicators based on the coefficient constraint range, introduces dimensional additional weights, uses neural networks for deep learning training, and obtains the weights of each causal indicator; inputs the target indicator change to obtain the impact coefficient of the target indicator change on the corresponding causal indicator, and integrates to generate the final attribution data.

[0074] This unit establishes a reliable attribution regression model for each indicator and its subordinate indicators with a single-layer causal relationship. This method not only covers the influencing factors of all relevant indicators, but also ensures the interpretability of the model, making it easy to understand and apply. In the process of model construction, efficient relationship mining is achieved by using the causal relationship between indicators provided by the knowledge graph. However, due to the large number of indicator dimensions involved, the model construction may cause complexity problems. In this regard, through a carefully designed processing solution, the representation and calculation of multi-dimensional data are simplified, making the model structure both efficient and concise, and significantly improving the efficiency of attribution analysis. In addition, for the attribution problem of multi-scale causal relationships, it not only supports the layer-by-layer decomposition analysis of multi-layer causal relationships, but also can identify and quantify the key causal factors in complex chains, thereby providing strong technical support for more comprehensive and accurate attribution analysis. This comprehensive solution makes the causal relationship modeling of knowledge graphs more practical in complex business environments and can better serve the decision-making needs of enterprises.

[0075] The attribution result output unit visualizes the attribution data of the target indicator movement to obtain the mutual influence of the financial indicator and the production indicator changes.

[0076] The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms. In addition, the functional units in the embodiments of the present invention can be all integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the above-mentioned integrated units can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0077] Those skilled in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0078] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-scale cross-domain indicator change attribution method based on knowledge enhancement, characterized in that: The following steps are involved: Use the knowledge base of the aquaculture production field and the data set with causal relationships to conduct two-stage training of the large language model, build an indicator causal knowledge graph through the trained large language model, and perform relationship verification on the indicator causal knowledge graph; Obtain an expert knowledge graph, and fuse the indicator causal knowledge graph with the expert knowledge graph to generate a fused knowledge graph; Construct a regression model, select multi-scale data from the fused knowledge graph as a training set, identify and extract causal relationship pairs, perform univariate Bayesian regression on the causal indicators in the causal relationship pairs, and read the coefficient constraint range; Based on the coefficient constraint range, a linear analysis of causal indicators is constructed, dimension additional weights are introduced, and deep learning training is performed using a neural network to obtain a regression model that integrates a knowledge-enhanced Bayesian network and a deep network based on prior constraints; Obtain target indicator changes from financial indicators and production indicators and import them into the regression model to obtain the impact coefficient of target indicator changes on corresponding causal indicators, and integrate them to generate final attribution data.

2. According to the knowledge-enhanced multi-scale cross-domain indicator change attribution method of claim 1, it is characterized in that: The large language model is trained in two stages using the knowledge base of the aquaculture production field and a dataset with causal relationships. Specifically: Initialize the large language model, obtain a first-stage training sample based on the professional knowledge base and corpus in the field of aquaculture production, use the first-stage training sample to train the large language model, process the first-stage training sample into a number of question-answer pairs through the prompt template of the large language model, and enable the large language model to understand the professional terms, complex semantic environment and internal logical relationships in the field of aquaculture through iterative training; Based on the data set with causal relationship annotations, two-stage training samples are obtained, and the large language model is trained again using the two-stage training samples. Multi-task training is adopted in the two-stage training process, and two task heads are set to be responsible for domain knowledge fine-tuning and causal relationship extraction respectively. Weight parameters are set to control the proportion of the two tasks. Through iterative training, the large language model has the ability of recognition and attribution analysis, and can capture the potential causal relationship between indicators.

3. According to the knowledge-enhanced multi-scale cross-domain indicator change attribution method of claim 1, it is characterized in that: The indicator causal knowledge graph is constructed through the trained large language model, and the relationship of the indicator causal knowledge graph is verified, specifically: Obtain a large language model that has mastered the knowledge in the field of aquaculture production after training, use the large language model to extract entities and causal relationships from documents related to the pre-selected projects, and use the entities and causal relationships as nodes and edge structures of the knowledge graph to construct an indicator causal knowledge graph; The actual data of the pre-selected items are obtained as the basis for verification, and the causal relationship existing in the indicator causal knowledge graph is verified using the Granger causality test. When the relationship verification meets the preset standards, the expression of the corresponding causal relationship in the indicator causal knowledge graph is strengthened; when the relationship verification does not meet the preset standards, the corresponding causal relationship is removed from the indicator causal knowledge graph.

4. According to the knowledge-enhanced multi-scale cross-domain indicator change attribution method of claim 1, it is characterized in that: Obtaining historical indicator change attribution instances, extracting expert-defined indicator relationship data involved in the historical indicator change attribution instances, preprocessing the extracted indicator relationship data, generating expert knowledge, and obtaining an expert knowledge graph based on the expert knowledge; The trained large language model is used to calculate the semantic similarity of entities. The indicator causal knowledge graph is aligned with the entities in the expert knowledge graph according to the semantic similarity. The relationships in the two knowledge graphs are aligned based on the expert knowledge graph, and a fused knowledge graph is generated through graph fusion.

5. According to the knowledge-enhanced multi-scale cross-domain indicator change attribution method of claim 1, it is characterized in that: Build a regression model, select multi-scale data from the fused knowledge graph as a training set, and identify and extract causal relationship pairs, specifically: The regression model is constructed by integrating the knowledge-enhanced Bayesian network and the deep network two-stage regression method based on prior constraints. Multi-scale data is constructed according to different time dimensions and spatial dimensions. The cause and effect indicators involved in the multi-scale data are extracted based on the fusion knowledge graph recognition. Data partitioning is introduced into the regression model, and the result indicator is segmented according to the fluctuation value. Different segments are modeled separately to establish a data-specific regression model. Data from adjacent time periods are selected as training sets to evaluate the impact of cause indicators on result indicators.

6. According to the knowledge-enhanced multi-scale cross-domain indicator change attribution method of claim 1, it is characterized in that: The causal indicators in the causal relationship pair are respectively subjected to univariate Bayesian regression, and the coefficient constraint range is read, specifically: For the automatically extracted causal relationship pairs, the cause indicator and the effect indicator are subjected to separate Bayesian regression, expressed as: y = α n ·x n , where y is the fluctuation of the fruit index, x n Due to the fluctuation of the indicator, α n is the influence coefficient of the nth factor indicator; Obtaining the influence coefficient α through Bayesian regression n The mean and variance of the regression model are used to obtain the coefficient constraint range of the subsequent regression network [μ-3*σ 2 ,μ+3*σ 2 ].

7. According to the knowledge-enhanced multi-scale cross-domain indicator change attribution method of claim 1, it is characterized in that: Based on the coefficient constraint range, a linear analysis of causal indicators is constructed, dimension additional weights are introduced, and deep learning training is performed using a neural network, specifically: After obtaining the coefficient constraint range, the cause indicator and the effect indicator are combined to construct a linear regression, which is expressed as: Among them, y is the fluctuation of the fruit index, x n Due to the fluctuation of the indicator, α n is the influence coefficient of the nth factor, and the influence coefficient is controlled within the coefficient constraint range each time the back propagation gradient descends; In addition, the multiple dimensions of the training samples are processed comprehensively, and the dimensions are encoded and represented using one-hot encoding to obtain the embedding vectors of the dimensions. The embedding vectors of multiple dimensions are imported into the fully connected layer output to generate dimension additional weights to characterize the influence differences of different dimensions. The dimension additional weight is combined with the linear regression to fit into the neural network, which is expressed as: α n ij Add weights to the i-th dimension and the j-th dimension. After iterative deep training, according to α n ij α n Characterizes the complete influence coefficient of the nth factor indicator.

8. A multi-scale cross-domain indicator change attribution system based on knowledge enhancement, characterized in that: Implementing the multi-scale cross-domain indicator change attribution method based on knowledge enhancement as described in any one of claims 1 to 7, the system includes a knowledge graph construction unit, an indicator change attribution unit and an attribution result output unit; The knowledge graph construction unit uses a large language model to construct an indicator causal knowledge graph, fuses the indicator causal knowledge graph with an expert knowledge graph based on business expert knowledge to generate a fused knowledge graph, and extracts causal association information between any indicator and other indicators based on the fused knowledge graph; The indicator change attribution unit integrates the knowledge-enhanced Bayesian network and the deep network two-stage regression method based on prior constraints to build a regression model, selects corresponding multi-scale data as a training set, identifies and extracts causal relationship pairs, performs univariate Bayesian regression on the causal indicators in the causal relationship pairs, and reads the coefficient constraint range; Construct a linear analysis of causal indicators based on the coefficient constraint range, introduce dimension additional weights, use neural networks for deep learning training, and obtain the weights of each causal indicator; Input the target indicator change to obtain the impact coefficient of the target indicator change on the corresponding cause indicator, and integrate to generate the final attribution data; The attribution result output unit visualizes the attribution data of the target indicator movement to obtain the mutual influence of the financial indicator and the production indicator changes.

Images