Knowledge-enhanced multi-scale cross-domain indicator anomaly attribution method and system

Abstract

The application discloses a kind of based on knowledge enhancement's multi-scale cross-domain index abnormal change attribution method and system, it is related to index abnormal change attribution technical field, including: using aquaculture production field knowledge base and the data set with causal relationship two-stage training is carried out to large language model, using large language model constructs index causal knowledge graph and the atlas fusion of expert knowledge graph, constructs the regression model of fusion knowledge enhanced bayesian network and priori constraint based on deep network, select multi-scale data as training set from fusion knowledge graph, identify and extract causal relationship pair, model training is carried out, obtains target index abnormal change in financial index and production index and imports the regression model after training, obtains the attribution data corresponding to target index abnormal change.The application solves the problem that positioning and analysis of abnormal index basically rely on artificial experience due to numerous data, complex dimension, is conducive to the efficient operation of enterprise and data-driven decision-making of enterprise.

Description

Technical Field

[0001] This invention relates to the field of index anomaly attribution technology, and more specifically, to a multi-scale, cross-domain index anomaly attribution method and system based on knowledge enhancement. Background Technology

[0002] In agricultural production, a data gap has always existed between finance and production. Building effective knowledge bridges to achieve cross-domain connections and integration is a major challenge facing agricultural enterprises. In particular, given the complex relationships between financial and production indicators in the livestock industry, comprehensively and clearly capturing the relationships between various indicators is crucial for identifying and addressing business problems, optimizing decisions, and adjusting operations.

[0003] However, traditional methods for attributing anomalies in indicators either rely entirely on manual analysis by experts, which is time-consuming and requires highly specialized expertise, or they depend entirely on explicit calculation formulas provided in simple operating environments. Once extended to complex operating environments where clear relationships cannot be provided, they become ineffective. For example, in the scenario addressed by this invention, anomalies between financial and production indicators, if addressed using traditional methods, suffer from the inability to combine domain business knowledge to identify relevant influencing indicators, thus failing to pinpoint the key risk indicators. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes a multi-scale, cross-domain indicator anomaly attribution method and system based on knowledge enhancement. Through the high integration of knowledge graphs and indicator attribution tasks, it spans from finance to production, helping enterprises monitor abnormal fluctuations in various business indicators in real time, quickly pinpoint the root cause of problems, and provide in-depth analysis and intelligent diagnosis to support decision optimization and business adjustments.

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

[0006] The large language model is trained in two stages using a knowledge base in the field of aquaculture production and a dataset with causal relationships. An indicator causal knowledge graph is constructed using the trained large language model, and the relationship of the indicator causal knowledge graph is verified.

[0007] Obtain the 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 the training set, identify and extract causal relationship pairs, perform univariate Bayes regression on the causal indicators in the causal relationship pairs respectively, and read the coefficient constraint range.

[0009] Based on the range of the coefficient constraints, a linear analysis of the causal index is constructed. Dimensional additional weights are introduced, and a 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] The target indicator fluctuations are obtained from financial and production indicators and imported into the regression model to obtain the impact coefficients of the target indicator fluctuations on the corresponding causal indicators. These are then integrated to generate the final attribution data.

[0011] In this approach, a knowledge base from the aquaculture production domain and a dataset containing causal relationships are used to train a large language model in two stages, specifically:

[0012] A large language model is initialized. Based on the professional knowledge base and corpus of the aquaculture production field, a first-stage training sample is obtained. The first-stage training sample is used to train the large language model. The first-stage training sample is processed into several question-answer pairs through the prompt template of the large language model. Through iterative training, the large language model can understand the professional terminology, complex semantic environment and internal logical relationship of the aquaculture field.

[0013] Two-stage training samples are obtained based on a dataset with causal relationship annotations. The large language model is then trained again using these two-stage training samples. During the two-stage training, multi-task training is adopted, with two task heads 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 is made capable of recognition and attribution analysis, and can capture the potential causal relationships between indicators.

[0014] In this scheme, a causal knowledge graph of indicators is constructed through a trained large language model, and the relationships of the causal knowledge graph of indicators are verified, specifically as follows:

[0015] After training, a large language model is acquired that has mastered knowledge in the field of aquaculture production. The large language model is used to extract entities and causal relationships from documents related to pre-selected projects. The entities and causal relationships are used 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. The Granger causality test is used to verify the causal relationships existing in the indicator causal knowledge graph. When the relationship verification meets the preset standard, the expression of the corresponding causal relationship in the indicator causal knowledge graph is strengthened. When the relationship verification does not meet the preset standard, the corresponding causal relationship is removed from the indicator causal knowledge graph.

[0017] In this scheme, historical indicator fluctuation attribution instances are obtained, expert-defined indicator relationship data involved in the historical indicator fluctuation attribution instances are extracted, the extracted indicator relationship data is 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 entity semantic similarity. Based on the semantic similarity, the entities in the indicator causal knowledge graph and the expert knowledge graph are aligned. Using the expert knowledge graph as a benchmark, the relationships in the two knowledge graphs are aligned. A fused knowledge graph is generated through graph fusion.

[0019] In this approach, a regression model is constructed, multi-scale data is selected from the fused knowledge graph as the training set, and causal relationship pairs are identified and extracted, specifically as follows:

[0020] A regression model is constructed by integrating a two-stage regression method that combines knowledge-enhanced Bayesian networks and prior-constrained deep networks. Multi-scale data is constructed based on different time and spatial dimensions, and causal and effect indicators involved in the multi-scale data are identified and extracted based on the integrated knowledge graph.

[0021] The regression model introduces data partitioning, dividing the result index into segments based on the magnitude of its fluctuation. Different segments are modeled separately to establish a data-specific regression model. Furthermore, data from adjacent time periods are selected as the training set to evaluate the impact of the cause index on the result index.

[0022] In this scheme, the causal indicators in the causal relationship pairs are subjected to univariate Bayes regression, and the coefficient constraint range is read, specifically as follows:

[0023] For automatically extracted causal pairs, perform separate Bayesian regression on the causal and effect indicators, representing it as: y = α n ·x n Where y represents the fluctuation of the result indicator, and x n Due to the fluctuation of the indicator, α n The influence coefficient of the nth factor indicator;

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

[0025] In this scheme, a linear analysis of causal indicators is constructed based on the aforementioned coefficient constraint range. Dimensional weights are introduced, and a neural network is used for deep learning training. Specifically:

[0026] After obtaining the coefficient constraint range, the causal and effect indicators are combined to construct a linear regression, which is expressed as: Where y represents the fluctuation of the result indicator, x n Due to the fluctuation of the indicator, α n The influence coefficient of the nth factor index is controlled within the coefficient constraint range during each backpropagation gradient descent.

[0027] In addition, the training samples are processed in multiple dimensions, and one-hot encoding is used to encode the dimensions to obtain the dimension embedding vectors. The embedding vectors of multiple dimensions are then imported into the fully connected layer output to generate dimension-added weights, which represent the differences in the influence of different dimensions.

[0028] By adding weights to the aforementioned dimensions and combining them with the linear regression, the data is fed into a neural network for fitting, as shown below: α n ij Weights are assigned to the i-th and j-th dimensions, and after iterative deep training, weights are assigned according to α. n ij ·α n The complete influence coefficient representing the nth factor indicator.

[0029] The second aspect of this 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, and then merges the indicator causal knowledge graph with an expert knowledge graph based on business expert knowledge to generate a fused knowledge graph. Based on the fused knowledge graph, causal relationship information between any indicator and other indicators is extracted.

[0031] The attribution unit for index anomalies integrates a two-stage regression method combining a knowledge-enhanced Bayesian network and a deep network based on prior constraints to construct a regression model. It selects corresponding multi-scale data as the training set, identifies and extracts causal relationship pairs, and performs univariate Bayesian regression on the causal indicators within each pair, reading the coefficient constraint range. Based on the coefficient constraint range, it constructs a linear analysis of the causal indicators, introduces dimensional weights, and uses a neural network for deep learning training to obtain the weights of each causal indicator. It inputs the target index anomaly to obtain the influence coefficients of the corresponding causal indicators, and integrates these to generate the final attribution data.

[0032] The attribution result output unit visualizes the attribution data of the target indicator movement, and obtains the mutual influence of the changes in financial indicators and production indicators.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention proposes and implements a data-driven knowledge modeling method for causal relationships of indicators in the aquaculture industry, as well as an automated construction method for an 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 anomaly attribution algorithm is proposed for aquaculture enterprises.

[0035] This invention, through a layer-by-layer penetrating analysis of abnormal indicators, highly integrates knowledge graphs and indicator attribution tasks. The system, from macro trends to micro details (multi-scale), and from finance to production (cross-domain), helps enterprises monitor abnormal fluctuations of various business indicators in real time, quickly locate the root cause of problems, and provide in-depth analysis and intelligent diagnosis to support decision optimization and business adjustments. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0037] Figure 1 A flowchart of a knowledge-enhanced multi-scale cross-domain indicator anomaly attribution method is shown;

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

[0039] Figure 3 A flowchart of the multi-scale index attribution modeling implementation is shown;

[0040] Figure 4A block diagram of a knowledge-enhanced multi-scale cross-domain indicator anomaly attribution system is shown. Detailed Implementation

[0041] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

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

[0044] S102, The large language model is trained in two stages using a knowledge base in the field of aquaculture production and a dataset with causal relationships. An indicator causal knowledge graph is constructed through the trained large language model, and the relationship of the indicator causal knowledge graph is verified.

[0045] S104, Obtain the expert knowledge graph, and perform graph fusion between the indicator causal knowledge graph and the expert knowledge graph to generate a fused knowledge graph;

[0046] S106, Construct a regression model, select multi-scale data from the fused knowledge graph as the training set, identify and extract causal relationship pairs, perform univariate Bayes regression on the causal indicators in the causal relationship pairs respectively, and read the coefficient constraint range.

[0047] S108, Based on the range of the coefficient constraints, construct a linear analysis of the causal index, introduce additional weights for the dimension, use a neural network for deep learning training, and obtain a regression model that integrates a Bayesian network with knowledge enhancement and a deep network based on prior constraints.

[0048] S110, Obtain the target indicator anomalies from financial indicators and production indicators, import them into the regression model, obtain the impact coefficients of the target indicator anomalies on the corresponding causal indicators, and integrate them to generate the final attribution data.

[0049] Leveraging advanced large-scale pre-modeling technology, this approach employs a two-stage learning process to master domain knowledge and causal relationship extraction capabilities. It can extract relationships between indicators from commonly used enterprise textual materials that are not covered by expert knowledge. In the first stage, the large language model systematically learns industry-specific professional knowledge from domain-specific unstructured corpora, ensuring the model can understand and identify core concepts and terms relevant to the enterprise domain. In the second stage, through causal relationship extraction, the model further acquires the ability to identify and perform attribution analysis, enabling it to efficiently and accurately capture potential causal relationships between indicators in practical applications. The large language model transcends the limitations of expert knowledge, automatically exploring and revealing complex and hidden indicator relationships within the domain, forming a more comprehensive knowledge system. This process not only reduces the cost of manual screening and annotation but also dynamically adapts to data changes, continuously optimizing the structure and content of the knowledge graph. Compared to traditional methods, this approach offers greater flexibility and scalability, providing enterprise managers with more accurate and in-depth decision support, thereby enhancing the overall intelligence level of data analysis and management.

[0050] Figure 2 A schematic diagram of the process for constructing a fusion knowledge graph in an embodiment is shown.

[0051] A large language model is initialized by obtaining initial training samples from a professional knowledge base and corpus in the livestock production field. These samples include domain monographs, research papers, and professional forum content. The large language model is trained using these initial training samples, which are then processed into question-answer pairs using prompt templates provided by the large language model. Iterative training enables the large language model to understand the professional terminology, complex semantic environments, and internal logical relationships within the livestock production field. The initial training not only involves adjusting the large language model parameters and improving the algorithm, but more importantly, it involves a deep learning and mastery of domain knowledge. The initial training samples need to cover a large amount of textual resources related to livestock production, such as the pig production process, information on the entire process from production to sales, and internal company documents. These initial training samples are input into the large language model as the training set. Through multiple rounds of iterative training, the large language model gradually masters the professional terminology, conceptual framework, and internal logical relationships of the field, building a deep understanding of livestock production to support more accurate causal reasoning and application needs.

[0052] To extract the elusive quantitative relationships between indicators, a second-stage training sample is obtained from a dataset labeled with causal relationships, such as the Causal-TimeBank dataset. This second-stage training sample is then used to train the large language model again. To ensure a perfect combination of domain knowledge fine-tuning and causal relationship extraction, multi-task training is employed during the second-stage training process. Two task heads are set up to handle domain knowledge fine-tuning and causal relationship extraction respectively, and weight parameters are set to control the proportion of the two tasks, forming a multi-task loss function. In each batch, data from both tasks is randomly sampled to ensure balanced learning data for both tasks. Through iterative training, the large language model acquires the ability to recognize and perform attribution analysis, enabling it to capture potential causal relationships between indicators.

[0053] Knowledge graphs are a method for structured representation of knowledge, capable of displaying different entities, concepts, and their relationships in the form of a graph. When there is a lack of explicit quantitative logical relationships between indicators, Large Language Models (LLMs) can be used to train knowledge graph construction models to accurately uncover the potential relationships between indicators.

[0054] A large language model, trained to acquire knowledge of the aquaculture production field, was used to extract entities and causal relationships from documents related to pre-selected projects. These documents included internal financial and production data. The entities and causal relationships were then used as nodes and edges in a knowledge graph. Based on this knowledge graph framework, an indicator-based causal knowledge graph was constructed. The domain knowledge and causal relationship extraction capabilities of the large language model played a crucial role in this construction, validating the necessity of the first two training stages. The model can identify technical terms and contextual logical relationships in the data, thus assisting in the construction of a more accurate and comprehensive knowledge graph.

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

[0056] The Granger causality test is based on the core idea that if a change in one variable X helps predict future changes in another variable Y, then X is a Granger cause of Y. Assume a causal relationship exists between X and Y in the knowledge graph, i.e., X is a cause of Y. Let's set a financial time series X... t and production time series Y tTest X t Does Granger cause Y? t Regression can be performed using the following two models.

[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] Where α, β1, and γ1 are linear parameters, ∈ t For the noise term, the key to the Granger test is comparing the goodness of fit between the two models. Granger causality tests are typically two-way. Besides testing X... t Did Granger cause Y? t Y also needs to be tested t Did Granger cause X? t Through two-way testing, a more comprehensive understanding of the financial time series X can be achieved. t and production time series Y t The dynamic relationship between them. If the model is significantly improved after introducing lag values ​​(i.e., γ1 is significantly non-zero), then X can be considered... t Granger caused Y t The representation of this relationship in the knowledge graph is preserved and strengthened. However, if γ1 is 0, it is considered to remove it from the knowledge graph to avoid invalid or weakly related information interfering with the graph's structure. The reliability and accuracy of the knowledge graph are further improved by calculating Granger causal relationships between nodes. 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, thus providing solid and reliable support for deeper analysis and applications.

[0060] Historical indicator anomaly attribution instances are obtained, and expert-defined indicator relationship data involved in these instances are extracted. The extracted indicator relationship data is preprocessed to generate expert knowledge, and an expert knowledge graph is obtained based on this knowledge. Entity semantic similarity is calculated using a trained large language model. Based on semantic similarity, entities in the indicator causal knowledge graph and the expert knowledge graph are aligned. Using the expert knowledge graph as a benchmark, relationships between the two knowledge graphs are aligned. After alignment, entities or relationships not covered by expert knowledge in the indicator causal knowledge graph are added to the expert knowledge graph to enhance its completeness. Finally, a fused knowledge graph is generated through graph fusion.

[0061] By introducing expert knowledge graphs, the system can integrate the insights and analytical results accumulated by experts over a long period of time, combining them with the graphs trained by the large model to obtain a more comprehensive and richer knowledge graph. The integrated knowledge graph not only possesses the potential causal relationship information of indicators learned by the large language model, but also retains the experts' deep understanding of the correlations between key indicators, better reflecting business logic and actual operations. This integration method has significant advantages in developing stable indicator relationship attribution models, providing a more reliable foundation for causal analysis, helping to improve the accuracy and robustness of attribution models, and meeting actual business needs. The introduction of expert knowledge not only enhances the scientific rigor and accuracy of the graph construction, but also ensures the rationality of 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 illustrating the multi-scale value index attribution modeling of an embodiment is shown.

[0063] A two-stage regression model is constructed by integrating knowledge-enhanced Bayesian networks and prior-constrained deep networks. Multi-scale data is built based on different time and spatial dimensions, supporting monthly, quarterly, semi-annual, and annual data scales, and various organizational levels. For example, when performing a five-level monthly attribution task, monthly data from the fifth level should be selected as the training set to ensure that the model training data is consistent with the target scale of the task, thereby improving the accuracy and effectiveness of attribution analysis. Based on the fusion of knowledge graphs, causal and effect indicators involved in the multi-scale data are identified and extracted, comprehensively evaluating the impact of each causal indicator on the effect indicator.

[0064] To address the issues of high volatility and significant changes in training data, a data partitioning approach is introduced into the regression model to ensure its stability and adaptability. The model is segmented based on the magnitude of fluctuations in the result indicator, with different segments modeled separately. This establishes a data-specific regression model, and data from adjacent time periods are selected as the training set to evaluate the impact of the cause indicator on the result indicator. This dynamic updating mechanism of data partitioning better reflects the timeliness of the data, ensuring the model remains effective and accurate in constantly changing real-world environments.

[0065] For automatically extracted causal pairs, perform separate Bayesian regression on the causal and effect indicators, representing it as: y = α n ·x n Where y represents the fluctuation of the result indicator, and x n Due to the fluctuation of the indicator, α n The influence coefficient of the nth factor indicator; the influence coefficient α is obtained through Bayesian regression. nThe mean and variance are used to obtain the coefficient constraint range [μ-3*σ] of the subsequent regression network of the regression model. 2 ,μ+3*σ 2 This approach can address the problem of data sparsity and limit the influence coefficients to prevent overfitting. For the problem of even more sparse data caused by partitioning, Bayesian methods are used to model the prior distribution of the sparse data. Then, constraints are applied to the network based on the prior distribution to achieve accurate modeling with a small number of samples.

[0066] After obtaining the coefficient constraint range, the causal and effect indicators are combined to construct a linear regression, which is expressed as: Where y represents the fluctuation of the result indicator, x n Due to the fluctuation of the indicator, α n The influence coefficient of the nth factor is controlled within the coefficient constraint range during each backpropagation gradient descent. During the regression model training process, the influence weight of each factor is learned using training samples, thereby quantifying its magnitude.

[0067] During the modeling process, although the causal relationships between indicators have been established through knowledge graphs, the challenge of excessive and fine-grained enterprise data dimensions still needs to be overcome. Multiple dimensions lead to insufficient training samples, affecting the sufficiency of model training and potentially causing redundant model parameters and increased complexity. To address this, multiple dimensions of the training samples are comprehensively processed, and one-hot encoding is used to represent these dimensions, obtaining dimension embedding vectors. These embedding vectors are then fed into a fully connected layer to generate additional weights for each dimension, representing the differences in influence between different dimensions. This allows features of different dimensions to influence the effect of "cause" indicators on "effect" indicators in an adaptive weighted manner, effectively simplifying the dimensionality problem. For scenarios with numerous dimensions, the fine granularity leads to a small number of samples, making training difficult. Vectorizing these dimensions not only effectively solves the problem of dimensional diversity but also maintains simplicity.

[0068] By adding weights to the aforementioned dimensions and combining them with the linear regression, the data is fed into a neural network for fitting, as shown below: α n ij Weights are assigned to the i-th and j-th dimensions, and after iterative deep training, weights are assigned according to α. n ij ·α n The complete influence coefficient of the nth factor indicator is represented. 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 accurate and stable modeling of indicator relationships in complex multidimensional enterprise data.

[0069] The method involves acquiring target indicator anomalies from multi-scale data corresponding to financial and production indicators, importing them into a regression model, obtaining the influence coefficients of target indicator anomalies on causal indicators, and integrating them to generate the final attribution data. Further, the regression model construction proposed in this embodiment is mainly applicable to single-hop causal relationships, where the effect indicator is directly affected by the causal indicator, for example, A directly causes B. In this case, the regression model can determine the impact of A's anomaly on B's anomaly. For multi-hop causal chain problems, a preferred embodiment of this invention utilizes a chain model to calculate and progressively derive the influence coefficients in multi-level causal relationships. Specifically, the regression model between B and C can yield the influence coefficient of B's ​​anomaly on C's anomaly, while the regression model between A and B can yield the influence coefficient of A's anomaly on B. Since the anomaly of B is the same in both regression models, the amount of B's ​​anomaly can be aligned with the influence coefficient of B on C to ensure the continuity of the causal chain. Next, the influence coefficient of A on B is proportionally adjusted to align with the influence coefficient of B on C, ensuring that the proportional relationship is maintained during the transmission process. This multi-hop derivation method allows for the flexible calculation of the transmission effects between multi-level indicators without modifying the original model, thus enabling effective modeling of complex causal chains. Through this approach, the system can more accurately capture the indirect influence coefficient of A on C, providing a simple and efficient solution for attribution analysis of complex causal chains, 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 index change attribution system based on knowledge enhancement. The system includes a knowledge graph construction unit 401, an index 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, and then merges the indicator causal knowledge graph with an expert knowledge graph based on business expert knowledge to generate a fused knowledge graph. Based on the fused knowledge graph, causal relationship information between any indicator and other indicators is extracted.

[0072] This unit first trains a large language model comprehensively based on a professional knowledge base and rich corpus in the aquaculture field, enabling it to accurately understand professional terminology, complex semantic environments, and inherent logical relationships within the aquaculture sector. Subsequently, the large language model further learns from a general causal relationship extraction corpus, significantly improving the accuracy and broad applicability of its causal relationship extraction. Validation with real-world data allows for the automatic assessment of the authenticity and validity of connections within the knowledge graph, ensuring the timeliness and reliability of the graph content. Furthermore, by incorporating a knowledge graph built upon expert knowledge, the reliability and credibility of the generated graph are further enhanced, providing scientific and systematic support for the correlation analysis of indicators in the aquaculture industry.

[0073] The attribution unit for index anomalies integrates a two-stage regression method combining a knowledge-enhanced Bayesian network and a deep network based on prior constraints to construct a regression model. It selects corresponding multi-scale data as the training set, identifies and extracts causal relationship pairs, and performs univariate Bayesian regression on the causal indicators within each pair, reading the coefficient constraint range. Based on the coefficient constraint range, it constructs a linear analysis of the causal indicators, introduces dimensional weights, and uses a neural network for deep learning training to obtain the weights of each causal indicator. It inputs the target index anomaly to obtain the influence coefficients of the corresponding causal indicators, and integrates these 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-level causal relationship. This method not only covers all influencing factors of relevant indicators but also ensures the interpretability of the model, making it easy to understand and apply. During model construction, efficient relationship mining is achieved by utilizing the causal relationships between indicators provided by knowledge graphs. However, the large number of indicator dimensions involved may introduce complexity issues into model construction. To address this, a carefully designed processing scheme simplifies the representation and calculation of multi-dimensional data, making the model structure both efficient and concise, significantly improving the efficiency of attribution analysis. Furthermore, for attribution problems involving multi-scale causal relationships, it not only supports layer-by-layer decomposition and analysis of multi-level causal relationships but also identifies and quantifies key causal factors in complex chains, thus providing strong technical support for more comprehensive and accurate attribution analysis. This comprehensive solution makes knowledge graph causal relationship modeling more practical in complex business environments, better serving the decision-making needs of enterprises.

[0075] The attribution result output unit visualizes the attribution data of the target indicator movement, and obtains the mutual influence of the changes in financial indicators and production indicators.

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

[0077] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-scale, cross-domain indicator anomaly attribution method based on knowledge enhancement, characterized in that, This method is applied to the processing and analysis of multi-source heterogeneous time-series data in enterprises, and includes the following steps: The large language model is trained in two stages using a knowledge base in the field of aquaculture production and a dataset with causal relationships. The trained large language model has mastered the knowledge of the field of aquaculture production. The large language model is used to extract entities and causal relationships from documents related to pre-selected projects. The entities and causal relationships are used as nodes and edge structures of the knowledge graph to construct an indicator causal knowledge graph. The actual data of the selected projects are obtained as the basis for verification. The actual data includes time-series data of financial indicators and time-series data of production indicators with timestamps. Granger causality test is used to verify the causal relationships existing in the indicator causal knowledge graph. When the relationship verification meets the preset standard, the expression of the corresponding causal relationship in the indicator causal knowledge graph is strengthened. When the relationship verification does not meet the preset standard, the corresponding causal relationship is removed from the indicator causal knowledge graph. Obtain the expert knowledge graph, and fuse the indicator causal knowledge graph with the expert knowledge graph to generate a fused knowledge graph; A regression model for quantifying impact is constructed. Multi-scale data is selected from the fused knowledge graph as the training set to identify and extract causal pairs. For automatically extracted causal pairs, the cause and effect indicators are subjected to separate Bayesian regressions, as follows: ,in For the fluctuation of the result indicator, Due to fluctuations in the indicators, The influence coefficient of the nth factor indicator; Obtaining the influence coefficient through Bayesian regression The mean and variance are used to obtain the coefficient constraint range of the subsequent regression network of the regression model. ; After obtaining the coefficient constraint range, the causal and effect indicators are combined to construct a linear regression, which is expressed as: ,in For the fluctuation of the result indicator, Due to fluctuations in the indicators, The influence coefficient of the nth factor index is controlled within the coefficient constraint range during each backpropagation gradient descent. In addition, the training samples are processed in multiple dimensions, and one-hot encoding is used to encode the dimensions to obtain the dimension embedding vector. The embedding vectors of multiple dimensions are then imported into the fully connected layer output to generate dimension-added weights, which represent the differences in the influence of different dimensions. By adding weights to the aforementioned dimensions and combining them with the linear regression, the data is fed into a neural network for fitting, as shown below: , For the first Dimensions and the first Each dimension is assigned a weight, and after iterative deep training and learning, based on... Characterize the complete influence coefficient of the nth factor index and obtain the regression model of the knowledge-enhanced Bayesian network and the deep network based on prior constraints; Obtain time-series data on the changes in target indicators related to the financial and production fields, import them into the regression model, obtain the impact coefficients of the changes in target indicators on the causal indicators, and integrate them to generate the final attribution data.

2. The multi-scale cross-domain indicator anomaly attribution method based on knowledge enhancement according to claim 1, characterized in that, The large language model is trained in two stages using a knowledge base from the aquaculture production field and a dataset containing causal relationships. Specifically: A large language model is initialized. Based on the professional knowledge base and corpus of the aquaculture production field, a first-stage training sample is obtained. The first-stage training sample is used to train the large language model. The first-stage training sample is processed into several question-answer pairs through the prompt template of the large language model. Through iterative training, the large language model can understand the professional terminology, complex semantic environment and internal logical relationship of the aquaculture field. Two-stage training samples are obtained based on a dataset with causal relationship annotations. The large language model is then trained again using these two-stage training samples. During the two-stage training, multi-task training is adopted, with two task heads 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 is made capable of recognition and attribution analysis, and can capture the potential causal relationships between indicators.

3. The multi-scale cross-domain indicator anomaly attribution method based on knowledge enhancement according to claim 1, characterized in that, Obtain historical indicator anomaly attribution instances, extract expert-defined indicator relationship data involved in the historical indicator anomaly attribution instances, preprocess the extracted indicator relationship data to generate expert knowledge, and obtain an expert knowledge graph based on the expert knowledge. The trained large language model is used to calculate entity semantic similarity. Based on the semantic similarity, the entities in the indicator causal knowledge graph and the expert knowledge graph are aligned. Using the expert knowledge graph as a benchmark, the relationships in the two knowledge graphs are aligned. A fused knowledge graph is generated through graph fusion.

4. The multi-scale cross-domain indicator anomaly attribution method based on knowledge enhancement according to claim 1, characterized in that, To construct a regression model, multi-scale data from the fused knowledge graph is selected as the training set to identify and extract causal relationship pairs, specifically: A regression model is constructed by integrating a two-stage regression method that combines knowledge-enhanced Bayesian networks and prior-constrained deep networks. Multi-scale data is constructed based on different time and spatial dimensions, and causal and effect indicators involved in the multi-scale data are identified and extracted based on the integrated knowledge graph. The regression model introduces data partitioning, dividing the result index into segments based on the magnitude of its fluctuation. Different segments are modeled separately to establish a data-specific regression model. Furthermore, data from adjacent time periods are selected as the training set to evaluate the impact of the cause index on the result index.

Images