bi index hierarchical attribution and abnormal deduction method based on metadata causal graph

Abstract

The application discloses a BI index hierarchical attribution and abnormal deduction method based on metadata causal graph, and relates to the technical field of BI data intelligent analysis; four-dimensional metadata systems containing index calculation dependence, dimension level, data blood relationship and business association are first constructed, index causal graphs are constructed relying on inherent business relations of metadata, traditional data correlation fitting is replaced, abnormality is positioned step by step through three-level hierarchical tracing, invalid interference is intelligently filtered in combination with metadata tags, a quantitative model is built based on metadata rules, the influence proportion and deviation amplitude of each abnormal factor are accurately calculated, and finally, causal deduction is completed under the constraint of business rules in combination with a large language model, and structured attribution results are automatically output; the BI index hierarchical attribution and abnormal deduction method based on the metadata causal graph is adopted, accurate attribution that is interpretable and quantifiable is realized, the accuracy is significantly improved, and the method is suitable for enterprise BI intelligent analysis scenes.

Description

Technical Field

[0001] This invention relates to the field of BI data intelligent analysis technology, and in particular to a method for hierarchical attribution and anomaly inference of BI indicators based on metadata causal graphs. Background Technology

[0002] With the improvement of enterprises' digital management systems, BI (Business Intelligence) systems have become the core carriers for monitoring business indicators, analyzing data, and supporting decision-making. Real-time monitoring and anomaly analysis of massive business indicators are key to achieving refined operations and proactive risk control. Currently, most mainstream BI indicator anomaly intelligent attribution solutions in the industry rely on machine learning and big data correlation analysis algorithms to trace anomalies, which have many inherent shortcomings in practical applications.

[0003] First, traditional AI attribution relies on probabilistic matching and deduction based on the correlation of data values. Lacking standardized business logic constraints, it can only uncover superficial numerical relationships, failing to identify the underlying business rules, computational logic, and hierarchical relationships. This easily leads to misattributions where "data is relevant but business is irrelevant," resulting in low accuracy and poor interpretability. Second, existing attribution algorithms lack hierarchical and systematic anomaly decomposition logic. They cannot systematically break down complex anomaly information from multiple dimensions such as company, region, channel, and product, resulting in disorganized and redundant analysis results with a very low proportion of effective information, making it difficult to quickly pinpoint the root cause of anomalies. Finally, current technologies only have the ability to locate anomalies; they cannot quantify the contribution of various influencing factors to core indicators, nor can they generate directly applicable structured analysis reports based on business semantics. They can only achieve passive problem discovery, failing to support quantitative decision-making and intelligent business interpretation, resulting in poor practicality for technology implementation.

[0004] In summary, existing BI metric attribution technologies suffer from industry pain points such as lack of business logic support, low attribution accuracy, chaotic anomaly decomposition, lack of precise quantification capabilities, and output results lacking business semantics. There is an urgent need for an intelligent attribution technology solution that is based on real metadata business rules, provides hierarchical and precise source tracing, is quantifiable, and can intelligently generate business inference results. Summary of the Invention

[0005] The purpose of this invention is to provide a hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs. This method aims to overcome the shortcomings of existing traditional BI indicator AI attribution technologies, which rely on probability matching, lack business logic, produce messy results, cannot be quantified, and lack intelligent business interpretation. It relies on the full four-dimensional metadata of BI to construct a business-oriented causal graph, and combines hierarchical step-by-step tracing and intelligent interference filtering mechanisms to achieve accurate attribution based on real business rules. It uses metadata rules to drive a quantitative model to complete anomaly quantification, and integrates a large language model to achieve intelligent causal reasoning and structured report output, which greatly improves the accuracy, interpretability, and business application value of attribution.

[0006] To achieve the above objectives, this invention provides a hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs, comprising the following steps: S1. Collect and standardize all metadata of the BI system to build a four-dimensional metadata system including indicator calculation dependency metadata, dimension hierarchy metadata, data lineage metadata, and business association metadata, and solidify the calculation caliber, dependency relationship, hierarchical affiliation, data flow, and core information of business rules of all indicators. S2. Based on the inherent business relationships of the four-dimensional metadata in S1, define standardized graph construction rules to generate a visual causal graph including indicator nodes, dimension nodes, data source nodes, business scenario nodes and corresponding business causal relationship edges. S3. Based on the dimensional hierarchy metadata in S1 and the causal graph in S2, a top-down three-level layer-by-layer tracing algorithm is launched to gradually converge the anomaly range from the company's overall dimension, the middle-level subdivided business dimension, and the smallest business unit dimension, thus completing the orderly decomposition and location of multi-dimensional anomalies. S4. Call the metadata interference labeling system to automatically identify invalid interference scenarios, downgrade and remove interference data fluctuations, and refine real business anomaly information. S5. Based on the metadata dependent on the indicator calculation and the metadata of the dimension level, construct an abnormal contribution measurement model to accurately calculate the impact value, contribution ratio and threshold deviation of each abnormal factor on the core indicator. S6, based on the causal graph of S2, the hierarchical tracing results of S3, the interference filtering results of S4, and the quantitative indicator data of S5, integrates the large language model to complete causal reasoning, generates a structured attribution report and business simulation results, and outputs them externally.

[0007] Preferably, the specific content of S1 is as follows: S11. The metadata of indicator calculation is comprehensively collected, including the original calculation methods, nested calculation formulas, field dependencies, aggregation logic, statistical periods, and operation priority rules of all core indicators, derived indicators, and basic indicators in the BI system. The mathematical constraint relationship between the total indicator and the sub-indicator is clarified, providing the underlying calculation basis for subsequent anomaly quantification models and contribution calculations. S12. Dimensional Hierarchical Metadata: This section outlines the entire business dimension system for enterprise operation analysis, including the hierarchical relationships, inclusion relationships, dimension weighting, and dimension splitting rules for dimensions such as company, region, channel, product, customer, and time. It constructs a standardized top-down dimension hierarchy tree to provide hierarchical topology basis for the three-level hierarchical tracing algorithm and dimension contribution quantification. S13. Data lineage metadata traces the entire chain of data flow for each indicator from the original data source collection, data cleaning, ETL processing, hierarchical aggregation, and report display. It records the data source, processing node, update frequency, and data dependency chain to distinguish whether the indicator fluctuation is a real business fluctuation or an abnormal fluctuation in the data chain. S14. Business-related metadata is bound to the business scenarios, operating rules, assessment logic, industry benchmarks, and business thresholds corresponding to each indicator, establishing the correspondence between indicators and business scenarios, business processes, and operating nodes.

[0008] Preferably, the specific content of S2 is as follows: S21. Based on the standardized four-dimensional metadata base output by S1, the four types of metadata are parsed, deduplicated, and normalized to form a structured rule source for graph modeling. S22. Based on the metadata structure parsed in S21, automatically generate all entity nodes of the graph; S23. Calculate the dependency relationship based on the S11 indicator to construct the computation dependency edge, construct the hierarchical dependency edge based on the dimension hierarchy, construct the data flow edge based on the data lineage flow relationship, and construct the business-driven edge based on the business scenario matching relationship, thus forming the causal topology core of the graph. S24. Based on the causal topology of S23, initiate a business verification mechanism for the graph relationship to conduct compliance screening; After S25 and S24 are verified, the final causal graph topology of the indicator is solidified, forming a stable and reusable BI indicator causal inference network.

[0009] Preferably, the specific content of S3 is as follows: S31. The first level is a company-wide screening, based on the fluctuation of time series data of top core indicators, combined with the top-level correlation of the causal graph, to determine the abnormal state of the overall indicators of the enterprise, to lock the first-level business segment to which the abnormality belongs, and to distinguish whether it is an overall systemic abnormality or a local business abnormality. S32. The second level is the mid-level segmented business dimension decomposition. Based on the mid-level dimension structure of the S1 dimension hierarchy metadata, the abnormal business segments are horizontally decomposed in multiple dimensions. The core segmented dimensions such as region, channel, product line, and customer type are traversed in turn. The highly correlated abnormal segmented dimensions are located by combining the causal graph correlation edge, and the abnormal range is converged. S33. The third level is the positioning of the smallest business unit at the end. It continues to sink down to the smallest statistical granularity business unit, including the end dimensions such as single region, single channel, single product, and single cycle. Relying on the indicator calculation dependence and hierarchical subordinate relationship of the causal graph, it accurately locks the finest granular root cause carrier that causes the top-level indicator anomaly and completes the qualitative positioning of the anomaly.

[0010] Preferably, the specific content of S4 is as follows: S41. Automatically retrieve the full metadata tags corresponding to the current analysis task, including four categories: time dimension tags, system operation and maintenance tags, data management tags, and business scenario tags. Accurately map and match the tags of the original abnormal data, and batch mark all potential interference data samples. S42. Based on the matching results of S41, compare with the preset eight categories of standardized interference scenarios to complete the accurate classification of interference types; S43. Perform differentiated processing for different types of interference. For strong interference factors, reduce the weight of their corresponding data fluctuation components by 100%. For weak interference factors, reduce the weight proportionally based on the interference coefficient preset in the metadata. S44. After completing the interference differentiation processing in S43, regenerate the purified indicator time series dataset and synchronously generate interference filtering detail logs. S45. Finally, output pure, real business anomaly data as the sole input data source for S5 anomaly quantification modeling.

[0011] Preferably, the specific content of S5 is as follows: S51. Based on the standard threshold and statistical caliber of the metadata that the indicator calculation depends on, the degree of abnormal deviation of each sub-indicator and each subdivided dimension indicator is calculated through the indicator benchmark deviation formula, and the severity level of indicator abnormality is quantified. S52. Based on the dimension weight coefficients solidified by the dimension hierarchy metadata, the contribution ratio of each abnormal factor to the top-level core indicator is calculated by the formula of single-factor abnormal contribution ratio, so as to accurately distinguish the core root cause and secondary disturbance factors. S53. By correcting the calculation formula of the multi-factor superposition index, the overall fluctuation effect after the superposition of multi-dimensional anomalies is restored, the superposition interference is accurately removed, and the true influence value of a single factor is obtained.

[0012] Preferably, the indicator benchmark deviation formula in S51 is used to quantify the degree of abnormal deviation of the current indicator relative to the standard business threshold. It is the basic formula for determining the level of abnormality, as follows: ; in, This refers to the degree of abnormal deviation of the indicator. These are real-time measured values ​​for the indicators. The standard threshold / benchmark value of the fixed metadata indicator; the positive and negative deviations correspond to positive changes and negative anomalies in the indicator, and the larger the absolute value, the more severe the anomaly.

[0013] Preferably, the formula for the proportion of single-factor anomaly contribution in S52 is based on the weights of each sub-indicator dimension stored in the metadata of the S1 dimension hierarchy, to accurately calculate the influence weight of a single anomaly factor on the core total indicator, as follows: ; in, For the first The contribution percentage of each abnormal factor; For the first Abnormal deviation of individual sub-indicators; The first time that the S1 dimension hierarchical metadata is solidified The business weight coefficient of each sub-indicator corresponds to the total indicator; This represents the total number of valid outliers participating in the calculation.

[0014] Preferably, the multi-factor superposition index correction calculation formula in S53 is used to restore the theoretical fluctuation value of the total index under multi-dimensional abnormal superposition, and accurately distinguish the abnormal effects of single factors and superimposed factors, as follows: ; in, This is the theoretical value of the index after the superposition of multiple factors. For the first Actual measured values ​​of individual sub-indicators; The calculation of the S1 metric depends on the metadata being solidified. Individual indicator standard values.

[0015] Preferably, the specific working content of S6 is as follows: S61. Automatically collect the standardized results from S2-S5 processes to construct a dedicated inference input dataset for large language models; S62. Using metadata business rules as the reasoning boundary and causal graph topology as the transmission logic, closed-loop reasoning is carried out by combining indicator calculation methods, dimensional hierarchical subordinate relationships, and quantitative contribution weights. S63. The large language model is based on historical time series data baseline, current anomaly quantification magnitude, contribution weight of each factor, and indicator linkage rules solidified by metadata to deduce the subsequent trend of core business scenarios and predict the fluctuation range, change trend and risk level of core indicators in the future cycle. S64. Based on the preset standard template, automatically integrate the reasoning conclusions and inference data to generate a complete structured attribution analysis report.

[0016] Therefore, the hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs, as described above, has the following advantages compared with existing technologies: 1. This application differs from the traditional data correlation analysis model without logic. It strictly relies on the inherent business logic relationships of four types of standardized metadata: indicator calculation dependence, dimension level, data lineage, and business association. It constructs an indicator causal graph with strict business constraints, upgrades the traditional probability matching attribution to a traceable and reproducible business rule logic deduction attribution, avoids invalid association and erroneous attribution from the root, and effectively improves the attribution accuracy. 2. This application relies on dimensional hierarchical metadata to construct a top-down three-level business dimension tracing and deduction logic of "global-segmentation-endpoint", and decomposes multi-dimensional mixed anomaly problems layer by layer; at the same time, it intelligently identifies and filters invalid interference factors such as holidays, data supplementation, and system fluctuations based on the metadata tag system, which solves the technical pain points of messy traditional attribution results and scarce effective information. 3. Abandoning the traditional AI black-box fitting weighting method, it fully relies on the fixed rules, weight coefficients, and statistical standards of the indicator calculation dependent on metadata and dimension-level metadata to build a quantitative model, accurately quantifying the influence ratio and deviation of each single factor on the overall indicator, and realizing the technical upgrade from "qualitatively finding problems" to "quantitatively judging weights". 4. Achieve deep integration of metadata quantification results and large-scale model causal reasoning. Based on structured tracing results, quantified data, and graph causal logic, complete business semantic-level causal reasoning, anomaly interpretation, and trend inference through a large language model. Automatically generate standardized structured attribution reports and realize a fully intelligent closed loop of "data calculation - logical attribution - business interpretation - result output".

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is the overall structure diagram of the hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs of this invention; Figure 2 This is a flowchart of the three-level hierarchical tracing algorithm of the BI indicator hierarchical attribution and anomaly inference method based on metadata causal graph of the present invention; Figure 3 This is an attribution report generation interface diagram of the BI indicator hierarchical attribution and anomaly inference method based on metadata causal graph of the present invention; Figure 4This is a comparison chart of the attribution effects of the hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs of this invention. Figure 5 This invention presents a large-scale model deduction and business interpretation diagram for a hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs. Detailed Implementation

[0019] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0020] Example like Figures 1-5 As shown, the hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs of the present invention includes the following steps: S1. Collect and standardize all metadata of the BI system to build a four-dimensional metadata system including indicator calculation dependency metadata, dimension hierarchy metadata, data lineage metadata, and business association metadata, and solidify the calculation caliber, dependency relationship, hierarchical affiliation, data flow, and core information of business rules of all indicators. The specific content of S1 is as follows: S11. The metadata of indicator calculation is comprehensively collected, including the original calculation methods, nested calculation formulas, field dependencies, aggregation logic, statistical periods, and operation priority rules of all core indicators, derived indicators, and basic indicators in the BI system. The mathematical constraint relationship between the total indicator and the sub-indicator is clarified, providing the underlying calculation basis for subsequent anomaly quantification models and contribution calculations. S12. Dimensional Hierarchical Metadata: This section outlines the entire business dimension system for enterprise operation analysis, including the hierarchical relationships, inclusion relationships, dimension weighting, and dimension splitting rules for dimensions such as company, region, channel, product, customer, and time. It constructs a standardized top-down dimension hierarchy tree to provide hierarchical topology basis for the three-level hierarchical tracing algorithm and dimension contribution quantification. S13. Data lineage metadata traces the entire chain of data flow for each indicator from the original data source collection, data cleaning, ETL processing, hierarchical aggregation, and report display. It records the data source, processing node, update frequency, and data dependency chain to distinguish whether the indicator fluctuation is a real business fluctuation or an abnormal fluctuation in the data chain. S14. Bind the business-related metadata to the business scenarios, operating rules, assessment logic, industry benchmarks, and business thresholds corresponding to each indicator, establish the correspondence between indicators and business scenarios, business processes, and operating nodes, and ensure that subsequent attribution and inference are consistent with the real business semantics. S2. Based on the inherent business relationships of the four-dimensional metadata in S1, define standardized graph construction rules to generate a visual causal graph that includes indicator nodes, dimension nodes, data source nodes, business scenario nodes and corresponding business causal relationship edges. All relationships originate from solidified business rules, eliminating random numerical relationships. The specific content of S2 is as follows: S21. Based on the standardized four-dimensional metadata base output by S1, the four types of metadata are parsed, deduplicated, and normalized to form a structured rule source for graph modeling. S22. Based on the metadata structure parsed in S21, all entity nodes of the graph are automatically generated. All node attributes fully inherit the metadata fields. The node types are fixed and standardized, and custom random nodes are not allowed. Metric Nodes: Cover all basic metrics, derived metrics, and core operational metrics of the BI system. Node attributes include metric ID, metric name, calculation method, statistical period, standard threshold, and parent metric. Business Dimension Nodes: Cover all analytical dimensions and hierarchical units. Node attributes include dimension type, hierarchical level, parent dimension, dimension weight, and splitting granularity. Data source nodes: correspond to the original data source of the metrics, ETL processing nodes, and intermediate aggregation tables. Node attributes include data source, update frequency, processing chain, and data dependency ID. Business scenario nodes: These correspond to the business scenario, assessment module, and business process to which the indicator belongs. Node attributes include scenario type, business rules, threshold constraints, and applicable period. After node generation, duplicates are automatically removed and entities with the same origin are merged to ensure that the graph entities are unique and the structure is regular. S23. Calculate the dependency relationship based on the S11 indicator to construct the computation dependency edge, construct the hierarchical dependency edge based on the dimension hierarchy, construct the data flow edge based on the data lineage flow relationship, and construct the business-driven edge based on the business scenario matching relationship, thus forming the causal topology core of the graph. Calculation dependency edge (indicator mathematical causality): It is generated by the sub-indicator node pointing to the parent indicator node, based on the indicator calculation dependency metadata. It represents the causal relationship that "the change of the sub-indicator value directly mathematically drives the fluctuation of the upper-level indicator", and is the core topological basis of the quantitative model.

[0021] Hierarchical dependent edge (dimensional structure causality): It points from the lower-level dimension node to the upper-level dimension node, and is generated based on the dimension hierarchy metadata. It represents the hierarchical convergence causality of "the sum of subdivided dimension changes forming the overall dimension change", supporting S3 three-level hierarchical tracing.

[0022] Data flow edge (data link causality): It points from the data source / processing node to the final indicator node. It is generated based on the data lineage metadata and represents the link causality of "underlying data processing anomalies propagating upwards to indicator anomalies". It is used to distinguish between business anomalies and data anomalies.

[0023] Business-driven edge (scenario rule causality): It is generated by business scenario nodes pointing to corresponding indicator nodes, based on business-related metadata, and represents the business semantic causality of "changes in business rules and changes in scenario conditions causing indicator fluctuations". S24. Based on the causal topology of S23, a business verification mechanism is initiated to screen the graph relationships for compliance, further ensuring the rigor of causality; The business verification mechanism is as follows: Cross-dimensional spurious association removal: Delete numerical similarity associations that are not supported by business rules; Causality direction verification: unify the direction of causal transmission and prevent incorrect reverse derivation; Redundant relationship deduplication: merge duplicate edges with the same origin and logic to simplify the topology; Interference Relationship Masking: Downgrade and mask the interference scenario associations in the metadata tags to prevent interference factors from entering the causal reasoning chain; After S25 and S24 verifications are completed, the final indicator causal graph topology is solidified, forming a stable and reusable BI indicator causal reasoning network. The graph has complete capabilities for positive causal transmission, reverse root cause tracing, and multi-dimensional path decomposition, and can directly support S3 hierarchical tracing, S5 quantitative contribution decomposition, and S6 large-scale model causal reasoning. Finally, a visualized causal graph is output, realizing full-link causal transparency and visibility of "indicators-dimensions-data-business", providing the underlying logical framework for intelligent attribution throughout the entire process. S3. Based on the dimensional hierarchy metadata in S1 and the causal graph in S2, a top-down three-level hierarchical tracing algorithm is launched to converge the anomaly range step by step from the company's global dimension, the mid-level subdivided business dimension, and the smallest business unit dimension. This completes the orderly decomposition and location of multi-dimensional anomalies, realizes the orderly decomposition and hierarchical convergence of multi-dimensional mixed anomalies, and solves the problems of chaotic attribution dimensions and divergent root causes in traditional attribution. The specific details of S3 are as follows: S31. The first level is a company-wide screening, based on the fluctuation of time series data of top core indicators, combined with the top-level correlation of the causal graph, to determine the abnormal state of the overall indicators of the enterprise, to lock the first-level business segment to which the abnormality belongs, and to distinguish whether it is an overall systemic abnormality or a local business abnormality. S32. The second level is the mid-level segmented business dimension decomposition. Based on the mid-level dimension structure of the S1 dimension hierarchy metadata, the abnormal business segments are horizontally decomposed in multiple dimensions. The core segmented dimensions such as region, channel, product line, and customer type are traversed in turn. The highly correlated abnormal segmented dimensions are located by combining the causal graph correlation edge, and the abnormal range is converged. S33. The third level is the positioning of the smallest business unit at the end. It continues to sink down to the smallest statistical granularity business unit, including the end dimensions such as single region, single channel, single product, and single cycle. Relying on the indicator calculation dependence and hierarchical subordinate relationship of the causal graph, it accurately locks the finest granular root cause carrier that causes the top-level indicator anomaly and completes the qualitative positioning of the anomaly. Through a three-level convergence mechanism, the anomaly can be located in an orderly manner from "global fuzzy fluctuations" to "precise unit root causes", ensuring that the source tracing process is clear in hierarchy, logically traceable, and dimensionally consistent. S4. Call the metadata interference labeling system to automatically identify invalid interference scenarios such as holidays, data supplementation, system fluctuations, and test data. Downgrade and remove interference data fluctuations to extract real business anomaly information. The specific content of S4 is as follows: S41. Automatically retrieves all metadata tags corresponding to the current analysis task, including four categories: time dimension tags, system operation and maintenance tags, data management tags, and business scenario tags. Time dimension tags include weekdays, holidays, end-of-month / quarter-end / year-end settlement cycles, and statistical cycle switching indicators; system operation and maintenance tags include system upgrades, version iterations, interface debugging, service restarts, and link fluctuation indicators; data management tags include data supplementation, data reruns, baseline correction, test data entry, and data delay indicators; business scenario tags include marketing activities, operational suspension, and special periods; accurately maps and matches the tags of the original abnormal data, and batch marks all potential interference data samples; S42. Based on the matching results of S41, compare with the preset eight categories of standardized interference scenarios to complete the accurate classification of interference types; The eight categories of interference scenarios have been pre-defined and embedded into the metadata interference tagging system as follows: Natural fluctuations during holidays, monthly / quarterly periodic data supplementation, system upgrade and iteration fluctuations, data interface latency jitter, test data entry interference, settlement node data disturbances, data rerun coverage fluctuations, and baseline update fluctuations; Meanwhile, interference factors are divided into two categories: strong interference and weak interference. Strong interference refers to data fluctuations that have no business analysis value, such as test data and system restart fluctuations. Weak interference refers to scenarios with small and reasonable fluctuations that require weighting, such as holiday fluctuations and periodic data supplementation fluctuations. S43. Implement differentiated processing for different interferences. For strong interference factors, reduce the weight of their corresponding data fluctuation components by 100% to completely remove the index fluctuations caused by the interference. For weak interference factors, retain a small amount of reasonable fluctuations and remove invalid disturbance components. Through the component stripping mechanism, accurately distinguish between passive fluctuations caused by interference and active fluctuations caused by business operations. S44. After completing the interference differentiation processing in S43, regenerate the purified indicator time series dataset and simultaneously generate interference filtering detail logs to record the interference type, interference period, downgrade weight, stripped fluctuation amount, and remaining effective fluctuation amount identified this time, so as to achieve full-link traceability and auditability of the filtering process. S45. Finally, it outputs pure, real business anomaly data as the sole input data source for S5 anomaly quantification modeling, ensuring the accuracy of subsequent quantification calculations and root cause analysis from the source. S5. Based on the metadata dependent on the indicator calculation and the metadata of the dimension level, construct an abnormal contribution measurement model to accurately calculate the impact value, contribution ratio and threshold deviation of each abnormal factor on the core indicator. The specific details of S5 are as follows: S51. Based on the standard threshold and statistical caliber of the metadata that the indicator calculation depends on, the degree of abnormal deviation of each sub-indicator and each subdivided dimension indicator is calculated through the indicator benchmark deviation formula, and the severity level of indicator abnormality is quantified. The indicator deviation formula is used to quantify the degree of abnormal deviation of the current indicator from the standard business threshold. It is the basic formula for determining the level of abnormality, as follows: ; in, This refers to the degree of abnormal deviation of the indicator. These are real-time measured values ​​for the indicators. The standard threshold / benchmark value of the indicator is fixed for metadata; the positive and negative deviations correspond to positive anomalies and negative abnormalities of the indicator, and the larger the absolute value, the more severe the abnormality. S52. Based on the dimension weight coefficients solidified by the dimension hierarchy metadata, the contribution ratio of each abnormal factor to the top-level core indicator is calculated by the formula of single-factor abnormal contribution ratio, so as to accurately distinguish the core root cause and secondary disturbance factors. The formula for the single-factor anomaly contribution ratio is based on the weights of each sub-indicator dimension stored in the S1 dimension-level metadata, accurately calculating the impact weight of a single anomaly factor on the core total indicator, as follows: ; in, For the first The contribution percentage of each abnormal factor; For the first Abnormal deviation of individual sub-indicators; The first time that the S1 dimension hierarchical metadata is solidified The business weight coefficient of each sub-indicator corresponds to the total indicator; The total number of valid outliers participating in the calculation; S53. By using the multi-factor superposition index correction calculation formula, the overall fluctuation effect after multi-dimensional abnormal superposition is restored, and the superposition interference is accurately removed to obtain the true influence value of a single factor. The multi-factor superposition index correction calculation formula is used to restore the theoretical fluctuation value of the total index under the superposition of multi-dimensional anomalies, and accurately distinguish the abnormal effects of single factors and superimposed factors, as follows: ; in, This is the theoretical value of the index after the superposition of multiple factors. For the first Actual measured values ​​of individual sub-indicators; The calculation of the S1 metric depends on the metadata being solidified. Standard values ​​for individual sub-indicators; S6, based on the causal graph of S2, the hierarchical tracing results of S3, the interference filtering results of S4, and the quantitative indicator data of S5, integrates the large language model to complete causal reasoning, generates a structured attribution report and business simulation results, and outputs them to external parties. The specific tasks of S6 are as follows: S61. Automatically collect the standardized results from S2-S5 processes to construct a dedicated inference input dataset for the large language model, including the node relationships and anomaly propagation topology of the S2 indicator causal graph; the complete investigation path, anomaly convergence process, and root cause localization results of the S3 three-level hierarchical tracing; the interference types, filtering details, and effective data ratio of the S4 interference filtering; and the deviation, contribution ratio, correction value, and anomaly level of each anomaly factor in the S5, ensuring that the large model's inference input data is complete, structured, and traceable. S62. Using metadata business rules as the reasoning boundary and causal graph topology as the transmission logic, combined with indicator calculation caliber, dimensional hierarchical subordinate relationship, and quantitative contribution weight, closed-loop reasoning is carried out to automatically sort out the complete transmission link of top-level indicator anomalies, explain the business mechanism of sub-indicator fluctuations being transmitted upwards, distinguish the three types of anomaly causes: data operation and maintenance disturbances, system abnormal fluctuations, and real business problems, and accurately determine the core root cause and secondary influencing factors based on quantitative weights, eliminating unfounded reasoning and semantic illusions. S63, the large language model is based on historical time series data baseline, current anomaly quantification magnitude, contribution weight of each factor, and indicator linkage rules solidified by metadata. It can deduce the subsequent trend of core business scenarios (anomaly continuously fermenting scenario, anomaly smoothly mitigating scenario, anomaly completely repairing scenario), predict the fluctuation range, change trend, and risk level of core indicators in the future cycle, and simulate the indicator improvement effect after single-dimensional strategy optimization, so as to realize risk prediction in advance and optimization effect simulation. S64. According to the preset standard template, automatically integrate the reasoning conclusions and inference data to generate a complete structured attribution analysis report. The report has eight fixed contents, including statistics on the overview of indicator anomalies, complete path of hierarchical tracing, explanation of invalid interference filtering, multi-dimensional quantitative contribution details, conclusions on the root causes of core anomalies, business mechanism of anomaly transmission, future trends and risk prediction, and targeted business optimization suggestions. The report is logically rigorous, the data is traceable, and the conclusions are relevant to business operations. It can be directly used for business review, data maintenance and rectification, and business strategy optimization, achieving the final business implementation of BI intelligent attribution, as shown in Table 1.

[0024] Table 1. Comparison of Attribution Schemes;

[0025] In the specific implementation process, the retail group's unified BI business analysis platform was selected as the implementation vehicle. The platform's core monitoring is the top-level indicator of monthly total sales, covering multiple regions, channels, and product categories, and integrating data from store POS, e-commerce, CRM, and data warehouse ETL. The company's original anomaly analysis could only identify indicator fluctuations, unable to distinguish between various disturbances such as real business fluctuations, system maintenance, marketing activities, and data processing. Manual investigation was time-consuming, lacked quantifiable root cause weights, and the analysis conclusions lacked business application guidance. This time, the steps S1-S6 were applied to the real anomaly event of a significant decline in the group's sales in June.

[0026] It unifies and collects all enterprise metrics, dimensions, data flow, and business rule information, standardizes and constructs a four-dimensional metadata foundation, and fully precipitates metric calculation formulas, hierarchical structures, full-link data lineage, and business scenario threshold tags. Based on the standardized metadata, it automatically generates a causal graph containing four types of entities: metrics, dimensions, data sources, and business scenarios. It generates four types of causal edges: calculation, hierarchy, data flow, and business drive, based on the native business relationships of the metadata. After business compliance verification to eliminate false numerical associations, it solidifies a reusable and visual causal reasoning network, completing the construction of the system's underlying logical skeleton.

[0027] When a significant deviation of the top-level indicator from the baseline threshold is detected, an anomaly analysis task is triggered. Based on the dimensional hierarchy topology and causal graph, a three-level convergence tracing is performed from top to bottom. First, the overall abnormal attributes of the enterprise are determined, and the first-level business segment to which the anomaly belongs is identified. Then, each middle-level subdivision is horizontally decomposed to quickly narrow the scope of the anomaly. Finally, the analysis is carried out down to the smallest statistical granularity business unit to accurately locate the underlying root cause carrier that caused the top-level indicator to change, thus completing the orderly decomposition of multi-dimensional mixed anomalies.

[0028] The system retrieves four types of metadata tags preset by the system: time, operation and maintenance, data management and control, and business scenario. It automatically matches and identifies eight standardized interference scenarios, such as holidays, system upgrades, data reruns, and test data. It distinguishes between strong and weak interference factors and performs differentiated weight attenuation processing. It completely removes false fluctuations caused by strong interference and weakens weak interference disturbance components proportionally. It generates filtered detailed logs and outputs pure business time-series data after removing interference, which serves as the only valid input for quantitative calculation.

[0029] Based on the fixed indicator benchmarks and dimensional business weights of metadata, a quantitative model is built. The deviation formula is used to quantify the severity of anomalies in each sub-indicator. The contribution ratio of each inducement to the core indicator is calculated based on the factor weight. Then, the multi-factor superposition correction formula is used to remove the multi-dimensional superposition interference, accurately restore the indicator fluctuation value caused by a single inducement, clearly distinguish the core root cause from the secondary disturbance factor, and realize the transition from qualitative anomaly identification to quantitative weight determination.

[0030] By integrating graph topology, hierarchical tracing paths, interference filtering logs, and quantitative calculation results to form a standardized input dataset, and using metadata business rules as constraints and causal transmission links as logic to execute closed-loop reasoning of a large language model, the complete anomaly transmission business mechanism is sorted out, and the causes of data-related, operation and maintenance-related, and business-related anomalies are distinguished. Combining historical time-series baselines, the future trends and risk levels of indicators in multiple scenarios are inferred. Finally, a complete structured business analysis report is automatically generated according to a fixed template, which includes anomaly overview, tracing path, interference description, quantitative details, root cause conclusions, trend predictions, and optimization suggestions. This report can be exported and delivered to business departments for implementation and rectification.

[0031] The business anomalies and causal relationships generated in this analysis are simultaneously updated to the four-dimensional metadata system and causal graph, completing the model iteration and solidification. Subsequent anomalies of similar indicators can be directly attributed to the existing graph and metadata rules, forming a sustainable and continuously optimized BI intelligent analysis capability.

[0032] Once the entire process is implemented, it completely solves the misjudgment problem of "data-related but business-irrelevant" in traditional correlation attribution, significantly improving attribution accuracy; it eliminates the need for manual layer-by-layer investigation, significantly improving the efficiency of anomaly analysis; it enables all anomaly factors to be quantified and traceable; and it automatically outputs standardized reports with business semantics and implementation suggestions, fully covering the intelligent closed loop of the entire process of "data monitoring - root cause identification - quantitative analysis - business interpretation - risk prediction", adapting to the enterprise's routine operational indicator monitoring scenarios.

[0033] Therefore, this invention adopts the above-mentioned method of hierarchical attribution and anomaly inference of BI indicators based on metadata causal graphs. It constructs a business-oriented causal graph based on the full four-dimensional metadata of BI, and combines hierarchical step-by-step tracing and intelligent interference filtering mechanism to achieve accurate attribution based on real business rules. It completes anomaly quantification by driving the quantitative model through metadata rules, and integrates a large language model to realize intelligent causal reasoning and structured report output, which greatly improves the accuracy, interpretability and business implementation value of attribution.

[0034] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs, characterized in that, Includes the following steps: S1. Collect and standardize all metadata of the BI system to build a four-dimensional metadata system including indicator calculation dependency metadata, dimension hierarchy metadata, data lineage metadata, and business association metadata, and solidify the calculation caliber, dependency relationship, hierarchical affiliation, data flow, and core information of business rules of all indicators. S2. Based on the inherent business relationships of the four-dimensional metadata in S1, define standardized graph construction rules to generate a visual causal graph including indicator nodes, dimension nodes, data source nodes, business scenario nodes and corresponding business causal relationship edges. S3. Based on the dimensional hierarchy metadata in S1 and the causal graph in S2, a top-down three-level layer-by-layer tracing algorithm is launched to gradually converge the anomaly range from the company's overall dimension, the middle-level subdivided business dimension, and the smallest business unit dimension, thus completing the orderly decomposition and location of multi-dimensional anomalies. S4. Call the metadata interference labeling system to automatically identify invalid interference scenarios, downgrade and remove interference data fluctuations, and refine real business anomaly information. S5. Based on the metadata dependent on the indicator calculation and the metadata of the dimension level, construct an abnormal contribution measurement model to accurately calculate the impact value, contribution ratio and threshold deviation of each abnormal factor on the core indicator. S6, based on the causal graph of S2, the hierarchical tracing results of S3, the interference filtering results of S4, and the quantitative indicator data of S5, integrates the large language model to complete causal reasoning, generates a structured attribution report and business simulation results, and outputs them externally.

2. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 1, characterized in that: The specific content of S1 is as follows: S11. The metadata of indicator calculation is comprehensively collected, including the original calculation methods, nested calculation formulas, field dependencies, aggregation logic, statistical periods, and operation priority rules of all core indicators, derived indicators, and basic indicators in the BI system. The mathematical constraint relationship between the total indicator and the sub-indicator is clarified, providing the underlying calculation basis for subsequent anomaly quantification models and contribution calculations. S12. Dimensional Hierarchical Metadata: This section outlines the entire business dimension system for enterprise operation analysis, including the hierarchical relationships, inclusion relationships, dimension weighting, and dimension splitting rules for dimensions such as company, region, channel, product, customer, and time. It constructs a standardized top-down dimension hierarchy tree to provide hierarchical topology basis for the three-level hierarchical tracing algorithm and dimension contribution quantification. S13. Data lineage metadata traces the entire chain of data flow for each indicator from the original data source collection, data cleaning, ETL processing, hierarchical aggregation, and report display. It records the data source, processing node, update frequency, and data dependency chain to distinguish whether the indicator fluctuation is a real business fluctuation or an abnormal fluctuation in the data chain. S14. Business-related metadata is bound to the business scenarios, operating rules, assessment logic, industry benchmarks, and business thresholds corresponding to each indicator, establishing the correspondence between indicators and business scenarios, business processes, and operating nodes.

3. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 2, characterized in that: The specific content of S2 is as follows: S21. Based on the standardized four-dimensional metadata base output by S1, the four types of metadata are parsed, deduplicated, and normalized to form a structured rule source for graph modeling. S22. Based on the metadata structure parsed in S21, automatically generate all entity nodes of the graph; S23. Calculate the dependency relationship based on the S11 indicator to construct the computation dependency edge, construct the hierarchical dependency edge based on the dimension hierarchy, construct the data flow edge based on the data lineage flow relationship, and construct the business-driven edge based on the business scenario matching relationship, thus forming the causal topology core of the graph. S24. Based on the causal topology of S23, initiate a business verification mechanism for the graph relationship to conduct compliance screening; After S25 and S24 are verified, the final causal graph topology of the indicator is solidified, forming a stable and reusable BI indicator causal inference network.

4. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 3, characterized in that: The specific details of S3 are as follows: S31. The first level is a company-wide screening, based on the fluctuation of time series data of top core indicators, combined with the top-level correlation of the causal graph, to determine the abnormal state of the overall indicators of the enterprise, to lock the first-level business segment to which the abnormality belongs, and to distinguish whether it is an overall systemic abnormality or a local business abnormality. S32. The second level is the mid-level segmented business dimension decomposition. Based on the mid-level dimension structure of the S1 dimension hierarchy metadata, the abnormal business segments are horizontally decomposed in multiple dimensions. The core segmented dimensions such as region, channel, product line, and customer type are traversed in turn. The highly correlated abnormal segmented dimensions are located by combining the causal graph correlation edge, and the abnormal range is converged. S33. The third level is the positioning of the smallest business unit at the end. It continues to sink down to the smallest statistical granularity business unit, including the end dimensions such as single region, single channel, single product, and single cycle. Relying on the indicator calculation dependence and hierarchical subordinate relationship of the causal graph, it accurately locks the finest granular root cause carrier that causes the top-level indicator anomaly and completes the qualitative positioning of the anomaly.

5. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 4, characterized in that: The specific content of S4 is as follows: S41. Automatically retrieve the full metadata tags corresponding to the current analysis task, including four categories: time dimension tags, system operation and maintenance tags, data management tags, and business scenario tags. Accurately map and match the tags of the original abnormal data, and batch mark all potential interference data samples. S42. Based on the matching results of S41, compare with the preset eight categories of standardized interference scenarios to complete the accurate classification of interference types; S43. Perform differentiated processing for different types of interference. For strong interference factors, reduce the weight of their corresponding data fluctuation components by 100%. For weak interference factors, reduce the weight proportionally based on the interference coefficient preset in the metadata. S44. After completing the interference differentiation processing in S43, regenerate the purified indicator time series dataset and synchronously generate interference filtering detail logs. S45. Finally, output pure, real business anomaly data as the sole input data source for S5 anomaly quantification modeling.

6. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 5, characterized in that: The specific details of S5 are as follows: S51. Based on the standard threshold and statistical caliber of the metadata that the indicator calculation depends on, the degree of abnormal deviation of each sub-indicator and each subdivided dimension indicator is calculated through the indicator benchmark deviation formula, and the severity level of indicator abnormality is quantified. S52. Based on the dimension weight coefficients solidified by the dimension hierarchy metadata, the contribution ratio of each abnormal factor to the top-level core indicator is calculated by the formula of single-factor abnormal contribution ratio, so as to accurately distinguish the core root cause and secondary disturbance factors. S53. By correcting the calculation formula of the multi-factor superposition index, the overall fluctuation effect after the superposition of multi-dimensional anomalies is restored, the superposition interference is accurately removed, and the true influence value of a single factor is obtained.

7. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 6, characterized in that: The deviation formula for the indicator benchmark in S51 is used to quantify the degree of abnormal deviation of the current indicator from the standard business threshold. It is the basic formula for determining the level of abnormality, as follows: ； in, This refers to the degree of abnormal deviation of the indicator. These are real-time measured values ​​for the indicators. The standard threshold / benchmark value of the fixed metadata indicator; the positive and negative deviations correspond to positive changes and negative anomalies in the indicator, and the larger the absolute value, the more severe the anomaly.

8. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 7, characterized in that: The formula for the single-factor anomaly contribution ratio in S52 is based on the weights of each sub-indicator dimension stored in the S1 dimension-level metadata. It accurately calculates the impact weight of a single anomaly factor on the core total indicator, as follows: ； in, For the first The contribution percentage of each abnormal factor; For the first Abnormal deviation of individual sub-indicators; The first time that the S1 dimension hierarchical metadata is solidified The business weight coefficient of each sub-indicator corresponds to the total indicator; This represents the total number of valid outliers participating in the calculation.

9. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 8, characterized in that: The multi-factor superposition index correction calculation formula in S53 is used to restore the theoretical fluctuation value of the total index under the superposition of multi-dimensional anomalies, and accurately distinguish the abnormal effects of single factors and superimposed factors, as follows: ； in, This is the theoretical value of the index after the superposition of multiple factors. For the first Actual measured values ​​of individual sub-indicators; The calculation of the S1 metric depends on the metadata being solidified. Individual indicator standard values.

10. The hierarchical attribution and anomaly inference method for BI indicators based on metadata causal graphs according to claim 9, characterized in that: The specific tasks of S6 are as follows: S61. Automatically collect the standardized results from S2-S5 processes to construct a dedicated inference input dataset for large language models; S62. Using metadata business rules as the reasoning boundary and causal graph topology as the transmission logic, closed-loop reasoning is carried out by combining indicator calculation methods, dimensional hierarchical subordinate relationships, and quantitative contribution weights. S63. The large language model is based on historical time series data baseline, current anomaly quantification magnitude, contribution weight of each factor, and indicator linkage rules solidified by metadata to deduce the subsequent trend of core business scenarios and predict the fluctuation range, change trend and risk level of core indicators in the future cycle. S64. Based on the preset standard template, automatically integrate the reasoning conclusions and inference data to generate a complete structured attribution analysis report.

Images