An index layer model construction analysis method and system based on artificial intelligence

Abstract

The present application relates to the technical field of data information processing, and more particularly to an index layer model construction analysis method and system based on artificial intelligence, wherein the method comprises the following steps: step 1: multi-source data synchronization and quality evaluation; step 2: business context analysis and intelligent division of subject domains, analyzing business document and data table field semantics; step 3: intelligent analysis of index requirements and generation of calculation logic, analyzing the index requirements input by the user through natural language processing technology, matching the business entities in the knowledge graph, generating initial calculation logic and optimizing execution efficiency; step 4: dynamic updating and consistency checking of index data, executing the optimized calculation logic to generate index data, tracking data blood relationship and triggering dynamic updating, and setting threshold according to historical fluctuation range to alarm abnormity. Through automatic and intelligent processing, the probability of human intervention and artificial error is reduced, and the efficiency and accuracy of data processing are improved.

Description

Technical Field

[0001] This invention relates to the field of data information processing technology, and in particular to an indicator layer model construction and analysis method and system based on artificial intelligence. Background Technology

[0002] In data processing fields such as banking, the construction and management of indicator-level models in existing technologies mainly rely on manually driven processes, which have the following significant drawbacks:

[0003] The problem of fragmented and redundant calculations of metrics: Existing systems typically define metric calculation rules based on the needs of independent reports, leading to the repeated development of the same metrics in different reports. For example, "average monthly transaction amount per customer" requires separate SQL logic to be written in the customer analysis report and the risk management report, lacking a unified standard. This problem stems from data silos between business departments and the lack of a global metric library, resulting in data redundancy, increased storage costs, and difficulty in ensuring consistency of metrics across reports.

[0004] Manual development is inefficient: the generation of indicator calculation logic relies heavily on developers manually parsing requirements and writing SQL scripts. Taking a bank's practice as an example, developing a single complex indicator (such as "cross-channel customer churn rate") requires an average of 3-5 person-days. Furthermore, due to the long processing chain and complex definitions, error localization accounts for over 40% of the time spent during data verification. In addition, manual coding is prone to introducing logical biases, such as misuse of aggregate functions or omission of correlation conditions, further increasing rework costs.

[0005] Insufficient business responsiveness: Frequent changes in banking business needs (such as adjustments to regulatory rules or the launch of new products) require existing technology to manually re-analyze the requirements and modify the calculation logic, resulting in an average response cycle of 1-2 weeks. For example, when a bank added a "green loan ratio" indicator due to regulatory requirements, it had to coordinate with multiple departments to redefine the definition, severely lagging behind the timeliness requirements of business operations. Summary of the Invention

[0006] To achieve the above objectives, this invention provides a method and system for constructing and analyzing indicator-level models based on artificial intelligence. The method for constructing and analyzing indicator-level models based on artificial intelligence includes the following steps:

[0007] Step 1: Multi-source data synchronization and quality assessment. Synchronize the full data from multiple heterogeneous databases to the distributed computing platform, assess the quality of the synchronized data, and generate cleaned data tables and quality scoring reports.

[0008] Step 2: Business context analysis and intelligent subject domain division. Parse the semantics of business documents and data table fields, construct a knowledge graph containing business entity relationships, and divide the data tables into preset subject domains based on the knowledge graph;

[0009] Step 3: Intelligent parsing and calculation logic generation of indicator requirements. Natural language processing technology is used to parse the indicator requirements input by the user, match them with business entities in the knowledge graph, generate initial calculation logic, and optimize execution efficiency.

[0010] Step 4: Dynamic update and consistency verification of indicator data. Execute the optimized calculation logic to generate indicator data, track data lineage and trigger dynamic updates, and set thresholds based on historical fluctuation ranges to issue anomaly alerts.

[0011] Preferably, the quality assessment includes the following sub-steps:

[0012] Perform statistical analysis on the numerical distribution of each field, build an anomaly detection model based on historical data distribution, and identify anomalous data points;

[0013] The anomaly detection model is constructed in the following manner:

[0014] Randomly sample field values ​​to generate multiple isolated trees, and each tree recursively partitions the data space;

[0015] The probability of an anomaly is calculated based on the path length of the data point in the isolated tree; the shorter the path length, the higher the probability of an anomaly.

[0016] The dynamic threshold is determined by comparing the anomaly probability with the dynamic threshold, which is achieved through the following method:

[0017] The path length distribution of normal data points in historical data is statistically analyzed, and the Pth percentile of the distribution is taken as the threshold benchmark.

[0018] Adjust the threshold benchmark based on the risk level of the current business scenario; reduce the P value in high-risk scenarios to tighten the threshold.

[0019] Preferably, the generation of the quality scoring report includes:

[0020] Define field integrity scoring rules: Calculate the field missing rate. A score of full marks is given if the missing rate is below the first threshold. For every 1% increase in the missing rate above the first threshold, a preset score is deducted.

[0021] Define uniqueness scoring rules: Calculate the proportion of duplicate records in the primary key field. If the proportion of duplicate records is lower than the second threshold, the score is full; if it exceeds the threshold, the score is deducted proportionally.

[0022] Define consistency scoring rules: compare the foreign key matching rate in the related tables, and trigger a data consistency alarm when the matching rate is lower than the third threshold;

[0023] The first threshold, the second threshold, and the third threshold are determined in the following manner:

[0024] Analyze the distribution of field missing rate, primary key duplication rate and foreign key matching rate in historical data, and take the highest tolerance value under normal business scenarios as the initial threshold.

[0025] The threshold is dynamically adjusted based on user feedback. If a user marks an alarm as a false alarm, the corresponding threshold is relaxed by a preset percentage.

[0026] Preferably, the construction of the knowledge graph includes:

[0027] The business documents are parsed using a pre-trained natural language processing model to extract entity names and relationships between entities. The entities include business objects, indicator types, and calculation rules.

[0028] The data table fields are semantically mapped to entities. The mapping process includes:

[0029] Perform word segmentation on the field names to generate word vectors;

[0030] Calculate the cosine similarity between word vectors and entity names in the knowledge graph. If the similarity exceeds the mapping threshold, establish the association between the field and the entity.

[0031] The mapping threshold is determined in the following way:

[0032] Based on historical manually annotated field-entity mapping data, the minimum similarity value of successful matches is calculated and used as the initial threshold;

[0033] The threshold is dynamically adjusted based on the mapping accuracy. If the mapping error rate of a newly added data table increases, the threshold is raised to enhance strictness.

[0034] Preferably, the subject domain division includes:

[0035] Feature extraction is performed on the data table fields to generate table-level feature vectors, specifically including:

[0036] The distribution ratio of field types in the statistical table, including numeric, character, and date types;

[0037] Extract the number of associations between fields in the table and entities in the knowledge graph, and use this as a semantic relevance feature;

[0038] The feature vectors are input into the classification model, which is trained in the following way:

[0039] The topic domain labels from historical data tables were used as the training set, and the gradient boosting decision tree algorithm was used to train the model.

[0040] During the model inference phase, the probability of the output data table belonging to each topic domain is calculated, and the topic domain with the highest probability is selected as the classification result.

[0041] When a new topic domain is added, the model is updated in the following way:

[0042] Collect sample data tables of newly added topic domains, extract their feature vectors, and manually label them;

[0043] An incremental learning algorithm is used to update only the weights of the last fully connected layer of the classification model.

[0044] Preferably, the indicator requirement analysis includes:

[0045] By using a pre-trained natural language understanding model to break down user request statements and identify the following elements:

[0046] Calculation target: Extracting core noun phrases from sentences through dependency parsing;

[0047] Indicator types: Identify calculation rules for "mean", "sum", and "growth rate" through keyword matching;

[0048] Time conditions: Extract the time range description using regular expressions and convert it into a standard time format;

[0049] When a requirement statement is ambiguous, it can be clarified in the following ways:

[0050] Generate multiple candidate parsing results and calculate the confidence level of each result;

[0051] The system provides the user with the top N most confident parsing results and receives the user's selection to determine the final logic.

[0052] Preferably, the computational logic optimization includes:

[0053] Collect historical structured query language execution logs and extract the following parameters:

[0054] Execution time, CPU utilization, and memory consumption;

[0055] The ratio of the resulting data volume to the source data volume;

[0056] Based on the above parameters, construct the environment state of the reinforcement learning model, and define the following optimization actions:

[0057] Adjust the table join order, prioritizing table joins with strict filtering conditions;

[0058] Add temporary indexes to speed up the retrieval of frequently queried fields;

[0059] Break down complex queries into multiple subqueries and execute them in parallel;

[0060] The model is trained using the Q-learning algorithm, and the reward function is defined as the weighted sum of the percentage increase in execution efficiency and the percentage decrease in resource consumption.

[0061] Preferably, the bloodline analysis includes:

[0062] The following metadata is recorded during the indicator calculation process:

[0063] Enter the name, version, and key fields of the data table;

[0064] The filtering conditions, aggregation functions, and relationships used in the computational logic;

[0065] When the upstream data table changes, the update is triggered in the following ways:

[0066] If the version number of the monitoring data table is updated and the affected field includes a field that the metric depends on, then it is marked as needing to be recalculated.

[0067] The task queue is updated based on the indicators generated according to blood relations, and the tasks are scheduled for execution according to priority.

[0068] The priority is determined in the following way:

[0069] The number of times a statistical indicator is cited in downstream reports indicates its priority; the more times it is cited, the higher its priority.

[0070] Priority weights are dynamically adjusted based on the service level agreement requirements set by the business departments.

[0071] Preferably, the setting of the abnormal alarm threshold includes:

[0072] The mean and standard deviation of the sliding window are calculated based on historical indicator data, and the window size is dynamically adjusted according to the business cycle.

[0073] Define exception detection rules:

[0074] If the current indicator value exceeds the mean ± K times the standard deviation, it is considered abnormal;

[0075] The ±K value is determined in the following way:

[0076] Analyze the fluctuation multiples of indicators from historical abnormal events, and take the minimum fluctuation multiple as the initial K value;

[0077] Based on the balance between false alarm rate and false negative rate, the optimal K value is determined using the receiver operating characteristic curve;

[0078] When the same indicator triggers alarms consecutively, the alarm level will be automatically escalated and the designated person in charge will be notified.

[0079] Accordingly, embodiments of the present invention also provide an artificial intelligence-based indicator layer model construction and analysis system for running the artificial intelligence-based indicator layer model construction and analysis method described in any one of the embodiments of the present invention, comprising the following modules:

[0080] Multi-source data synchronization and quality assessment module: includes a data synchronization unit and a quality assessment unit;

[0081] The data synchronization unit is connected to multiple heterogeneous databases and is used to synchronize all data to the distributed computing platform.

[0082] The quality assessment unit is connected to the data synchronization unit, receives synchronized data, performs anomaly detection and quality scoring, and outputs a cleaned data table and a quality scoring report.

[0083] The business knowledge graph construction module includes a business document parsing unit and a semantic mapping unit;

[0084] The business document parsing unit is connected to an external business document library and extracts business entities and relationships through a natural language processing model to generate an initial knowledge graph.

[0085] The semantic mapping unit is connected to the quality assessment module, receives the cleaned data table field information, associates the fields with knowledge graph entities through word vector similarity calculation, and outputs the bound knowledge graph.

[0086] The topic domain intelligent segmentation module includes a feature extraction unit and a classification model unit;

[0087] The feature extraction unit is connected to the business knowledge graph construction module to extract the field type distribution and semantic relevance features of the data table and generate table-level feature vectors.

[0088] The classification model unit is connected to the feature extraction unit, and outputs the topic domain classification results of the data table based on the feature vector and historical topic domain label data.

[0089] Indicator requirement analysis and calculation logic generation module: including natural language understanding unit and logic optimization unit;

[0090] The natural language understanding unit receives the indicator requirement text input by the user, and outputs the calculation object, indicator type and time conditions through dependency parsing and keyword matching.

[0091] The logic optimization unit is connected to the business knowledge graph construction module and the topic domain intelligent segmentation module. It generates initial calculation logic based on the matched fields and template library, optimizes the execution path through a reinforcement learning model, and outputs an optimized structured query language script.

[0092] Indicator data management and verification module: includes calculation execution unit, lineage tracing unit and anomaly alarm unit;

[0093] The computation execution unit is connected to the logic optimization unit and executes a structured query language script to generate an indicator data table.

[0094] The lineage tracing unit is connected to the calculation execution unit, records the upstream tables and fields on which the indicator data depends, and links with the data change monitoring unit to trigger dynamic update tasks;

[0095] The abnormal alarm unit is connected to the bloodline tracking unit, and sets a dynamic threshold based on the historical index fluctuation range to generate alarm notifications for abnormal data.

[0096] The beneficial effects of this invention are:

[0097] By combining natural language processing and knowledge graphs, the system can quickly and accurately understand user needs, automatically generate computational logic, and ensure real-time updates and consistency checks of the calculation results. This not only improves decision-making efficiency but also reduces human intervention, significantly enhancing the automation and accuracy of business analysis. Through automated and intelligent processing, the probability of human intervention and human error is reduced, improving the efficiency and accuracy of data processing. Especially in a big data environment, it can track data changes in real time, automatically update indicators, generate intelligent reports, and provide timely warnings when anomalies occur, thereby helping enterprises to identify problems and make decisions promptly, ultimately improving the accuracy of data analysis and the agility of business operations. Attached Figure Description

[0098] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0099] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0100] Figure 2 A flowchart illustrating the steps for defining optimization actions in the method of this invention;

[0101] Figure 3 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0102] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0103] Please see Figures 1-3This invention provides an artificial intelligence-based indicator layer model construction and analysis method. In the first step, the system synchronizes all data from different data sources to a distributed computing platform. Due to the heterogeneous data sources, a unified quality assessment is required to ensure the accuracy and completeness of the data. The system uses an automated data quality assessment model to detect outliers, missing values, and inconsistencies, and generates cleaned data tables and quality scoring reports. This process not only improves data reliability but also provides high-quality foundational data for subsequent indicator calculations.

[0104] After the data quality assessment is completed, the system begins to parse the semantics of business documents and data table fields, constructing a knowledge graph of relationships between business entities. Through natural language processing technology, business documents are analyzed to identify entities such as business objects and metric types, and data tables are categorized into predefined subject areas based on their business content. This step enables different data tables to be automatically classified according to their business semantics, thus providing a clear business background and context for subsequent metric requirement analysis.

[0105] After receiving the user's input of indicator requirements, the system uses natural language processing technology to parse the user's needs, automatically identifying elements such as the calculation object, indicator type, and time conditions, and matching them with business entities in the knowledge graph to generate preliminary calculation logic. Subsequently, the system optimizes the calculation logic to improve execution efficiency. In this way, manual intervention can be reduced, and efficient and accurate calculation logic can be generated quickly, greatly improving the automation and accuracy of indicator generation.

[0106] When executing optimized calculation logic, the system tracks the source and changes of data based on data lineage, ensuring that relevant indicators are dynamically updated when upstream data changes. The system also sets anomaly alarm thresholds based on historical data fluctuation ranges and promptly alerts when anomalies occur. Through this mechanism, the system ensures data consistency, timely updates and verification of indicators, and avoids errors in analysis results caused by data fluctuations.

[0107] In one possible implementation, firstly, a statistical analysis is performed on the numerical distribution of each data field to understand the historical data distribution characteristics of each field. This step, through detailed analysis of historical data, identifies the range and trend of normal values ​​for each field, providing basic information for anomaly detection. Next, based on the distribution of historical data, an anomaly detection model is constructed specifically to identify outliers in the data. This method can effectively discover potential biases or errors in the data, preventing them from negatively impacting subsequent indicator calculations.

[0108] The core of the anomaly detection model is the isolation tree algorithm. Specifically, the value of each data field is randomly sampled to generate multiple isolation trees. Each isolation tree recursively partitions the data space, thus isolating data points. The shorter the path length, the easier it is for a data point to be isolated, resulting in a higher probability and indicating that the data point is more likely to be an anomaly. By calculating the path length of a data point in the isolation tree and then calculating the anomaly probability based on the path length, the system can effectively identify anomalous data that does not conform to the overall data distribution.

[0109] Preferably, given data point x i And an isolated tree, path length h(x) i The calculation formula for ) is as follows:

[0110] h(x i )=Number of edges traversed to isolate x i ;

[0111] Data point x is obtained by recursively partitioning multiple isolated trees. i The path length distribution can be used to calculate the probability of anomalies.

[0112] Anomaly probability p(x) i The path length can be calculated by relating it to the total path length distribution of the isolated tree, using the following formula:

[0113]

[0114] Where h(x) i ) is the data point x i In an isolated tree, the path length c(n) is a normalization constant with respect to the dataset size n, typically c(n) = 2 × (log₂(n-1) + γ), where γ is an adjustment coefficient. The anomaly probability p(x) is obtained from the anomaly detection model. i This data point is compared to a dynamic threshold to determine whether it is abnormal. This dynamic threshold is obtained by statistically analyzing the path length distribution of normal data points in historical data.

[0115] To more accurately determine whether data points are anomalies, the system identifies anomalies by comparing them with a dynamic threshold. The dynamic threshold is first set by statistically analyzing the path length distribution of normal data points in historical data, and then using the P-th percentile of this distribution as the default threshold benchmark. This allows the threshold benchmark to adapt to the distribution characteristics of different datasets. Subsequently, the system adjusts this threshold based on the risk level of the current business scenario. For high-risk business scenarios, the system lowers the P-value to tighten the threshold, making anomaly detection more sensitive and able to capture more subtle anomaly data points; while in low-risk scenarios, the threshold may be appropriately relaxed to reduce false alarms.

[0116] Preferably, the path length distribution of all normal data points in historical data is statistically analyzed; the P-th percentile value is extracted from the path length distribution and used as the default anomaly detection threshold.

[0117] Threshold=Percentile(p,P);

[0118] Where P is a preset percentile, such as 95%, indicating that 95% of the data points in the path length distribution are considered normal data points.

[0119] In one possible implementation, the field integrity scoring rule first assesses data integrity primarily by statistically analyzing the field missing rate. The specific steps are as follows:

[0120] Calculate the percentage of missing data for each field. If the missing rate is below a set first threshold, the field receives full marks, indicating good data integrity. When the field's missing rate exceeds the first threshold, a pre-defined score is deducted for every 1% exceeding the threshold. For example, if the first threshold is 5% and the field's missing rate is 6%, the corresponding score is deducted.

[0121] By analyzing historical data, the missing rate distribution of each field under normal business scenarios is obtained, thereby determining the first threshold. This threshold is set as the maximum tolerable missing rate to ensure data quality.

[0122] Uniqueness scoring rules primarily assess data uniqueness by statistically analyzing the proportion of duplicate records in the primary key field. The specific steps are as follows:

[0123] The system calculates the percentage of duplicate records in the primary key field. If the percentage is below a second threshold, the field receives a perfect uniqueness score. When the percentage exceeds the second threshold, points are deducted proportionally based on the excess portion. This means that data duplication issues directly impact the data quality score.

[0124] By analyzing historical data and the duplication rate of primary key fields under normal business scenarios, a second threshold is determined. This threshold represents the maximum acceptable duplication rate; exceeding this rate indicates a data quality problem.

[0125] The consistency scoring rule evaluates data consistency by comparing the foreign key matching rate in the related tables. The specific steps are as follows:

[0126] Compare the foreign keys in the related table with the primary keys in the parent table. If the foreign key match rate is below the third threshold, a data consistency alert is triggered, indicating a potential data association problem.

[0127] By analyzing historical data, we can determine the distribution of foreign key matching under normal business scenarios and extract the highest tolerable matching rate as a third threshold. If the foreign key matching rate is lower than this threshold, it indicates an anomaly in the relationship between the data.

[0128] Determining and dynamically adjusting the first, second, and third thresholds are key steps in generating the entire quality scoring report.

[0129] Specifically, by analyzing the distribution of field missing rate, primary key duplication rate, and foreign key matching rate in historical data, first, second, and third thresholds are set. These thresholds represent the highest tolerable values ​​under normal data quality conditions. For example, the first threshold might be set to 5%, the second threshold to 2%, and the third threshold to 95%.

[0130] These thresholds are dynamically adjusted based on user feedback. If a user marks an alert as a false alarm after a given event, the system will accordingly relax the threshold setting to reduce the occurrence of similar alerts in the future. For example, if the missing rate of a certain field is frequently considered a false alarm, the system may relax the first threshold from 5% to 6%.

[0131] The user feedback mechanism influences threshold adjustments by indicating whether alarms are accurate. The system dynamically corrects the thresholds based on historical feedback to adapt to different business scenarios and user needs.

[0132] In one possible implementation, the business document is first parsed using a pre-trained natural language processing model, and relevant entity names and relationships between entities are extracted. These entities typically include:

[0133] Business objects: such as core objects related to business, like customers, orders, and products.

[0134] Indicator types: such as sales revenue, customer growth rate, inventory level, etc., which are metrics for measuring business performance.

[0135] Calculation rules: such as statistical methods or rules for average, maximum, minimum, etc.

[0136] These entities and the relationships between them will form the basic structure of the knowledge graph, helping to clarify the business semantics and interactions of different entities.

[0137] Next, semantic mapping is performed between the data table fields and the knowledge graph entities. The mapping process includes the following steps:

[0138] First, the field names in the data table are segmented into words and converted into word vectors. Then, natural language processing techniques, such as Word2Vec or BERT models, are used to generate word vector representations of the field names.

[0139] Calculate the cosine similarity between these word vectors and entity names in the knowledge graph. Cosine similarity is a common method for measuring the similarity between two vectors; the closer the value is to 1, the more similar the two vectors are. This method can be used to determine the semantic similarity between field names and entity names.

[0140] If the cosine similarity between a field and an entity name exceeds a set mapping threshold, an association is established between the field and the entity, indicating that the field in the data table corresponds to an entity in the knowledge graph.

[0141] To ensure accurate mapping between fields and entities, setting the mapping threshold is crucial. The specific steps are as follows:

[0142] By mapping historical manually labeled data table fields to entities, the minimum similarity score of these successful matches is calculated and used as the initial mapping threshold. This ensures that a mapping relationship is established only when the similarity between a field and an entity name reaches a certain level.

[0143] The system dynamically adjusts the threshold based on the accuracy of the mapping. If the mapping error rate of a newly added data table increases, the system will raise the mapping threshold to increase the strictness of the matching, thereby reducing erroneous mapping relationships. This dynamic adjustment mechanism can automatically optimize the mapping accuracy based on continuous changes in business and the characteristics of different data tables.

[0144] In one possible implementation, a table-level feature vector is first generated by extracting features from the data table fields. This step extracts key features of the data table through the following dimensions:

[0145] Statistical analysis of the distribution of different field types in a data table, including numeric, character, and date types, helps to gain a preliminary understanding of the table's basic structure. Different field types often correspond to different business scenarios and analytical methods; therefore, statistical analysis of the distribution of these field types can aid in the initial understanding of the table's fundamental structure.

[0146] Extract the number of associations between fields in the data table and entities in the knowledge graph. This feature reflects the strength of the relationship between fields and entities at the business semantic level, thus demonstrating the importance of fields in different business subject domains.

[0147] These features are extracted to form a table-level feature vector, which serves as input data for subsequent classification models to train and infer.

[0148] The generated feature vectors are input into the classification model, which is then trained in the following way:

[0149] The training set uses the topic domain labels of historical data tables, which have already been manually labeled to determine their respective topic domains. This allows the model to learn how to predict the topic domains of a table based on its feature vectors.

[0150] To train the model, the Gradient Boosting Decision Tree (GBDT) algorithm was employed. This algorithm improves classification accuracy by constructing multiple weak classifiers and combining their predictions. GBDT is suitable for handling problems with complex feature spaces and nonlinear data relationships, thus effectively enhancing the model's predictive power.

[0151] During the model inference phase, the classification model outputs the probability that a data table belongs to each topic domain. Each data table will obtain a probability distribution for different topic domains, representing its likelihood of belonging to each topic domain. Then, by selecting the topic domain with the highest probability, the model determines the final classification result for the data table.

[0152] When a new topic domain is added, the classification model needs to be updated to maintain its timeliness and accuracy.

[0153] The update steps are as follows:

[0154] First, data tables related to the newly added topic domains are collected and manually labeled to determine their topic domain labels. Next, feature vectors are extracted from these new data tables. An incremental learning algorithm is used to update the model, specifically by updating only the weights of the last fully connected layer of the classification model. Incremental learning ensures that the introduction of new data does not disrupt the stability of the original model, and this method allows the model to make accurate predictions about the newly added topic domains without retraining the entire model.

[0155] In one possible implementation, when parsing user requests, it is first necessary to break down the user's natural language request statement and extract the key information from it, which involves the following steps:

[0156] The computational object is typically a core noun phrase in the requirement statement, representing the business element that needs to be calculated. Examples include "sales revenue" and "number of employees." Dependency parsing techniques are used to identify the subject or object of the statement—the core noun phrase—to extract the computational object. The core of this step is understanding which entity in the statement is the subject of the calculation.

[0157] Indicator types typically include calculation rules such as "mean," "sum," and "growth rate," which define how to perform numerical calculations on the object being calculated. By using keyword matching, specific keywords in the user's request statement, such as "mean," "sum," and "growth rate," are identified to determine the type of indicator to be calculated. At this point, the algorithm needs to accurately identify the calculation rules related to the object being calculated to ensure the accuracy of the calculation logic.

[0158] Time conditions refer to the time range or points in time involved in user requirements, usually expressed as "last year," "the past three months," etc. Regular expressions are used to extract the time range descriptions from the statements and convert them into a standard time format. Regular expressions can effectively identify different time format expressions and standardize them for subsequent processing.

[0159] In some cases, user requests may be ambiguous or unclear. For example, does "sum" refer to the sum over a specific time period or a long-term cumulative sum? To handle these ambiguities, the system generates multiple candidate parsing results and calculates the confidence level of each result.

[0160] By using a natural language understanding model, the system can generate multiple possible parsing results, each representing a different understanding of the same requirement statement. For example, the statement "total sales over the past three months" might be parsed into two different time ranges: "the past three months" or "the past 12 months".

[0161] For each candidate parsing result, the system calculates its confidence score, reflecting how well the parsing result matches the user's needs. Typically, the confidence score is calculated based on the model's depth of understanding of the statement, the keyword matching degree, and the reasonableness of the context.

[0162] When a requirement statement is ambiguous, the system will provide the user with the top N parsing results with the highest confidence levels. The user can then select the most suitable parsing result based on their actual needs, ultimately determining the computational logic for the requirement.

[0163] The system displays multiple candidate resolution results to the user through a user interface or interactive method, and labels the confidence level of each resolution result. Users can clarify ambiguities by selecting the appropriate resolution result.

[0164] The system determines the final indicator calculation logic based on the user's selection. If the user selects a certain analysis result, the system will use that result to generate the corresponding calculation model.

[0165] In one possible implementation, to optimize query execution logic, the system first needs to collect historical SQL execution logs, which record various important performance metrics during execution. By analyzing historical logs, the system can extract the following parameters:

[0166] Execution time: Records the execution time of SQL queries, used to measure the overall performance of the query.

[0167] Central Processing Unit (CPU) utilization: This indicates the level of CPU resource consumption during query execution, reflecting the query's demand for computing resources.

[0168] Memory consumption: This indicates the amount of memory used during query execution, reflecting the memory usage efficiency of the query.

[0169] The ratio of result data volume to source data volume: By analyzing the relationship between the data volume of the query results and the source data, we can understand the complexity of data processing in the query.

[0170] The collected parameters provide important input data for subsequent optimization models, helping the system identify which queries have performance bottlenecks and then optimize them accordingly.

[0171] By analyzing the collected historical execution logs, the system constructs the environment state of the reinforcement learning model based on the above parameters. Environment state refers to the resource consumption and performance characteristics faced by the system when executing a query under given conditions. For example, queries with long execution times, high CPU usage, large memory consumption, and an inappropriate ratio of result data to source data will be considered targets for optimization.

[0172] The construction of the environment status includes a comprehensive assessment of the resource consumption of different query executions. Each parameter can be assigned a certain weight in a weighted manner to comprehensively describe the execution status of the query.

[0173] In reinforcement learning, optimization actions refer to the adjustments a system makes under specific environmental conditions to optimize query execution efficiency and resource consumption. This method defines the following optimization actions:

[0174] In multi-table join queries, the system analyzes the impact of different join orders on query performance and prioritizes joins with strict filtering conditions. This reduces the amount of intermediate data and thus improves query efficiency.

[0175] For frequently queried fields, the system can identify hot fields based on historical execution logs and automatically add temporary indexes. Temporary indexes can accelerate retrieval of these fields, reduce data scanning time, and significantly improve query efficiency.

[0176] For complex SQL queries, the system can optimize query efficiency by breaking them down into multiple subqueries and executing them in parallel. Parallel processing can make full use of multi-core CPUs and improve query execution speed.

[0177] Q-learning is an algorithm in reinforcement learning that learns and optimizes decision-making strategies through interaction between the agent and its environment. Specifically:

[0178] The agent continuously adjusts its optimization strategy by interacting with the database to execute queries. The system provides feedback on the optimized execution results of each query, and adjusts the optimization strategy based on this feedback.

[0179] The reward function is used to evaluate the effectiveness of optimization actions. It is defined as the weighted sum of the percentage increase in execution efficiency and the percentage decrease in resource consumption. By calculating the reduction in query execution time and the reduction in resource consumption (CPU, memory), corresponding rewards are given to incentivize the system to take optimization actions that improve performance and reduce resource consumption.

[0180] In one possible implementation, during the indicator calculation process, the system needs to record some important metadata, which forms the basis of lineage analysis, and mainly includes:

[0181] Record detailed information about the source data tables that each metric calculation depends on. This includes the table name, version number, and key fields. This information helps track changes to the data source and allows for the timely identification of potential dependency issues.

[0182] Record all logical details related to the indicator calculation, including the filtering conditions used, aggregate functions, and relationships between tables. This information helps in understanding how each part of the calculation process functions and how they depend on each other.

[0183] By recording this metadata, the system can accurately understand the source, calculation logic, and data dependencies of each metric, constructing a detailed data lineage diagram. This is a prerequisite for subsequent updates and adjustments to the metrics.

[0184] A crucial aspect of lineage analysis is the ability to promptly identify changes in data sources, particularly updates to data table version numbers. The system triggers updates in the following ways:

[0185] When the version number of the input data table changes, the system automatically detects this change and assesses whether the change affects the indicator calculation. If the changed field is a field that the current indicator calculation depends on, the system will mark the indicator as "needs recalculation" to ensure the accuracy of the indicator data.

[0186] This monitoring mechanism ensures that relevant indicators can be updated in a timely manner when the data source changes, thus avoiding errors in calculation results caused by data source updates.

[0187] When upstream data tables change, the system generates a queue of metric update tasks based on lineage. These tasks represent the metrics that need to be updated, and their execution order and priority determine the efficiency of the metric updates.

[0188] Each metric's calculation relies on a series of data tables and fields. Based on lineage analysis, the system generates a task queue representing the metrics that need to be recalculated. The task queue is generated strictly according to the dependencies between metrics, ensuring that each metric is recalculated promptly after its upstream data is updated.

[0189] To ensure efficient and timely updates of metrics, the system schedules task execution according to certain priority rules. The priority is determined as follows:

[0190] The system tracks the number of times each metric is referenced in downstream reports. A metric with more references indicates a greater impact on downstream business operations, and therefore a higher priority. This ensures that key metrics are updated first, guaranteeing that business departments' data needs are met preferentially.

[0191] Based on the needs of different business units and their Service Level Agreements (SLAs), the system can flexibly adjust the priority of metric updates. For example, some departments may have higher requirements for data timeliness and need to update the metrics they rely on quickly, while other departments can wait a little longer. This ability to dynamically adjust priorities ensures that the metric update process can meet the needs of different business units.

[0192] In one possible implementation, the setting of the anomaly alarm threshold for this method primarily relies on the analysis of historical indicator data. To accurately detect anomalies in a dynamic environment, the historical fluctuations of the indicators need to be considered; therefore:

[0193] By setting a sliding window, the mean and standard deviation of the indicator within the current time period are calculated. The size of this window is dynamically adjusted according to the business cycle to better adapt to different business scenarios. A larger window reflects data fluctuations over a longer period, while a smaller window can capture short-term fluctuations more promptly. In this way, a more stable benchmark level can be obtained, reducing the impact of short-term abnormal fluctuations on the judgment results.

[0194] This sliding window-based calculation method helps to dynamically adjust the abnormal alarm threshold and avoid misjudgments caused by periodic changes or external disturbances.

[0195] Next, the system defines the rules for determining anomalies based on the calculated mean and standard deviation:

[0196] When the current value of an indicator exceeds the range of the sliding window mean ± K times the standard deviation, the indicator is considered to have an anomaly. This method can capture significant fluctuations and help to promptly identify deviations from the norm.

[0197] The selection of the K value is crucial to the effectiveness of anomaly detection. If the K value is too small, it may lead to missed detections; if the K value is too large, it may lead to false positives. Therefore, the K value needs to be precisely adjusted based on historical data.

[0198] By analyzing the fluctuation multiples of indicators from historical anomalies, the system can take the minimum multiple of all historical fluctuations as the initial K value. This initial value helps the system set a reasonable benchmark threshold, avoiding overly strict or overly lenient judgment rules.

[0199] By analyzing the balance between false positive and false negative rates, the system uses the Receiver Operating Characteristic (ROC) curve to determine the optimal K value. The ROC curve is a tool for measuring classifier performance, helping to determine the balance between accuracy and recall at different thresholds. By optimizing the K value, the system can achieve the best anomaly detection performance, reduce unnecessary alarms, and ensure that genuine anomalies are detected in a timely manner.

[0200] This K-value optimization method can dynamically adjust the alarm threshold, ensuring reliable anomaly detection capabilities in different business scenarios.

[0201] To ensure timely response to abnormal events, the system has an alarm level escalation mechanism:

[0202] When the same metric triggers multiple consecutive alarms, the system will automatically escalate the alarm level. Consecutive alarms usually indicate a more serious or persistent problem, thus requiring greater attention.

[0203] When an alarm level is escalated, the system will automatically notify the designated responsible person so that timely action can be taken to avoid greater impact due to delays. Notification to the responsible person can be delivered through various means (such as SMS, email, or in-system alerts) to ensure timely and effective information delivery.

[0204] Accordingly, embodiments of the present invention also provide an artificial intelligence-based indicator layer model construction and analysis system for running the artificial intelligence-based indicator layer model construction and analysis method described in any one of the embodiments of the present invention, comprising the following modules:

[0205] Multi-source data synchronization and quality assessment module: includes a data synchronization unit and a quality assessment unit;

[0206] The data synchronization unit is connected to multiple heterogeneous databases and is used to synchronize all data to the distributed computing platform.

[0207] The quality assessment unit is connected to the data synchronization unit, receives synchronized data, performs anomaly detection and quality scoring, and outputs a cleaned data table and a quality scoring report.

[0208] The business knowledge graph construction module includes a business document parsing unit and a semantic mapping unit;

[0209] The business document parsing unit is connected to an external business document library and extracts business entities and relationships through a natural language processing model to generate an initial knowledge graph.

[0210] The semantic mapping unit is connected to the quality assessment module, receives the cleaned data table field information, associates the fields with knowledge graph entities through word vector similarity calculation, and outputs the bound knowledge graph.

[0211] The topic domain intelligent segmentation module includes a feature extraction unit and a classification model unit;

[0212] The feature extraction unit is connected to the business knowledge graph construction module to extract the field type distribution and semantic relevance features of the data table and generate table-level feature vectors.

[0213] The classification model unit is connected to the feature extraction unit, and outputs the topic domain classification results of the data table based on the feature vector and historical topic domain label data.

[0214] Indicator requirement analysis and calculation logic generation module: including natural language understanding unit and logic optimization unit;

[0215] The natural language understanding unit receives the indicator requirement text input by the user, and outputs the calculation object, indicator type and time conditions through dependency parsing and keyword matching.

[0216] The logic optimization unit is connected to the business knowledge graph construction module and the topic domain intelligent segmentation module. It generates initial calculation logic based on the matched fields and template library, optimizes the execution path through a reinforcement learning model, and outputs an optimized structured query language script.

[0217] Indicator data management and verification module: includes calculation execution unit, lineage tracing unit and anomaly alarm unit;

[0218] The computation execution unit is connected to the logic optimization unit and executes a structured query language script to generate an indicator data table.

[0219] The lineage tracing unit is connected to the calculation execution unit, records the upstream tables and fields on which the indicator data depends, and links with the data change monitoring unit to trigger dynamic update tasks;

[0220] The abnormal alarm unit is connected to the bloodline tracking unit, and sets a dynamic threshold based on the historical index fluctuation range to generate alarm notifications for abnormal data.

[0221] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0222] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing and analyzing indicator-level models based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Multi-source data synchronization and quality assessment. Synchronize the full data from multiple heterogeneous databases to the distributed computing platform, assess the quality of the synchronized data, and generate cleaned data tables and quality scoring reports. The quality assessment includes: statistical analysis of the numerical distribution of each field, constructing an anomaly detection model based on historical data distribution, and identifying abnormal data points; the anomaly detection model is constructed in the following way: randomly sampling field values ​​to generate multiple isolated trees, each tree recursively dividing the data space, calculating the anomaly probability based on the path length of the data point in the isolated tree, with shorter path lengths resulting in higher anomaly probabilities; the dynamic threshold is determined by comparing the anomaly probability with a dynamic threshold: statistically analyzing the path length distribution of normal data points in historical data, taking the P-th percentile of the distribution as the threshold benchmark, and adjusting the threshold benchmark according to the risk level of the current business scenario, reducing the P-value in high-risk scenarios to tighten the threshold; The generation of the quality score report includes: defining field integrity scoring rules, calculating the field missing rate, with full marks awarded for a missing rate below a first threshold, and deducting a preset score for every 1% exceeding the first threshold; defining uniqueness scoring rules, calculating the proportion of duplicate records in the primary key field, with full marks awarded for a duplicate proportion below a second threshold, and deducting points proportionally for proportions exceeding the second threshold; defining consistency scoring rules, comparing the foreign key matching rate in the associated table, and triggering a data consistency alarm when the matching rate is below a third threshold; the first, second, and third thresholds are determined as follows: analyzing the distribution of field missing rate, primary key duplication rate, and foreign key matching rate in historical data, taking the highest tolerance value under normal business scenarios as the initial threshold, and dynamically adjusting the thresholds based on user feedback; if a user marks an alarm as a false alarm, the corresponding threshold is relaxed by a preset percentage. Step 2: Business context analysis and intelligent subject domain division. Parse the semantics of business documents and data table fields, construct a knowledge graph containing business entity relationships, and divide the data tables into preset subject domains based on the knowledge graph; Step 3: Intelligent parsing of indicator requirements and generation of calculation logic. Natural language processing technology is used to parse the indicator requirements input by the user, match them with business entities in the knowledge graph, generate the initial calculation logic, and optimize the execution efficiency. The indicator requirement parsing includes: splitting user requirement statements using a pre-trained natural language understanding model to identify the following elements: the calculation object, extracting core noun phrases from the statement through dependency parsing; the indicator type, identifying calculation rules for "mean," "sum," and "growth rate" through keyword matching; and the time condition, extracting the time range description through regular expressions and converting it into a standard time format. When the requirement statement is ambiguous, it is clarified in the following ways: generating multiple candidate parsing results, calculating the confidence level of each result, providing feedback to the user with the top N parsing results with the highest confidence levels, and receiving the user's selection to determine the final logic. Step 4: Dynamic update and consistency verification of indicator data. Execute the optimized calculation logic to generate indicator data, track data lineage and trigger dynamic updates, and set thresholds based on historical fluctuation ranges to issue anomaly alerts. Lineage analysis includes recording the following metadata during metric calculation: the name, version, and key fields of the input data table; the filtering conditions, aggregate functions, and relationships used in the calculation logic; when an upstream data table changes, an update is triggered in the following ways: monitoring changes in the data table version number; if the version number is updated and the affected fields include metric dependency fields, it is marked as needing recalculation; a metric update task queue is generated based on lineage, and execution is scheduled according to priority; the priority is determined by: counting the number of times the metric is referenced in downstream reports; the more times it is referenced, the higher the priority; and dynamically adjusting the priority weight according to the service level agreement requirements set by the business department. The setting of the abnormal alarm threshold includes: calculating the mean and standard deviation of the sliding window based on historical indicator data, with the window size dynamically adjusted according to the business cycle; defining anomaly judgment rules: if the current indicator value exceeds the mean ± K times the standard deviation, it is judged as abnormal; the ± K value is determined in the following way: analyzing the indicator fluctuation multiple of historical abnormal events, taking the minimum fluctuation multiple as the initial K value, and determining the optimal K value based on the balance between false alarm rate and false negative rate using the receiver operation characteristic curve; when the same indicator triggers alarms continuously, the alarm level is automatically upgraded and the designated responsible person is notified.

2. The method for constructing and analyzing an indicator layer model based on artificial intelligence according to claim 1, characterized in that, The construction of the knowledge graph includes: The business documents are parsed using a pre-trained natural language processing model to extract entity names and relationships between entities. The entities include business objects, indicator types, and calculation rules. The data table fields are semantically mapped to entities. The mapping process includes: Perform word segmentation on the field names to generate word vectors; Calculate the cosine similarity between word vectors and entity names in the knowledge graph. If the similarity exceeds the mapping threshold, establish the association between the field and the entity. The mapping threshold is determined in the following way: Based on historical manually annotated field-entity mapping data, the minimum similarity value of successful matches is calculated and used as the initial threshold; The threshold is dynamically adjusted based on the mapping accuracy. If the mapping error rate of a newly added data table increases, the threshold is raised to enhance strictness.

3. The method for constructing and analyzing an indicator layer model based on artificial intelligence according to claim 1, characterized in that, The subject area division includes: Feature extraction is performed on the data table fields to generate table-level feature vectors, specifically including: The distribution ratio of field types in the statistical table, including numeric, character, and date types; Extract the number of associations between fields in the table and entities in the knowledge graph, and use this as a semantic relevance feature; The feature vectors are input into the classification model, which is trained in the following way: The topic domain labels from historical data tables were used as the training set, and the gradient boosting decision tree algorithm was used to train the model. During the model inference phase, the probability of the output data table belonging to each topic domain is calculated, and the topic domain with the highest probability is selected as the classification result. When a new topic domain is added, the model is updated in the following way: Collect sample data tables of newly added topic domains, extract their feature vectors, and manually label them; An incremental learning algorithm is used to update only the weights of the last fully connected layer of the classification model.

4. The method for constructing and analyzing an indicator layer model based on artificial intelligence according to claim 1, characterized in that, The computational logic optimization includes: Collect historical structured query language execution logs and extract the following parameters: Execution time, CPU utilization, and memory consumption; The ratio of the resulting data volume to the source data volume; Based on the above parameters, construct the environment state of the reinforcement learning model, and define the following optimization actions: Adjust the table join order, prioritizing table joins with strict filtering conditions; Add temporary indexes to speed up the retrieval of frequently queried fields; Break down complex queries into multiple subqueries and execute them in parallel; The model is trained using the Q-learning algorithm, and the reward function is defined as the weighted sum of the percentage increase in execution efficiency and the percentage decrease in resource consumption.

5. An artificial intelligence-based indicator layer model construction and analysis system, used to run the artificial intelligence-based indicator layer model construction and analysis method according to any one of claims 1-4, characterized in that, Includes the following modules: Multi-source data synchronization and quality assessment module: includes a data synchronization unit and a quality assessment unit; The data synchronization unit is connected to multiple heterogeneous databases and is used to synchronize all data to the distributed computing platform. The quality assessment unit is connected to the data synchronization unit, receives synchronized data, performs anomaly detection and quality scoring, and outputs a cleaned data table and a quality scoring report. The business knowledge graph construction module includes a business document parsing unit and a semantic mapping unit; The business document parsing unit is connected to an external business document library and extracts business entities and relationships through a natural language processing model to generate an initial knowledge graph. The semantic mapping unit is connected to the quality assessment module, receives the cleaned data table field information, associates the fields with knowledge graph entities through word vector similarity calculation, and outputs the bound knowledge graph. Intelligent Subject Domain Partitioning Module: Includes feature extraction units and classification model units; The feature extraction unit is connected to the business knowledge graph construction module to extract the field type distribution and semantic relevance features of the data table and generate table-level feature vectors. The classification model unit is connected to the feature extraction unit, and outputs the topic domain classification results of the data table based on the feature vector and historical topic domain label data. Indicator requirement analysis and calculation logic generation module: including natural language understanding unit and logic optimization unit; The natural language understanding unit receives the indicator requirement text input by the user, and outputs the calculation object, indicator type and time conditions through dependency parsing and keyword matching. The logic optimization unit is connected to the business knowledge graph construction module and the topic domain intelligent segmentation module. It generates initial calculation logic based on the matched fields and template library, optimizes the execution path through a reinforcement learning model, and outputs an optimized structured query language script. Indicator data management and verification module: includes calculation execution unit, lineage tracing unit and anomaly alarm unit; The computation execution unit is connected to the logic optimization unit and executes a structured query language script to generate an indicator data table. The lineage tracing unit is connected to the calculation execution unit, records the upstream tables and fields on which the indicator data depends, and links with the data change monitoring unit to trigger dynamic update tasks; The abnormal alarm unit is connected to the bloodline tracking unit, and sets a dynamic threshold based on the historical index fluctuation range to generate alarm notifications for abnormal data.

Images