Public accumulation fund management system based on multivariate heterogeneous data fusion and control method thereof

Abstract

The invention discloses an accumulation fund management system based on multivariate heterogeneous data fusion and a control method thereof, and relates to the field of accumulation fund management systems, and the system comprises a data access layer which is used for receiving various heterogeneous data sources of a bank end, an enterprise end and an accumulation fund management center end; the data fusion layer is used for performing data cleaning operation, data matching operation and data model construction operation on the heterogeneous data sources; the business processing layer is used for processing the accumulation fund core business; the data analysis layer and the business analysis layer are used for performing data analysis operation based on database storage data and user input data to generate an evaluation result; the security management layer is used for performing authority control operation on the access user, and the security management layer is also used for performing desensitization and encryption operation on the sensitive information; and the user service layer is used for providing mobile phone APP service, official account service and Web portal service.

Description

Technical Field

[0001] The present specification relates to the field of provident fund management systems, and more specifically, the present application relates to a provident fund management system based on multi-heterogeneous data fusion and a control method thereof. Background Art

[0002] Provident Fund business involves banks, enterprises and provident fund management centers. The data types from different sources are diverse. Due to the scattered data formats and sources, the system lacks an efficient data integration mechanism, leading to data island phenomenon.

[0003] Provident Fund policies such as loan amounts and contribution ratios often change. The traditional system has a long business logic adjustment cycle and lacks a dynamic rule engine, making it difficult to respond to policy adjustments quickly, reducing the system's adaptability and efficiency.

[0004] Therefore, it is necessary to propose a provident fund management system and its control method based on multi-heterogeneous data fusion to at least solve some of the above problems and to improve the capabilities of data integration, business processing and user experience. Summary of the invention

[0005] A series of simplified concepts are introduced in the Summary of the Invention section, which will be further described in detail in the Detailed Description of the Invention section. The Summary of the Invention section of this application does not mean to attempt to define the key features and essential technical features of the claimed technical solution, nor does it mean to attempt to determine the scope of protection of the claimed technical solution.

[0006] In the first aspect, the present application proposes a provident fund management system based on multi-heterogeneous data fusion, the system comprising: Data access layer, the data access layer is used to receive various heterogeneous data sources from banks, enterprises and provident fund management centers, the various heterogeneous data sources include structured data, unstructured data, semi-structured data and mixed data; A data fusion layer, wherein the data fusion layer is used to perform data cleaning operations, data matching operations, and data model building operations on the heterogeneous data sources; A business processing layer, which is used to process the core business of the provident fund. The business processing layer includes a dynamic rule engine, which adjusts its corresponding business logic in real time based on policy changes; Data analysis layer: the business analysis layer is used to generate evaluation results by performing data analysis operations based on database stored data and user input data; The security management layer is used to control the access rights of users and to desensitize and encrypt sensitive information. User service layer: The user service layer is used to provide mobile APP services, public account services and web portal services.

[0007] In a second aspect, the present application proposes a control method for the provident fund management system based on multi-heterogeneous data fusion in the first aspect, comprising: Control the data access layer to receive multiple heterogeneous data sources from the bank side, the enterprise side and the provident fund management center side, wherein the multiple heterogeneous data sources include structured data, unstructured data and semi-structured data; The data fusion layer is controlled to perform data cleaning operations, data matching operations and data model building operations on the heterogeneous data sources to obtain provident fund fusion data.

[0008] In a feasible implementation, the data fusion layer includes a data cleaning module, a data matching module and a data model building module; The above-mentioned control of the above-mentioned data fusion layer performs data cleaning operations, data matching operations and data model building operations on the above-mentioned heterogeneous data sources to obtain provident fund combined data, including: Controlling the data cleaning module to perform data cleaning operations on the heterogeneous data to obtain cleaned heterogeneous data; Control the data matching module to perform a data matching operation on the cleaned heterogeneous data based on the intelligent matching model to obtain matched data; The data model building module is controlled to perform a data model building operation on the matched data to obtain provident fund aggregate data, wherein the data model building module includes an aggregation model, a prediction model and a classification model.

[0009] In a feasible implementation manner, the above intelligent matching model includes a feature extraction unit, a similarity calculation unit, a matching decision unit and a data output unit; The feature extraction unit is used to extract different data features corresponding to different data types based on different feature extraction methods; The similarity calculation unit is used to calculate the comprehensive similarity based on the similarity of each feature type and the weight coefficient corresponding to the similarity of each feature type; The matching decision unit is used to obtain data matching results based on comprehensive similarity; The data output unit is used to generate matched data based on the data matching result.

[0010] In a feasible implementation manner, the feature extraction unit performs feature extraction based on the first feature extraction method, the second feature extraction method, the third feature extraction method and the fourth feature extraction method; The first feature extraction method is used to extract the first feature of the structured data, the second feature extraction method is used to extract the second feature of the unstructured data, the third feature extraction method is used to extract the third feature of the semi-structured data, and the fourth feature extraction method is used to extract the fourth feature of the mixed data; The first feature extraction method is implemented by performing a time series feature extraction operation based on the feature combination of important fields, the second feature extraction method is implemented based on the feature extraction methods corresponding to different unstructured data types, the third feature extraction method is implemented based on field parsing operations, nested structure parsing operations and statistical feature extraction operations, and the fourth feature extraction method is implemented based on hierarchical extraction operations and multi-modal feature learning operations.

[0011] In a feasible implementation, the above comprehensive similarity Determined based on the following formula:

[0012] in, is the first feature similarity, is the weight coefficient corresponding to the first feature similarity above, is the second feature similarity, is the weight coefficient corresponding to the second feature similarity, is the third feature similarity, is the weight coefficient corresponding to the third feature similarity mentioned above, is the fourth feature similarity, is the weight coefficient corresponding to the fourth feature similarity; When the first feature is a significant time series feature, the similarity of the first feature is determined based on the following formula: :

[0013] In the formula, are the first eigenvectors of two different data, Indicates the calculation result of dynamic time warping; When the first feature is a non-significant time series feature, the similarity of the first feature is determined based on the following formula: :

[0014] In the formula, and is the first eigenvector The corresponding i-th component; In the case where the second feature is a feature corresponding to text data, the second feature similarity is determined based on the cosine similarity of the vectors corresponding to the word frequency and the inverse document frequency. ; In the case where the second feature is a feature corresponding to the image data, the second feature similarity is determined based on the convolution feature similarity. ; The third feature similarity is calculated based on the following formula :

[0015] in, To summarize field similarity, is the nested field similarity, To calculate the similarity of comprehensive field statistics, is the first weight coefficient, is the second weight coefficient, is the third weight coefficient; Similarity of the above summary fields Based on the following calculation:

[0016] in, For each field The corresponding basic similarity, Indicates the total number of fields; The above basic similarity Based on the following calculation:

[0017] In the formula, For different data source fields The value of is the text field similarity; Similarity of the above nested fields Based on the following calculation:

[0018] is the total number of nested fields, For the The weight of the nested fields; The similarity of the above comprehensive field statistical features Based on the following calculation:

[0019] is the statistical feature vector of the semi-structured data in the first record, is the statistical feature vector of the semi-structured data in the second record; The similarity of the fourth feature Determined based on the following formula:

[0020] vector Determined by the following formula:

[0021] in, is the fourth eigenvector of the first record, is the fourth eigenvector of the second record, is the ensemble learning weight, represents the first eigenvector mentioned above, represents the second eigenvector mentioned above, represents the third eigenvector mentioned above.

[0022] In a feasible implementation manner, the specific steps performed by the matching decision unit include: Calculate dynamic thresholds based on feature priority and data type; A comparison operation is performed based on the above comprehensive similarity and the above dynamic threshold to obtain a data matching result.

[0023] In a feasible implementation, it also includes: Control the business processing layer to process the core business of the provident fund based on the dynamic rule engine to generate business feedback data, wherein the dynamic rule engine adjusts its corresponding business logic in real time based on policy changes; The data analysis layer is controlled to perform data analysis operations based on the database storage data and the user input data to generate evaluation results.

[0024] In a feasible implementation, the specific steps of the above-mentioned dynamic rule engine adjusting its corresponding business logic in real time based on policy changes include: Obtain policy data, where the policy data comes from API data interface and manually input data; Automatically parse the key clauses of policy documents based on natural technology to extract key elements, where the key clauses include the conditions, restrictions and execution details of the clauses, and the key elements include the scope of use, rule conditions and execution requirements; Perform matching operations based on the above key elements and rule templates to generate a new version file; Perform virtual data testing based on the above-mentioned new version file to obtain virtual test results; If the above virtual test results are qualified, the above new version file will be released and enabled based on the execution time corresponding to the above policy data.

[0025] In a feasible implementation manner, the controlling the data analysis layer to perform data analysis operations based on the database storage data and the user input data to generate an evaluation result includes: The above evaluation result R is determined based on the following formula:

[0026] In the formula, is the regularization term for the evaluation results, represents the total number of features, is the comprehensive scoring function of the evaluation results, is the feature interaction term, For the The weight coefficient of the feature, is the regularization strength coefficient, is the feature interaction term weight coefficient, For the first The feature values ​​include age, income and loan demand, Indicates the historical data stored in the above database storage data eigenvalues, for The corresponding weight coefficient is, for The corresponding weight coefficient is, is the penalty coefficient, is the kernel function of feature interaction, The first eigenvalues, The first eigenvalues, represents the square difference of the interaction between two features, is the smoothing parameter.

[0027] In summary, the provident fund management system based on the fusion of multiple heterogeneous data proposed in this application supports the access and integration of multiple heterogeneous data, eliminates data islands, and improves the overall utilization of business data. Through the dynamic rule engine, it can quickly adapt to policy changes and ensure the legitimacy and flexibility of the system's business logic. Through comprehensive permission control and data encryption, the privacy of user data and the security of the system are ensured. Provide multiple access methods (APP, official account, Web) to improve user convenience. It is suitable for government provident fund management agencies to integrate data resources, handle policy adjustments, and improve service efficiency. It is also suitable for the enterprise side to efficiently submit employee contribution information and communicate with banks and provident fund management centers. It is also suitable for individual users to check account balances, apply for withdrawals or loans anytime and anywhere.

[0028] The present application proposes a provident fund management system and a control method thereof based on multi-heterogeneous data fusion. Other advantages, objectives and features of the present application will be partially reflected through the following description, and will also be partially understood by technical personnel in the field through research and practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present specification. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 A structural schematic diagram of a provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application; Figure 2 A structural schematic diagram of another provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application; Figure 3 A flow chart of a control method corresponding to a provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application; Figure 4 A flow chart of specific processing steps of a matching decision unit provided in an embodiment of the present application; Figure 5 A flow chart of a control method corresponding to another provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application; Figure 6 A process diagram of specific steps for a dynamic rule engine to adjust its corresponding business logic in real time based on policy changes provided in an embodiment of the present application. DETAILED DESCRIPTION

[0030] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0031] See also Figure 1 and Figure 2 , which is a structural schematic diagram 10 of a provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application, which may specifically include: Data access layer 11, the data access layer is used to receive multiple heterogeneous data sources from the bank side, the enterprise side and the provident fund management center side, the multiple heterogeneous data sources include structured data, unstructured data, semi-structured data and mixed data; The data fusion layer 12 is used to perform data cleaning operations, data matching operations and data model building operations on the heterogeneous data sources; The business processing layer 13 is used to process the core business of the provident fund. The business processing layer includes a dynamic rule engine. The dynamic rule engine adjusts its corresponding business logic in real time based on policy changes. Data analysis layer 14, the business analysis layer is used to perform data analysis operations based on database storage data and user input data to generate evaluation results; The security management layer 15 is used to control the access rights of users and to desensitize and encrypt sensitive information. The user service layer 16 is used to provide mobile APP services, public account services and Web portal services.

[0032] Exemplarily, the data access layer 11 is responsible for receiving a variety of heterogeneous data sources from the bank side, the enterprise side and the provident fund management center side, and providing data input support for the entire system. The heterogeneous data types supported by this system include but are not limited to structured data, unstructured data, semi-structured data and mixed data. Structured data includes database table data, and structured data has fixed fields and formats (such as transaction records, basic user information, etc.). Unstructured data includes text files, images, audio, etc. (such as contract scans, policy documents, etc.). Semi-structured data includes JSON and XML, and the format of semi-structured data fields is not fixed (such as dynamic data returned by API). Mixed data includes a combination of multiple data types (such as a record that contains both user information and image attachments). The data access layer 11 receives and uniformly manages data input from different sources to ensure the integrity and accuracy of data flow to subsequent modules.

[0033] The data fusion layer 12 cleans, matches and models the accessed heterogeneous data to form a unified high-quality data resource. The data fusion layer includes data cleaning operations, data matching operations and data model building operations. Data cleaning operations are used to remove redundant data, fill missing values, and correct erroneous values. Unify data formats (such as standardization of date and amount fields). Data matching operations are based on intelligent matching models to associate and match related data in different data sources. For example, match transaction records on the bank side with salary records on the enterprise side to the same user. Data model building operations are used to build a unified data model (such as user account model, deposit record model, etc.) that meets business needs, and provide standardized data interfaces that can be used by the business processing layer and the data analysis layer. The data fusion layer 22 eliminates data silos by fusing heterogeneous data sources and provides consistent data for business processing and analysis.

[0034] The business processing layer 13 handles the core business logic related to provident funds and is the business operation center of the system. The business processing layer 13 contains a dynamic rule engine that can adjust the business logic in real time according to policy changes. The dynamic rule engine dynamically adjusts business rules according to changes in national and local provident fund policies (such as loan amount adjustments, new deposit ratio requirements, etc.). For example: a certain policy adjusts the maximum amount of provident fund loans, and the rule engine updates the loan calculation logic in real time. The business processing layer supports a variety of business processes, such as provident fund withdrawals, loan approvals, account transfers, etc. The business processing layer 13 can ensure that the system responds quickly to policy changes and improve business processing efficiency and accuracy.

[0035] The data analysis layer 14 analyzes the provident fund data and generates evaluation results to support decision-making. The historical data stored in the system (such as user deposit records and loan records) is used for trend analysis. For example, analyzing a user's past repayment behavior to predict loan risks. The user's current input (such as application amount, repayment plan, etc.) is received and combined with historical data for real-time evaluation. For example, based on the user's income level and repayment ability, a loan feasibility evaluation result is generated. The data analysis layer 14 provides real-time analysis results (such as loan evaluation, repayment suggestions, etc.) to the user service layer, and also provides data support for the optimization rules of the business layer.

[0036] The security management layer 15 ensures the security of the system and data and prevents unauthorized access or data leakage. It can set different access rights according to user roles (such as ordinary users, administrators, and enterprise operators) to ensure that users can only view and operate authorized data. It can desensitize key data (such as user ID numbers and bank account information) (such as using "*" to cover part of the data). It uses data encryption technology (such as AES or RSA encryption) to store and transmit sensitive information. It ensures the compliance of the system and enhances users' trust in the system.

[0037] The user service layer 16 is an interactive layer for users, providing users with a convenient service entrance. Service types include but are not limited to mobile APP services, public account services and web portal services. Through the mobile APP service, users can check their account balances, apply for loans, submit withdrawal applications, etc. through the mobile APP. At the same time, the mobile APP can provide push notifications (such as deposit reminders and policy changes). The public account service can be based on the services of platforms such as WeChat, supporting lightweight functions such as quick query and message push. The web portal service provides comprehensive functional services suitable for enterprises and managers. For example, enterprises import deposit information in batches and administrators approve loans. The user service layer 16 provides multi-channel user access methods to meet the usage habits and scenario requirements of different users.

[0038] In summary, the provident fund management system based on the fusion of multiple heterogeneous data proposed in this application supports the access and integration of multiple heterogeneous data, eliminates data islands, and improves the overall utilization of business data. Through the dynamic rule engine, it can quickly adapt to policy changes and ensure the legitimacy and flexibility of the system's business logic. Through comprehensive permission control and data encryption, the privacy of user data and the security of the system are ensured. Provide multiple access methods (APP, official account, Web) to improve user convenience. It is suitable for government provident fund management agencies to integrate data resources, handle policy adjustments, and improve service efficiency. It is also suitable for the enterprise side to efficiently submit employee contribution information and communicate with banks and provident fund management centers. It is also suitable for individual users to check account balances, apply for withdrawals or loans anytime and anywhere.

[0039] In a second aspect, the present application proposes a control method for the provident fund management system based on multi-heterogeneous data fusion in the first aspect. Figure 3 A flow chart of a control method corresponding to a provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application, the method comprising: S110, controlling the data access layer to receive multiple heterogeneous data sources from a bank side, an enterprise side, and a provident fund management center side, wherein the multiple heterogeneous data sources include structured data, unstructured data, and semi-structured data; For example, the data access layer can provide comprehensive access to data related to the provident fund system, ensuring that the system can collect complete data from different sources.

[0040] Data sources include but are not limited to banks, enterprises and provident fund management centers. Banks can provide users’ bank transaction records, account information, repayment records, etc. Enterprises can provide employees’ salary records, deposit records, work information, etc. Provident fund management centers can provide users’ provident fund account status, deposit history, withdrawal records, etc.

[0041] The types of heterogeneous data sources include but are not limited to structured data, unstructured data, and semi-structured data. Structured data includes tabular data in the database, such as basic user information (name, ID number), salary records, repayment amounts, etc. Structured data has fixed fields, clear formats, and is easy to process.

[0042] Unstructured data includes text, images, audio, etc., such as scanned policy documents, user-uploaded extraction credentials, contracts, etc. Unstructured data has various formats and requires additional parsing and feature extraction.

[0043] Semi-structured data includes data in JSON and XML formats (data returned by APIs, dynamic data uploaded in batches by enterprises). Semi-structured data has a certain structure, but the fields are not fixed.

[0044] Access methods for various heterogeneous data sources include but are not limited to API interfaces, batch file import, and real-time stream processing.

[0045] The API interface achieves real-time data synchronization with the bank and enterprise ends through a preset data interface.

[0046] Batch file import is to batch access employee provident fund payment records, policy documents, etc. uploaded by the enterprise.

[0047] Real-time stream processing is for high-frequency transaction data, and supports real-time stream processing (such as Kafka and Flume).

[0048] S120, controlling the data fusion layer to perform data cleaning operations, data matching operations, and data model building operations on the heterogeneous data sources to obtain provident fund fusion data.

[0049] Exemplarily, the data fusion layer cleans, matches and models the accessed heterogeneous data to form a unified high-quality data resource. The data fusion layer includes data cleaning operations, data matching operations and data model building operations. Data cleaning operations are used to remove redundant data, fill missing values, and correct erroneous values. Unify data formats (such as standardization of date and amount fields). Data matching operations are based on intelligent matching models to associate and match related data in different data sources. For example, match transaction records on the bank side with salary records on the enterprise side to the same user. Data model building operations are used to build a unified data model (such as a user account model, a deposit record model, etc.) that meets business needs, and provide standardized data interfaces that can be used by the business processing layer and the data analysis layer. The data fusion layer 22 eliminates data silos by fusing heterogeneous data sources and provides consistent data for business processing and analysis.

[0050] In summary, this control method forms a high-quality provident fund data resource management process through the multi-access of the data access layer and the efficient processing of the data fusion layer. This method not only solves the problem of access and fusion of multi-source heterogeneous data, but also significantly improves the quality of data, the efficiency of business processing and the accuracy of decision support, providing strong support for the intelligence and automation of the provident fund management system.

[0051] In some examples, the data fusion layer includes a data cleaning module, a data matching module, and a data model building module; The above-mentioned control of the above-mentioned data fusion layer performs data cleaning operations, data matching operations and data model building operations on the above-mentioned heterogeneous data sources to obtain provident fund combined data, including: Controlling the data cleaning module to perform data cleaning operations on the heterogeneous data to obtain cleaned heterogeneous data; Control the data matching module to perform a data matching operation on the cleaned heterogeneous data based on the intelligent matching model to obtain matched data; The data model building module is controlled to perform a data model building operation on the matched data to obtain provident fund aggregate data, wherein the data model building module includes an aggregation model, a prediction model and a classification model.

[0052] Exemplarily, the data fusion layer is used to process diversified data from multiple heterogeneous data sources, which come from banks, enterprises and provident fund management centers, including structured, unstructured and semi-structured data. The data fusion layer includes three modules: data cleaning module, data matching module and data model building module, each module is responsible for different stages of data processing to achieve high-quality data fusion.

[0053] The data cleaning module is used to clean and standardize data from different sources to ensure data quality and consistency.

[0054] Specific steps may include: 1. Remove redundant data: Multiple data sources may contain duplicate data entries. For example, a user's basic information may exist in the data of banks, enterprises, and provident fund management centers at the same time. The cleaning module removes these redundant data to avoid repeated calculations in subsequent processing.

[0055] 2. Fill missing values: fill missing fields (such as part of a user's history). You can use average filling (for numerical data) or interpolation (for time series data) to handle missing values.

[0056] 3. Correcting erroneous values: Detect and correct abnormal values ​​and erroneous values ​​in the data. For example, a negative deposit amount is an illegal value and needs to be corrected to a reasonable positive number or marked as data that requires manual intervention.

[0057] 4. Unify data formats: Unify the formats of data from different sources, such as unifying the date format to "YYYY-MM-DD" and unifying the amount field to the same unit. This ensures data consistency during subsequent processing.

[0058] After the data cleaning operation of the data cleaning module, high-quality heterogeneous data can be obtained, providing standardized data input for subsequent matching and modeling operations.

[0059] The data matching module is used to associate and match cleaned data from different sources to form a complete user data view. Use unique identifiers (such as user ID number, provident fund account number) as primary keys to match data from different sources to the same user. For example, the transaction records on the bank side, the salary records on the enterprise side, and the payment history on the provident fund management center side are associated through the user's ID number. For fields that are not completely consistent, an intelligent matching model is used for matching. The model uses a variety of similarity calculation methods to improve the accuracy of matching, such as text similarity, numerical similarity, etc.

[0060] Multi-source data association is used to solve the problem of many-to-many matching. For example, if the same user has multiple deposit records in different companies, the matching module will associate the data by time sequence or priority to ensure the integrity and correctness of the data. After matching, the fused data is obtained, and all relevant information of the user is unified into the same complete data record.

[0061] By building a data model, we can provide structured, standardized, high-quality data for subsequent business processing and data analysis. The aggregation model summarizes and aggregates the matched data. For example, we build a user account model to summarize the user's deposit amount, withdrawal record, account balance and other information. These data are used for subsequent business operations and queries. The prediction model can predict the user's future deposit trend, loan repayment ability, etc. For example, based on the user's salary growth, deposit history and repayment record, predict whether he will have a high repayment ability in the future. The classification model can classify users and divide them into different categories (such as low-risk users, high-risk users) based on specific characteristics (such as income level, deposit frequency, repayment record, etc.). This classification information can be used in the dynamic rule engine to make business decisions, such as determining the loan amount and repayment plan.

[0062] The data fusion layer processes multi-source heterogeneous data through cleaning, matching, and model building operations, and ultimately generates consistent and high-quality fused data. Cleaning operations remove noise data and fill in missing information to make the data more complete and reliable. Through data matching, data from multiple sources are associated with the same user to form a complete data view for the user, eliminating the problem of data silos. Through data model building, the fused data is standardized and structured to provide efficient data support for subsequent business processing and data analysis.

[0063] For example, it can be applied to provident fund payment management. Through the matching module, the payment records on the enterprise side, the transaction information on the bank side and the account information of the provident fund management center are integrated to establish a unified user account view and support the automated processing of provident fund payment business. It can also be applied to provident fund loan approval. Through the prediction model and classification model, combined with the user's historical data, loan evaluation and risk classification are generated to support automated loan approval at the business processing layer. It can also be applied to policy response and business adjustment. Based on the integrated data, the dynamic rule engine can adjust the business logic in real time according to the evaluation results and classification provided by the model to meet policy changes or user needs.

[0064] This embodiment solves the problem of heterogeneous data fusion through data cleaning, matching and model building, ensures the integrity and consistency of data, provides a high-quality fusion data foundation for the provident fund management system, and enhances the system's intelligence, automation and business processing capabilities.

[0065] In some examples, the intelligent matching model includes a feature extraction unit, a similarity calculation unit, a matching decision unit, and a data output unit; The feature extraction unit is used to extract different data features corresponding to different data types based on different feature extraction methods; The similarity calculation unit is used to calculate the comprehensive similarity based on the similarity of each feature type and the weight coefficient corresponding to the similarity of each feature type; The matching decision unit is used to obtain data matching results based on comprehensive similarity; The data output unit is used to generate matched data based on the data matching result.

[0066] Exemplarily, features suitable for matching are extracted from different data types, and the data is converted into comparable feature representations. For multiple data types (structured data, unstructured data, semi-structured data), the characteristics of each data type are identified and classified. For example, structured data includes bank account information and corporate salary data, unstructured data includes uploaded pictures, contract texts, etc., and semi-structured data is dynamic data in JSON or XML format.

[0067] Different feature extraction methods are used for different data types. Structural data can extract numerical features, statistical features (such as salary amount, number of deposits, account balance) and time features (such as repayment period).

[0068] For unstructured data, natural language processing (NLP) technology can be used to extract keywords or semantic features in the text (such as TF-IDF, Word2Vec), and convolutional neural networks (CNN) can be used to extract visual features for image data.

[0069] Semi-structured data can parse nested fields and extract logical relationships or statistical features between fields. The extracted features are converted into vector form so that different types of data can be compared in a unified feature space. For example, text features can be converted into word vectors and numerical data can be normalized into feature vectors.

[0070] The similarity calculation unit is used to calculate the similarity between different data according to the extracted features.

[0071] The matching decision unit determines whether the data matches based on the comprehensive similarity and decides how to merge. If the comprehensive similarity of the user is greater than the set threshold T, the system determines that the bank and enterprise data belong to the same user and matches them.

[0072] Generate and output the matched data based on the matching results to ensure that the matched data can be further used by the system. For successfully matched data entries, merge information from different sources to generate a complete user data view. For example, unify the transaction records on the bank side, the salary data and provident fund deposit information on the enterprise side into the same user account. Data output includes necessary field cleaning and standardization of the merged data to ensure that it conforms to the unified data format of the system.

[0073] The merged matching data is stored in the system database for the business processing layer and the analysis layer to call. The matching results are output to the business caller to ensure that they can be used for further business processing, decision analysis, etc.

[0074] In some examples, the feature extraction unit performs feature extraction based on a first feature extraction method, a second feature extraction method, a third feature extraction method, and a fourth feature extraction method; The first feature extraction method is used to extract the first feature of the structured data, the second feature extraction method is used to extract the second feature of the unstructured data, the third feature extraction method is used to extract the third feature of the semi-structured data, and the fourth feature extraction method is used to extract the fourth feature of the mixed data; The first feature extraction method is implemented by performing a time series feature extraction operation based on the feature combination of important fields, the second feature extraction method is implemented based on the feature extraction methods corresponding to different unstructured data types, the third feature extraction method is implemented based on field parsing operations, nested structure parsing operations and statistical feature extraction operations, and the fourth feature extraction method is implemented based on hierarchical extraction operations and multi-modal feature learning operations.

[0075] Exemplarily, the first feature extraction method extracts the first feature based on structured data. Structured data usually comes from a table in a database, and contains data with fixed fields, such as user account information, salary records, deposit amounts, etc. For example: a user's monthly provident fund deposit records and salary payment records.

[0076] The feature extraction method is implemented by performing time series feature extraction operations based on feature combinations of important fields, specifically including important field extraction operations and time series feature extraction operations.

[0077] The important field extraction operation is to select fields from the table that have a significant impact on business decisions, such as salary amount, account balance, etc.

[0078] The time series feature extraction operation builds time series features based on these important fields and analyzes the trend of data changes over time. Time series features include but are not limited to trend features, cycle features, and statistical features. Trend features include the growth rate of account balances. Cycle features include, for example, the monthly deposit frequency. Statistical features include, for example, the average deposit amount and variance in the last six months.

[0079] The second feature extraction method extracts the second feature based on the unstructured data, thereby extracting semantics, keywords and other features from the unstructured data to effectively represent data such as images and texts. Unstructured data includes data in unfixed formats such as text, pictures, and audio, such as scanned copies of contracts, policy documents, and image certificates uploaded by users.

[0080] The second feature extraction method includes feature extraction of text data, feature extraction of image data, and extraction of audio data. Text data can use Natural Language Processing (NLP) tools such as TF-IDF (erm Frequency-Inverse Document Frequency). You can also use Word2Vec or BERT models to extract semantic features of text to represent important information in the text.

[0081] Image data can be used to extract visual features of images using Convolutional Neural Network (CNN). CNN can extract edge features, color histograms, etc. in images to ensure that image data can be effectively converted into feature vectors.

[0082] Audio data can be characterized by extracting the spectral characteristics of the audio signal or by extracting acoustic features (such as Mel-Frequency Cepstral Coefficients, MFCC).

[0083] The third feature extraction method extracts the third feature based on the semi-structured data, thereby performing field analysis, nested analysis and statistical feature extraction on the semi-structured data to extract deep-level information of the data.

[0084] Semi-structured data includes data in JSON and XML formats. The structure of this type of data is not fixed, but it has certain marking characteristics, such as data returned by APIs and employee information uploaded in batches by companies.

[0085] The third feature extraction method includes field parsing operation, nested structure parsing operation and statistical feature extraction. The field parsing operation extracts all available data attributes by parsing each field in the semi-structured data layer by layer. For example, parse each key-value pair from JSON. The nested structure parsing operation parses the nested part and extracts the relationship between the data. For example, extract nested fields such as user account information and salary history in a nested JSON. The statistical feature extraction operation is used to extract statistical information, such as the number of records contained in a field, the average value and maximum value of the field value, etc., to describe the overall characteristics of the data. For example, calculate the average salary and minimum salary of employees in a company.

[0086] The fourth feature extraction method extracts the fourth feature based on mixed data, performs hierarchical extraction and multimodal feature learning on mixed data containing multiple data types, and comprehensively extracts different types of data features.

[0087] Mixed data is a combination of multiple types of data, which may include structured data (such as tables), unstructured data (such as text), and semi-structured data (such as XML). For example, a provident fund application record may include the user's basic information, a text description of the application reason, and uploaded supporting materials.

[0088] The fourth feature extraction method includes a hierarchical extraction operation and a multimodal feature learning operation. The hierarchical extraction operation is used to divide the mixed data into multiple levels (such as structured parts and unstructured parts), and respectively apply the first, second, and third feature extraction methods mentioned above. Feature extraction is performed on the data at each level to ensure the effective expression of each part of the data. The multimodal feature learning operation can use a multimodal deep learning model (such as a multimodal neural network) to jointly learn different types of data and extract comprehensive features from the mixed data. For example, by jointly learning images and texts, multimodal features describing a user are extracted, and image and text features are mapped to the same feature space.

[0089] Through a variety of feature extraction methods, we ensure that different data types can be effectively represented and processed, enhancing the integrity and consistency of the data. Time series feature extraction of structured data provides strong support for trend analysis and prediction, such as the changing trend of user repayment ability. Feature extraction of unstructured data (such as text and images) allows more information to be included in the matching process, increasing the comprehensiveness and accuracy of matching. Hierarchical extraction and multimodal learning of mixed data enable heterogeneous data in complex business scenarios to be uniformly processed, forming an effective matching basis.

[0090] In some examples, the above comprehensive similarity Determined based on the following formula:

[0091] in, is the first feature similarity, is the weight coefficient corresponding to the first feature similarity above, is the second feature similarity, is the weight coefficient corresponding to the second feature similarity, is the third feature similarity, is the weight coefficient corresponding to the third feature similarity mentioned above, is the fourth feature similarity, is the weight coefficient corresponding to the fourth feature similarity; When the first feature is a significant time series feature, the similarity of the first feature is determined based on the following formula: :

[0092] In the formula, are the first eigenvectors of two different data, Indicates the calculation result of dynamic time warping; When the first feature is a non-significant time series feature, the similarity of the first feature is determined based on the following formula: :

[0093] In the formula, and is the first eigenvector The corresponding i-th component; In the case where the second feature is a feature corresponding to text data, the second feature similarity is determined based on the cosine similarity of the vectors corresponding to the word frequency and the inverse document frequency. ; In the case where the second feature is a feature corresponding to the image data, the second feature similarity is determined based on the convolution feature similarity. ; The third feature similarity is calculated based on the following formula :

[0094] in, To summarize field similarity, is the nested field similarity, To calculate the similarity of comprehensive field statistics, is the first weight coefficient, is the second weight coefficient, is the third weight coefficient; Similarity of the above summary fields Based on the following calculation:

[0095] in, For each field The corresponding basic similarity, Indicates the total number of fields; The above basic similarity Based on the following calculation:

[0096] In the formula, For different data source fields The value of is the text field similarity; Similarity of the above nested fields Based on the following calculation:

[0097] is the total number of nested fields, For the The weight of the nested fields; The similarity of the above comprehensive field statistical features Based on the following calculation:

[0098] is the statistical feature vector of the semi-structured data in the first record, is the statistical feature vector of the semi-structured data in the second record; The similarity of the fourth feature Determined based on the following formula:

[0099] vector Determined by the following formula:

[0100] in, is the fourth eigenvector of the first record, is the fourth eigenvector of the second record, is the ensemble learning weight, represents the first eigenvector mentioned above, represents the second eigenvector mentioned above, represents the third eigenvector mentioned above.

[0101] It should be noted that the difference between significant time series features and insignificant time series features lies mainly in the degree of influence of time series on data matching, that is, whether the time dependency or time trend in the data has an important impact on the matching and analysis results in a specific scenario. The distinction between the two can be based on the time dimension characteristics of the data, the changing trend of time information, and the degree of impact on the business.

[0102] Exemplarily, this embodiment provides comprehensive support for structured data, unstructured data, semi-structured data and mixed data, and covers the processing requirements of various heterogeneous data through different feature extraction methods and similarity calculation methods. Different similarity calculation methods are adopted for different features (such as whether the time series features are significant or not) to ensure the accuracy of the matching results. The Dynamic Time Warping (DTW) method is used to cope with the dynamic changes of time series data, and the text features are calculated by cosine similarity, which fully considers the characteristics of different types of features. By combining the similarities of each feature according to the weight, the final similarity is generated, which provides a flexible and accurate measurement standard for data matching. The mixed features are extracted using the multimodal feature learning method to ensure the fusion of data features in complex scenarios and improve the comprehensiveness of matching. Supports the analysis of nested structure data and complex fields, especially the detailed analysis of the relationship between fields in semi-structured data, which improves the accuracy of matching. Nested field similarity and statistical feature similarity calculation ensure high matching accuracy when processing multi-level data.

[0103] In some examples, Figure 4 A schematic diagram of the specific processing steps of a matching decision unit provided in an embodiment of the present application is provided, and the specific steps of the matching decision unit include: S210, calculating a dynamic threshold based on feature priority and data type; S220: Perform a comparison operation based on the comprehensive similarity and the dynamic threshold to obtain a data matching result.

[0104] For example, the goal of the matching decision unit is to determine whether data from different sources match by comprehensive similarity, and how to merge the matched data. The core is to calculate the dynamic threshold based on feature priority and data type, and then compare the comprehensive similarity with the dynamic threshold to determine whether the data matches.

[0105] According to the importance of each feature and the characteristics of the data type, a matching threshold suitable for the current scenario is dynamically generated to enhance the flexibility of matching decisions. Feature priority is defined based on business needs and the importance of data features, determining which features have a greater impact on the final matching results during the matching process. Different features may have different impacts on the matching results. For example, in provident fund matching, user ID and account number usually have a higher priority, while some text features may have a lower priority. Different data types (such as structured, unstructured, semi-structured, and mixed data) require different processing methods and thresholds in matching. For example, the similarity of structured data can highly reflect the accuracy of matching, and the threshold can also be set higher, while for text data, a lower matching standard may need to be set based on semantic similarity. The dynamic threshold is determined based on the priority and data type of each feature. If the features of a piece of data include user ID (high priority), account balance (medium priority), and text description (low priority), the dynamic threshold will rely more on the similarity of user ID, while the similarity of text description has a relatively small impact on the threshold.

[0106] The comprehensive similarity is compared with the dynamic threshold to determine whether there is a match. If the comprehensive similarity is higher than the dynamic threshold, it means that the data from different sources are similar enough to meet the matching conditions, and it is marked as a successful match.

[0107] Dynamic thresholds can be automatically adjusted as business environments or matching requirements change. For example, when business importance is high, the threshold can be appropriately increased to increase the strictness of matching; when more data needs to be covered, the threshold can be lowered to relax the matching criteria.

[0108] In some examples, Figure 5 A flow chart of a control method corresponding to another provident fund management system based on multi-heterogeneous data fusion provided in an embodiment of the present application, wherein the method further includes: S310, controlling the business processing layer to process the core business of the provident fund based on the dynamic rule engine to generate business feedback data, wherein the dynamic rule engine adjusts its corresponding business logic in real time based on policy changes; S320, controlling the data analysis layer to perform data analysis operations based on the database storage data and the user input data to generate evaluation results.

[0109] Exemplarily, a dynamic rule engine is used based on the business processing layer to respond to policy changes in real time, ensuring flexible adjustment of business logic to better meet policy requirements and user needs.

[0110] The dynamic rule engine is an intelligent business processing component that can dynamically update and adjust business logic according to policy changes to adapt to different business scenarios and needs. When national or local policies change, the rule engine can automatically adjust without large-scale modifications to the system. For example, when the maximum amount of provident fund loans is adjusted, the dynamic rule engine can automatically apply the new loan limit and modify the relevant business logic.

[0111] The dynamic rule engine is responsible for handling the core business of provident funds, including provident fund deposits, withdrawals, loan approval, etc. For example, when processing provident fund loan applications, the system will determine whether the loan conditions are met based on the applicant's income, deposit status, credit record and other information, combined with the business logic of the dynamic rule engine, and generate an approval result.

[0112] Business feedback data is the output of provident fund business processing, indicating the processing results. For example, in the loan approval process, business feedback data may include approval results (pass / fail), loan amount, repayment plan, etc. When the dynamic rule engine processes provident fund withdrawal applications, it will also generate corresponding feedback data, such as withdrawal amount, account remaining balance, etc.

[0113] The data analysis layer conducts in-depth analysis based on the stored data in the database and the data entered by the user, generates user evaluation results, and provides support for decision-making and business processing. The historical data stored in the database and the real-time data entered by the user are combined to form a complete user data set for further analysis.

[0114] Based on the user's income, deposit record, and credit history, the system generates loan assessment results, such as whether the loan is approved, the recommended loan amount, and the repayment period. Based on the user's historical behavior data and current input, it generates risk levels, such as low risk, medium risk, and high risk, to support business personnel in making further judgments. Based on the analysis results, it provides users with personalized suggestions, such as recommending the best loan plan, or reminding them to increase the provident fund deposit amount to improve the loan approval success rate.

[0115] In some examples, Figure 6 A schematic diagram of a process for a dynamic rule engine to adjust its corresponding business logic in real time based on policy changes provided in an embodiment of the present application. The specific steps of the dynamic rule engine to adjust its corresponding business logic in real time based on policy changes include: S410, obtaining policy data, wherein the policy data is derived from an API data interface and manually input data; S420, automatically parsing the key clause content of the policy document based on natural technology to extract key elements, wherein the key clause content includes the conditions of the clause, the restrictions of the clause and the execution details of the clause, and the key elements include the scope of use, rule conditions and execution requirements; S430, performing a matching operation based on the key elements and the rule template to generate a newly added version file; S440, performing a virtual data test based on the newly added version file to obtain a virtual test result; S450: When the virtual test result is qualified, the newly added version file is released and enabled based on the execution time corresponding to the policy data.

[0116] For example, data information on policy changes is collected to provide a basis for subsequent automated processing and rule updates. Policy data comes from API data interfaces and manually input data. Obtaining policy data from the government or industry's policy release system, the latest policy content and related changes can be obtained in real time through the API interface. For example, notifications on adjustments to provident fund loan policies can be obtained through the government's open API. Some policy updates may not be available through the automatic API interface, and policy content must be manually entered by business personnel, especially for regional differences or individual special policies.

[0117] Use natural language processing technology to automatically parse policy content and extract key information for subsequent rule adjustments. Automatically parse policy documents to extract key clauses, such as policy execution conditions, restrictions, and specific execution details. Use natural language processing technologies, such as named entity recognition (NER) and dependency parsing, to identify important entities and relationships in the policy from the text. The above key clauses include clause conditions, clause restrictions, and clause execution details. The above key elements include scope of use, rule conditions, and execution requirements.

[0118] The following is a possible implementation method, for example: Terms and Conditions: The execution conditions described in the policy, such as "You can apply for withdrawal only if your monthly income exceeds 5,000 yuan."

[0119] Restrictive terms: such as "only 50% of the balance can be withdrawn".

[0120] Implementation details: such as “the policy will take effect from January 1, 2024.”

[0121] Convert the clause content into key elements, which will be used for subsequent rule adjustment and generation: Scope of application: The scope to which the policy applies, such as applicable to certain specific regions or specific groups of people.

[0122] Rule conditions: conditions that need to be met, such as income level, account balance, etc.

[0123] Implementation requirements: How the rules are implemented, such as withdrawal limits or application procedures.

[0124] Automatically parse a paragraph in the policy document: "From January 1, 2024, the withdrawal limit for users with deposit balances less than 100,000 yuan will be adjusted to 30%", extract the scope of use (deposit balance less than 100,000 yuan), rule conditions (withdrawal limit adjusted to 30%) and execution requirements (execution time is January 1, 2024).

[0125] Match the extracted policy elements with the existing rule templates in the system to generate a version file containing the new policy content. The rule template is a standardized business logic structure used to manage provident fund business, which contains conditions, parameters, restrictions and other content of different business types, which is convenient for flexible configuration and updating. According to the extracted key elements, match and configure with the rule templates in the system to generate a new version file. For example, match "adjust the withdrawal limit to 30%" with the extraction rule template, update the relevant fields in the template, and generate a rule file containing the new policy content. Update the key element of "adjusting the withdrawal limit" to 30% extracted from the policy document to the extraction rule template to generate a new version file for adjusting the business logic.

[0126] Before releasing new rules, conduct virtual data testing to ensure the correctness and business feasibility of the new rule version. Use virtual user data or historical data to simulate the newly added rule version to verify the correctness of the rule logic. Simulate various scenarios, such as high-income and low-income users, different withdrawal amounts and times, to ensure the correct application of the rules under different conditions.

[0127] If the virtual test results meet expectations, it means that the rule logic is correct. If the virtual test results do not meet expectations, the rules are adjusted and the templates or key elements are modified again until the test passes.

[0128] On the premise that the test passes, the new rule version is released and enabled in the system to make it a formal business logic. According to the execution time described in the policy, the new rule version is set to be automatically released or manually enabled at a specific time. For example, if the policy execution time is January 1, 2025, the new rule version is automatically released on that date.

[0129] After the new version file is officially activated, the dynamic rule engine will integrate the new rules into the business logic of the provident fund management system to ensure that all business operations are performed in accordance with the latest rules.

[0130] The dynamic rule engine can automatically obtain policy data and generate new rules through natural language parsing and matching operations to achieve automatic response to policy changes. It reduces manual adjustments caused by policy changes and improves the flexibility and response speed of the business system. It verifies the correctness of the newly added version file through virtual data testing to ensure that the published business logic can be correctly applied. When the virtual test fails, it can be adjusted in time to avoid business risks caused by rule errors. The dynamic rule engine can not only adjust the rules, but also automatically publish new rules according to the execution time of the policy, making the system highly intelligent. It ensures that the new policy can take effect at the scheduled time point without manual intervention, reducing problems caused by untimely operation.

[0131] In some examples, controlling the data analysis layer to perform a data analysis operation based on the database stored data and the user input data to generate an evaluation result includes: The evaluation result R is determined based on the following formula: ; In the formula, is the regularization term for the evaluation results, represents the total number of features, is the comprehensive scoring function of the evaluation results, is the feature interaction term, For the The weight coefficient of the feature, is the regularization strength coefficient, is the feature interaction term weight coefficient, The first The feature values ​​include age, income and loan demand, Indicates the first historical data stored in the database storage data eigenvalues, for The corresponding weight coefficient is, for The corresponding weight coefficient is, is the penalty coefficient, is the kernel function of feature interaction, The first eigenvalues, The first eigenvalues, represents the square difference of the interaction between two features, is the smoothing parameter.

[0132] Exemplarily, by evaluating the multi-feature combination of user input data and historical data, the system can comprehensively consider the impact of various features (such as age, income, and loan demand). Each feature has its corresponding weight and scoring function to ensure that the importance of each feature is reflected in the evaluation, thereby generating more accurate evaluation results. The feature interaction term considers the complex interaction between user input data and historical data through the kernel function, capturing nonlinear patterns that are difficult to represent with linear models. This interaction between features makes the evaluation model more flexible and can better reflect the association between different features in reality. The regularization term is used to smooth the evaluation model to avoid overfitting of certain features and ensure that the model still has good generalization ability when processing new data. Through regularization, the evaluation results are more robust and the model is more stable. The data analysis layer combines user input data and historical storage data to make the evaluation results personalized and more targeted. For example, in the approval of provident fund loans, the user's current income level and historical repayment record can be combined to tailor a loan plan for the user to improve the user experience.

[0133] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A provident fund management system based on multi-heterogeneous data fusion, characterized in that: include: A data access layer, which is used to receive multiple heterogeneous data sources from the bank side, the enterprise side and the provident fund management center side, wherein the multiple heterogeneous data sources include structured data, unstructured data, semi-structured data and mixed data; A data fusion layer, the data fusion layer is used to perform data cleaning operations, data matching operations and data model building operations on the heterogeneous data sources; A business processing layer, which is used to process the core business of the provident fund. The business processing layer includes a dynamic rule engine, which adjusts its corresponding business logic in real time based on policy changes; The data analysis layer is used to generate evaluation results by performing data analysis operations based on database storage data and user input data; The security management layer is used to perform permission control operations on access users and is also used to desensitize and encrypt sensitive information; User service layer, which is used to provide mobile APP services, public account services and Web portal services.

2. A control method, used for the provident fund management system based on multi-heterogeneous data fusion according to claim 1, characterized in that: include: Controlling the data access layer to receive multiple heterogeneous data sources from a bank side, an enterprise side, and a provident fund management center side, wherein the multiple heterogeneous data sources include structured data, unstructured data, and semi-structured data; The data fusion layer is controlled to perform data cleaning operations, data matching operations and data model building operations on the heterogeneous data sources to obtain provident fund fusion data.

3. The control method according to claim 2, characterized in that: The data fusion layer includes a data cleaning module, a data matching module and a data model building module; The controlling the data fusion layer to perform data cleaning operations, data matching operations, and data model building operations on the heterogeneous data sources to obtain provident fund fusion data includes: Controlling the data cleaning module to perform a data cleaning operation on the heterogeneous data to obtain cleaned heterogeneous data; Controlling the data matching module to perform a data matching operation on the cleaned heterogeneous data based on an intelligent matching model to obtain matched data; The data model building module is controlled to perform a data model building operation on the matched data to obtain provident fund aggregate data, wherein the data model building module includes an aggregation model, a prediction model and a classification model.

4. The control method according to claim 3, characterized in that: The intelligent matching model includes a feature extraction unit, a similarity calculation unit, a matching decision unit and a data output unit; The feature extraction unit is used to extract different data features corresponding to different data types based on different feature extraction methods; The similarity calculation unit is used to calculate the comprehensive similarity based on the similarity of each feature type and the weight coefficient corresponding to the similarity of each feature type; The matching decision unit is used to obtain a data matching result based on the comprehensive similarity; The data output unit is used to generate matched data based on the data matching result.

5. The control method according to claim 4, characterized in that: The feature extraction unit performs feature extraction based on a first feature extraction method, a second feature extraction method, a third feature extraction method and a fourth feature extraction method; The first feature extraction method is used to extract the first feature of the structured data, the second feature extraction method is used to extract the second feature of the unstructured data, the third feature extraction method is used to extract the third feature of the semi-structured data, and the fourth feature extraction method is used to extract the fourth feature of the mixed data; The first feature extraction method is implemented by performing a time series feature extraction operation based on the feature combination of important fields, the second feature extraction method is implemented based on feature extraction methods corresponding to different unstructured data types, the third feature extraction method is implemented based on field parsing operations, nested structure parsing operations and statistical feature extraction operations, and the fourth feature extraction method is implemented based on hierarchical extraction operations and multi-modal feature learning operations.

6. The control method according to claim 5, characterized in that: The comprehensive similarity Determined based on the following formula: ； in, is the first feature similarity, is the weight coefficient corresponding to the first feature similarity, is the second feature similarity, is the weight coefficient corresponding to the second feature similarity, is the third feature similarity, is the weight coefficient corresponding to the third feature similarity, is the fourth feature similarity, is the weight coefficient corresponding to the fourth feature similarity; When the first feature is a significant time series feature, the first feature similarity is determined based on the following formula: : ； In the formula, and are the first eigenvectors of two different data, Indicates the calculation result of dynamic time warping; When the first feature is a non-significant time series feature, the first feature similarity is determined based on the following formula: : ； In the formula, and is the first eigenvector and The corresponding i-th component; In the case where the second feature is a feature corresponding to text data, the second feature similarity is determined based on the cosine similarity of the vectors corresponding to the word frequency and the inverse document frequency. ; In the case where the second feature is a feature corresponding to the image data, the second feature similarity is determined based on the convolution feature similarity ; The third feature similarity is calculated based on the following formula : ； in, To summarize field similarity, is the nested field similarity, To calculate the similarity of comprehensive field statistics, is the first weight coefficient, is the second weight coefficient, is the third weight coefficient; The summary field similarity Based on the following calculation: ； in, For each field The corresponding basic similarity, Indicates the total number of fields; The basic similarity Based on the following calculation: ； In the formula, and For different data source fields The value of is the text field similarity; The nested field similarity Based on the following calculation: ； is the total number of nested fields, For the The weight of the nested fields; The comprehensive field statistical feature similarity Based on the following calculation: ； is the statistical feature vector of the semi-structured data in the first record, is the statistical feature vector of the semi-structured data in the second record; The fourth feature similarity Determined based on the following formula: ； vector Determined by the following formula: ； in, is the fourth eigenvector of the first record, is the fourth eigenvector of the second record, is the ensemble learning weight, represents the first eigenvector, represents the second eigenvector, represents the third eigenvector.

7. The control method according to claim 4, characterized in that: The specific steps performed by the matching decision unit include: Calculate dynamic thresholds based on feature priority and data type; A comparison operation is performed based on the comprehensive similarity and the dynamic threshold to obtain a data matching result.

8. The control method according to claim 2, characterized in that: Also includes: Controlling the business processing layer to process the core business of the provident fund based on the dynamic rule engine to generate business feedback data, wherein the dynamic rule engine adjusts its corresponding business logic in real time based on policy changes; The data analysis layer is controlled to perform data analysis operations based on the database storage data and the user input data to generate an evaluation result.

9. The control method according to claim 8, characterized in that: The specific steps of the dynamic rule engine adjusting its corresponding business logic in real time based on policy changes include: Obtaining policy data, wherein the policy data is derived from an API data interface and manually input data; Automatically parse the key clauses of the policy document based on natural technology to extract key elements, where the key clauses include the conditions, restrictions and execution details of the clauses, and the key elements include the scope of use, rule conditions and execution requirements; Performing a matching operation based on the key elements and the rule template to generate a newly added version file; Perform virtual data testing based on the newly added version file to obtain a virtual test result; When the virtual test result is qualified, the newly added version file is released and enabled based on the execution time corresponding to the policy data.

10. The control method according to claim 9, characterized in that: The controlling the data analysis layer to perform a data analysis operation based on the database storage data and the user input data to generate an evaluation result includes: The evaluation result R is determined based on the following formula: ； In the formula, is the regularization term for the evaluation results, represents the total number of features, is the comprehensive scoring function of the evaluation results, is the feature interaction term, For the The weight coefficient of the feature, is the regularization strength coefficient, is the feature interaction term weight coefficient, The first The feature values ​​include age, income and loan demand, Indicates the first historical data stored in the database storage data eigenvalues, for The corresponding weight coefficient is, for The corresponding weight coefficient is, is the penalty coefficient, is the kernel function of feature interaction, The first eigenvalues, are the j-th eigenvalues ​​of user input data and historical input data, respectively. represents the square difference of the interaction between two features, is the smoothing parameter.

Images