Self-service modeling management method based on large model

Abstract

The invention discloses a self-service modeling management method based on a large model, which is realized by combining a big data technology with a page interaction mode and is characterized in that auditor demands are analyzed by operating a page to interact with a demand side, and a system recommends an auditing model capable of quickly coping with related demands from a deployed model library through calculation. Or a new model is generated directly according to the rule and threshold adjustment scheme, and self-service modeling of the demand side is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous modeling, and in particular to a method for self-service modeling management based on a large model. Background Art

[0002] The label modeling method is a data modeling technology that models data distributed in different databases in a networked manner around entities, relationships and labels. This method achieves a structured description of data by establishing label models, processing label rules, and synchronizing label scheduling, helping to more deeply understand the characteristics, behaviors and needs of the object. Label modeling is not only applicable to multiple analysis scenarios in different data domains such as users, goods, equipment, suppliers, stores, and corporate archives, but is also widely used in big data applications such as precision marketing, personalized recommendations, user portraits, and credit scores that require calculations based on detailed data. The core of label modeling is to provide a unified label data directory and label call interface for upper-level applications, and to achieve high-value label sharing and reuse through customized label rule functions to form a label operation ecosystem. This method allows for a wide range of analysis scenarios in different data fields such as users, goods, equipment, suppliers, stores, and corporate archives. By establishing a label description method around multiple entity objects and their relationships, deep mining and application of data can be achieved.

[0003] Compared with the traditional modeling process, label modeling is to build a logical model directly on top of the existing physical data / model. Through the proxy analysis of each data service, users can perform various calculations directly on the view without the need for large-scale processing of physical data in advance. It can be calculated immediately after use. However, it should be clear that each label still needs to correspond to the physical table to ensure the accuracy and consistency of the data. This modeling method not only improves data processing efficiency, but also provides rich data resources and flexible data application capabilities for upper-level applications.

[0004] However, the traditional label modeling method has the following disadvantages:

[0005] 1) Lack of effective management of models:

[0006] Existing label modeling methods only focus on the process of model building, and lack systematic and continuous evaluation of indicators such as the usage and effectiveness of the model after it is established.

[0007] There is insufficient management of the model's entire life cycle, and there is a lack of means for regular verification and updating, which makes it impossible to guarantee the model's effectiveness and accuracy.

[0008] 2) Long optimization and development cycle:

[0009] The process of optimizing and developing models requires repeated communication to convert business requirements into IT requirements. The model is then modeled by IT auditors and verified and confirmed by business auditors. The entire process consumes a lot of time and resources.

[0010] The development cost is high and timeliness cannot be guaranteed.

[0011] 3) Over-reliance on auditors’ level:

[0012] Existing methods have high requirements on the capabilities of auditors, and require a lot of time to understand the data and master data analysis, processing, and processing techniques. They are unable to quickly obtain target data, labels, and models.

[0013] Relying on purely manual analysis of data and labels is prone to errors and omissions, and cannot accurately complete modeling.

[0014] 4) Limited monitoring capabilities:

[0015] Existing modeling technologies mainly rely on pre-set warning rules and monitoring thresholds, and have limited ability to identify new risk points and abnormal situations caused by adjustments to focus points / areas.

[0016] Poor response and adaptability, unable to promptly and effectively warn of changing risks and emerging new problems.

[0017] 5) Lack of flexibility:

[0018] Traditional label modeling methods often rely on specific algorithm frameworks and data processing processes, which may not fully consider future variability and scalability when designed. As the business environment and data characteristics change, these fixed frameworks may become no longer applicable or difficult to adjust.

[0019] In summary, the shortcomings of the existing technology can be classified as lack of effective management of models, long optimization and development cycle, limited monitoring capabilities, and over-reliance on the level of auditors. This patent aims to address these shortcomings by introducing big data and machine AI big model technology to provide an efficient, accurate, and self-service risk warning modeling method, overcome the limitations of the existing technology, and provide an effective way to manage and develop models. Summary of the invention

[0020] In view of the deficiencies in the prior art, the present invention provides a self-service modeling management method based on a large model, which solves the problems mentioned in the above background technology.

[0021] To achieve the above objectives, the present invention is implemented through the following technical solutions: a self-service modeling management method based on a large model, specifically including the following contents:

[0022] First of all, the self-service modeling management method is comprehensive, which is reflected in the following aspects:

[0023] This method has the comprehensiveness of audit content, which is manifested in:

[0024] Covering all aspects of business management: The audit model should cover all aspects of business management, including strategy, market, finance, operations and other business activities. This ensures that the audit can comprehensively review the business operations without missing any important links.

[0025] Focus on core business processes and key success factors: The audit model should focus on the core business processes and key success factors of the enterprise on the basis of comprehensiveness. This will help the audit to more accurately identify key issues and potential risks in the enterprise's operations.

[0026] The application of audit models should support a variety of audit objectives and needs of enterprises. Whether it is annual financial report audits, special audit investigations, or continuous monitoring and risk warnings, audit models can provide strong support. By flexibly configuring and adjusting model parameters, audit models can meet the needs of different audit scenarios.

[0027] This method has comprehensive evaluation indicators, which are as follows:

[0028] Combination of financial indicators and non-financial indicators: The evaluation indicators in the audit model include not only financial indicators such as revenue, profit, debt-to-asset ratio, etc., but also non-financial indicators such as customer satisfaction, market share, innovation capability, etc. This combination helps to more comprehensively evaluate the overall performance of the enterprise.

[0029] Combining quantitative indicators with qualitative indicators: Quantitative indicators are expressed in data and are highly objective, but may ignore certain complexities and uncertainties; qualitative indicators can reflect the complexity and uncertainty of business activities and are a supplement to quantitative indicators. The audit model should include both types of indicators to draw more comprehensive and accurate evaluation conclusions.

[0030] This method has the comprehensiveness of the audit method, which is manifested in the following aspects:

[0031] Use multiple audit methods simultaneously: The audit model should support the use of multiple audit methods simultaneously, such as traditional audit methods, modern audit methods (such as risk-oriented audit, data-based audit, etc.), and audit big models based on big data and artificial intelligence technology, etc. This will help auditors choose appropriate audit methods according to specific circumstances and improve the efficiency and effectiveness of audit work.

[0032] Continuous optimization and innovation: With the continuous changes in the audit environment and the continuous development of audit technology, the audit model should be continuously optimized and innovated. This includes the introduction of new audit concepts, methods and tools, as well as the improvement and perfection of the existing audit model to ensure that it always remains comprehensive and advanced.

[0033] This method has the comprehensiveness of the audit process, which is manifested in:

[0034] Full process coverage from data collection to result output: The audit model should be able to cover the entire process from data collection, data preprocessing, model training, model evaluation to result output. This ensures the continuity and integrity of the audit work and helps auditors fully understand the company's operating conditions and potential risks.

[0035] Support continuous audit and risk monitoring: The audit model should be risk-oriented and achieve continuous audit and risk monitoring by real-time monitoring of the audited entity’s business management dynamics. This helps to timely discover and respond to potential risks and ensure the sound operation of the enterprise.

[0036] This method has the comprehensiveness of audit model maintenance and update, which is manifested in:

[0037] Regular evaluation and verification: Regularly evaluate and verify the audit model to check the accuracy and effectiveness of the model. Based on the evaluation results, timely adjust and optimize the model parameters and algorithms to ensure that the model always maintains the best status.

[0038] Timely update and iteration: With the development of corporate business and changes in the external environment, audit models also need to be continuously updated and iterated. New data sources and algorithms should be introduced in a timely manner, and model structures and parameters should be adjusted to adapt to new audit needs and risk characteristics.

[0039] Compliance review: During the model update and iteration process, compliance review is conducted to ensure that the model complies with relevant laws, regulations and regulatory requirements, thus avoiding compliance risks faced by enterprises due to model problems.

[0040] Therefore, the problem of lack of effective management of models in the prior art can be overcome.

[0041] In addition to being comprehensive, the self-service modeling management method also has modeling autonomy, specifically:

[0042] By using big data technology and the interactive interface of the system, the demand is analyzed in an interactive way through the operation interface. The system can quickly recommend and generate application models based on existing basic data, or recommend the best model that meets the demand, the rules of the model, and the adjustment plan of the model threshold, so as to realize autonomous modeling by the demand side.

[0043] The specific implementation is as follows:

[0044] Identify the risk warning areas that the system needs to support (such as financial fraud, supply chain risk, network security, etc.), and determine the specific questions and types of needs that the demand side may raise; first, clarify the specific needs and goals of risk warning modeling, which includes determining the types of risks that need to be monitored; (such as financial fraud, compliance violations, operational risks, etc., the first step is to obtain needs); then collect and analyze business needs to ensure that the model design meets the needs of actual work; and use big data technology in combination with the system's interactive interface to communicate in depth with the demand side through input or clicks on the system operation interface. This is manifested in operating and interacting with the demand side through the system's interactive interface to gain an in-depth understanding of their specific needs and problems.

[0045] Select components and tools of the graphical user interface GUI framework (Java's Swing and JavaFX, Python's Tkinter and PyQt, C#'s WinForms and WPF, etc.) that suit the project requirements to quickly generate the operation interface.

[0046] Then use the interface design tools that come with the IDE (such as Eclipse's WindowBuilder plug-in, IntelliJIDEA's GUI design function, Visual Studio's WinForms and WPF design tools, etc.) to quickly generate the interface. Build the system's interactive interface so that it can understand and analyze the user needs and habits of the demand side; design interactive interfaces, including guidance interfaces, request interfaces, error handling, etc., to ensure that the modeling needs and software requirements of the demand side are accurately captured.

[0047] The interactive system connected to the model library can understand and parse the input requirements of the demand side. The implementation algorithm is as follows: the input text is segmented into single words or phrases, and a large amount of labeled data is used for model training to achieve real-time interaction with the demand side;

[0048] Establish a relevant model library to provide decision support for the interactive interface of the system. The process of establishing the model in the model library is: collecting the data required for the model, preprocessing and labeling the data, designing, generating and training the model, optimizing and evaluating, deploying and integrating. The specific corresponding steps are as follows:

[0049] Collecting the data required for the model includes the following steps: According to the model requirements, collect basic data and label data. Basic data includes corporate financial statements, transaction records, and system logs, while label data is the risk event identifier that has been verified by manual review or historical data;

[0050] We can use the following formula to describe this process:

[0051] D base,labe=fcollect(R,S)D base ,D label =f collect (R,S)

[0052] Here, the function fcollect(R,S) collects and possibly preprocesses the basic data D from the data source S according to the model requirements R. base and label data D label ; Indicates that data collection can be completed through this function.

[0053] The preprocessing and labeling of data includes the following steps: cleaning, deduplication, normalization and other preprocessing operations on the previously collected data to improve data quality. At the same time, the label data is labeled and classified in detail for use in model training.

[0054] Data preprocessing usually includes steps such as cleaning, deduplication and normalization. These steps can be expressed as raw A series of operations:

[0055] D clean =f clean (D raw )

[0056] D unique =f dedup (D clean )

[0057] D normalized =f normalize (D unique )

[0058] Label annotation is the process of converting the original label dataset Lraw into the labeled label dataset Lannotated. This usually involves detailed classification, verification and confirmation of the labels:

[0059] L annotated =f annotate (L raw )

[0060] D raw : Raw dataset, containing the collected unprocessed data.

[0061] D clean : The cleaned data set has removed noise, missing values, etc.

[0062] D unique : The deduplicated data set removes duplicate records.

[0063] D normalized : After normalization, all features are scaled to the same dimension.

[0064] L raw : Original label dataset, containing labels that are not fully annotated.

[0065] L annotated : Labeled dataset, the labels have been classified and confirmed in detail.

[0066] f clean : Data cleaning function, used to remove noise, fill missing values, etc.

[0067] f dedup : Data deduplication function, used to remove duplicate records in a data set.

[0068] f normalize : Data normalization function, used to scale the data to a specific range (such as 0 to 1).

[0069] f annotate : Label marking function, used to classify and confirm labels in detail.

[0070] Design, generate and train the model; use the preprocessed and labeled data to train and optimize the model. First, set up the training environment and configure appropriate computing resources and deep learning frameworks; select the algorithm for model training, including data loading, model building, loss function definition, and optimizer setting; use the training set to train the model, calculate the loss through forward propagation, and update the model parameters through back propagation, monitor indicators such as loss value and accuracy during training, and adjust the training strategy in time; use the validation set to validate the model and evaluate the generalization ability and performance of the model; adjust the model structure and parameters according to the validation results.

[0071] The partial algorithm design is as follows:

[0072] Forward propagation: where σ is the activation function.

[0073] Loss calculation:

[0074] Parameter update:

[0075] This section describes a typical process of model training using gradient descent (here is stochastic gradient descent SGD) in deep learning. It includes three main steps: forward propagation, loss calculation, and parameter update. The following is a detailed understanding of these three steps:

[0076] Forward propagation is the process by which the model calculates the input data through each layer and finally obtains the output. In the formula you gave:

[0077]

[0078] x is the input data.

[0079] W is the weight of the model, which is the parameter connecting different neurons or layers.

[0080] b is the bias term, which is used to adjust the output of the activation function.

[0081] σ is the activation function, which transforms the output of the previous layer into a non-linear representation, helping the model learn complex patterns.

[0082] is the predicted output of the model.

[0083] This process is calculated one by one through each layer of the model, starting from the input layer and ending at the output layer.

[0084] Loss calculation is the process of evaluating the difference between the model prediction results and the actual labels. Here is the formula:

[0085]

[0086] L is the average loss over the entire dataset.

[0087] N is the sample size.

[0088] is the true label of the i-th sample (0 or 1 for a binary classification problem).

[0089] This loss function encourages the model to output predictions that are closer to the true labels.

[0090] Parameter updating is the process of calculating the gradient of the loss function with respect to the model parameters and updating these parameters in the opposite direction of the gradient in the hope of reducing the value of the loss function. In stochastic gradient descent (SGD), this process is based on small batches of data (or mini-batch), and its formula is:

[0091]

[0092] W and b are the model parameters to be updated.

[0093] η is the learning rate, which controls the step size of the parameter update.

[0094] B is the mini-batch size, i.e. the number of samples used in each update.

[0095] and are the gradients of the loss function with respect to W and b, respectively, calculated and averaged over each sample in the mini-batch.

[0096] This process is repeated until a stopping condition is met, such as when the model performance no longer improves significantly or when a preset number of training epochs is reached. In each iteration, forward propagation is performed to calculate the predicted output, loss calculation is performed to evaluate the predicted performance, and parameter updates are performed to optimize the model parameters.

[0097] When implementing, you need to build a training environment suitable for deep learning. This includes selecting computing resources (such as GPU servers) and installing deep learning frameworks. For computing resources, use servers equipped with NVIDIA GPUs to ensure that there is enough memory and storage to process large data sets. Choose popular deep learning frameworks such as TensorFlow or PyTorch. As we can see from the previous article, we have completed the preprocessing and labeling of data. Now we have basic data sets and labeled data sets. Then use PyTorch's DataLoader to batch load data for efficient training. Define a neural network model suitable for risk warning to complete the model building work. When encountering classification problems, use the cross entropy loss function. Select Adam as the optimizer and set the learning rate. Use the training set to train the model. Calculate the loss through forward propagation, and update the model parameters through back propagation. Use the validation set to evaluate the generalization ability and performance of the model. According to the validation results, you may need to adjust the model structure (such as adding hidden layers, changing activation functions, etc.) or hyperparameters (such as learning rate, adjusting model structure and parameters is an indispensable part of machine learning projects, which helps improve the performance and generalization ability of the model.)

[0098] Training model Optimization and evaluation: Apply the trained model to the test set to generate prediction results, evaluate indicators based on the prediction results and actual conditions, analyze the evaluation results, and understand the strengths and weaknesses of the model; model optimization is the process of adjusting and improving the model based on the evaluation results to improve its performance.

[0099] Deployment and integration: The deployment and integration of the self-service risk warning modeling method based on big model technology can follow the following steps:

[0100] System deployment includes the following steps:

[0101] Determine the hardware resources required for deployment, such as servers and GPUs, and ensure that they meet the operating requirements of large AI models; configure the software environment, including operating systems, programming languages ​​(such as Python), AI frameworks (such as TensorFlow, PyTorch), and necessary dependent libraries and tools;

[0102] Ensure that the basic data and label data have been sorted, preprocessed and formatted as necessary to meet the model input requirements, and store the data in an accessible storage system;

[0103] Prepare the trained large model files (such as weight files, configuration files, etc.) and ensure that their versions are compatible with the deployment environment;

[0104] Use the API or tools provided by the graphical user interface (GUI) framework to load the trained model into the deployment environment, perform necessary initialization and configuration on the model; encapsulate the model into an API or microservice that can provide services to the outside world, configure the service interface so that it can accept requests from auditors and return the corresponding model prediction results;

[0105] System integration includes the following steps:

[0106] Develop an interactive interface for auditors to interact with the demand side through a browser, mobile device, or system operation interface; the interface should include functional modules such as model requirements, model recommendations, model generation, model evaluation, and provide clear operational instructions and prompts.

[0107] Integrate the front-end interface with the back-end service to ensure that the auditor's request can be correctly passed to the model service and receive the response from the model service; implement data encryption transmission and verification mechanism to ensure the security and integrity of the data transmission process.

[0108] After deployment is complete, perform comprehensive testing and verification to ensure that the model service can run normally and meet business requirements, and fix any problems found during the testing process.

[0109] The interactive system extracts relevant features from the integrated data based on the requirements entered by the demander. The features generate corresponding labels, and the most suitable model is matched from the model library based on the specific requirements and labels of the demander. If there are multiple candidate models, they are ranked and recommended based on historical performance, computing resource consumption and other factors.

[0110] Assume that the model library is M, and each model m i ∈M has a label set compatibility function C i (L), represents the model m i The degree of match with the given label set L. In addition, we define P i For model m i Performance score on historical data, R i For model m i The computational resource consumption required (can be some measure of time, memory, energy, etc.).

[0111] In order to select the most suitable model from the model library, we can define a comprehensive scoring function S i , which takes into account the model's compatibility, historical performance, and computing resource consumption:

[0112] S i =α·C i (L)+β·P i -γ·R i

[0113] Among them, α, β, and γ are weight parameters, which are adjusted according to the preferences and needs of the demand side (for example, if the demand side is very concerned about computing efficiency, the value of γ can be increased).

[0114] The system's interactive interface allows the demand side to adjust the model's rules and thresholds based on actual business conditions, and provides real-time feedback on the performance changes of the adjusted model to help the demand side find the optimal configuration.

[0115] Deploy the adjusted model to the production environment, ensure seamless integration with existing systems, and configure the computing and storage resources required to run the model.

[0116] Apparently, the interactive system can automatically complete the process of building new models based on the guidance and data provided by auditors, overcoming the problem of long optimization and development cycles.

[0117] S3: Adaptability of the problem.

[0118] The system collects data related to focus points and areas in real time or periodically, uses data analysis technology to identify anomalies and trend changes in the data, analyzes changes in the field, and evaluates their potential impact on the model application scenarios. Combined with the model's evaluation results, it evaluates whether the model still meets current business needs; if auditors find that the existing model has deficiencies and cannot accurately audit existing business scenarios or risk scenarios, auditors can also use existing basic data and label data to quickly generate new audit models through the system operation interface, and can also optimize existing models to respond to new risks and problems.

[0119] Overcome the limited monitoring capabilities of existing technologies through regular updates and monitoring.

[0120] S4: Closed-loop nature of model iterative optimization.

[0121] This method implements iterative optimization of the label model through automated evaluation and feedback mechanism, continuous optimization and iteration mechanism, iterative strategy mechanism, and label library management and expansion mechanism.

[0122] Specifically, it can be understood that a new model has been generated above and compared with the existing model. If the new model is better than the existing model, the old model is iterated and the new model and labels are used.

[0123] The above process of updating the model and iterating can be implemented using the following function:

[0124] First, we define several key variables and functions:

[0125] Model performance indicators (M_Performance): indicators used to quantitatively evaluate the quality of the model, such as accuracy, recall, F1 score, etc.

[0126] New model generation function (GenerateNewModel(Data,Labels)): A function that generates a new model based on new data and label data.

[0127] Model comparison function (CompareModels(ModelA, ModelB)): A function that compares the performance of two models and returns a Boolean value indicating whether model A is better than model B.

[0128] Model update function (UpdateModel(OldModel,NewModel)): A function that replaces the old model with the new model.

[0129] Feedback cycle time (T_Cycle): The cycle time for model evaluation and update.

[0130] Based on the above definition, we can construct a simplified formula to describe the closed-loop nature of the iterative optimization of the model:

[0131] Suppose we have two models: CurrentModel (current model) and NewModel (newly generated model). We can set a conditional statement to represent the iterative update process of the model:

[0132] ifCompareModels(NewModel,CurrentModel)=TruethenUpdateModel(CurrentModel,NewModel)

[0133] We use a time series index t to represent the state of each cycle, and use loops or recursive structures to describe ongoing processes:

[0134] For eacht∈{T0,T0+T_Cycle,T0+2T_Cycle,…}

[0135] The system automatically iterates and updates, which can overcome the problem of lack of flexibility of existing technologies compared to existing technologies.

[0136] Finally, the key parameters and formulas in the whole process can be linked to each other to form a closed-loop system: the results of data collection (D base ,D label ) directly affects the input of data preprocessing and annotation. The preprocessed and annotated data (D clean ,D unique ,D normalized ,D annotated ) is the basis of model training. The results of model training (using SGD or Adam optimizer to update weights W and bias b) are used for services in the deployment and integration stages. The performance of the deployed model is fed back to the model iteration optimization stage through evaluation to generate a new model and update the old model.

[0137] Beneficial Effects

[0138] The present invention provides a self-service modeling management method based on a large model. Compared with the prior art, it has the following beneficial effects:

[0139] 1) Improve model effectiveness and accuracy:

[0140] Implement a comprehensive model management strategy. The system regularly evaluates and maintains models from multiple dimensions such as frequency of use and application effect, manages model classifications by level, and provides optimization suggestions for inefficient models to improve the overall effectiveness and accuracy of the model.

[0141] 2) Implement a self-service risk warning modeling method:

[0142] By utilizing big data technology in combination with the system operation interface to analyze demand through interaction with the demand side, the system can quickly generate existing basic data and labels for audit models through calculations and recommendations, or directly recommend the best model that meets the demand and its rules and threshold adjustment plans, thereby achieving self-service modeling by the demand side.

[0143] Reduce over-reliance on manual analysis, data selection and labels, reduce the impact of selection errors or errors, and achieve self-service fast and accurate modeling.

[0144] 3) Strengthen risk identification capabilities:

[0145] Based on the model evaluation results, the system will continuously monitor changes in focus points / areas and provide targeted model optimization suggestions. It will no longer rely on pre-set rules and thresholds, but will continuously optimize the model to respond to new risks and problems.

[0146] This dynamic adjustment strategy can effectively improve risk identification capabilities and ensure that the model can adapt to changing environments and challenges.

[0147] 4) Label model optimization / iteration (evaluation mechanism).

[0148] Through mechanisms such as automated evaluation and feedback, continuous optimization and iteration, iterative strategies, and label library management and expansion, we ensure the continuous optimization and iteration of the label model, improve the accuracy and adaptability of the model, and provide users with more accurate and efficient modeling services.

[0149] To sum up, the purpose of the invention of this patent is to provide an improved risk warning modeling method based on big data technology, which improves the accuracy and effectiveness of modeling through regular evaluation and model grading, enhances risk identification capabilities through dynamic optimization and real-time feedback mechanisms, strengthens the ability to adapt to and respond to new risks, and improves the overall efficiency and performance of modeling. At the same time, it simplifies the modeling process and realizes self-service modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0150] Figure 1 It is a flow chart of the operating steps of the present invention. DETAILED DESCRIPTION

[0151] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0152] Embodiment 1

[0153] We use a self-service modeling management method based on a large model to implement it in a real scenario to illustrate the actual operation process of this method:

[0154] In the specific implementation, we first have to build the corresponding model library, set up a deep learning training environment, select a high-performance GPU server, install deep learning frameworks such as PyTorch, and build a model library containing a variety of basic network structures based on existing research results and industry standards, such as LSTM, CNN, Transformer, etc.

[0155] Then the user (or the demander) directly operates the system operation interface of the interactive system and inputs specific business requirements. The interactive system can recommend the most suitable model and its parameter configuration from the model library we have established based on the user's needs. At this time, the demander only needs to input or import the required data according to the data interface corresponding to the recommended model and parameters.

[0156] The system will automatically collect data related to audit focus and areas from various data sources (such as business systems, external databases, sensors, etc.) based on set time intervals (such as daily, weekly, monthly, etc.) or based on specific event triggers (such as market changes, policy adjustments, etc.). These data may include but are not limited to transaction records, user behaviors, market trends, policy documents, etc., and use statistical analysis, machine learning, data mining and other data analysis techniques to conduct in-depth analysis of pre-processed data to identify anomalies, trend changes and correlations in the data. These analyses can help auditors understand the status of business operations and discover potential risks and problems; by monitoring key indicators and patterns in the data, the system can promptly discover trend changes and anomalies, which may be harbingers of business risks, fraud or system vulnerabilities, etc.

[0157] Auditors combine data analysis results and changes in the field (such as policy adjustments, changes in the market environment, etc.) to evaluate whether the current audit model still meets business needs. This includes comprehensive consideration of the model's accuracy, efficiency, stability, and adaptability. If it is found that the existing model is insufficient and cannot accurately audit existing business scenarios or risk scenarios, auditors can use existing basic data and label data to quickly generate new audit models through interface interaction. Here, auditors can modify thresholds to generate new rules to achieve this, or modify labels and thresholds to generate new rules to achieve this. From here, it can be clearly seen that the interactive system can automatically complete the construction process of the new model based on the guidance and data provided by the auditor (without repeating the entire modeling process). This greatly shortens the model development cycle and improves the efficiency and flexibility of audit work.

[0158] At this time, the newly generated model needs to be verified and tested by the auditor to ensure its accuracy and stability. After verification, the model can be deployed to the actual production environment to start performing audit tasks. The new model here is optimized based on the previous model according to the new scenario and rules, so it will replace the original model in the model library, and the iteration of the model library is completed here.

[0159] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0160] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0161] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A self-service modeling management method based on a large model, characterized by: Specifically include the following: S1: The model’s management approach is comprehensive: The comprehensiveness of the audit content; Comprehensiveness of evaluation indicators; the comprehensiveness of the audit methodology; The comprehensiveness of the audit process; Audit the comprehensiveness of model maintenance and updates. S2: Autonomy of modeling. By using big data modeling technology and page interactive operations, the system analyzes the needs in the form of an operation interface. The system can quickly recommend and generate labels for application models (such as audit models) based on existing basic data, or recommend the best model that meets the needs, model rules, and model threshold adjustment plans to achieve autonomous modeling on the demand side. S3: Adaptability of the problem. Based on the results of the model assessment, the method will continue to provide targeted model optimization suggestions for changes in the monitored "focus points" and / or "areas". This method does not rely on pre-set rules and thresholds, but instead continuously and autonomously optimizes the model to respond to new risks and issues. S4: Closed-loop nature of model iterative optimization. This method implements iterative optimization of the label model through automated evaluation and feedback mechanism, continuous optimization and iteration mechanism, iterative strategy mechanism, and label library management and expansion mechanism.

2. The method for self-service modeling management based on a large model according to claim 1, characterized in that: The management method of the S1 model is comprehensive and specifically manifested as follows: Comprehensiveness of audit content: The business scope of the audit model covers all aspects of enterprise operation and management, and the model focuses on core business processes and key success factors; the application of the audit model should support the enterprise's various audit objectives and needs; Comprehensiveness of evaluation indicators: Combination of financial indicators with non-financial indicators, and combination of quantitative indicators with qualitative indicators; Comprehensiveness of audit methods: using multiple audit methods simultaneously, and continuously optimizing and updating them; Comprehensiveness of the audit process: Covering the entire process from data collection to result output, supporting continuous auditing and risk monitoring; Comprehensiveness of audit model maintenance and updates: regular evaluation and verification, timely updates and iterations, and compliance reviews.

3. The method for self-service modeling management based on a large model according to claim 1, characterized in that: The content of S2 includes: By using big data modeling technology and page interactive operations, the system analyzes the needs in the form of an operation interface. The system can quickly recommend and generate labels for application models (such as audit models) based on existing basic data, or recommend the best model that meets the needs, model rules, and model threshold adjustment plans to achieve autonomous modeling on the demand side. The specific implementation is as follows: Clarify the risk warning areas that the system needs to support, determine the specific questions and types of needs that the demand side may raise; and use big data technology to interact with the page and communicate deeply with the demand side through interactive operation interfaces to obtain specific needs; Clarify the functions that the interface needs to achieve, and determine the overall layout and interaction mode of the interface based on these functions, so that it can understand and analyze the usage habits of the demander, so that it can be easily operated during the interaction; Establish a relevant model library to provide decision support for the operation interface. The process of establishing the model in the model library is: collect the data required for the model, preprocess and annotate the data, design, generate and train the model, optimize and evaluate, deploy and integrate. The specific corresponding steps are as follows: Collecting the data required for the model includes the following steps: collecting basic data and label data according to the model requirements; The data preprocessing and labeling includes the following steps: cleaning, deduplication, normalization and other preprocessing operations on the collected data to improve the data quality; at the same time, detailed labeling and classification of the label data for use in model training; Designing, generating and training a model includes the following steps: using preprocessed and labeled data to train and optimize the model, first setting up the training environment and configuring appropriate computing resources and deep learning frameworks; selecting an algorithm for model training; using the training set to train the model, calculating the loss through forward propagation, and updating the model parameters through back propagation, monitoring indicators such as loss value and accuracy during training, and adjusting the training strategy in a timely manner; using the validation set to validate the model and evaluate the model's generalization ability and performance; and adjusting the model structure and parameters based on the validation results. Optimization and evaluation include the following steps: applying the trained model to the test set, generating prediction results, evaluating indicators based on the prediction results and actual conditions, and analyzing the evaluation results; while model optimization is the process of adjusting and improving the model based on the evaluation results to improve its performance. The deployment and integration includes the following steps: The deployment and integration of the self-service risk early warning modeling method based on the big model technology can follow the following steps: System deployment includes the following steps: Determine the hardware resources required for deployment, configure the software environment, and the necessary dependent libraries and tools; Ensure that the basic data and label data have been sorted, preprocessed and formatted as necessary to meet the model input requirements, and store the data in an accessible storage system; Prepare the trained large model file and ensure that its version is compatible with the deployment environment; Use the API or tools provided by the graphical user interface framework to load the trained model into the deployment environment, and perform necessary initialization and configuration on the model; encapsulate the model into an API or microservice that can provide services to the outside world, and configure the service interface so that it can accept requests from auditors and return the corresponding model prediction results; System integration includes the following steps: Develop an interactive interface for auditors to enable them to interact with AI robots through a browser or mobile device; the interface should include functional modules such as model recommendation, model generation, and model evaluation, and provide clear operational instructions and prompts. Integrate the front-end interface with the back-end service to ensure that the auditor's request can be correctly passed to the model service and receive the response from the model service; implement data encryption transmission and verification mechanism to ensure the security and integrity of the data transmission process. After deployment is complete, perform comprehensive testing and verification to ensure that the model service can run normally and meet business requirements, and fix any problems found during the testing process. The system extracts relevant features from the integrated data based on the requirements input by the demand side. The features generate corresponding labels, and the most suitable model is matched from the model library based on the specific requirements and labels of the demand side. If there are multiple candidate models, they are ranked and recommended based on historical performance, computing resource consumption and other factors. The system's interactive interface allows the demand side to adjust the model's rules and thresholds based on actual business conditions, and provides real-time feedback on the performance changes of the adjusted model to help the demand side find the optimal configuration. Deploy the adjusted model to the production environment, ensure seamless integration with existing systems, and configure the computing and storage resources required to run the model.

4. The method for self-service modeling management based on a large model according to claim 1, characterized in that: In S3, based on the model evaluation results, the system will continuously monitor changes in focus points and areas and provide targeted model optimization suggestions: The system collects data related to the focus and domain in real time or regularly, uses data analysis technology to identify anomalies and trend changes in the data, analyzes changes in the domain, and evaluates their potential impact on the model application scenario. Combined with the model's evaluation results, it evaluates whether the model still meets the current business needs; If auditors find that the existing model is insufficient and cannot accurately audit existing business scenarios or risk scenarios, auditors can also use the existing basic data and label data to quickly generate new audit models through the system's interactive interface, or optimize the existing model to respond to new risks and problems.

5. The method for self-service modeling management based on a large model according to claim 1, characterized in that: In S4, evaluation criteria and methods for optimization iteration of the label model are formulated, and the model is iterated according to the evaluation results to form a closed-loop model optimization iteration system.

Images