Method and system for guaranteeing calculation consistency and reliability of risk control model

Abstract

The invention relates to the technical field of data processing, in particular to a method and system for guaranteeing risk control model calculation consistency and reliability. The system comprises a decision engine core service which is used for analyzing and executing a decision process file and calling a strategy rule and a model; the unified data evaluation framework is used for executing dependency analysis, recursive evaluation, cache read-write and model calculation service calling after receiving a model score acquisition request, and returning consistent model scores; the model full life cycle management platform is used for completing model management; the system field management center is used for automatically generating a globally unique and named normative model associated field according to preset technical constraints after receiving a model registration event, and synchronously updating the field cache; and the metadata blood relationship analysis module is used for establishing and maintaining a blood relationship graph in real time. According to the technical scheme, the consistency of data calculation can be ensured, the data calculation efficiency is improved, and therefore the reliability and stability of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method and system for ensuring the consistency and reliability of risk control model calculations. Background Technology

[0002] In the field of credit risk control, the decision engine is the core system. It calculates a customer's credit score (model score) by executing a series of strategy rules and risk models, and makes credit decisions accordingly. Currently, the execution of strategy rules and credit limit calculation in the decision engine are highly dependent on the calculation results (i.e., model scores) of one or more risk models.

[0003] The workflow orchestration of the decision engine primarily relies on manual intervention. Deployers must use their experience and expertise to manually and precisely place the "model calculation nodes" before all "nodes using the model score." This approach can barely maintain normal operation in simple workflow scenarios. For example, when the workflow structure is simple, the number of nodes is small, and the logical relationships are not complex, deployment personnel can relatively clearly understand the order of model score calculation and usage, ensuring the smooth execution of the workflow.

[0004] However, with the continuous changes in financial markets and the increasing complexity of banking operations, modern risk control decision-making processes are exhibiting highly complex characteristics. Typical scenarios include main processes embedding common sub-processes, which increases process layers and makes logical relationships intricate; the existence of policy set forwarding mechanisms further increases the dynamism and uncertainty of the process; the collaborative work of multiple component modules requires close cooperation and frequent data interaction between modules; and the use of complex conditional branches makes the process direction have multiple possibilities and is difficult to predict.

[0005] In such complex scenarios, the manual orchestration of decision engine workflows exposes numerous problems. On one hand, manually analyzing the dependencies between model computation and usage becomes extremely difficult and prone to errors. Due to the high complexity of the process and the large number of nodes, deployment personnel struggle to comprehensively and accurately grasp the dependencies between each node, easily leading to duplicate orchestration of model nodes. This not only results in redundant model calculations, wasting significant computing resources, but also reduces system efficiency and impacts the timeliness of credit decisions. Summary of the Invention

[0006] The purpose of this invention is to propose a method to ensure the consistency and reliability of risk control model calculations. This technical solution can ensure the consistency of data calculations, improve data calculation efficiency, and thus improve the reliability and stability of the system.

[0007] To achieve the above objectives, in a first aspect, the present invention proposes a method for ensuring the consistency and reliability of risk control model calculations, comprising: The core service of the decision engine is used to parse and execute decision process files, and call strategy rules and models during execution; it has a built-in process parser and a unified data evaluation framework. The unified data evaluation framework provides a unified interface for obtaining model scores using the chain of responsibility pattern. Upon receiving a model score request, it performs dependency resolution, recursive evaluation, cache read / write, and model calculation service calls, and returns a consistent model score. The model lifecycle management platform is used to manage models, including model creation, training, validation, deployment and real-time monitoring. Before deployment, it performs input parameter type validation, output field checking, dependency loop detection and hierarchical depth validation through the model validator. The system field management center is used to automatically generate globally unique and named model association fields according to preset technical constraints after receiving a model registration event, and to update the field cache synchronously. The metadata lineage analysis module is used to build and maintain a field-model-decision process lineage diagram in real time, and to prevent deletion or modification operations when a model or field is detected to be in use.

[0008] Beneficial effects of the basic solution: This solution relies on the unified data evaluation framework built into the core service of the decision engine, and automates the entire process of model score acquisition using a chain of responsibility model, covering key aspects such as dependency resolution, recursive evaluation, cache read / write, and model calculation service calls. This automation mechanism completely eliminates the tedious operation of manually arranging the order of model nodes in traditional risk control systems. On the one hand, it significantly reduces the technical threshold and workload of deployment personnel, allowing them to complete system configuration without needing to deeply understand the complex dependencies between models; on the other hand, it fundamentally eliminates system failures caused by manual sorting errors and missed dependency resolution, significantly improving model deployment efficiency and system stability. Simultaneously, the system field management center automatically generates globally unique association fields that conform to specifications and caches them synchronously during model registration, further reducing the workload of manual field definition and maintenance, achieving automated control of the model access process, and overall promoting a reduction in risk control system operation and maintenance costs and an improvement in operational efficiency.

[0009] This solution ensures the consistency and availability of model calculations through multiple mechanisms. First, a unified data evaluation framework ensures standardized processing of model score acquisition requests. Regardless of the complexity of dependencies, recursive evaluation and dependency resolution output consistent model scores, avoiding discrepancies in model calculation results across different calling scenarios. Second, an intelligent dependency guarantee mechanism proactively identifies the preconditions for model calculations, ensuring that model scores are calculated before use in the decision-making process. This fundamentally avoids policy failures and misjudgments caused by missing or unready model scores. Third, the model lifecycle management platform performs multi-dimensional verification before model release, including input parameter type verification and output field checks, ensuring the standardization and reliability of deployed models and improving model calculation quality from the source. These mechanisms collectively guarantee the accuracy and consistency of risk control decisions.

[0010] This solution achieves efficient utilization of system resources. When processing model score retrieval requests, the unified data evaluation framework first performs dependency resolution to clarify the calling order and relationships of each model, avoiding resource waste caused by disordered calls. Simultaneously, the framework's integrated caching read / write mechanism caches already calculated model results. When subsequent decision-making processes require calling the same model, the result is directly retrieved from the cache, eliminating the need to repeatedly call the model calculation service. This significantly reduces resource consumption from redundant calculations and lowers the overall system load. Optimized resource utilization directly improves the execution efficiency of the decision-making process, shortens the response time for risk control decisions, and enables rapid response to high-concurrency transaction scenarios. It ensures timely risk assessment even under massive transaction requests, enhancing user experience while further strengthening the timeliness of risk interception.

[0011] This solution constructs a comprehensive observable and auditable system through a metadata lineage analysis module and full lifecycle management log recording. The metadata lineage analysis module establishes and maintains a lineage diagram of fields, models, and decision-making processes in real time, clearly presenting the correlation links between the three and enabling full-chain traceability from raw data fields to model calculations and decision results. Simultaneously, this module can detect the usage status of models or fields, preventing deletion or modification operations when they are detected as being in use, thus avoiding system anomalies caused by changes in key elements. Furthermore, the model lifecycle management platform records information throughout the entire process of model creation, training, deployment, and monitoring, and the core service of the decision engine also retains detailed execution logs. This information complements the lineage diagram, forming a complete audit chain. Through full-chain visual traceability capabilities, it not only facilitates technical personnel in quickly locating system faults and troubleshooting problems but also meets the compliance requirements of internal audits and regulatory agencies, ensuring the transparency and traceability of the risk control decision-making process and improving the organization's compliance management level.

[0012] As a feasible and preferred solution, a unified data evaluation framework includes: The caching service is used to store the calculation results of the model's associated fields and set expiration policies; Model metadata service, used to provide the model's input parameter definitions and dependencies; The model computation service is used to perform model inference and return scores after all input parameters are available; The controller is called recursively. When it is found that the input parameters depend on the fields associated with other models, it recursively triggers its own process until all the underlying values ​​are obtained.

[0013] As a feasible and preferred solution, a unified data evaluation framework directly returns the cached value when the cache is hit, and executes the following steps sequentially when the cache is missed: Call the model metadata service to obtain the input parameter definitions; Recursively parse and calculate all dependent fields; The model score is obtained by calling the model calculation service. Write the model to the cache and return it.

[0014] As a feasible preferred solution, the recursive parsing algorithm is as follows: if the input parameters of the target model's associated field still depend on another model's associated field, the framework first recursively calculates the value of the other model's associated field until all underlying dependent variables or model components are calculated.

[0015] As a feasible and preferred solution, the naming rule for model-related fields generated by the system field management center is: S_F_FM98{model name}_{model version}, and the model UUID, version number and output field mapping relationship are recorded in the field attributes.

[0016] As a feasible preferred solution, the metadata lineage analysis module dynamically updates the lineage diagram by listening to model registration events and decision process definition change events, and immediately returns a prohibition signal and outputs alarm information when it receives a deletion / modification request for a referenced model or field.

[0017] As a feasible preferred solution, the process parser built into the core service of the decision engine converts the decision process file into a directed acyclic graph structure and executes each node in the topological order to ensure that the model call order and input data are consistent when the same decision process is executed multiple times.

[0018] As a feasible preferred solution, the technical constraints include: model input parameters are only allowed to come from basic variables or other model outputs, the model dependency depth does not exceed a preset level, and circular dependencies between models are not allowed.

[0019] As a feasible and preferred solution, the decision engine core service, model lifecycle management platform, system field management center, unified data evaluation framework and metadata lineage analysis module communicate with each other through an asynchronous message mechanism for event notification.

[0020] Secondly, this invention also proposes a method to ensure the consistency and reliability of risk control model calculations, which utilizes the aforementioned system for ensuring the consistency and reliability of risk control model calculations, including the model release stage, the decision-making process execution stage, and the system monitoring and maintenance stage. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture and component interactions of a system that ensures the consistency and reliability of risk control model calculations. Detailed Implementation

[0022] To make the technical solution and advantages of this application clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only some embodiments of the present invention, and are only used to explain this application, not to limit it. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated; they can be combined with each other to achieve better technical effects. The same reference numerals appearing in the accompanying drawings of the following embodiments represent the same features or components, and can be applied to different embodiments.

[0023] Furthermore, unless otherwise defined, the technical or scientific terms used in this invention description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.

[0024] The present invention will now be described in further detail with reference to the accompanying drawings.

[0025] Reference Figure 1 This disclosure provides a system that ensures the consistency and reliability of risk control model calculations. It adopts a microservice architecture design, with each component communicating through Dubbo RPC and using an asynchronous message mechanism for event notification.

[0026] The system mainly includes a decision engine core service, a model lifecycle management platform, a system field management center, a unified data evaluation framework, and a metadata lineage analysis module.

[0027] The core service of the decision engine serves as the system's execution hub, responsible for parsing and executing the decision-making process. It incorporates a unified data evaluation framework, designed using the chain of responsibility pattern, supporting the unified acquisition and processing of various field values. It receives requests from external sources, invokes corresponding strategy rules and models for calculation according to the preset decision-making process, and ultimately returns the decision result.

[0028] The core service of the decision engine has a built-in process parser that can read and parse decision process definition files. These files, in XML or JSON format, detail the various nodes in the decision process, the connections between them, and the strategy rules or models executed by each node. The process parser then breaks the file down into its internal data structure for subsequent execution.

[0029] After parsing the decision-making process, the core service of the decision engine sequentially calls the corresponding strategy rules and models according to the node definitions in the process. For model calls, the core service of the decision engine sends a request to the unified data evaluation framework to obtain the model score. After processing the request, the unified data evaluation framework returns the corresponding model score, and the core service of the decision engine continues to execute subsequent nodes.

[0030] The model lifecycle management platform is used for the creation, training, validation, deployment, and monitoring of risk control models, ensuring the quality and availability of models throughout their entire lifecycle.

[0031] When a new model passes validation and is ready for release, the model validator on the model lifecycle management platform performs multi-dimensional checks, including input parameter type validation, output field checks, dependency cycle detection, and hierarchical depth validation. After successful validation, the platform sends a model registration event notification to the system field management center, the event content of which includes the model's complete metadata information.

[0032] For model creation and training, the model lifecycle management platform provides model creation tools, allowing users to define model structure and parameters visually. It also integrates various machine learning algorithm libraries, such as TensorFlow and PyTorch, facilitating model training. For example, users can select the logistic regression algorithm using the platform's interface, set the corresponding feature parameters and training dataset, and start the model training process. During training, the platform monitors the progress and performance metrics, such as accuracy and recall, in real time and displays the information.

[0033] For model validation, the model lifecycle management platform validates the model after training. Multiple validation methods are employed, such as cross-validation and hold-out, to evaluate the model's performance and generalization ability. For example, cross-validation divides the dataset into k subsets, using each subset sequentially as the test set and the remaining subsets as the training set for model training and evaluation, ultimately calculating metrics such as average accuracy. If the model's validation results do not meet preset standards, the platform will prompt the user to adjust or retrain the model.

[0034] For model deployment, once the model has passed validation, a model deployment request can be initiated on the model lifecycle management platform. The platform's model validator will perform multi-dimensional checks, including input parameter type validation, output field checks, dependency cycle detection, and hierarchical depth validation.

[0035] Input parameter type validation: Checks whether the input parameter types of the model meet the preset requirements. For example, if the model requires numeric input parameters, but the actual input parameters provided are string types, the validation will fail.

[0036] Output field check: Verify that the model's output fields conform to the specifications. For example, check whether the names and data types of the output fields meet the system requirements.

[0037] Dependency cycle detection: Detects whether there are circular dependencies between models. For example, if model A depends on model B, and model B depends on model A, then a circular dependency exists, and the check fails.

[0038] Depth-of-dependency validation: Checks whether the dependency depth of a model exceeds a preset depth. For example, if the preset depth is 3, and model A depends on model B, model B depends on model C, and model C depends on model D, then the dependency depth of model A is 3. If it depends on other models, it exceeds the preset depth, and the validation fails.

[0039] After successful verification, the model lifecycle management platform sends a model registration event notification to the system field management center. The event content includes the model's complete metadata information.

[0040] For model monitoring, the platform will monitor the model's operation in real time after the model is published. Monitoring metrics include the number of model calls, computation time, and distribution of output results. For example, alarm rules can be set so that when the model's computation time exceeds a preset threshold (such as 1 second) or the distribution of output results shows abnormal changes, the platform will promptly issue an alarm to notify relevant personnel to handle the situation.

[0041] When modifying or deprecating a published model, the model lifecycle management platform also performs strict metadata verification. For example, when modifying a model, it checks whether the modified model still complies with the technical constraints of "automatic calculation," including whether the input parameter types and dependencies have changed and whether they are within the allowed range. If the modification causes the model dependency depth to exceed the preset level or causes problems such as circular dependencies, the verification will fail, and the model modification operation will be blocked.

[0042] When a model is delisted, the lineage analysis module checks whether the model is currently in use. If it is, the user is prompted that the model cannot be delisted to prevent errors in the ongoing decision-making process due to model delisting.

[0043] The system field management module is used to manage all fields in the system, including ordinary business variables, input fields, and model-related fields.

[0044] Upon receiving a model registration event, the field generator built into the field management module will automatically create the corresponding model-related fields according to predefined business rules.

[0045] Specifically, a globally unique model association field is automatically generated for each risk control model that meets specific technical constraints. These constraints include: model input parameters are only allowed to originate from basic variables or output scores from other models; model dependency depth does not exceed a preset level (e.g., the preset level is 3 levels, meaning a model can depend on a maximum of 3 other models); and circular dependencies between models are not allowed.

[0046] The naming of model-related fields follows strict rules to ensure the uniqueness and readability of field names, such as S_F_FM98{model name}_{model version}. For example, for a model named "CreditScoreModel" with version number "1.0", the generated model-related field name would be S_F_FM98CreditScoreModel_1.0.

[0047] The field attributes record detailed metadata about the associated model, including the model's unique identifier (UUID), version number, and output field mapping relationship. For example, the model UUID is "12345678 - 1234 - 5678 - 1234 -567812345678", the version number is "1.0", and the output field mapping relationship is "the model output score corresponds to the field name 'score'".

[0048] After a new field is generated, the management center will update the field cache synchronously and notify the lineage analysis module to update the lineage diagram.

[0049] The unified data evaluation framework is the core framework for handling model score requests. Designed using the chain of responsibility pattern, it supports the unified acquisition and processing of various field values, enabling automated calculation of model scores. It employs a recursive parsing algorithm to achieve intelligent dependency resolution and recursive evaluation. When a strategy rule or other model requires the value of a field associated with a particular model, it automatically intervenes, coordinating multiple services to complete the model score calculation.

[0050] Specifically, when the policy rule executor sends a request to the unified data evaluation framework to obtain the model-related field value (for example, in a credit decision process, the policy rule needs to decide whether to approve a loan based on the output score of a certain risk model, at which point the policy rule executor will send a request to obtain the model-related field value corresponding to that model).

[0051] The framework first queries the cache service to check if there is a valid cached value for the field. If there is a valid cached value, the framework directly returns the value to the policy rule executor, and the process ends. For example, if the value of the model-related field has been calculated in a previous calculation and stored in the cache, and the cache has not expired, the framework will directly retrieve the value from the cache and return it; if the cached value of the field is invalid or empty, the next step of processing will continue.

[0052] The unified data evaluation framework calls the model metadata service to obtain the complete input parameter definition of the model associated with the field. If a model's input parameter depends on other model-related fields, the framework identifies the dependency and recursively calls its own evaluation process to query the values ​​of the dependent fields. For example, suppose when retrieving the value of a model-related field X, it finds that an input parameter of its associated model M depends on a model-related field Y. The framework will then query the cache service for the value of field Y. If field Y is also not cached, the framework will continue to perform dependency resolution and recursive evaluation for field Y. The unified data evaluation framework will retrieve its input parameter definition and check whether the input parameter depends on other fields until all underlying dependent model scores and basic variables have been calculated.

[0053] After collecting all input parameters (including model scores and basic variables), the unified data evaluation framework calls the model calculation service to perform model calculations and obtain the values ​​of the basic variables. For example, if the input parameters of model M include basic variables such as customer age and income, the framework will call the variable calculation service to obtain the values ​​of these variables from the corresponding data sources.

[0054] The model calculation service performs calculations based on preset algorithms and parameters, and returns the model score. The unified data evaluation framework writes the obtained model score to the cache service for later use, and finally returns the model score to the original requester—the policy rule executor.

[0055] The metadata lineage analysis module is used to track the usage of model-related fields in real time, establish a complete lineage graph between fields, models, and decision flows, and ensure that model fields currently in use cannot be arbitrarily deleted or modified, thus guaranteeing the stability and reliability of the system.

[0056] Establishing lineage relationships: When the system field management center generates new model-related fields or the model's metadata changes, it notifies the metadata lineage analysis module. Based on this information and the definition of the decision-making process, the module establishes the lineage relationships between fields and models, and between models and decision flows. For example, it records which model calculated a certain "model-related field" and in which decision flows that model was used.

[0057] Using the lineage diagram, the module can monitor the usage of model-related fields in real time. When a model field currently in use needs to be modified or deleted, the module will issue an alert to prevent the operation, thus ensuring the stability and reliability of the system. For example, if a decision-making process is using a model-related field, and a user attempts to delete the model corresponding to that field, the module will prompt the user that the model is currently in use and cannot be deleted.

[0058] The metadata lineage analysis module provides complete lineage tracing and detailed execution log recording capabilities. This information enables end-to-end visualization, tracing back from the decision outcome to the original model calculation, meeting internal audit and regulatory compliance requirements. For example, when an erroneous risk decision occurs, the lineage diagram and execution logs can be used to quickly pinpoint which model's calculation went wrong, and what the model's input data and calculation process were.

[0059] This disclosure provides a method for ensuring the consistency and reliability of risk control model calculations, which utilizes the aforementioned system for ensuring the consistency and reliability of risk control model calculations, and includes the following steps.

[0060] Model Deployment Phase: After completing model creation, training, and validation, the model lifecycle management platform initiates a model deployment request. The platform's built-in model validator performs multi-dimensional checks on the model. Upon successful validation, it sends a model registration event notification to the system field management center. Based on the notification, the system field management center automatically creates the corresponding "model association fields" and updates the field cache. Simultaneously, it notifies the metadata lineage analysis module to update the lineage graph.

[0061] Decision-making process execution phase: After receiving an external decision request, the core service of the decision engine parses the decision process definition file and sequentially calls the corresponding strategy rules and models. When it needs to obtain the value of a certain "model-related field," it sends a request to the unified data evaluation framework. Following the request processing flow described above, the unified data evaluation framework obtains the model score through dependency resolution and recursive evaluation mechanisms and returns it to the core service of the decision engine. Based on the model score and other strategy rules, the core service of the decision engine makes the final decision result.

[0062] System monitoring and maintenance phase: The model lifecycle management platform monitors the operation of published models in real time. When performance degradation or anomalies are detected, relevant personnel are promptly notified for handling. The metadata lineage analysis module continuously tracks the usage of associated fields in the model to ensure system consistency and stability. When model upgrades or field modifications are required, all modules work together to ensure the legality and security of the operation.

[0063] This disclosure also proposes a system for ensuring the consistency and reliability of risk control model calculations, which utilizes the aforementioned method for ensuring the consistency and reliability of risk control model calculations.

[0064] The above content is merely an embodiment of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can improve and implement this solution based on the guidance provided in this application and their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A system for ensuring the consistency and reliability of risk control model calculations, characterized in that, include: The core service of the decision engine is used to parse and execute decision process files, and call strategy rules and models during the execution process; Built-in process parser and unified data evaluation framework; The unified data evaluation framework provides a unified interface for obtaining model scores using the chain of responsibility pattern. Upon receiving a model score request, it performs dependency resolution, recursive evaluation, cache read / write, and model calculation service calls, and returns a consistent model score. The model lifecycle management platform is used to manage models, including model creation, training, validation, deployment and real-time monitoring. Before deployment, it performs input parameter type validation, output field checking, dependency loop detection and hierarchical depth validation through the model validator. The system field management center is used to automatically generate globally unique and named model association fields according to preset technical constraints after receiving a model registration event, and to update the field cache synchronously. The metadata lineage analysis module is used to build and maintain a field-model-decision process lineage diagram in real time, and to prevent deletion or modification operations when a model or field is detected to be in use.

2. The system for ensuring the consistency and reliability of risk control model calculations according to claim 1, characterized in that, The unified data evaluation framework includes: The caching service is used to store the calculation results of the model's associated fields and set expiration policies; Model metadata service, used to provide the model's input parameter definitions and dependencies; The model computation service is used to perform model inference and return scores after all input parameters are available; The controller is called recursively. When it is found that the input parameters depend on the fields associated with other models, it recursively triggers its own process until all the underlying values ​​are obtained.

3. The system for ensuring the consistency and reliability of risk control model calculations according to claim 2, characterized in that, A unified data evaluation framework returns the cached value directly when the cache is hit, and executes the following steps sequentially when the cache is missed: Call the model metadata service to obtain the input parameter definitions; Recursively parse and calculate all dependent fields; The model score is obtained by calling the model calculation service. Write the model to the cache and return it.

4. The system for ensuring the consistency and reliability of risk control model calculations according to claim 3, characterized in that, The recursive parsing algorithm is as follows: if the input parameters of the target model's associated field still depend on another model's associated field, the framework first recursively calculates the value of the other model's associated field until all underlying dependent variables or model scores are calculated.

5. A system for ensuring the consistency and reliability of risk control model calculations according to claim 1, characterized in that, The naming rule for model-related fields generated by the system field management center is: S_F_FM98{model name}_{model version}, and the model UUID, version number, and output field mapping relationship are recorded in the field attributes.

6. The system for ensuring the consistency and reliability of risk control model calculations according to claim 1, characterized in that, The metadata lineage analysis module dynamically updates the lineage graph by listening to model registration events and decision process definition change events. When it receives a request to delete or modify a referenced model or field, it immediately returns a prohibition signal and outputs an alarm message.

7. The system for ensuring the consistency and reliability of risk control model calculations according to claim 1, characterized in that, The process parser built into the core service of the decision engine converts the decision process file into a directed acyclic graph structure and executes each node in the topological order to ensure that the model call order and input data are consistent when the same decision process is executed multiple times.

8. The system for ensuring the consistency and reliability of risk control model calculations according to claim 1, characterized in that, Technical constraints include: model input parameters are only allowed to come from basic variables or other model outputs, model dependency depth does not exceed a preset level, and circular dependencies between models are not allowed.

9. The system for ensuring the consistency and reliability of risk control model calculations according to claim 1, characterized in that, The decision engine core service, model lifecycle management platform, system field management center, unified data evaluation framework and metadata lineage analysis module communicate with each other through an asynchronous message mechanism to notify each other of events.

10. A method for ensuring the consistency and reliability of risk control model calculations, employing a system for ensuring the consistency and reliability of risk control model calculations as described in any one of claims 1-9, comprising a model release phase, a decision process execution phase, and a system monitoring and maintenance phase.

Images