Rule-based and machine learning-based auditing methods, systems, devices, and storage media

By combining rule feature vectors and model feature vectors, automated review of merchant information changes, financial transaction anti-fraud, and enterprise qualification verification has been achieved, solving the problems of long processing time and unstable quality of manual review, and improving review efficiency and consistency.

CN123118015APending Publication Date: 2026-11-17GUIYANG SHIJIHENGTONG TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611262839.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In manual review, the process of manually comparing each field and searching across systems results in long review times for a single entry. The review standards rely on personal experience and subjective judgment, leading to low efficiency and inconsistent quality.

Method used

By extracting rule feature vectors and model feature vectors, automated judgment is achieved through rule matching and violation probability prediction. The dual joint judgment is performed by combining the matching judgment of rule feature vectors and the violation probability prediction of model feature vectors, replacing manual field-by-field comparison and subjective judgment.

Benefits of technology

It has achieved fully automated judgment from data to review conclusion, improving review efficiency and quality, and reducing reliance on personal experience and subjective judgment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118015A_ABST
    Figure CN123118015A_ABST
Patent Text Reader

Abstract

This invention proposes a rule-based and machine learning-based auditing method, system, device, and storage medium, relating to the field of information auditing. The method includes: acquiring multi-source data to be audited; obtaining rule feature vectors and model feature vectors from the data to be audited, representing the conditional judgment results of each field and semantic and business statistical information related to violation judgment, respectively; matching the rule feature vectors with a preset rule base to obtain a compliant or non-compliant rule judgment result; inputting the model feature vectors into a preset classification model to predict the violation probability; determining a violation when the rule judgment result is non-compliant and the violation probability is greater than or equal to a first preset probability, and determining compliance when the rule judgment result is compliant and the violation probability is less than or equal to a second preset probability. This fully automated auditing method, using both rule matching and violation probability determination, replaces manual auditing, avoiding the influence of auditors' personal experience and subjective judgment on the audit, and improving auditing efficiency and quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of information review technology, and more specifically, to a review method, system, device, and storage medium based on rules and machine learning. Background Technology

[0002] In business scenarios such as merchant information change review, financial transaction anti-fraud review, and enterprise qualification review, the review process is generally completed manually. After receiving a task to be reviewed, the auditor manually logs into the relevant business system, retrieves the data to be reviewed item by item, and compares the original values ​​of each field with the new values ​​after the changes (such as enterprise name, unified social credit code, legal representative's ID number, business scope, etc.). During the comparison process, the auditor needs to switch back and forth between multiple independent systems or documents, including historical review records, external data sources, and the latest policy documents, to confirm whether the format of each data item is compliant, whether the content is consistent, and whether the changes are reasonable. Finally, the auditor synthesizes all the above information and, based on their own experience, makes a review conclusion of approval, rejection, or requiring supplementary materials, and enters the review conclusion and brief remarks into the system.

[0003] However, this manual review method has the following problems in actual operation: manual comparison of each field and cross-system lookup lead to long review times for a single item, and the review efficiency is difficult to keep up with the continuous growth of business volume; the review standards rely entirely on the reviewer's personal experience and subjective judgment, and the consistency of conclusions between different reviewers or even the same reviewer at different time periods is poor, resulting in large fluctuations in review quality; thus, the review work relies heavily on personal experience and personal ability, and the efficiency and quality of information review need to be improved. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a rule-based and machine learning-based auditing method, system, device, and storage medium to solve the technical problem that auditing work heavily relies on personal experience and ability, and the efficiency and quality of information auditing need to be improved.

[0005] To achieve the above objectives, in a first aspect, embodiments of the present invention provide a review method based on rules and machine learning. The method includes: acquiring multi-source data to be reviewed; extracting rule feature vectors and model feature vectors from the data to be reviewed, wherein the rule feature vectors represent the conditional judgment results of each field in the data to be reviewed, and the model feature vectors represent semantic information and business statistical information related to violation judgment in the data to be reviewed; matching the rule feature vectors with a preset rule base to obtain a rule judgment result, wherein the rule judgment result includes whether the data to be reviewed is compliant or non-compliant; inputting the model feature vectors into a preset classification model to predict the probability of non-compliance, thereby obtaining the probability of non-compliance of the data to be reviewed; determining that the data to be reviewed is non-compliant when the rule judgment result is non-compliant and the probability of non-compliance is greater than or equal to a first preset probability; and determining that the data to be reviewed is compliant when the rule judgment result is compliant and the probability of non-compliance is less than or equal to a second preset probability, wherein the second preset probability is less than the first preset probability.

[0006] In this embodiment, rule feature vectors and model feature vectors are extracted from multi-source data to be reviewed. The rule feature vectors are used to directly match and determine whether the data is compliant or non-compliant. The model feature vectors are then used to calculate the probability of non-compliance. This dual determination of compliance and non-compliance is achieved through both rule matching and non-compliance probability assessments. Thus, rule feature vector matching replaces manual field-by-field comparison, model feature vector non-compliance probability prediction replaces manual cross-system searches and subjective judgment, and the combined determination of rule matching and non-compliance probability replaces the single decision-making process dominated by the auditor's personal experience. This fully automates the process from data to review conclusion, avoiding the influence of auditor's personal experience and subjective judgment on the review results, and improving the efficiency and quality of information review.

[0007] In some embodiments, extracting rule feature vectors and model feature vectors from the data to be reviewed includes: extracting keyword features, numerical threshold features, and logical association features from the data to be reviewed to form the rule feature vector; extracting semantic features from the text data of the data to be reviewed, and obtaining statistical features based on historical business data associated with the data to be reviewed to form the model feature vector.

[0008] This setup ensures that each feature in the rule feature vector corresponds to the judgment criterion of the condition field in a rule, enabling rule matching to align item by item and make accurate comparisons; it also allows the model feature vector to be input into a preset classification model, where the probability of violation is calculated based on the semantic information of the data itself and business statistical information.

[0009] In some embodiments, matching the rule feature vector with a preset rule base to obtain a rule determination result includes: comparing the condition determination result of the rule feature vector with the conditions of each rule in the preset rule base; each rule in the preset rule base contains conditions and a corresponding determination conclusion; determining the rule whose conditions are met as the triggered rule; and obtaining the rule determination result based on the determination conclusion corresponding to the triggered rule.

[0010] This setup ensures that every rule determination has a clear triggering basis and conclusion source by comparing the condition judgment results in the rule feature vector with the conditions of each rule in the preset rule base, thus avoiding the problem of rule application standards varying from person to person under manual review.

[0011] In some embodiments, the step of inputting the model feature vector into a preset classification model to predict the probability of violation and obtain the probability of violation of the data to be reviewed includes: performing a weighted operation on the values ​​of each feature in the model feature vector and the corresponding parameters in the preset classification model to obtain a judgment score; and mapping the judgment score to obtain the probability of violation of the data to be reviewed.

[0012] This setup calculates the probability of violation by analyzing the values ​​of each feature in the model's feature vector, and then comprehensively judges the probability of violation of the data to be reviewed. This replaces the subjective evaluation of multi-dimensional information by human reviewers, thus improving the accuracy of the review.

[0013] In some embodiments, the rule determination result further includes suspected violation, and the method further includes: when the rule determination result is a violation and the violation probability is less than the first preset probability, the rule determination result is compliant and the violation probability is greater than the second preset probability, or the rule determination result is a suspected violation, a weighted calculation is performed on the determination value corresponding to the rule determination result and the violation probability based on a preset association strength and a preset scenario weight to obtain a determination score; the preset association strength characterizes the frequency of the rule triggered in the rule determination result co-occurring with each feature of the model feature vector, and the preset scenario weight characterizes the respective weights of the rule determination result and the violation probability in the current review scenario; when the determination score is greater than or equal to a third preset probability, the data to be reviewed is determined to be in violation, the third preset probability being less than the first preset probability and greater than the second preset probability; when the determination score is less than the third preset probability and greater than the second preset probability, the data to be reviewed is determined to be a suspected violation.

[0014] This setup allows for the calculation of a comprehensive judgment score based on the suspected violation result and the violation probability through preset association strength and preset scenario weights when a rule judgment result is found to be suspected of being a violation. This score is then compared with a third preset probability to further determine whether the audited data is a violation or a suspected violation. If it is suspected of being a violation, it is sent to a human review team for further judgment on the suspected violation.

[0015] In some embodiments, the method further includes: selecting target audit data from the audited data that has been audited, where the violation probability is greater than or equal to a first preset probability and the rule judgment result is compliant or suspected of being non-compliant; determining high-frequency feature combinations from the features included in the rule feature vector of the target audit data, wherein the high-frequency feature combinations represent multiple feature value conditions that occur together in the target audit data at a frequency greater than a preset frequency threshold; wherein each feature value condition is a judgment condition in which the value of a feature satisfies a preset value range; generating candidate audit rules by using the multiple feature value conditions as triggering conditions and violations as judgment conclusions; and supplementing the candidate audit rules to the preset rule library.

[0016] This setup allows unknown violation patterns discovered by the preset classification model to be automatically converted into executable audit rules. The preset rule base can be continuously expanded from historical audit data, overcoming the shortcomings of the rule base update lag and limited coverage under the manual rule writing method.

[0017] In some embodiments, the method further includes: from the audited audited data, selecting audited data whose rule judgment result is illegal and has not been rejected by manual review as first audited data; selecting audited data whose rule judgment result is compliant and whose illegal probability is less than a third preset probability and greater than a second preset probability as second audited data; and selecting audited data whose rule judgment result is illegal but has been confirmed as compliant by manual review as third audited data; and incrementally training the preset classification model based on the first audited data, the second audited data, and the third audited data to obtain an updated preset classification model.

[0018] This setup allows the high-confidence judgment results of the preset rule base to drive the parameter updates of the preset classification model, enabling the preset classification model to continuously learn from the trial experience of the preset rule base and reduce its dependence on manually labeled samples.

[0019] Secondly, embodiments of the present invention provide a review system based on rules and machine learning. The system includes: an acquisition module for acquiring multi-source data to be reviewed; extracting rule feature vectors and model feature vectors from the data to be reviewed, wherein the rule feature vectors represent the conditional judgment results of each field in the data to be reviewed, and the model feature vectors represent semantic information and business statistical information related to violation judgment in the data to be reviewed; a processing module for matching the rule feature vectors with a preset rule base to obtain a rule judgment result, wherein the rule judgment result includes whether the data to be reviewed is compliant or non-compliant; inputting the model feature vectors into a preset classification model to predict the probability of non-compliance, thereby obtaining the probability of non-compliance of the data to be reviewed; and a judgment module for determining that the data to be reviewed is non-compliant when the rule judgment result is non-compliant and the probability of non-compliance is greater than or equal to a first preset probability; and determining that the data to be reviewed is compliant when the rule judgment result is compliant and the probability of non-compliance is less than or equal to a second preset probability, wherein the second preset probability is less than the first preset probability.

[0020] Thirdly, embodiments of the present invention provide an electronic device, including a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor can execute the computer program to implement the rule-based and machine learning-based auditing method as described in the first aspect.

[0021] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the rule-based and machine learning-based auditing method as described in the first aspect.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating the steps of a rule-based and machine learning-based auditing method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the rule-based and machine learning-based review method provided in this embodiment of the invention; Figure 3The flowcharts for forward and reverse feedback iterative optimization provided in the embodiments of the present invention; Figure 4 A schematic diagram of the functional modules of the rule-based and machine learning-based review system provided in an embodiment of the present invention; Figure 5 A block diagram of an electronic device provided in an embodiment of the present invention.

[0025] Icons: 1000 - Rule-based and machine learning-based auditing system; 1100 - Acquisition module; 1200 - Processing module; 1300 - Judgment module; 2000 - Electronic device; 2100 - Processor; 2200 - Memory; 2300 - Bus; 2400 - Communication interface. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0029] As described in the background section, manual review methods have the following problems in actual operation: manual comparison of each field and cross-system queries result in long review times for each item, and the review efficiency is difficult to keep up with the continuous growth of business volume; the review standards rely entirely on the reviewer's personal experience and subjective judgment, and the consistency of conclusions between different reviewers or even the same reviewer at different times is poor, resulting in large fluctuations in review quality; thus, the review work relies heavily on personal experience and ability, and the efficiency and quality of information review need to be improved.

[0030] To address this, this invention provides a rule-based and machine learning-based auditing method. It replaces manual field-by-field comparison with rule feature vector matching, replaces manual cross-system review and subjective judgment with violation probability prediction using model feature vectors, and replaces the single decision-making process dominated by the auditor's personal experience with a dual joint determination of rule matching conclusions and violation probabilities. This achieves a fully automated judgment process from data to audit conclusion, avoiding the influence of auditor's personal experience and subjective judgment on the audit results, and improving the efficiency and quality of information auditing. (See also...) Figure 1 and Figure 2 , Figure 1 This invention provides a step diagram of a rule-based and machine learning-based auditing method. Figure 2 The flowchart of the rule-based and machine learning-based auditing method provided in the embodiment of the present invention includes steps S100 to S600.

[0031] S100: Obtain multi-source data awaiting review.

[0032] In this embodiment, the sources of the multi-source data to be reviewed can include change requests pushed in real time by the business system via a RESTful API, batch review tasks transmitted asynchronously via a message queue, and third-party verification data synchronized from external data sources. Specifically, the data types of the data to be reviewed include text data, structured business data, and semi-structured data. Next, the text data needs to undergo noise reduction, word segmentation, stop word removal, and homophone / variant word normalization; the structured data needs to undergo missing value imputation, outlier detection, and format standardization; and the image data needs to undergo feature extraction (such as basic features of convolutional neural networks).

[0033] For example, taking the scenario of merchant information change review as an example, text data can be a textual description and remarks of the business scope, structured business data can be fields such as enterprise name, unified social credit code, legal person name, legal person ID number and registered capital, and semi-structured data can be JSON format comparison information of fields before and after the change.

[0034] S200. Extract rule feature vectors and model feature vectors from the data to be reviewed. The rule feature vectors represent the condition judgment results of each field in the data to be reviewed, and the model feature vectors represent the semantic information and business statistics related to the violation judgment in the data to be reviewed.

[0035] In this embodiment, the process of extracting rule feature vectors and model feature vectors from the data to be reviewed is performed simultaneously. For the rule feature vectors, condition judgment results are extracted from each field of the data to be reviewed. For example, for the legal person's ID number field, the process is to extract whether the field has been changed and whether the changed format is legal; for the declared amount field, the process is to extract whether the amount exceeds the preset budget limit and whether the approval process has been completed. These condition judgment results are discrete and enumerable, and each condition judgment result directly relates to the condition field of the corresponding rule in the preset rule base.

[0036] For the model feature vector, semantic features corresponding to semantic information are extracted from the text data of the data to be reviewed. Simultaneously, statistical features corresponding to business statistics are obtained based on historical business data associated with the data to be reviewed. Semantic features include, for example, encoding the business scope text into a high-dimensional semantic vector using a pre-trained model. Statistical features include, for example, the number of times the object to be reviewed has submitted changes in the past six months, the number of times it has been rejected historically and the rejection rate, and the total number of fields involved in the current change. The extracted rule feature vector and model feature vector serve as inputs for rule matching and violation probability prediction, respectively.

[0037] S300. Match the rule feature vector with the preset rule base to obtain the rule judgment result, which includes whether the data to be reviewed is compliant or non-compliant.

[0038] In this embodiment, the preset rule base stores multiple predefined audit rules. Each rule includes a condition part and a corresponding judgment conclusion. The condition part is a combination of several condition judgment requirements, and the judgment conclusion is compliant or non-compliant.

[0039] The matching process is as follows: Each condition judgment result in the rule feature vector is compared with the conditions of each rule in the preset rule library. When all conditions of a rule are met by the condition judgment results of the current data to be reviewed, that rule is triggered. The rule judgment result is derived based on the judgment conclusions corresponding to the triggered rules. When all triggered rules point to a violation conclusion, the rule judgment result is a violation; when all triggered rules point to a compliance conclusion, the rule judgment result is compliance. Rules in the preset rule library can be loaded according to priority, with higher-priority rules being matched first. The matching process for subsequent rules can be terminated after a clear conclusion is found, improving matching efficiency.

[0040] S400. Input the model feature vector into the preset classification model to predict the probability of violation and obtain the probability of violation of the data to be reviewed.

[0041] In this embodiment, the preset classification model is a pre-trained classification model whose internal parameters are trained based on historical audit data labeled with compliance or violation conclusions. After the model feature vector is input into the preset classification model, the values ​​of each feature in the model feature vector are weighted with the corresponding parameters in the model to obtain the judgment score, and then the judgment score is mapped to a violation probability between 0 and 1.

[0042] For example, in the model feature vector of a piece of data to be reviewed, the semantic features are encoded into a high-dimensional vector by the pre-trained model and concatenated with statistical features. After weighted operation with the model parameters, an original judgment score is obtained. This original judgment score is converted into a violation probability of 0.92 through a mapping function, indicating that the preset classification model believes that the data to be reviewed has a 92% probability of being in violation. The closer the violation probability is to 1, the more inclined the model is to judge it as in violation; the closer it is to 0, the more inclined the model is to judge it as compliant.

[0043] S500. If the rule judgment result is a violation and the violation probability is greater than or equal to the first preset probability, the data to be reviewed is determined to be a violation.

[0044] In this embodiment, the first preset probability is a relatively high probability threshold, such as 0.9. When the rule judgment result is a violation and the violation probability is greater than or equal to the first preset probability, it means that the rule side determines that the data to be reviewed is a violation, and the preset classification model side also determines that the data to be reviewed is a violation with a very high confidence level. The conclusions of the two channels are consistent, and their respective confidence levels are both at a high level. At this time, the data to be reviewed is directly determined to be a violation without entering a further collaborative calculation process. This scenario handles samples that have been determined to be violations by the preset rule base and the preset classification model with a high degree of confidence, outputting the conclusion with the shortest judgment path.

[0045] S600. If the rule determination result is compliant and the probability of violation is less than or equal to the second preset probability, the data to be reviewed is determined to be compliant; the second preset probability is less than the first preset probability.

[0046] In this embodiment, the second preset probability is a low probability threshold, such as 0.1. When the rule judgment result is compliant and the probability of violation is less than or equal to the second preset probability, it means that the rule side determines that the data to be reviewed is compliant, and the preset classification model side also determines that the data to be reviewed is compliant with an extremely low probability of violation. Both channels conclude that the data is compliant and their respective confidence levels are at a high level. At this time, the compliance of the data to be reviewed is directly determined, and there is no need to enter a further collaborative calculation process. A middle interval is formed between the first preset probability and the second preset probability, and this middle interval contains situations where there may be suspected violations.

[0047] In some embodiments, for step S200, one possible implementation of the present invention is that step S200 includes sub-steps S210~S220: S210. Extract keyword features, numerical threshold features, and logical association features from the data to be reviewed to form a rule feature vector.

[0048] In this embodiment, keyword features, numerical threshold features, and logical association features are three types of conditional judgment results extracted from the fields of the data to be reviewed. Keyword features indicate whether each field in the data to be reviewed contains specific keywords, such as whether the business scope field contains "finance" or "credit," or whether the company name field contains restrictive words. Numerical threshold features indicate whether the value of each field exceeds a preset threshold, such as whether the registered capital exceeds the preset upper limit for the corresponding industry, or whether the number of fields involved in the change exceeds a preset quantity threshold. Logical association features indicate whether multiple fields satisfy a preset logical relationship, such as a change in the legal representative's name but not their ID number constituting one logical association feature, or a change in the business scope to a high-risk industry without a corresponding increase in registered capital constituting another logical association feature. These three types of features make conditional judgments on each field of the data to be reviewed from different dimensions, together forming a rule feature vector.

[0049] S220. Extract semantic features from the text data of the data to be reviewed, and obtain statistical features based on the historical business data associated with the data to be reviewed, so as to form the model feature vector.

[0050] In this embodiment, semantic features are extracted from the text data of the data to be reviewed. The extraction method can be to input the text data into a pre-trained large model (such as a text BERT model or an image ViT model) and take the output vector of the model as the semantic features. For example, the textual description of the business scope can be input into the BERT model, and the output vector at a special marked position can be taken as the deep semantic representation of the text. This vector can capture the degree of proximity of the text to high-risk industries or abnormal business activities in the semantic space.

[0051] Statistical features are obtained based on historical business data associated with the data to be reviewed. This data is acquired by using the object identifier (such as merchant ID or applicant ID) in the data to be reviewed as an index to query historical behavioral statistics of that object from the historical business database. Examples include the number of change requests in the past six months, the number of historical rejections, the rejection rate, and the average number of fields involved in a single change. Semantic features and statistical features are concatenated to form the model's feature vector, which serves as the input for the pre-defined classification model, providing information on the deep semantics of the text and the statistical patterns of behavior.

[0052] In some embodiments, for step S300, one possible implementation of the present invention is that step S300 includes sub-steps S310~S330: S310. Compare the condition judgment result of the rule feature vector with the conditions of each rule in the preset rule base; each rule in the preset rule base contains conditions and corresponding judgment conclusions.

[0053] In this embodiment, each condition judgment result in the rule feature vector is compared item by item with the conditions of the rule currently being compared in the preset rule base. Each rule in the preset rule base adopts an IF-THEN structure, where the IF part is the triggering condition of the rule, which is composed of several condition judgment requirements combined with AND logic, and the THEN part is the judgment conclusion corresponding to the rule. For example, one rule is IF legal person's name changed AND legal person's ID number changed AND ID number format valid THEN compliant, and another rule is IF legal person's name changed AND legal person's ID number not changed THEN non-compliant.

[0054] Rules in the preset rule base can be pre-written by business experts or automatically generated and added to the base by subsequent feedback mechanisms. Rules can be categorized and managed as basic rules, temporary rules, and adaptive optimization rules. Different types of rules can be assigned different priorities, and they are loaded in priority order during matching.

[0055] S320. The rule whose condition is met is identified as the triggering rule.

[0056] In this embodiment, for the rule currently being compared in the preset rule base, if all conditions of the rule feature vector are satisfied, the rule is determined to be triggered. For example, if the condition of the data to be reviewed is that the legal representative's name has changed, the legal representative's ID number has also changed, and the ID number format is valid, then all conditions of the rule in the rule base that states "IF legal representative's name changed AND legal representative's ID number changed AND ID number format is valid" are satisfied, and this rule is determined to be triggered. If any condition of a rule is not satisfied, the rule is not triggered, and the comparison continues with the next rule.

[0057] S330. Based on the judgment conclusion corresponding to the triggered rule, the rule judgment result is obtained.

[0058] In this embodiment, after all preset rules have been compared, the judgment conclusions corresponding to all triggered rules are collected. If the judgment conclusions of all triggered rules are compliant, the rule judgment result is compliant; if at least one triggered rule has a judgment conclusion of non-compliance, the rule judgment result is non-compliance. When multiple rules are triggered simultaneously and their conclusions are inconsistent, the final rule judgment result can be determined according to the priority of the rules, with the conclusion of the higher-priority rule taking precedence. The rule judgment result can also include a list of identifiers of the triggered rules and matching details for each triggered rule, providing a basis for manual review and decision-making.

[0059] In some embodiments, for step S400, one possible implementation of the present invention is that step S400 includes sub-steps S410~S420: S410. The values ​​of each feature in the model feature vector are weighted and calculated with the corresponding parameters in the preset classification model to obtain the judgment score.

[0060] In this embodiment, the preset classification model stores multiple pre-trained parameters, each corresponding one-to-one with a feature in the model's feature vector. The weighted calculation process involves multiplying the values ​​of each feature in the model's feature vector by their corresponding parameters, summing the results, and then adding the bias parameter to obtain a judgment score. The magnitude of the judgment score reflects the preset classification model's assessment of the data to be reviewed for potential violations; a higher judgment score indicates that the model considers the data more likely to be in violation.

[0061] Furthermore, the preset classification model can be a tree model or a neural network model trained based on a supervised learning algorithm. The parameters inside the model are learned during the training phase through a large amount of historical audit data labeled with compliance or violation conclusions. What is learned is the statistical mapping pattern between the values ​​of various features and the final audit conclusion.

[0062] S420. Map the judgment score to obtain the probability of violation of the data to be reviewed.

[0063] In this embodiment, the judgment score is mapped to a violation probability between 0 and 1 after being transformed by a mapping function. The purpose of the mapping function is to compress the range of judgment scores to the interval between 0 and 1, so that the output result can be interpreted as a probability value. The closer the violation probability is to 1, the more likely the preset classification model is to judge the data to be reviewed as violation; the closer it is to 0, the more likely it is to judge as compliant. The violation probability can also be accompanied by the model's decision-making basis, that is, the contribution information of each feature to the violation probability, making the prediction result of the violation probability interpretable.

[0064] In some embodiments, the rule determination result also includes suspected violations. For the rule-based and machine learning-based auditing method, one possible implementation of the present invention includes steps S700-S720: S700. When the rule judgment result is a violation and the violation probability is less than the first preset probability, the rule judgment result is compliant and the violation probability is greater than the second preset probability, or the rule judgment result is a suspected violation, the judgment value and violation probability corresponding to the rule judgment result are weighted and calculated based on the preset association strength and preset scenario weight to obtain the judgment score. The preset association strength represents the frequency of the occurrence of the rule triggered in the rule judgment result and various features of the model feature vector. The preset scenario weight represents the respective weight of the rule judgment result and the violation probability in the current review scenario.

[0065] In this embodiment, when the rule determination result is a violation but the violation probability does not reach the first preset probability, the rule determination result is compliant but the violation probability exceeds the second preset probability, or the rule determination result itself is a suspected violation, the rule determination result and the violation probability determination cannot be unified, thus entering the suspected violation situation. At this time, the rule determination result is first mapped to a determination value. The mapping method is as follows: violation is mapped to a first determination value, such as 1; compliance is mapped to a second determination value, such as 0; and suspected violation is mapped to an intermediate determination value, such as 0.5. The purpose of the determination value is to transform the rule determination result, originally a category label, into a numerical form that can be mathematically calculated with the violation probability.

[0066] The preset scenario weights are pre-configured based on the current review scenario. The proportions of rule judgment results and violation probability in the weighted calculation can differ under different review scenarios. For example, in the financial transaction anti-fraud review scenario, the rules are mostly hard thresholds clearly stipulated by regulators, so the weight of the rule judgment result can be set to 0.7, and the weight of the violation probability can be 0.3. In the text content review scenario, violation patterns are diverse and difficult to exhaustively list with a limited number of rules, so the weight of the violation probability can be set to 0.7, and the weight of the rule judgment result can be 0.3.

[0067] The preset association strength characterizes the frequency with which triggered rules and features in the model feature vector co-occur in historical review data. This preset association strength can be quantified and stored using an association matrix. The rows of the matrix correspond to rule identifiers in the preset rule base, and the columns correspond to model features in the model feature vector. The element values ​​in the matrix represent the frequency with which the corresponding rule and model feature are simultaneously triggered in historical review data. Higher element values ​​indicate a high degree of co-occurrence between the rule and the model feature in historical records, indicating strong consistency in their judgment logic. In weighted calculations, this can be used as a positive adjustment factor.

[0068] Furthermore, the specific method of weighted calculation is as follows: multiply the judgment value by the weight of the rule judgment result in the preset scenario weight to obtain the first term, multiply the violation probability by the weight of the violation probability in the preset scenario weight to obtain the second term, multiply the preset association strength by the preset association strength adjustment coefficient to obtain the third term, and add the three terms together to obtain the judgment score.

[0069] For example, P_final = W_scene * P_rule + (1 - W_scene) * P_model + α * Corr; where P_rule (judgment value) is mapped from the rule result (violation = 1, suspected = 0.5, compliance = 0), P_model is the violation probability output by the model, α is the preset association strength adjustment coefficient, and W_scene is the preset scene weight.

[0070] Final judgment: If P_final ≥ preset threshold (e.g., 0.6), 0.6-0.9 is judged as "violation"; otherwise, 0.1-0.6 is judged as "suspected violation" and pushed to manual review with P_final, Corr and other information.

[0071] S710. If the score is greater than or equal to the third preset probability, the data to be reviewed is determined to be in violation, where the third preset probability is less than the first preset probability but greater than the second preset probability.

[0072] In this embodiment, the third preset probability is the judgment threshold after weighted calculation. The third preset probability is less than the first preset probability and greater than the second preset probability, for example, 0.6. When the judgment score is greater than or equal to the third preset probability, although the rule judgment result and the violation probability fail to achieve the high confidence consistency required by the deterministic final judgment condition in the original conclusion, the comprehensive judgment score after weighted calculation has reached a level sufficient to identify it as a violation. At this time, the data to be reviewed is determined to be a violation. This addresses the situation where the rule judges a violation but the preset classification model shows that the violation probability is not high enough, and the conclusion is determined to be a violation by weighting the association strength and scenario weight. It also addresses the situation where the rule judges compliance but the preset classification model shows that the violation probability is very high, and the unilateral judgment bias of the rule is corrected by weighted calculation.

[0073] S720. If the score is less than the third preset probability but greater than the second preset probability, the data to be reviewed is suspected of being in violation.

[0074] In this embodiment, when the judgment score is lower than the third preset probability but higher than the second preset probability (between 0.1 and 0.6), and the weighted comprehensive score is in the middle range, it is insufficient to confirm a violation but not low enough to be directly considered compliant. At this point, the data to be reviewed is determined to be suspected of being non-compliant, and the judgment score, the preset correlation strength value, the triggered rule identifier, and the violation probability derived from the preset classification model are pushed to manual review. The manual reviewer can make a final review decision with the assistance of the aforementioned quantitative information, and the manual review conclusion can further serve as a data source for subsequent rule judgments and preset classification model feedback iterations.

[0075] In some embodiments, to make the rule determination more accurate, positive feedback is provided, using a preset classification model to adjust the preset rule base. (See [reference needed]). Figure 3 , Figure 3 The flowchart for the positive and negative feedback iterative optimization provided in this embodiment of the invention, and a possible implementation of the review method based on rules and machine learning, include steps S800-S830: S800. From the pending audit data that has been audited, select target audit data with a violation probability greater than or equal to the first preset probability and a rule judgment result of compliance or suspected violation.

[0076] In this embodiment, the pending review data that has completed the review refers to historical pending review data for which a final review conclusion has been generated. By periodically scanning the accumulated historical review data, review data with a violation probability greater than or equal to a first preset probability is selected as target review data. That is, the preset classification model identifies it as a violation with high confidence, but the rule judgment result is compliance or suspected violation, meaning that the preset rule base has failed to intercept it. Target review data essentially represents new types of violation cases that the preset classification model has accurately identified but the preset rule base does not yet have the ability to intercept. New judgment rules can be generated from the selected target review data to compensate for the deficiencies of the preset rule base.

[0077] S810. Determine high-frequency feature combinations from the features contained in the rule feature vector of the target audit data. The high-frequency feature combinations represent multiple feature value conditions that appear together in the target audit data at a frequency greater than a preset frequency threshold. Among them, each feature value condition is a judgment condition that the value of a feature meets the preset value range.

[0078] In this embodiment, the frequency of multiple feature values ​​appearing together in the target review data is statistically analyzed from the features contained in the rule feature vectors of each target review data. For example, in a batch of selected target review data, the feature value "user registration duration less than 7 days" appears frequently, "number of transactions per day greater than 10" also appears frequently, and "all transaction objects are unknown accounts" also appears frequently. Furthermore, if the frequency of these three feature values ​​appearing together in the same batch of target review data exceeds a preset frequency threshold (e.g., 80%), then these three feature values ​​are identified as a high-frequency feature combination. The preset value ranges for each feature value are automatically determined or manually configured during the statistical process based on the distribution of each feature value in the target review data. For example, the preset value range for registration duration is less than 7 days, and the preset value range for the number of transactions is greater than 10.

[0079] S820. Generate candidate review rules by taking multiple feature values ​​as trigger conditions and violations as judgment conclusions.

[0080] In this embodiment, each feature value condition in the high-frequency feature combination corresponds to an IF sub-condition in a rule. Therefore, multiple feature value conditions constitute the complete triggering condition of the rule, and the judgment conclusion is set as a violation. For example, the candidate review rule generated by combining the three feature value conditions in S810 is: IF User registration time less than 7 days AND Daily transaction count greater than 10 AND Transaction objects are all unknown accounts THEN Violation. Moreover, in the process of generating candidate review rules, the high-frequency feature combination can be further optimized by using a decision tree generation algorithm to extract a rule structure with clear branch logic from the feature combination, thereby improving the high accuracy of the candidate review rules.

[0081] S830. Add candidate review rules to the preset rule library.

[0082] In this embodiment, the generated candidate review rules are added to the preset rule base after manual review or automatic confirmation. Low-risk candidate review rules can be set to be directly added to the preset rule base after automatic confirmation; high-risk candidate review rules or those involving sensitive business scenarios require manual review before being added to the preset rule base. When a candidate review rule is added to the preset rule base, it is assigned an initial priority and categorized into adaptive optimization rules. Once added to the preset rule base, the candidate review rule takes effect in the next rule matching process (i.e., step S300) and participates in the rule determination of new data to be reviewed, enabling the preset rule base to continuously expand from historical review data.

[0083] In some embodiments, to make the model's judgments more accurate, reverse feedback is provided, using a preset rule base to adjust the preset classification model. (See [reference needed]). Figure 3 One possible implementation of the present invention, a rule-based and machine learning-based auditing method, includes steps S900-S910: S900. From the pending data that has been reviewed and approved, the pending data that is determined to be non-compliant by the rules and has not been rejected by manual review is taken as the first review data; the pending data that is determined to be compliant by the rules and has a non-compliance probability that is less than the third preset probability and greater than the second preset probability is taken as the second review data; and the pending data that is determined to be non-compliant by the rules but has been confirmed to be compliant by manual review is taken as the third review data.

[0084] In this embodiment, by comparing the rule-based judgment results with the manual review conclusions, three categories of review data with different training values ​​are automatically divided from the historical pending review data that has already been reviewed. The first review data consists of pending review data where the rule judgment result is a violation and the manual review did not reject it; these are samples where the rule judgment is accurate and serve as positive samples for incremental training. The second review data consists of pending review data where the rule judgment result is compliant and the probability of violation is at a low level (less than the third preset probability and greater than the second preset probability); these are samples that both the rule and the preset classification model consider compliant and serve as negative samples for incremental training. The third review data consists of pending review data where the rule judgment result is a violation but is confirmed as compliant after manual review; these are samples that the rule previously mistakenly blocked and serve as difficult examples for incremental training. These difficult examples reflect the weak recognition of marginal cases by the preset classification model and are given higher training weights in incremental training.

[0085] S910. Based on the first review data, the second review data, and the third review data, incremental training is performed on the preset classification model to obtain the updated preset classification model.

[0086] In this embodiment, incremental training employs a transfer learning strategy. It loads the currently running pre-defined classification model and freezes the parameters of the bottom layers responsible for general feature extraction, keeping them unchanged. Only the parameters of the top classification layer responsible for classification decisions are fine-tuned. The reason for freezing the bottom-level parameters is that they learn general feature representations across scenarios, such as text semantic encoding and basic statistical patterns, which should not change significantly due to the addition of incremental samples. Fine-tuning only the top-level parameters allows the pre-defined classification model to quickly adapt to changes in the business scenario reflected by newly added labeled samples.

[0087] Furthermore, during training, the third-level review data is assigned a higher training weight than the first and second-level review data. This forces the pre-defined classification model to focus on difficult scenarios where rules are prone to misinterpretation when updating parameters. After incremental training, a new version of the pre-defined classification model is generated. The new version can be compared and verified with the judgment effect of the old version in actual review traffic. After successful verification, it is smoothly switched to online, resulting in the updated pre-defined classification model.

[0088] In some other embodiments, the rule definition format for configuring and managing the preset rule base is not limited to JSON format; it can also be XML format, YAML format, or stored in a relational database table. Rule matching can also be achieved through database query statements. Besides directly writing configuration files, rules can be generated via a graphical interface using drag-and-drop, or rules can be written in a domain-specific language. For rule matching execution, in addition to the methods described in the above embodiments, any execution framework capable of dynamic rule loading and condition comparison can be used as an alternative, such as other open-source rule engines or a self-developed rule executor based on decision trees or finite state machines. Rule base version management can be achieved by recording the rule version number, effective time, and expiration time in a database table to implement rule timeline management, or by using blue-green deployment or canary release mechanisms to achieve smooth rule upgrades and rollbacks.

[0089] In some other embodiments, for the integration and deployment of the preset classification model, in addition to the deployment architecture described in the above embodiments, the preset classification model inference service can be deployed entirely on the backend to complete feature extraction and inference computation via API calls. Alternatively, lightweight models can be deployed on the edge or device side to achieve local inference and combine with cloud collaboration. A function computing architecture can also be used to call the model inference service on demand. Besides online real-time inference, the preset classification model inference method can combine offline batch processing to perform periodic batch predictions for non-real-time review tasks. Real-time incremental inference of the model can also be achieved through a streaming computing framework. In addition to single-model prediction, multi-model integration methods can employ model integration strategies to fuse the prediction results of multiple models. Dynamic routing and switching of models under different review scenarios can also be achieved through a model gateway.

[0090] In some other embodiments, the machine learning algorithm used for the preset classification model, besides tree models, can be trained using classification algorithms such as support vector machines, logistic regression, or Naive Bayes, or deep learning models such as multilayer perceptrons, convolutional neural networks, recurrent neural networks, or graph neural networks. For semantic feature extraction of text data, in addition to pre-trained language models, word vector models or text feature representations based on word frequency and inverse document frequency can be used. For situations where multimodal review data coexists, early fusion, late fusion, or hybrid fusion methods can be used to effectively integrate heterogeneous features.

[0091] In some other embodiments, besides the approach of combining scene weights with association strength as described in the above embodiments, the weighted calculation can be implemented by using Bayesian model averaging, evidence theory, or fuzzy logic reasoning to achieve a comprehensive judgment of the rule determination result and the probability of violation. For the quantitative assessment of uncertainty in the probability of violation prediction result, in addition to probability threshold determination, confidence interval estimation or Monte Carlo methods can be used for reliability assessment. Regarding the generation method of decision records, in addition to structured logs, decision paths can be stored using knowledge graphs.

[0092] In some other embodiments, for the high-frequency feature combination mining algorithm, in addition to the methods described in the above embodiments, other association rule mining algorithms or frequent pattern tree algorithms can be used. For sample labeling, besides automatic labeling, an active learning strategy can be used to prioritize high-value samples for manual labeling, or the output of the rule judgment results can be used as weak labels to reduce labeling costs. For updating the preset classification model, besides incremental training, model distillation can be used to compress new knowledge into a lightweight model, or a model replay mechanism can be used to perform periodic full training combining historical and new data. For triggering the feedback process, besides scheduled task scanning, an event-driven architecture can be used to automatically trigger it when the rule false interception rate or model uncertainty exceeds a threshold.

[0093] Based on the above method, embodiments of the present invention also provide a system corresponding to the above method, such as... Figure 4 As shown, Figure 4 This is a schematic diagram of the functional modules of the rule-based and machine learning-based review system provided in this embodiment of the invention. It should be noted that the basic principle and technical effects of the rule-based and machine learning-based review system 1000 provided in this embodiment are the same as those in the above method embodiments. For the sake of brevity, parts not mentioned in this embodiment can be referred to the corresponding content in the method embodiments.

[0094] In this embodiment, the rule-based and machine learning-based auditing system 1000 includes an acquisition module 1100, a processing module 1200, and a judgment module 1300.

[0095] The acquisition module 1100 acquires multi-source data to be reviewed; it extracts rule feature vectors and model feature vectors from the data to be reviewed. The rule feature vectors represent the condition judgment results of each field in the data to be reviewed, and the model feature vectors represent the semantic information and business statistics related to violation judgment in the data to be reviewed. It can be understood that the acquisition module 1100 is used to perform the above steps S100~S200.

[0096] Processing module 1200 is used to match the rule feature vector with a preset rule base to obtain a rule judgment result, which includes whether the data to be reviewed is compliant or non-compliant; and to input the model feature vector into a preset classification model to predict the probability of non-compliance, thereby obtaining the probability of non-compliance of the data to be reviewed. It can be understood that processing module 1200 is used to execute the above steps S300~S400.

[0097] The judgment module 1300 is used to determine that the data to be reviewed is in violation when the rule judgment result is violation and the violation probability is greater than or equal to a first preset probability; and to determine that the data to be reviewed is compliant when the rule judgment result is compliant and the violation probability is less than or equal to a second preset probability; where the second preset probability is less than the first preset probability. It can be understood that the judgment module 1300 is used to execute the above steps S500~S600.

[0098] In some embodiments, the acquisition module 1100 is used to extract keyword features, numerical threshold features, and logical association features from the data to be reviewed to form a rule feature vector; extract semantic features from the text data of the data to be reviewed, and obtain statistical features based on the historical business data associated with the data to be reviewed to form a model feature vector. It can be understood that the acquisition module 1100 is used to perform the above steps S210~S220.

[0099] In some embodiments, the processing module 1200 is used to compare the condition determination result of the rule feature vector with the conditions of each rule in the preset rule base; each rule in the preset rule base contains conditions and corresponding determination conclusions; the rule whose conditions are met is determined as the triggered rule; based on the determination conclusions corresponding to the triggered rule, the rule determination result is obtained. It can be understood that the processing module 1200 is used to perform the above steps S310~S330.

[0100] In some embodiments, the processing module 1200 is used to perform a weighted operation on the values ​​of each feature in the model feature vector and the corresponding parameters in the preset classification model to obtain a judgment score; and to map the judgment score to obtain the probability of violation of the data to be reviewed. It can be understood that the processing module 1200 performs the above steps S410~S420.

[0101] In some embodiments, the rule determination result also includes suspected violation. The determination module 1300 is used to perform a weighted calculation of the determination value and violation probability corresponding to the rule determination result based on a preset association strength and a preset scenario weight to obtain a determination score when the rule determination result is a violation and the violation probability is less than a first preset probability, the rule determination result is compliant and the violation probability is greater than a second preset probability, or the rule determination result is suspected violation. The preset association strength represents the frequency of the occurrence of the rule triggered in the rule determination result and various features of the model feature vector. The preset scenario weight represents the respective weights of the rule determination result and the violation probability in the current review scenario. When the determination score is greater than or equal to a third preset probability, the data to be reviewed is determined to be in violation, where the third preset probability is less than the first preset probability and greater than the second preset probability. When the determination score is less than the third preset probability and greater than the second preset probability, the data to be reviewed is determined to be suspected of being in violation. It can be understood that the determination module 1300 is used to perform the above steps S700~S720.

[0102] In some embodiments, the processing module 1200 is used to filter target audit data from the audited data that has been audited, where the probability of violation is greater than or equal to a first preset probability and the rule judgment result is compliant or suspected of violation; determine high-frequency feature combinations from the features contained in the rule feature vector of the target audit data, where the high-frequency feature combinations represent multiple feature value conditions that appear together in the target audit data at a frequency greater than a preset frequency threshold; wherein each feature value condition is a judgment condition where the value of a feature satisfies a preset value range; generate candidate audit rules by using multiple feature value conditions as trigger conditions and violation as the judgment conclusion; and supplement the candidate audit rules to the preset rule library. It can be understood that the processing module 1200 is used to perform the above steps S800~S830.

[0103] In some embodiments, the processing module 1200 is used to select, from the completed audited audited data, the audited data that is determined to be non-compliant by rules but has not been rejected by manual review as the first audited data, the audited data that is determined to be compliant by rules and has a non-compliance probability less than a third preset probability but greater than a second preset probability as the second audited data, and the audited data that is determined to be non-compliant by rules but has been confirmed as compliant by manual review as the third audited data; based on the first audited data, the second audited data, and the third audited data, the preset classification model is incrementally trained to obtain an updated preset classification model. It can be understood that the processing module 1200 is used to perform the above steps S900~S910.

[0104] Based on the same inventive concept disclosed above, the present invention also provides a block diagram of an electronic device 2000 performing the above method. Please refer to... Figure 5 , Figure 5This is a block diagram of an electronic device provided in an embodiment of the present invention. The electronic device 2000 includes a processor 2100, a memory 2200, a bus 2300, and a communication interface 2400. The processor 2100 and the memory 2200 are connected via the bus 2300, and the processor 2100 communicates with external devices via the communication interface 2400.

[0105] Processor 2100 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed through integrated logic circuits in the hardware of processor 2100 or through software instructions. The processor 2100 may be a general-purpose processor 2100, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0106] The memory 2200 is used to store computer programs. For example, the rule-based and machine learning-based auditing system 1000 in this embodiment of the invention includes at least one software function module that can be stored in the memory 2200 in the form of software or firmware. After receiving the execution instruction, the processor 2100 executes the program to implement the rule-based and machine learning-based auditing method in this embodiment of the invention.

[0107] The memory 2200 may include high-speed random access memory (RAM) or non-volatile memory. Optionally, the memory 2200 may be a storage device built into the processor 2100 or a storage device independent of the processor 2100.

[0108] Bus 2300 can be ISA bus 2300, PCI bus 2300 or EISA bus 2300, etc. Figure 5 It is indicated by only one double-headed arrow, but does not mean that there is only one bus 2300 or one type of bus 2300.

[0109] Electronic devices 2000 can be mobile phones, tablets, laptops, desktop computers, and other computer devices.

[0110] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon. When executed by processor 2100, the computer program implements the rule-based and machine learning-based auditing method described above. The computer-readable storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A rule-based and machine learning-based auditing method, characterized in that, The method includes: Obtain data from multiple sources awaiting review; From the data to be reviewed, rule feature vectors and model feature vectors are extracted. The rule feature vectors represent the condition judgment results of each field in the data to be reviewed, and the model feature vectors represent the semantic information and business statistics related to the violation judgment in the data to be reviewed. The rule feature vector is matched with a preset rule base to obtain a rule judgment result, which includes whether the data to be reviewed is compliant or non-compliant. The model feature vector is input into a preset classification model to predict the probability of violation, thereby obtaining the probability of violation of the data to be reviewed. If the rule determination result is a violation and the violation probability is greater than or equal to the first preset probability, the data to be reviewed is determined to be a violation. If the rule determination result is compliant and the violation probability is less than or equal to the second preset probability, the data to be reviewed is determined to be compliant; the second preset probability is less than the first preset probability.

2. The method according to claim 1, characterized in that, The step of extracting rule feature vectors and model feature vectors from the data to be reviewed includes: Keyword features, numerical threshold features, and logical association features are extracted from the data to be reviewed to form the rule feature vector; Semantic features are extracted from the text data of the data to be reviewed, and statistical features are obtained based on the historical business data associated with the data to be reviewed, so as to form the model feature vector.

3. The method according to claim 1, characterized in that, The step of matching the rule feature vector with a preset rule base to obtain the rule determination result includes: The condition determination result of the rule feature vector is compared with the conditions of each rule in the preset rule base; each rule in the preset rule base contains conditions and corresponding determination conclusions. The rule that is met is identified as the triggering rule; Based on the judgment conclusion corresponding to the triggered rule, the rule judgment result is obtained.

4. The method according to claim 1, characterized in that, The step of inputting the model feature vector into a preset classification model to predict the probability of violation, and obtaining the probability of violation of the data to be reviewed, includes: The values ​​of each feature in the model feature vector are weighted and calculated with the corresponding parameters in the preset classification model to obtain the judgment score. The judgment score is mapped to obtain the probability of violation of the data to be reviewed.

5. The method according to claim 1, characterized in that, The rule determination result also includes suspected violations, and the method further includes: When the rule judgment result is a violation and the violation probability is less than the first preset probability, the rule judgment result is compliant and the violation probability is greater than the second preset probability, or the rule judgment result is a suspected violation, a judgment score is obtained by weighting the judgment value corresponding to the rule judgment result and the violation probability based on preset association strength and preset scenario weight. The preset association strength represents the frequency with which the rule triggered in the rule judgment result and each feature of the model feature vector co-occur. The preset scenario weight represents the respective weights of the rule judgment result and the violation probability in the current review scenario. If the determination score is greater than or equal to the third preset probability, the data to be reviewed is determined to be in violation, wherein the third preset probability is less than the first preset probability and greater than the second preset probability; If the determination score is less than the third preset probability but greater than the second preset probability, the data to be reviewed is determined to be suspected of being in violation.

6. The method according to claim 1, characterized in that, The method further includes: From the data to be reviewed that has been reviewed, select target data for review where the probability of violation is greater than or equal to a first preset probability and the rule judgment result is compliant or suspected of being non-compliant; High-frequency feature combinations are determined from the features contained in the rule feature vector of the target review data. The high-frequency feature combinations represent multiple feature value conditions that occur together in the target review data at a frequency greater than a preset frequency threshold. Each feature value condition is a judgment condition that the value of a feature satisfies a preset value range. The candidate review rules are generated by using the multiple feature values ​​as triggering conditions and violations as the judgment conclusion. The candidate review rules are added to the preset rule library.

7. The method according to claim 1, characterized in that, The method further includes: From the pending data that has been reviewed, the pending data that is determined to be in violation by the rule and has not been rejected by manual review is taken as the first review data; the pending data that is determined to be compliant by the rule and whose probability of violation is less than the third preset probability and greater than the second preset probability is taken as the second review data; and the pending data that is determined to be in violation by the rule but has been confirmed to be compliant by manual review is taken as the third review data. Based on the first review data, the second review data, and the third review data, the preset classification model is incrementally trained to obtain the updated preset classification model.

8. A rule-based and machine learning-based auditing system, characterized in that, The system includes: The acquisition module acquires multi-source data to be reviewed; it extracts rule feature vectors and model feature vectors from the data to be reviewed. The rule feature vectors represent the condition judgment results of each field in the data to be reviewed, and the model feature vectors represent the semantic information and business statistics related to violation judgment in the data to be reviewed. The processing module is used to match the rule feature vector with a preset rule base to obtain a rule judgment result, the rule judgment result including whether the data to be reviewed is compliant or non-compliant; and to input the model feature vector into a preset classification model to predict the probability of non-compliance, thereby obtaining the probability of non-compliance of the data to be reviewed. The judgment module is used to determine that the data to be reviewed is in violation when the rule judgment result is a violation and the violation probability is greater than or equal to a first preset probability; and to determine that the data to be reviewed is compliant when the rule judgment result is compliant and the violation probability is less than or equal to a second preset probability; wherein the second preset probability is less than the first preset probability.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program that can be executed by the processor to implement the rule-based and machine learning-based auditing method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rule-based and machine learning-based auditing method as described in any one of claims 1-7.