Audit supervision method and system based on machine learning

Abstract

The invention discloses an audit supervision method and system based on machine learning, belongs to the technical field of intelligent supervision, and the method forms a whole-process closed loop of pre-event risk identification, in-event situation monitoring and post-event quality evaluation. An auditing subject, a customer and environment characteristics are used as input information in a risk identification model for modeling analysis, and an auditing opinion deviation evaluation result is introduced in a post-event stage for backtracking verification and feedback linkage, so that a unified analysis system penetrating through different supervision stages is formed, the problems of data splitting and staged fault in traditional research are eliminated, and the risk identification efficiency is improved. And continuity of audit risk supervision circulation is realized. Specifically, through combination of machine learning and an interpretability analysis method, organic fusion of audit risk identification, quality evaluation and supervision feedback is realized, the scientificity, systematicness and real-time performance of audit supervision can be improved while the operability of the model is ensured, and the method has significant innovation value and wide popularization prospects.

Description

Technical Field

[0001] This invention belongs to the field of intelligent supervision technology, and more specifically, relates to an audit supervision method and system based on machine learning. Background Technology

[0002] Against the backdrop of policies continuously promoting the digitalization and intelligentization of the supervision of the certified public accountant industry, regulatory authorities are gradually introducing data analysis and modeling methods to implement more systematic and proactive supervision of certified public accountant firms, branches, and audit projects. The regulatory model is evolving from experience-driven to data-driven, model-driven, and process-linked approaches.

[0003] Before undertaking an audit project, regulatory authorities typically conduct targeted oversight of accounting firms, branch offices, and signing auditors through methods such as qualification reviews, risk warnings, and the establishment of key monitoring lists. The aim is to identify audit entities that may pose high audit risks at the outset. During the audit process, regulatory authorities usually conduct quality checks through methods such as on-site sampling inspections, review of audit working papers, and special inquiries. After the audit report is released, regulatory authorities typically conduct quality evaluations and trace accountability based on ex-post results such as financial restatements, administrative penalties, and regulatory notices.

[0004] However, current regulatory practices show that relevant technical means are mostly focused on a single regulatory link or a single risk dimension, and have not yet formed a unified regulatory framework covering the entire process of audit projects, including pre-audit, during-audit, and post-audit. In situations involving multi-level audit entities and complex client characteristics, the existing regulatory technology system struggles to provide a structured explanation of the causes of audit failures and to dynamically assess and provide closed-loop feedback on audit quality, thus limiting the forward-looking nature, targeted approach, and systemic synergy of regulatory measures. Summary of the Invention

[0005] In response to the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides an audit supervision method and system based on machine learning. Its purpose is to solve the technical problem that existing supervision technology systems are unable to achieve a structured explanation of the causes of audit failures and dynamic evaluation and closed-loop feedback of audit quality.

[0006] To achieve the above objectives, according to one aspect of the present invention, a machine learning-based audit supervision method is provided, comprising:

[0007] S1: Before auditing the current project, the marginal contribution of each individual feature of the current project in the risk identification model is quantitatively interpreted to obtain the contribution analysis results of the individual features and the corresponding list of high-risk single-factor features; the individual features are the audit entity, audit client, or audit environment; the contribution analysis results of the individual features are aggregated at the entity level to obtain the entity risk score, and a pre-emptive entity risk warning is given based on the range in which the entity risk score is located; the risk identification model is trained using a large number of audit entities' historical practice samples; S2: During the audit of the current project, the joint marginal contribution between any two features of the current project in the risk identification model is quantitatively interpreted to obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features. Based on the interaction feature contribution analysis results, in-process project risk warning is carried out. S3: After the current project is audited, an audit opinion prediction model is constructed using the high-risk single-factor feature list and the high-risk interactive factor feature list. The model prediction model is used to obtain the predicted opinion for the current project. The predicted opinion is compared with the actual opinion issued to obtain the audit opinion deviation index. Post-audit quality risk warning is issued based on the range of the audit opinion deviation index.

[0008] Furthermore, preceding S1, the following is also included: The multidimensional data related to the audit history practice sample were formatted to obtain a three-dimensional fusion dataset with variables including the audit entity, audit client, and audit environment; The multi-level features within the audit subject in the three-dimensional fusion dataset are dimensionality-reduced in their respective dimensions to obtain the principal component feature vector set; The risk identification model is trained by taking the principal component feature vector set, audit client characteristics, and audit environment characteristics as inputs, and the audit failure probability of the historical audit practice sample as output.

[0009] Furthermore, the marginal contribution of each individual feature of the current project in the risk identification model is quantitatively interpreted to obtain the single feature contribution analysis results and the corresponding list of high-risk single-factor features, including: The risk identification model Decomposed into ; This is the baseline prediction value of the risk identification model for the current sample, used to represent the basic risk level without considering the influence of specific features. The number of features participating in the prediction of the risk identification model. This represents the marginal contribution value of the j-th feature to the prediction result of the risk identification model under the current sample, which is used to characterize the degree of influence of this feature on the risk level; Extract the marginal contribution of each variable to the predicted probability of the sample. The results of the single feature contribution analysis are obtained; the marginal contribution values ​​of each variable are sorted, and the top-ranked variables are included in the list of high-risk single-factor features.

[0010] Furthermore, based on the range in which the subject's risk score falls, pre-emptive subject risk warnings are issued, including: Using formula This represents the risk score of subject i. Standardize it to obtain the risk level of subject i ; The number of audit entities falling within the e-th risk range. This refers to the set of audit entities that belong to the e-th risk interval in the risk level classification. This is a set of risk characteristics at the audit entity level. This represents the marginal contribution of the j-th audit subject characteristic to the prediction result of the risk identification model corresponding to subject i; This represents the comprehensive risk score for the e-th risk interval, obtained by summarizing the contribution of the risk characteristics of the main entities. This is the average risk score of the audit entity obtained based on historical audit sample statistics. This is the standard deviation of the risk scores of the audited entities in the historical sample, used to standardize the risk scores; When the risk level of subject i exceeds the upper limit threshold When the risk level of subject i is within the range, a warning will be issued for that subject using the first method; At that time, the entity will be given a warning in the second manner.

[0011] Furthermore, the quantitative interpretation of the joint marginal contribution between any two features of the current project in the risk identification model yields the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features, including: The risk identification model Decomposed into The risk identification model The second decomposition is as follows: ; This is the baseline prediction value of the risk identification model for the current sample, used to represent the basic risk level without considering the influence of specific features. Representing variables With variables The interaction effect of the two measures their joint marginal contribution in predicting the probability of audit failure. Extract the joint marginal contribution of any two variables to the predicted probability of the sample. The single feature contribution analysis results are obtained; the joint marginal contribution values ​​corresponding to any two variables are sorted, and the top-ranked values ​​are included in the high-risk interaction factor feature list.

[0012] Furthermore, the step of conducting in-process project risk warning based on the interaction feature contribution analysis results includes: when the overall risk intensity is high... When the risk exceeds the first risk trigger threshold, the project will be identified as being in a high-risk interactive scenario, and a project risk warning signal will be automatically triggered. This represents the joint marginal contribution of the j-th and k-th features in audit project p to the prediction of audit failure probability in the risk identification model. It is used to measure the intensity of the superimposed impact of multi-feature interaction on project risk. The subscript p represents the audit project index, and the subscripts j and k represent the j-th and k-th features involved in the interaction, respectively.

[0013] Furthermore, the step of conducting in-process project risk warning based on the interaction feature contribution analysis results includes: when local extreme risk contribution When the risk exceeds the second risk trigger threshold, it is determined that the project has experienced a single-point feature-driven amplification of abnormal risks, and a project risk warning signal is automatically triggered. This represents the joint marginal contribution of the j-th and k-th features in audit project p to the prediction of audit failure probability in the risk identification model. It is used to measure the intensity of the superimposed impact of multi-feature interaction on project risk. The subscript p represents the audit project index, and j and k represent the j-th and k-th features involved in the interaction, respectively.

[0014] Furthermore, the post-audit risk warning based on the interval of the audit opinion deviation index includes: When the audit opinion deviation index of a certain audit project Exceeding the upper limit threshold When this happens, the project will be marked as the primary display item, and prompts for key review and accountability will be automatically pushed to it. when Located in the interval When necessary, the project will be marked as the second display method for enhanced review or included in the subsequent continuous tracking list; and The preset risk threshold; when Below the lower threshold When the third display method is selected, it is considered that the audit quality is normal.

[0015] Furthermore, the audit opinion deviation index is expressed as: ; in, Indicates sample In the The actual audit opinion obtained in the year; when the index is positive and the value is large, it indicates that the predicted opinion deviates significantly from the actual opinion, and there is a potential audit quality risk; conversely, it indicates that the audit opinion has a high degree of consistency with the model benchmark. The audit opinion prediction model is expressed as follows: , This represents the feature vector in the high-risk single-factor feature list and the high-risk interaction factor feature list.

[0016] According to another aspect of the present invention, a machine learning-based audit and oversight system is provided, comprising: The pre-audit module is used to quantify the marginal contribution of each individual feature of the current project in the risk identification model before the current project is audited, to obtain the contribution analysis results of the individual features and the corresponding list of high-risk single-factor features, to aggregate the contribution analysis results of the individual features at the subject level to obtain the subject risk score, and to issue a pre-audit risk warning based on the interval in which the subject risk score is located; the risk identification model is trained using a large number of audit subjects' audit history practice samples. The in-process early warning module is used to quantify and interpret the joint marginal contribution between any two features of the current project in the risk identification model during the audit of the current project, obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features, and conduct in-process project risk early warning based on the interaction feature contribution analysis results; The post-audit warning module is used to construct an audit opinion prediction model using the high-risk single-factor feature list and the high-risk interactive factor feature list after the current project has been audited; to obtain the model prediction opinion for the current project using the audit opinion prediction model; to compare the model prediction opinion with the actual opinion issued to obtain the audit opinion deviation index; and to issue a post-audit quality risk warning based on the range of the audit opinion deviation index.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) The machine learning-based audit supervision method provided by this invention forms a closed loop throughout the entire process of pre-event risk identification, in-event situation monitoring, and post-event quality evaluation. It uses the characteristics of the audit subject, client, and environment as input information in the risk identification model for modeling and analysis, and introduces the audit opinion deviation evaluation results in the post-event stage for retrospective verification and feedback linkage, forming a unified analysis system that runs through different regulatory stages, eliminating the problems of data fragmentation and stage-based discontinuity in traditional research, and realizing the continuity of audit risk supervision flow. Specifically, by combining machine learning with interpretability analysis methods, it realizes the organic integration of audit risk identification, quality evaluation, and regulatory feedback, which can improve the scientificity, systematicness, and real-time nature of audit supervision while ensuring the operability of the model. It has significant innovative value and broad application prospects, and is especially suitable for risk-oriented regulatory decision-making under conditions of limited regulatory resources.

[0018] (2) This invention has multi-level, structured risk identification and modeling capabilities. It extracts core structural features at three levels: accounting firm, branch office, and auditor, and combines machine learning methods to identify the risk interaction relationship between the audit entity and the client, thereby achieving refined risk characterization and high-risk scenario identification at the audit project level and improving the stability and generalization ability of the risk prediction model.

[0019] (3) The model in this invention is interpretable and can directly support regulatory decisions. The system uses the SHAP method to output the marginal impact and direction of each feature on the risk prediction results, enabling regulators to clearly identify the source and formation path of risks, thereby solving the problem that traditional machine learning models cannot be used for regulation due to "uninterpretable results".

[0020] (4) This invention constructs a post-event quality retrospective and feedback mechanism based on the “excellent audit opinion benchmark”. The system learns from historical high-quality audit samples and constructs an audit opinion benchmark model. By comparing the actual audit opinion of a new project with the model’s predicted opinion, an audit opinion deviation index is generated to identify potential quality anomalies.

[0021] (5) This invention enables the visualization of audit supervision results and intelligent operation support. The system generates red, yellow and green risk markers according to the audit opinion deviation index and displays the risk distribution of accounting firms, branch offices and auditors in layers on the supervision interface, so that supervisors can identify high-risk objects and trigger measures such as pre-interview, in-process spot checks or post-interview accountability, thereby improving the efficiency of regulatory resource allocation and the accuracy of supervision. Attached Figure Description

[0022] Figure 1 This is a flowchart of an audit supervision method based on machine learning provided in an embodiment of the present invention; Figure 2This is a schematic diagram of an audit and supervision system based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Example 1 The machine learning-based audit and supervision method provided in this embodiment, such as Figure 1 As shown, it includes: S1-S3.

[0025] S1: Before auditing the current project, the marginal contribution of each individual feature of the current project in the risk identification model is quantitatively interpreted to obtain the contribution analysis results of the individual features and the corresponding list of high-risk single-factor features; the individual features are the audit entity, audit client, or audit environment; the contribution analysis results of the individual features are aggregated at the entity level to obtain the entity risk score, and the entity risk warning is given in advance based on the range of the entity risk score; the risk identification model is trained using a large number of audit entities' audit history practice samples.

[0026] S2: During the audit of the current project, the joint marginal contribution between any two features of the current project in the risk identification model is quantitatively interpreted to obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features. Based on the interaction feature contribution analysis results, in-process project risk warnings are issued.

[0027] S3: After the current project is audited, construct an audit opinion prediction model using a list of high-risk single-factor characteristics and a list of high-risk interactive factors characteristics. Use the audit opinion prediction model to obtain the model prediction opinion for the current project. Compare the model prediction opinion with the actual opinion issued to obtain the audit opinion deviation index. Conduct post-event quality risk warning based on the range of the audit opinion deviation index.

[0028] Specifically, regarding pre-audit risk warning signals: Pre-audit risk assessment and screening of key regulatory targets, based on historical practice samples of audit entities, aggregates the contribution results of entity-type characteristics at the entity level to construct an audit entity risk score. As an optional implementation method, this score can be used to generate tiered risk warning signals, thereby calculating a risk index.

[0029] in, For the auditing entity A collection of historical samples; For the sample In features The absolute value of the contribution; This is the feature set of the main layer. For Standardization yields the relative risk level:

[0030] in, This represents the average risk intensity of all auditing entities. The standard deviation of the risk intensity of all audited entities is given. The standardized index can be used for cross-entity comparisons and to classify entity risk levels accordingly. Quantile division rules:

[0031] in, and This is a risk threshold, which regulatory authorities can dynamically adjust based on regulatory intensity and actual needs. When an entity's risk level exceeds the upper limit threshold ( When the risk level is within a certain range, the entity can be marked as red and a "Entity Risk Warning Signal" can be issued. This serves as a preliminary reminder during the audit project acceptance phase and as a constraint on project allocation. When this happens, the project can be marked with a yellow light, and it is recommended to strengthen the review or add it to the subsequent continuous tracking list.

[0032] (2) In-process project risk trigger signals: During the audit project execution phase, identify high-risk interaction scenarios under the characteristic combination of "audit entity × audit client" and output project risk trigger signals to support in-process supervision measures such as on-site spot checks and key reviews. First, retrieve the interaction SHAP value. ,in Indicates the current audit project. Indicates the characteristics of the auditing entity. This indicates the characteristics of the audit client. Triggers can be based on two risk intensity criteria, which can be used individually or simultaneously: First, triggered by cumulative interaction intensity. This involves considering the set of interaction characteristics of all possible stakeholders in the project – the customer. The overall risk intensity of the project is obtained by performing a cumulative weighted summation. :

[0033] in, Used to filter negative or risk-free contributions, retaining only the risk amplification portion. When When the risk threshold is exceeded (which can be set according to the actual situation of regulatory resources), the project will be judged as being in a high-risk interactive superposition scenario, and the project risk warning signal will be automatically triggered.

[0034] Second, further examine whether the combination of characteristics of a single entity and its customers contributes to localized extreme risks, and calculate:

[0035] when When the risk threshold is exceeded (which can be set according to the actual situation of regulatory resources), it is determined that the project has shown abnormal risk amplification driven by a single point feature, and a high-risk warning for the project is triggered.

[0036] When any of the above signals is triggered, the project will be automatically included in the project risk warning list. For example, it will be visualized as a yellow or red signal in the "Intelligent Audit Failure Supervision Panel" and simultaneously pushed to the supervisory personnel to support in-process supervision and handling strategies such as key spot checks, remote review or temporary on-site verification.

[0037] (3) Post-audit quality deviation signals. The audit opinion deviation index is retrieved and used as a basis for project quality retrospective analysis, key reviews, and liability determination. To facilitate a more intuitive presentation of audit failure risks, it can be based on... Set risk level:

[0038] in, and Full sample available The threshold for quantiles can be dynamically adjusted by regulatory authorities based on the intensity of regulation and actual needs. When an audit project... Exceeding the upper limit threshold ( When this happens, the project can be marked as red, and a notification for key review and accountability will be automatically sent; when Located in the interval When this happens, the project can be marked with a yellow light, indicating a need for enhanced review or inclusion in a follow-up monitoring list. Below the lower threshold ( When a signal is detected, it can be marked as green, indicating normal audit quality. These three types of signals are integrated and presented in the "Intelligent Audit Failure Monitoring Panel" in a three-column structure: the left column corresponds to audit entity risk warnings, the middle column to project risk triggers, and the right column to post-audit quality deviation feedback. It supports multi-dimensional searching, filtering, and exporting based on "accounting firm—branch office—auditor—audit client—year," and can be integrated with existing regulatory processes via a standardized API interface to achieve closed-loop, interconnected monitoring of risk identification, process control, and quality retrospective analysis.

[0039] As an optional implementation, the audit opinion deviation index is expressed as: ; in, Indicates sample In the The actual audit opinion obtained in the year; when the index is positive and the value is large, it indicates that the predicted opinion deviates significantly from the actual opinion, and there is a potential audit quality risk; conversely, it indicates that the audit opinion has a high degree of consistency with the model benchmark. The audit opinion prediction model is expressed as follows: , This represents the feature vectors in the high-risk single-factor feature list and the high-risk interactive factor feature list.

[0040] Based on the high-risk feature set obtained from the previous screening, an "audit opinion prediction model" is constructed to predict the probability of whether a non-standard audit opinion will be issued for the sample in year t. The prediction function is expressed as follows:

[0041] in This indicates that the high-risk features, after feature reduction, single-factor screening, and interaction effect identification, are used as the input feature vector. This represents a nonlinear prediction function trained using the gradient boosting tree method. Given that this model aims to characterize the standard audit judgment benchmark under a given risk exposure structure, to avoid interference from abnormal audit behavior on the model learning process, this invention selects only historical samples not identified as audit failures as the training set during the model training phase to learn the auditor's opinion issuance pattern under normal professional practice conditions; audit failure samples are only used for subsequent comparison and verification of the deviation index. Five-fold cross-validation is used during model training, combined with undersampling and early stopping mechanisms to prevent overfitting. After completing the model prediction, the predicted opinions are compared with the actual opinions issued to construct the "Audit Opinion Deviation Index" (…). ), which is defined as:

[0042] in, For the sample In the The index represents the actual audit opinions obtained in a given year. A positive and large index indicates a significant deviation between the predicted and actual opinions, suggesting potential audit quality risks; conversely, a negative index indicates a high degree of consistency between the audit opinions and the model benchmark. This index measures the degree of deviation of audit opinions from the standard audit judgment benchmark under risk structure constraints. By comparing deviation indices across different audit projects, audit entities, and time dimensions, continuous and comparable measurement of audit quality can be achieved, providing a quantitative basis for tiered monitoring, post-audit tracing, and accountability identification in subsequent audit supervision.

[0043] In summary, this invention uses audit projects as the analytical unit, integrating characteristic information from accounting firms, branch offices, auditors, and clients to form a top-down, hierarchical risk signal path. Supervisory personnel can start with the overall risk distribution at the accounting firm level, drilling down to the branch office and individual auditor levels to ultimately identify high-risk projects. By transforming the structural patterns identified by the model into an operational risk identification and response mechanism, it achieves a connection from static analysis to dynamic supervision, constructing a fully intelligent audit supervision system encompassing "discovery—early warning—feedback—accountability." In other words, this invention does not directly use the probability of a single audit failure risk prediction as the basis for regulatory decisions. Instead, it transforms the continuous risk information identified by the model into multidimensional, interpretable, and operational regulatory signals through feature contribution analysis, feature interaction effect identification, and risk grading rule mapping of the model output results. This is because a single risk prediction probability is insufficient to reflect the source of risk, the level of responsibility, and the intervention path, and cannot directly support hierarchical supervision and differentiated handling decisions. Based on this, the risks of the audit entity, the interactive risks of the audit project, and the audit quality deviation signals are uniformly integrated and presented in a visual manner on the intelligent supervision panel, forming a closed-loop supervision process covering pre-audit project identification, in-process monitoring, and post-audit feedback.

[0044] Furthermore, prior to S1, the process includes: formatting the multidimensional data related to the audit history practice samples to obtain a three-dimensional fusion dataset with variables including the audit subject, audit client, and audit environment; performing dimensionality reduction on the multi-level features within the audit subject in the three-dimensional fusion dataset to obtain the principal component feature vector set; and training the risk identification model using the principal component feature vector set, audit client features, and audit environment features as inputs, and the audit failure probability of the audit history practice samples as output.

[0045] It should be noted that the process of obtaining historical audit practice samples involves: collecting multidimensional data related to audit projects, and then cleaning, matching, standardizing, and structuring the collected data to construct a unified audit project sample dataset that can be used for model analysis. The data includes at least audit entity data, audit client data, and external environment data. All types of data are linked and time-aligned using a unified audit project identifier, forming a three-dimensional fusion dataset of "audit entity—audit client—audit environment," providing a data foundation for subsequent audit failure identification and audit quality evaluation.

[0046] The specific steps are as follows: (1) Data source collection step: By establishing data connection interfaces with public information disclosure channels, regulatory information platforms and commercial or open source database systems, data related to the audit project is automatically collected. Among them, the relevant information of the audit subject is used to characterize the professional characteristics of the three levels of accounting firms, branches and signing certified public accountants; the relevant information of the audited entity is used to reflect its governance structure, financial operation status, business risks and internal control characteristics; the relevant information of the external environment is used to characterize external factors such as the macroeconomic environment, industry regulatory intensity, legal litigation environment and market sentiment. (2) Data security and access control step: A hierarchical management mechanism is established for the collected data according to the data type and sensitivity, and corresponding access control and usage record rules are set to ensure that different data are only accessed and used within the authorized scope, thereby ensuring the compliance and security of the data collection, storage and analysis process. (3) Data cleaning and standardization step: Data from different data sources are formatted, missing values ​​are processed and outliers are corrected to eliminate the differences in data scope and time dimension between different data sources. (4) Based on a unified audit project identifier, the audit entity data, audit client data, and external environment data are correlated and matched, and aligned in terms of time dimension to construct a structured audit project data relationship. Among them, "signing certified public accountant - audit client - audit year" is used as the core matching key, and through the hierarchical mapping relationship between the certified public accountant and its branch office and firm, the unified alignment of multi-level information of the audit entity is achieved, and finally a three-dimensional fusion dataset of "audit entity - audit client - external environment" is constructed to provide a consistent, comparable, and computable data foundation for the subsequent risk identification model. (5) The data structuring and storage step constructs the processed data into a structured audit project dataset that can be used for subsequent feature extraction, model training, and risk analysis.

[0047] After obtaining the full sample data, multicollinearity tests were performed on the features of each dimension. To avoid model instability caused by feature redundancy and multicorrelation, this invention introduces principal component analysis to reduce the dimensionality of highly correlated variables in the audit entity within their respective dimensions, extracting principal component feature vectors: ;in, This is the standardized audit entity characteristic matrix. The first part is obtained from the eigenvalue decomposition of the covariance matrix. The feature load matrix consists of eigenvectors. This is the set of principal component features after dimensionality reduction. To reduce the impact of multicollinearity and improve the stability of feature representation, this invention performs grouped principal component analysis on the features of different levels of the audit subject, extracting representative principal component features within each level. Based on the proportion of explained variance, the features with the cumulative explained rate reaching a preset threshold are selected. Each principal component serves as the core feature representation of the audit subject. Finally, the dimensionality-reduced principal component features of the audit subject are integrated with the original features of the audit client and the external environment to form a unified comprehensive feature set of "audit subject - audit client - audit environment," which serves as the unified input for subsequent training of the audit failure identification model and regulatory risk identification.

[0048] Furthermore, the marginal contribution of each individual feature of the current project in the risk identification model is quantitatively interpreted, resulting in the analysis results of the contribution of each individual feature and its corresponding list of high-risk single-factor features, including: the risk identification model Decomposed into ; This is the baseline prediction value of the risk identification model for the current sample, used to represent the basic risk level without considering the influence of specific features. The number of features used in the risk identification model prediction. This represents the marginal contribution of the j-th feature to the risk identification model's prediction result in the current sample, used to characterize the degree of influence of this feature on the risk level. The marginal contribution values ​​of each variable to the sample's predicted probability are extracted from this value. The results of the single feature contribution analysis are obtained; the marginal contribution values ​​of each variable are ranked, and the top-ranked variables are included in the list of high-risk single-factor features.

[0049] Specifically, based on the constructed dimensionality-reduced comprehensive feature set and using historical audit project sample data as the training foundation, a "risk identification model" is built using the Gradient Boosting Tree (XGBoost) algorithm. This model aims to meet the risk characterization needs of audit and regulatory data, which commonly exhibits high-dimensional features, nonlinear relationships, and the coexistence of multiple feature interactions. The model's prediction function is as follows:

[0050] in Indicates sample In the The predicted probability of an audit failure occurring in a given year. In this invention, "audit failure" refers to a situation where an audit client commits a major financial violation that is subsequently determined by regulatory authorities within a certain accounting year, and the certified public accountant responsible for auditing the financial statements for that year issues a standard unqualified audit opinion on the financial statements for that year. A value of 1 indicates that an audit failure event has occurred, and a value of 0 indicates that it has not occurred. This is the comprehensive feature vector of the audit "entity-customer-environment" after feature engineering. (Model) This represents the nonlinear prediction function obtained by training on the training set to learn data patterns.

[0051] After the model training is completed, the risk identification model is applied to audit project samples that did not participate in the model training or to newly occurring audit projects, and outputs the corresponding audit failure risk prediction probability as the basis for subsequent risk interpretation and regulatory analysis. In the model training and prediction process, the following strategies are adopted to improve the model identification effect and stability: (1) an undersampling and early stopping mechanism is introduced to alleviate the impact of sample class imbalance on model training and prevent overfitting; (2) by adjusting hyperparameters such as learning rate, tree depth, and subsample ratio, the model structure and training process are adaptively optimized; (3) a 5-fold cross-validation is used to evaluate the generalization performance of the model to ensure the consistency and robustness of the model's predictions on different sample sets. The trained and applied risk identification model provides a unified and interpretable risk prediction output and basic model support for subsequent single-factor feature importance analysis, feature interaction effect analysis, and audit quality quantitative evaluation.

[0052] Furthermore, the joint marginal contribution between any two features of the current project in the risk identification model is quantitatively interpreted, resulting in the analysis results of the interaction feature contribution and the corresponding list of high-risk interaction factor features, including: Risk identification model Decomposed into Risk identification model The second decomposition is as follows: ; This is the baseline prediction value of the risk identification model for the current sample, used to represent the basic risk level without considering the influence of specific features. Representing variables With variables The interaction effect is measured to assess the joint marginal contribution of both variables in predicting the probability of audit failure. The joint marginal contribution of any two variables to the predicted probability of the sample is extracted from this. The results of single feature contribution analysis are obtained; the joint marginal contribution values ​​of any two variables are sorted, and the top-ranked ones are included in the list of high-risk interaction factor features.

[0053] Specifically, the trained model and prediction results incorporate the SHAP (Shapley Additive Explanations) method to quantify the marginal contribution of each individual feature in the prediction model. Based on game theory's Shapley values, this method characterizes the independent influence of features on the prediction results of audit failure risk by calculating the changes in the predicted output caused by features under different input conditions. Based on SHAP single-factor decomposition, the prediction model can be expressed as:

[0054] in, The model represents the sample In the The predicted probability of an audit failure occurring in a given year; This represents the baseline value, which is equal to the expected predicted value of all samples under featureless conditions. Representing variables The marginal contribution value of the predicted probability for this sample, with the sign reflecting the direction of influence (positive risk amplification or negative risk mitigation). After predicting all samples, the average absolute value of the SHAP for each feature is calculated (i.e., This involves measuring the global importance of the feature within the overall sample. Subsequently, the marginal contribution values ​​of the audit entity, client, and external environment features are ranked to obtain a list of high-risk single-factor features.

[0055] Unlike existing regulatory technologies that primarily rely on rule thresholds or human experience for risk assessment, this invention introduces a model-driven feature contribution analysis mechanism based on a risk identification model. This mechanism structures complex machine learning predictions into interpretable and rankable single-factor risk indicators oriented towards regulatory feature levels. By calculating and ranking the SHAP contribution values ​​of each feature in the prediction model, and based on the hierarchical division of audit supervision objects and the application scenarios of audit supervision business processes, this invention achieves a mapping transformation from "audit failure risk probability" to "regulatory feature level risk signals." This allows regulatory authorities to directly identify key factors with significant marginal impact on audit failure risk without needing to understand or access the model's internal parameter structure. The resulting list of single-factor high-risk features can be directly embedded into audit project acceptance and pre-audit processes to support the selection of key regulatory objects, risk-oriented resource allocation, and the formulation of differentiated regulatory strategies, thereby effectively transforming model outputs into regulatory decision-making elements.

[0056] Based on the results of the single-factor feature contribution analysis, to identify the interactive effects of audit entity, client, and external environment characteristics in the risk formation process, this invention further introduces SHAP interaction values ​​to perform second-order decomposition of the model output. The model can be expressed as:

[0057] in, Representing variables With variables The interaction effect measures the combined marginal contribution of both factors in predicting the probability of audit failure, with its sign reflecting the amplification or mitigation effect of the interaction on risk. The interaction SHAP values ​​of all historical samples are statistically summarized to construct a "High-Risk Audit Failure Feature Interaction Matrix," and feature combinations with significant interaction contributions at the sample level are identified using the mean. Subsequently, this matrix can be visualized in the form of a heatmap or interaction network to reveal typical coupling scenarios between different risk factors, highlighting risk coupling scenarios such as "high-risk firm characteristics × high-risk client characteristics" and amplified risk patterns such as "specific auditor characteristics × client financial or governance characteristics." This step outputs a list of high-risk interaction combinations for dynamic monitoring and key audit target identification during the audit process.

[0058] Based on the aforementioned audit failure risk identification results and risk structure analysis information, an audit opinion deviation index is constructed. This index measures the consistency between the audit opinion actually issued by the auditor in a specific audit project and the audit opinion decision-making model obtained through statistical learning based on historical high-quality audit samples under given risk exposure conditions. This enables dynamic and comparable evaluation of audit quality and provides quantitative basis for model backtesting. From a regulatory application perspective, the audit opinion deviation index characterizes the degree of deviation between the audit opinions issued in different audit projects and the conventional audit judgment model under similar risk conditions. This helps identify potential audit quality anomalies and provides quantitative reference for post-audit review, accountability, and regulatory feedback.

[0059] The present invention has the following beneficial effects in addressing the problems existing in the prior art: First, there is a lack of dimensions for characterizing the risks of audit entities before auditing. Current pre-audit methods rely heavily on historical violations, qualification information, and publicly disclosed indicators, failing to reflect the hierarchical differences in personnel capabilities, business expertise, resource allocation, and client portfolio stability among different levels of audit entities—from the firm as a whole to its branches to individual auditors. This makes it difficult to quantify the risk exposure of audit entities, potentially leading to the failure of high-risk audit entities during the project acceptance phase. This invention aims to systematically characterize high-risk characteristics by quantifying the importance of characteristics at each level to audit failure.

[0060] Second, the amplifying effect of interactive risk scenarios in in-process auditing is difficult to identify. Audit failures are often triggered by the combined effects of the audit entity's characteristics, client operational risks, project complexity, and other factors, exhibiting significant combinatorial and nonlinear characteristics. Existing in-process supervision mainly relies on sampling and phased verification, making it difficult to identify high-risk coupling scenarios between "specific entities and specific clients," thus hindering timely intervention. This invention identifies high-risk combinations of audit entity and client characteristics, systematically presenting risk exposure patterns under different characteristic combinations. This provides data support for regulatory authorities to identify potential risk audit projects during the in-process stage and provides an analytical basis for achieving pre-emptive risk identification and process intervention based on model output.

[0061] Third, this invention addresses the lack of quantifiable benchmarks for post-audit quality evaluation. Current audit quality evaluation methods largely rely on ex-post results such as financial restatements and administrative penalties, exhibiting significant time lags. Furthermore, macro-level indicators like firm rankings fail to reflect the professional differences between different branch offices and individual auditors, resulting in highly subjective, non-retrospective, and non-quantifiable quality judgments. This invention integrates the judgment characteristics of high-quality audit samples to construct a reference benchmark (which can also be seen as a quantitative expression of collective audit judgment experience) based on statistical learning results from historical high-quality audit samples. This benchmark quantifies the deviation between actual audit opinions and the benchmark, providing regulatory authorities with more timely risk warning references.

[0062] Fourth, there is a lack of a systematic analysis and closed-loop supervision mechanism that spans the pre-event, during-event, and post-event stages. In current audit supervision practices, pre-event entity screening, during-event project monitoring, and post-event quality evaluation are often carried out separately by different business modules. This results in fragmented regulatory information, inconsistent analytical standards, and difficulty in forming a unified risk perception framework and continuous feedback mechanism. Even if regulatory authorities identify abnormal signals at a certain stage, it is difficult to promptly transmit them to other stages and use them to dynamically adjust regulatory strategies, leading to fragmented and reactive regulatory measures. This invention constructs a unified data foundation and analysis model that spans the pre-event, during-event, and post-event stages, organically linking "risk identification—situation analysis—quality evaluation—regulatory handling," thereby transforming audit supervision from fragmented judgment to a systematic, closed-loop, and forward-looking approach.

[0063] Example 2 This embodiment provides an audit and supervision system based on machine learning, including: a pre-event warning module, a real-time warning module, and a post-event warning module. The system interface layout and visualization method are shown in the attached figure. Figure 2 As shown.

[0064] The pre-audit module is used to quantify the marginal contribution of each individual feature of the current project in the risk identification model before the current project is audited, to obtain the contribution analysis results of the individual features and the corresponding list of high-risk single-factor features, to aggregate the contribution analysis results of the individual features at the subject level to obtain the subject risk score, and to provide pre-audit subject risk warning based on the range in which the subject risk score is located; the risk identification model is trained using a large number of audit subjects' audit history practice samples.

[0065] The in-process early warning module is used to quantify the joint marginal contribution between any two features of the current project in the risk identification model during the audit of the current project, obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features, and conduct in-process project risk early warning based on the interaction feature contribution analysis results.

[0066] The post-audit warning module is used to construct an audit opinion prediction model after the current project has been audited, using a list of high-risk single-factor features and a list of high-risk interactive factors. The model prediction model is used to obtain the predicted opinion for the current project. The predicted opinion is compared with the actual opinion issued to obtain the audit opinion deviation index. The post-audit quality risk warning is issued based on the range of the audit opinion deviation index.

[0067] This invention provides a machine learning-based audit supervision system (hereinafter referred to as "the System"), applicable to the full-process, intelligent supervision of accounting firms, branch offices, and audit projects by regulatory agencies in the certified public accountant industry. The System adopts a layered architecture design and can be deployed on the regulatory agency's private cloud or server cluster. It consists of three core parts: a front-end visual interaction layer, a back-end intelligent analysis engine, and an audit project-themed database. Each module communicates and exchanges data through an internal service bus and standardized APIs, ensuring the system's scalability and stability. The overall implementation and operation of the system follow the principles of "data-driven, model interpretation, tiered early warning, and closed-loop feedback." The specific implementation process and module functions are as follows: (1) System Deployment and Data Initialization: Before deploying this system, it is necessary to complete the research and connection with the regulatory data environment. The system connects with the China Certified Public Accountant Industry Management Information System, the Listed Company Information Disclosure Platform, the Enterprise Credit Information Disclosure System, and third-party public opinion / business information databases through the configured data interface to realize the automatic collection of multi-source data. During the initialization phase, the system will clean, align, and standardize the historical data according to the preset unified primary key rule of "Individual Auditor Code - Branch Office Code - Firm Code - Audit Year - Project Code" to build an initial structured audit project sample library, laying the data foundation for model training and benchmark establishment.

[0068] (2) Core Model Training and Release: After the data is ready, the system initiates the core model training process: First, the risk identification model: The XGBoost algorithm is used, and the model is trained using the label "whether an audit failure occurred" in the historical samples. During the training process, overfitting is controlled through cross-validation and sample balancing strategies to improve the model's generalization ability and robustness. The trained model and its metadata (including feature list, importance ranking, and performance report) will be stored in the system's model library for subsequent identification services. Second, the audit opinion deviation identification model: To construct an audit quality evaluation benchmark, the system selects a "high-quality audit sample set" (samples that have not been identified as audit failures) from the historical samples. This set is used to train another XGBoost model to learn the decision-making rules of auditors issuing audit opinions based on the "entity-client-environment" characteristics under normal practice conditions. This model is released independently as a post-event quality evaluation benchmark.

[0069] (3) Routine Supervision and Operation Process: After the system enters routine operation, new audit projects will be processed according to the following process: First, data synchronization and feature engineering: New project data is automatically and periodically integrated, executing the same data cleaning and feature engineering process as the initialization phase. This includes grouping and performing principal component analysis (PCA) to reduce the dimensionality of audit entity features, forming usable feature vectors for the model. Second, risk prediction and interpretation: The processed features are input into the published risk identification model to obtain project-level risk probabilities. Simultaneously, the system calls the SHAP interpretation engine to automatically calculate the marginal contribution value (SHAP value) of each feature and the interaction values ​​between features. Third, generation of tiered early warning signals: Based on preset rules, the system transforms model output into actionable regulatory signals: Pre-event entity early warning: It aggregates the feature contributions of auditors or firms' historical projects, calculates their standardized risk scores, and classifies them into red (high), yellow (medium), and green (low) risk levels according to quantiles; In-event project trigger: It identifies combinations where the interaction value of "audit entity × audit client" exceeds a dynamic threshold (e.g., the 90th quantile of the historical interaction value distribution) and generates project-level risk alerts; Post-event quality evaluation: It inputs project features into the excellent audit opinion benchmark model to obtain predicted opinion probabilities, compares them with actual opinions to calculate the "audit opinion deviation index," and classifies quality levels according to the index distribution.

[0070] (4) Regulatory Application and Operation: All early warning signals and risk analysis results are integrated into the "Audit Intelligent Regulatory Dashboard" (front-end visual panel). Regulatory personnel can: First, panoramic monitoring: drill down from the firm overview to branch offices, auditors, and individual projects to view risk distribution, trend changes, and key feature contributions. Second, early warning handling: for "red light" or "yellow light" early warnings marked by the system, tasks can be initiated with one click, such as adding to the inspection list, initiating the interview procedure, or issuing a letter of concern. Third, report export: supports filtering and exporting risk analysis reports, high-risk lists, and regulatory recommendations by multiple dimensions.

[0071] (5) Closed-loop feedback and system self-optimization: To ensure the continuous effectiveness of the system, the following closed-loop mechanism has been established: First, feedback of handling results: Regulatory personnel fill in the handling conclusions of the early warning projects (such as problems found during inspections, penalty results, and rectification status) through system forms or interfaces, which are linked to the original early warning records. Second, model iteration and update: The system is set to automatically trigger the model retraining process every 12 months or when the accumulated sample growth exceeds a certain proportion (such as 30%), and optimize the model using the feedback data. At the same time, it supports regulatory experts to adjust the risk classification threshold and regulatory rule parameters according to policy changes. Third, performance monitoring dashboard: A model performance monitoring dashboard is provided to display key indicators such as model prediction accuracy and feedback coverage, which helps to evaluate system effectiveness and guide the direction of optimization.

[0072] Through the above implementation methods, this system organically embeds machine learning capabilities into the audit and supervision business process, realizing a complete closed loop from data perception, intelligent analysis, risk warning to decision support, handling feedback, and model evolution, providing a feasible technical solution for building an intelligent, precise, and forward-looking modern audit and supervision system.

[0073] Example 3 The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0074] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0075] Example 4 This invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method described in the above embodiments of this invention.

[0076] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" in this invention are intended to illustrate the invention and are not intended to limit the invention.

[0077] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An audit supervision method based on machine learning, characterized in that, include: S1: Before auditing the current project, quantify the marginal contribution of each individual feature of the current project as described in the risk identification model to obtain the contribution analysis results of the individual feature and the corresponding list of high-risk single-factor features. The single feature is the audit subject, audit client, or audit environment; the contribution analysis results of the single feature are aggregated at the subject level to obtain the subject risk score, and the subject risk warning is given in advance based on the interval of the subject risk score; The risk identification model was trained using a large number of audit entities' historical practice samples; S2: During the audit of the current project, the joint marginal contribution between any two features of the current project in the risk identification model is quantitatively interpreted to obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features. Based on the interaction feature contribution analysis results, in-process project risk warning is carried out. S3: After the current project is audited, an audit opinion prediction model is constructed using the high-risk single-factor feature list and the high-risk interactive factor feature list. The model prediction model is used to obtain the predicted opinion for the current project. The predicted opinion is compared with the actual opinion issued to obtain the audit opinion deviation index. Post-audit quality risk warning is issued based on the range of the audit opinion deviation index.

2. The audit and supervision method based on machine learning as described in claim 1, characterized in that, Before S1, it also includes: The multidimensional data related to the audit history practice sample were formatted to obtain a three-dimensional fusion dataset with variables including the audit entity, audit client, and audit environment; The multi-level features within the audit subject in the three-dimensional fusion dataset are dimensionality-reduced in their respective dimensions to obtain the principal component feature vector set; The risk identification model is trained by taking the principal component feature vector set, audit client characteristics, and audit environment characteristics as inputs, and the audit failure probability of the historical audit practice sample as output.

3. The audit and supervision method based on machine learning as described in claim 1, characterized in that, The marginal contribution of each individual feature of the current project in the risk identification model is quantitatively interpreted to obtain the single feature contribution analysis results and the corresponding list of high-risk single-factor features, including: The risk identification model Decomposed into ; This is the baseline prediction value of the risk identification model for the current sample, used to represent the basic risk level without considering the influence of specific features. The number of features participating in the prediction of the risk identification model. This represents the marginal contribution value of the j-th feature to the prediction result of the risk identification model under the current sample, which is used to characterize the degree of influence of this feature on the risk level; Extract the marginal contribution of each variable to the predicted probability of the sample. The results of the single feature contribution analysis are obtained; the marginal contribution values ​​of each variable are sorted, and the top-ranked variables are included in the list of high-risk single-factor features.

4. The machine learning-based audit and supervision method as described in claim 3, characterized in that, Pre-emptive risk warnings are issued based on the risk score range of the subject, including: Using formula This represents the risk score of subject i. Standardize it to obtain the risk level of subject i ; The number of audit entities falling within the e-th risk range. This refers to the set of audit entities that belong to the e-th risk interval in the risk level classification. This is a set of risk characteristics at the audit entity level. This represents the marginal contribution of the j-th audit subject characteristic to the prediction result of the risk identification model corresponding to subject i; This represents the comprehensive risk score for the e-th risk interval, obtained by summarizing the contribution of the risk characteristics of the main entities. This is the average risk score of the audit entity obtained based on historical audit sample statistics. This is the standard deviation of the risk scores of the audited entities in the historical sample, used to standardize the risk scores; When the risk level of subject i exceeds the upper limit threshold In such cases, the entity will be given a warning using the first method. When the risk level of subject i is in the range At that time, the entity will be given a warning in the second manner.

5. The machine learning-based audit and supervision method as described in claim 1, characterized in that, The method quantifies the joint marginal contribution between any two features of the current project in the risk identification model to obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features, including: The risk identification model Decomposed into The risk identification model The second decomposition is as follows: ; This is the baseline prediction value of the risk identification model for the current sample, used to represent the basic risk level without considering the influence of specific features. Representing variables With variables The interaction effect of the two measures their joint marginal contribution in predicting the probability of audit failure. Extract the joint marginal contribution of any two variables to the predicted probability of the sample. The single feature contribution analysis results are obtained; the joint marginal contribution values ​​corresponding to any two variables are sorted, and the top-ranked values ​​are included in the high-risk interaction factor feature list.

6. The machine learning-based audit and supervision method as described in claim 5, characterized in that, The step of conducting in-process project risk warning based on the interaction feature contribution analysis results includes: When the overall risk intensity is superimposed When the risk exceeds the first risk trigger threshold, the project will be identified as being in a high-risk interactive scenario, and a project risk warning signal will be automatically triggered. This represents the joint marginal contribution of the j-th and k-th features in audit project p to the prediction of audit failure probability in the risk identification model. It is used to measure the intensity of the superimposed impact of multi-feature interaction on project risk. The subscript p represents the audit project index, and the subscripts j and k represent the j-th and k-th features involved in the interaction, respectively.

7. The machine learning-based audit and supervision method as described in claim 5, characterized in that, The step of conducting in-process project risk warning based on the interaction feature contribution analysis results includes: When local extreme risks contribute When the risk exceeds the second risk trigger threshold, it is determined that the project has experienced a single-point feature-driven amplification of abnormal risks, and a project risk warning signal is automatically triggered. This represents the joint marginal contribution of the j-th and k-th features in audit project p to the prediction of audit failure probability in the risk identification model. It is used to measure the intensity of the superimposed impact of multi-feature interaction on project risk. The subscript p represents the audit project index, and j and k represent the j-th and k-th features involved in the interaction, respectively.

8. The machine learning-based audit and supervision method as described in claim 5, characterized in that, The post-audit quality risk warning based on the interval of the audit opinion deviation index includes: When the audit opinion deviation index of a certain audit project Exceeding the upper limit threshold When this happens, the project will be marked as the primary display item, and prompts for key review and accountability will be automatically pushed to it. when Located in the interval When necessary, the project will be marked as the second display method for enhanced review or included in the subsequent continuous tracking list; and The preset risk threshold; when Below the lower threshold When the third display method is selected, it is considered that the audit quality is normal.

9. The machine learning-based audit and supervision method as described in claim 8, characterized in that, The audit opinion deviation index is expressed as follows: ； in, Indicates sample In the The actual audit opinion obtained in the year; when the index is positive and the value is large, it indicates that the predicted opinion deviates significantly from the actual opinion, and there is a potential audit quality risk; conversely, it indicates that the audit opinion has a high degree of consistency with the model benchmark. The audit opinion prediction model is expressed as follows: , This represents the feature vector in the high-risk single-factor feature list and the high-risk interaction factor feature list.

10. An audit and supervision system based on machine learning, characterized in that, include: The pre-audit module is used to quantify and interpret the marginal contribution of each single feature of the current project in the risk identification model before the current project is audited, to obtain the single feature contribution analysis results and the corresponding high-risk single factor feature list, to aggregate the single feature contribution analysis results at the subject level to obtain the subject risk score, and to conduct pre-audit of subject risk based on the interval of the subject risk score; The risk identification model was trained using a large number of audit entities' historical practice samples; The in-process early warning module is used to quantify and interpret the joint marginal contribution between any two features of the current project in the risk identification model during the audit of the current project, obtain the interaction feature contribution analysis results and the corresponding list of high-risk interaction factor features, and conduct in-process project risk early warning based on the interaction feature contribution analysis results; The post-audit warning module is used to construct an audit opinion prediction model using the high-risk single-factor feature list and the high-risk interactive factor feature list after the current project has been audited; to obtain the model prediction opinion for the current project using the audit opinion prediction model; to compare the model prediction opinion with the actual opinion issued to obtain the audit opinion deviation index; and to issue a post-audit quality risk warning based on the range of the audit opinion deviation index.

Images