Method and system for constructing credibility evaluation model based on Bayesian network

Abstract

The invention discloses a credibility evaluation model construction method and system based on a Bayesian network, and relates to the field of auditing evidence credibility evaluation, and the method comprises the following steps: carrying out the classification preprocessing of multi-source collection auditing evidences; fusing data mining and special identification extraction feature setting variables; performing expert analysis to construct a Bayesian network structure; calculating a conditional probability table according to a large amount of historical data; and performing cross validation evaluation on operations such as optimization model, modulation-containing structure, change probability and variable feature, so as to ensure the accurate evaluation credibility of the model. According to the model construction method and system, the accuracy and efficiency of auditing evidence credibility evaluation are improved, the evidence relation is accurately analyzed, the misjudgment risk is reduced, auditing is accelerated through the automatic process, the method and system adapt to multiple scenes, a reliable basis is provided for auditing decision making, and auditing work is helped to be efficiently carried out and the quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of audit evidence credibility assessment models, and in particular to a credibility assessment model construction method and system based on a Bayesian network. Background Art

[0002] In today's complex and ever-changing business environment, audit work faces unprecedented challenges and pressures. With the continuous expansion of corporate scale, increasing diversification of business and frequent transactions, the amount of information that needs to be processed in the audit process has exploded. Traditional audit evidence evaluation methods have gradually exposed many limitations in the face of such massive information.

[0003] Traditional auditing relies on the professional experience and subjective judgment of auditors to analyze and evaluate audit evidence. This approach may work to a certain extent when faced with a small amount of simple audit evidence, but when the amount of evidence is large and the types are complex, the efficiency and accuracy of manual evaluation will be greatly reduced. On the one hand, it is difficult for auditors to conduct a comprehensive and in-depth review of all evidence within a limited time, and it is easy to miss key information, resulting in increased audit risks. On the other hand, subjective judgments are highly uncertain, and different auditors may come to different conclusions on the same evidence. There is a lack of unified and objective evaluation standards, making it difficult to ensure the consistency and reliability of audit results.

[0004] At the same time, existing audit tools and technologies have obvious deficiencies in evidence integration and correlation analysis. Audit evidence is often scattered in multiple data sources such as financial systems, business databases, and contract documents. There is a lack of effective integration mechanisms between data sources, making it difficult for auditors to grasp the internal connections between evidence as a whole. For example, abnormal fluctuations in financial data may be related to specific events in business operations, but due to the lack of automated correlation analysis methods, it is difficult for auditors to discover this potential causal relationship in a timely manner, and thus it is impossible to accurately judge the credibility of audit evidence and the degree of impact on audit conclusions.

[0005] In addition, with the rapid development of information technology, the degree of informatization of enterprises has been continuously improved, and the storage and transmission methods of data have also undergone tremendous changes. Electronic data has become one of the main sources of audit evidence, but the characteristics of electronic data such as easy modification and easy forgery have also brought new challenges to audit work. Traditional audit methods are difficult to effectively deal with the authenticity and integrity verification of electronic data, and cannot meet the strict requirements of the information age for audit quality and security.

[0006] In summary, in order to improve the quality and efficiency of audit work and reduce audit risks, an innovative audit evidence credibility assessment method and system is urgently needed. The credibility assessment model construction method and system based on Bayesian network came into being. It can make full use of the advantages of Bayesian network in uncertainty reasoning and probability analysis, and conduct scientific, objective and efficient assessment of massive and complex audit evidence, providing strong support for audit decision-making and adapting to the needs of modern audit development. Summary of the invention

[0007] The present invention proposes a Bayesian network-based credibility assessment model construction method and system to solve the problems mentioned in the above-mentioned prior art.

[0008] In order to achieve the above object, the present invention adopts the following technical solution: a method for constructing a credibility evaluation model based on a Bayesian network, comprising the following steps: Audit evidence collection and classification steps: Collect financial statements, business transactions, and contract documents as multi-source audit evidence, and classify them into finance, business operations, compliance, and related parties based on type, source, and business field. Pre-process the evidence using semantic cleaning technology; Feature extraction and variable determination steps: Integrate data mining and expert experience to extract features, determine Bayesian network variables based on the features, divide discrete variables into multiple categories, and discretize continuous variables according to equal frequency or equal interval method; suppose the continuous variable X has a value range of [a, b], and the interval width of equal interval discretization is , where n is the number of discrete intervals, the discretized interval is ; Steps for constructing the Bayesian network structure: construct a Bayesian network structure based on the relationship between the audit process and the logic analysis of evidence by at least 5 experts; experts score 1-10 points, and those above 6 points are considered strong associations. Specifically, let the expert score for the relationship between evidence A and evidence B be S AB , relationship strength , when R AB ≥0.6, strong association; Steps to determine the conditional probability table: Collect more than 500 historical case data within 3 years, use the maximum likelihood estimation method for discrete variables, and assume that the discrete variable Y has m values ​​y1, y2, ···, y m , under the parent node value combination C, Y=y i The conditional probability of ,in For combination C, Y=y i The number of occurrences of N(C) is the number of occurrences of combination C; continuous variables are estimated using the kernel density method and are organized into a table and compared with the actual audit results for verification; Model evaluation and optimization steps: Use the cross-validation method to divide the training set and the test set into 8:2, and evaluate through accuracy and recall. If the model evaluation indicators do not meet the requirements, optimize the model. Suppose the number of positive samples in the test set is TP+FN, the number of samples predicted to be positive and actually positive is TP, the number of samples predicted to be negative and actually positive is FN, the number of samples predicted to be positive and actually negative is FP, and the number of samples predicted to be negative and actually negative is TN. Then the accuracy , recall rate , , F1 represents the indicator used to comprehensively evaluate the performance of the audit evidence credibility assessment model. The higher the F1 value, the higher the balance between precision and recall. The calculation formula of precision is: ; Model optimization methods include adjusting the Bayesian network structure, modifying the conditional probability table, and increasing or decreasing feature variables.

[0009] Furthermore, in the audit evidence collection and classification steps, for the collection of external regulatory information, web crawler technology combined with manual screening is used; the web crawler captures information according to the preset list of regulatory agency websites and keyword lists, and the crawling frequency is once a day; during manual screening, the auditors screen the captured information according to the needs of the audit project.

[0010] Furthermore, in the feature extraction and variable determination steps, when extracting customer satisfaction features from business operation evidence, a combination of questionnaire survey and data analysis is adopted; during data analysis, factor analysis methods are used to reduce the dimension of questionnaire survey data and extract main factors as customer satisfaction features.

[0011] Furthermore, in the Bayesian network structure construction step, when determining the strength and direction of the relationship between evidence, a data-driven association rule mining algorithm is used for auxiliary analysis. The mined association rules are compared and verified with the expert judgment results, and inconsistent association rules are analyzed and confirmed again by experts.

[0012] Furthermore, in the step of determining the conditional probability table, when the maximum likelihood estimation method is used to calculate the conditional probability of discrete variables, the Bayesian estimation method is used for supplementary calculations when the amount of data is lower than the threshold. The prior distribution of the Bayesian estimation method is determined based on historical experience data and domain knowledge, and the parameter selection of the prior distribution makes the calculation results more robust.

[0013] Furthermore, in the model evaluation and optimization step, when adjusting the Bayesian network structure, a structural learning algorithm is used, including a scoring search-based algorithm or a constraint-based algorithm. The scoring search-based algorithm uses the Bayesian Information Criterion BIC as a scoring function to search for the optimal network structure; the constraint-based algorithm constructs a network structure according to the conditional independence relationship between variables.

[0014] Furthermore, a system for implementing a method for constructing a credibility assessment model based on a Bayesian network includes: Evidence collection and classification module: It is equipped with multiple data interfaces to connect to financial systems, business databases, document management systems, and external network data sources. The collection scope includes financial statements, business records, contract documents, and regulatory information. Feature extraction and variable determination module: It has a built-in data mining algorithm library and an audit domain knowledge base. It extracts features from the audit evidence in the evidence base based on the rules in the knowledge base and the data mining algorithms in the algorithm library, including decision tree algorithms and clustering algorithms, and determines the variables of the Bayesian network based on the features. Network structure construction module: It has a human-computer interaction interface. Audit experts are invited to input the relationship judgment information between evidences through the interface. Combined with the data-driven association analysis module, the Bayesian network structure is constructed based on the expert opinions and data association analysis results. The constructed network structure is displayed in a graphical way. Conditional probability table generation module: connected to the historical audit case database, used to read historical data, calculate the conditional probability of each node variable in the Bayesian network, generate a conditional probability table, and automatically select the calculation method during the calculation process, including maximum likelihood estimation method or kernel density estimation method; Model evaluation and optimization module: Use cross-validation technology to evaluate the constructed audit evidence credibility assessment model, and automatically determine whether the model needs to be optimized based on the evaluation results. If optimization is required, automatically call the structural learning algorithm, probability adjustment algorithm, etc. to optimize the model. The optimized model is evaluated again until the model performance meets the requirements.

[0015] Furthermore, in the evidence collection and classification module, anonymization and sentiment analysis technology are used to collect employee feedback records. The anonymization processing complies with privacy regulations, and sentiment analysis identifies positive, neutral, and negative feedback with an accuracy rate of over 90%.

[0016] Furthermore, in the feature extraction and variable determination module, the audit knowledge graph has a regular update and maintenance mechanism, and is connected with data from authoritative audit research institutions and industry associations, with more than 50 knowledge nodes and 100 related edges updated monthly.

[0017] Furthermore, in the network structure construction module, the visual interactive interface supports multi-language switching and multi-user collaborative operation. Multi-language switching meets the needs of cross-border audits, and multi-user collaboration supports multiple experts to build network structures simultaneously in different locations.

[0018] Compared with the prior art, the present invention has the following beneficial effects: In terms of assessment accuracy, by widely collecting audit evidence from multiple sources and scientifically classifying and deeply preprocessing it, combining data mining and expert experience to accurately extract features and determine variables, and constructing a Bayesian network structure and precise conditional probability table that reflects the complex logic of evidence, it can comprehensively consider various types of evidence and their interrelationships, reduce human judgment errors and information omissions, make the credibility assessment of audit evidence more accurate and reliable, and effectively reduce the risk of audit misjudgment.

[0019] From the perspective of evaluation efficiency, the system's automated data collection, feature extraction, network construction and evaluation processes have greatly shortened the time for audit evidence processing. Compared with traditional manual audit methods, large-scale evidence analysis can be completed in a shorter time, significantly improving audit work efficiency, enabling audit projects to be advanced more quickly, timely discovering problems and providing decision-making basis, and enhancing the ability of enterprises to cope with risks and seize opportunities.

[0020] In terms of risk prevention and control, accurate credibility assessment helps to identify potential audit risk points in advance. Through quantitative analysis of evidence credibility, we can conduct targeted in-depth audits of high-risk areas, timely discover financial fraud, compliance loopholes and other issues, provide key support for enterprises to formulate risk response strategies, effectively protect corporate asset security, and maintain corporate reputation and market image.

[0021] In terms of adaptability, the model can be flexibly adjusted according to the characteristics and needs of different audit projects. Whether it is financial audit, compliance audit or special audit, it can be adapted by optimizing the network structure, adjusting characteristic variables and conditional probabilities, etc. It is widely used in various enterprises and audit scenarios, has strong versatility and scalability, promotes technological progress and development in the audit industry, and helps audit work move towards intelligence and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic block diagram of a Bayesian network-based credibility assessment model construction method proposed in the present invention; Figure 2 This is a schematic block diagram of the audit evidence credibility evaluation system based on Bayesian network proposed by the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0025] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0026] Reference Figure 1-2 :A method for constructing a credibility evaluation model based on a Bayesian network, comprising the following steps: Audit evidence collection and classification steps: Audit evidence is collected extensively from a variety of audit data sources, including financial statement data, business transaction records, contract documents, internal management system documents, interview records of relevant personnel, and external regulatory information. The collected audit evidence is classified into financial evidence, business operation evidence, compliance evidence, related party evidence, etc. according to the type of evidence, source channel, and business field involved. The number of classifications is not less than 5, and the number of samples of each type of evidence is not less than 100, to ensure that each type of evidence has sufficient representativeness and analytical value.

[0027] Each type of audit evidence is preprocessed to remove duplicate data, erroneous data, and ambiguous data. Text evidence is processed using text cleaning technology based on semantic understanding to remove irrelevant words and redundant sentences, making the text content concise and clear. The effective information extraction rate of the processed text data is not less than 90%; for numerical evidence, data standardization is performed to unify the data into a preset numerical range or data format to ensure data consistency and comparability. The error rate of the data after standardization does not exceed ±5%.

[0028] Feature extraction and variable determination steps: Extract the features of audit evidence by combining data mining technology with domain expert knowledge. Extract financial indicator features from financial evidence, such as debt-to-asset ratio, gross profit margin, net profit margin, current ratio, etc., with a feature count of no less than 8; extract business process key node features, business volume change features, customer satisfaction features, etc. from business operation evidence, with a feature count of no less than 10; extract compliance clause compliance features, violation risk level features, etc. from compliance evidence, with a feature count of no less than 6.

[0029] Determine the variables of the Bayesian network based on the extracted features. Set the key features as network node variables. The variable types include discrete variables and continuous variables. The value categories of discrete variables are determined according to the actual situation of the features. For example, the compliance with compliance clauses can be divided into three categories: full compliance, partial compliance, and non-compliance. Continuous variables are discretized using equal frequency or equal distance discretization methods. The number of discrete intervals is not less than 5 to ensure that the variables can accurately reflect the characteristics of audit evidence and facilitate subsequent probability calculations and model construction. Suppose a continuous variable X, whose value range is [a, b], and the interval width of equal distance discretization is [a, b]. , where n is the number of discrete intervals, the discretized interval is .

[0030] Bayesian network structure construction steps The Bayesian network structure is constructed based on the inherent logical relationship between the audit business process and the evidence. Invite at least 5 experts with rich audit experience to analyze and judge the causal relationship and dependency relationship between the evidence, and determine the strength and direction of the relationship between the evidence through the expert scoring method. The scoring range is 1-10 points, and the relationship strength threshold is set at 6 points. When the expert score exceeds the threshold, it is determined that there is a strong correlation. Suppose the expert's score for the relationship between evidence A and evidence B is S. AB , relationship strength , when R AB When ≥0.6, it is considered a strong correlation.

[0031] Based on the results of expert judgment, a graphical modeling tool is used to draw the Bayesian network structure. The nodes in the network represent the audit evidence variables, and the edges represent the relationship between the variables. The complexity of the network structure is determined according to the scale and complexity of the audit project. The number of nodes is not less than 20, and the number of edges is not less than 30, to ensure that the network structure can fully and accurately reflect the complex relationship between audit evidence.

[0032] Steps to determine the conditional probability table: Collect a large amount of historical audit case data, with a time span of no less than 3 years and no less than 500 cases. Calculate the conditional probability of each node variable in the Bayesian network under different parent node value combinations based on historical data. Use the maximum likelihood estimation method to calculate the conditional probability. For discrete variables, calculate the frequency of occurrence of each value as the conditional probability estimate; for continuous variables, use the kernel density estimation method to calculate the conditional probability distribution function, and control the estimation error within ±10%. Suppose the discrete variable Y has m values ​​y1, y2, ···, y m , under the parent node value combination C, Y=y i The conditional probability of ,in For combination C, Y=y i N(C) is the number of occurrences of combination C).

[0033] The calculated conditional probabilities are organized into a conditional probability table. The accuracy of the conditional probability table is evaluated by comparing it with the actual audit results. The number of samples for comparison and verification is not less than 100, and the accuracy rate is required to be no less than 85%. If the accuracy rate does not meet the requirements, the calculation method is readjusted or additional data is used for calculation until the accuracy of the conditional probability table meets the requirements.

[0034] Model evaluation and optimization steps: Use the cross-validation method to evaluate the constructed audit evidence credibility assessment model, divide the historical audit case data into a training set and a test set in a ratio of 8:2, use the training set data to build the model, and use the test set data to verify the model. Accuracy, recall, and F1 value are used as evaluation indicators. The accuracy rate is required to be no less than 80%, the recall rate is required to be no less than 75%, and the F1 value is required to be no less than 0.78. If the model evaluation indicators do not meet the requirements, the model is optimized. Suppose the number of positive samples in the test set is TP+FN, the number of samples predicted to be positive and actually positive is TP, the number of samples predicted to be negative and actually positive is FN, the number of samples predicted to be positive and actually negative is FP, and the number of samples predicted to be negative and actually negative is TN. The accuracy rate is , recall rate , , F1 represents the indicator used to comprehensively evaluate the performance of the audit evidence credibility assessment model. The higher the F1 value, the higher the balance between precision and recall. The calculation formula of precision is: .

[0035] Model optimization methods include adjusting the Bayesian network structure, modifying the conditional probability table, increasing or decreasing characteristic variables, etc. By analyzing the cases of misclassification in the model evaluation results, we can find out the problems and deficiencies of the model, make targeted optimization adjustments, and re-evaluate the optimized model until the model evaluation indicators meet the requirements, ensuring that the model has good performance and reliability and can accurately evaluate the credibility of audit evidence.

[0036] In the present invention, in the audit evidence collection and classification step, the collection of external regulatory information adopts a web crawler technology combined with manual screening. The web crawler crawls information according to the preset regulatory agency website list and keyword list, with a crawling frequency of once a day, and the integrity of the crawled information is not less than 95%; during manual screening, professional auditors screen the captured information according to the audit project requirements, and the screening accuracy is not less than 98%, ensuring the accuracy and relevance of external regulatory information.

[0037] In the present invention, in the feature extraction and variable determination steps, when extracting the customer satisfaction features of business operation evidence, a combination of questionnaire survey and data analysis is adopted. The number of samples for the questionnaire survey is not less than 200, and the questionnaire design covers customer satisfaction evaluations on product quality, service attitude, delivery timeliness, etc.; when analyzing the data, the factor analysis method is used to reduce the dimension of the questionnaire survey data, and the main factors are extracted as customer satisfaction features. The factor explained variance ratio is not less than 80%, so that the customer satisfaction features can more accurately reflect the business operation status.

[0038] In the present invention, in the Bayesian network structure construction step, in addition to the expert scoring method, a data-driven association rule mining algorithm is also used for auxiliary analysis when determining the strength and direction of the relationship between evidence. The minimum support of the association rule mining algorithm is set to 0.2, and the minimum confidence is set to 0.6. The mined association rules are compared and verified with the expert judgment results. Inconsistent association rules are analyzed and confirmed again by experts to ensure the scientificity and accuracy of the Bayesian network structure construction.

[0039] In the present invention, in the conditional probability table determination step, when the maximum likelihood estimation method is used to calculate the conditional probability of discrete variables, the Bayesian estimation method is used for supplementary calculations when the amount of data is small. The prior distribution of the Bayesian estimation method is determined based on historical experience data and domain knowledge, and the parameter selection of the prior distribution makes the calculation result more robust, which can improve the accuracy of the conditional probability estimation and reduce the estimation error when the amount of data is limited.

[0040] In the present invention, in the model evaluation and optimization step, when adjusting the Bayesian network structure, a structural learning algorithm, such as a scoring search-based algorithm or a constraint-based algorithm, is used. The scoring search-based algorithm uses the Bayesian Information Criterion (BIC) as a scoring function to search for the optimal network structure; the constraint-based algorithm constructs a network structure based on the conditional independence relationship between variables, and by comparing the optimization effects of different algorithms, the most suitable structural adjustment method is selected to improve the model optimization efficiency.

[0041] The present invention also discloses an audit evidence credibility evaluation system based on Bayesian network, comprising: Evidence collection and classification module: It is equipped with multiple data interfaces and can connect to data sources such as financial systems, business databases, document management systems, and external networks to realize the automatic collection of audit evidence. The collection scope covers various types such as financial statements, business records, contract documents, and regulatory information. The collection speed is not less than 100 copies per second, and the collected evidence can be classified and preprocessed according to the method described in claim 1. The processed evidence is stored in a dedicated evidence library. The storage capacity of the evidence library is not less than 10TB, which can meet the storage needs of a large amount of audit evidence.

[0042] Feature extraction and variable determination module: Built-in data mining algorithm library and audit field knowledge base, based on the rules in the knowledge base and the data mining algorithms in the algorithm library, such as decision tree algorithm, clustering algorithm, etc., to extract features of audit evidence in the evidence library. It can extract features of financial indicators, business processes, compliance status, etc., and determine the variables of the Bayesian network based on the features. The accuracy of variable determination is not less than 95%, ensuring a solid foundation for the subsequent construction of the Bayesian network.

[0043] Network structure construction module: It has a human-computer interaction interface, and audit experts can be invited to input the relationship judgment information between evidence through the interface. At the same time, combined with the data-driven association analysis module, the Bayesian network structure is constructed based on expert opinions and data association analysis results. The construction process is visualized, and experts can view and modify the network structure in real time. The constructed network structure is displayed in a graphical way, which is easy to understand and analyze. The construction time of the network structure does not exceed 30 minutes.

[0044] Conditional probability table generation module: connected to the historical audit case database, it can read a large amount of historical data, calculate the conditional probability of each node variable in the Bayesian network, and generate a conditional probability table. During the calculation process, it can automatically select the appropriate calculation method, such as maximum likelihood estimation or kernel density estimation, and verify the accuracy of the calculation results. The verification error does not exceed ±8%, ensuring the reliability of the conditional probability table.

[0045] Model evaluation and optimization module: Use cross-validation technology to evaluate the constructed audit evidence credibility evaluation model, and automatically determine whether the model needs to be optimized based on the evaluation results. If optimization is required, the structural learning algorithm, probability adjustment algorithm, etc. can be automatically called to optimize the model. The optimized model is evaluated again until the model performance meets the requirements. The number of model evaluation and optimization cycles does not exceed 5 times to ensure that the model can quickly and effectively reach the optimal state.

[0046] In the present invention, in the evidence collection and classification module, anonymization and sentiment analysis technology are used to collect employee feedback records. Anonymization meets the requirements of privacy regulations, and sentiment analysis identifies positive, neutral, and negative feedback with an accuracy rate of over 90%, providing a multi-angle evidence perspective for audits and protecting employee privacy.

[0047] In the present invention, in the feature extraction and variable determination module, the audit knowledge graph is regularly updated and maintained. It connects with authoritative audit research institutions and industry associations, updates more than 50 knowledge nodes and more than 100 associated edges per month, ensures the timeliness and accuracy of the knowledge graph, and provides reliable knowledge support for feature extraction and variable determination.

[0048] In the present invention, in the network structure construction module, the visual interactive interface supports multi-language switching and multi-user collaborative operation. Multi-language switching meets the needs of transnational audits, and multi-user collaboration allows multiple experts to build network structures remotely and simultaneously, improving collaborative efficiency by more than 40%, promoting global audit resource sharing and collaboration.

[0049] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a credibility assessment model based on a Bayesian network, characterized in that: The following steps are involved: Audit evidence collection and classification steps: Collect financial statements, business transactions, and contract documents as multi-source audit evidence, and pre-process the evidence through semantic cleaning technology; Feature extraction and variable determination steps: extract features by integrating data mining and expert experience, determine Bayesian network variables based on the features, divide discrete variables into multiple categories, and discretize continuous variables using equal frequency or equal interval methods; Steps for constructing the Bayesian network structure: construct a Bayesian network structure based on the relationship between the audit process and the logic analysis of evidence by at least 5 experts; experts score 1-10 points, and those above 6 points are considered strong associations. Specifically, let the expert score for the relationship between evidence A and evidence B be S AB , relationship strength , when R AB ≥0.6, strong association; Steps to determine the conditional probability table: Collect more than 500 historical case data within 3 years, use the maximum likelihood estimation method for discrete variables, and assume that the discrete variable Y has m values ​​y1, y2, ···, y m , under the parent node value combination C, Y=y i The conditional probability of ,in For combination C, Y=y i The number of occurrences of N(C) is the number of occurrences of combination C; continuous variables are estimated using the kernel density method and are organized into a table and compared with the actual audit results for verification; Model evaluation and optimization steps: Use the cross-validation method to evaluate the accuracy and recall rate. If the model evaluation indicators do not meet the requirements, optimize the model.

2. The method for constructing a credibility assessment model based on a Bayesian network according to claim 1, characterized in that: In the audit evidence collection and classification steps, for the collection of external regulatory information, web crawler technology combined with manual screening is used; the web crawler captures information according to the preset list of regulatory agency websites and keyword lists, and the crawling frequency is once a day; during manual screening, the auditors screen the captured information based on the requirements of the audit project.

3. The method for constructing a credibility assessment model based on a Bayesian network according to claim 1, characterized in that: In the feature extraction and variable determination steps, when extracting customer satisfaction features of business operation evidence, a combination of questionnaire survey and data analysis is used; when analyzing data, factor analysis is used to reduce the dimension of questionnaire survey data and extract the main factors as customer satisfaction features; In the feature extraction and variable determination steps, suppose the continuous variable X has a value range of [a, b] and the interval width of the equidistant discretization is , where n is the number of discrete intervals, the discretized interval is .

4. The method for constructing a credibility assessment model based on a Bayesian network according to claim 1, characterized in that: In the Bayesian network structure construction step, a data-driven association rule mining algorithm is also used for auxiliary analysis. The mined association rules are compared and verified with the expert judgment results, and inconsistent association rules are analyzed and confirmed again by experts.

5. The method for constructing a credibility assessment model based on a Bayesian network according to claim 1, characterized in that: In the step of determining the conditional probability table, when the maximum likelihood estimation method is used to calculate the conditional probability of discrete variables, the Bayesian estimation method is used for supplementary calculations when the amount of data is lower than the threshold. The prior distribution of the Bayesian estimation method is determined based on historical experience data and domain knowledge, and the parameter selection of the prior distribution makes the calculation results more robust.

6. The method for constructing a credibility assessment model based on a Bayesian network according to claim 1, characterized in that: In the model evaluation and optimization step, when adjusting the Bayesian network structure, a structural learning algorithm is used, including a score search-based algorithm or a constraint-based algorithm. The score search-based algorithm uses the Bayesian Information Criterion BIC as a scoring function to search for the optimal network structure; the constraint-based algorithm constructs the network structure according to the conditional independence relationship between variables; In the model evaluation and optimization step, the model optimization method includes adjusting the Bayesian network structure, modifying the conditional probability table, and increasing or decreasing the feature variables. The specific optimization formula used is: assuming that the number of positive samples in the test set is TP+FN, the number of samples predicted to be positive and actually positive is TP, the number of samples predicted to be negative and actually positive is FN, the number of samples predicted to be positive and actually negative is FP, and the number of samples predicted to be negative and actually negative is TN, then the accuracy is , recall rate , , F1 represents the index used to comprehensively evaluate the performance of the audit evidence credibility assessment model. The higher the F1 value, the higher the balance between precision and recall; The calculation formula of precision is: .

7. A system for implementing the Bayesian network-based credibility assessment model construction method according to any one of claims 1 to 6, characterized in that: include: Evidence collection and classification module: It is equipped with multiple data interfaces to connect to financial systems, business databases, document management systems, and external network data sources. The collection scope includes financial statements, business records, contract documents, and regulatory information. Feature extraction and variable determination module: It has a built-in data mining algorithm library and an audit domain knowledge base. It extracts features from the audit evidence in the evidence base based on the rules in the knowledge base and the data mining algorithms in the algorithm library, including decision tree algorithms and clustering algorithms, and determines the variables of the Bayesian network based on the features. Network structure construction module: It has a human-computer interaction interface. Audit experts are invited to input the relationship judgment information between evidences through the interface. Combined with the data-driven association analysis module, the Bayesian network structure is constructed based on the expert opinions and data association analysis results. The constructed network structure is displayed in a graphical way. Conditional probability table generation module: connected to the historical audit case database, used to read historical data, calculate the conditional probability of each node variable in the Bayesian network, generate a conditional probability table, and automatically select the calculation method during the calculation process, including maximum likelihood estimation method or kernel density estimation method; Model evaluation and optimization module: Use cross-validation technology to evaluate the constructed audit evidence credibility assessment model, and automatically determine whether the model needs to be optimized based on the evaluation results. If optimization is required, automatically call the structural learning algorithm, probability adjustment algorithm, etc. to optimize the model. The optimized model is evaluated again until the model performance meets the requirements.

8. The system for constructing a credibility assessment model based on a Bayesian network according to claim 7, characterized in that: In the evidence collection and classification module, anonymization and sentiment analysis technology are used to collect employee feedback records. Anonymization complies with privacy regulations, and sentiment analysis identifies positive, neutral, and negative feedback.

9. The system for constructing a credibility assessment model based on a Bayesian network according to claim 7, characterized in that: In the feature extraction and variable determination module, the audit knowledge graph is regularly updated and maintained, and connected with the data of audit research institutions and industry associations, with no less than 50 knowledge nodes and 100 related edges updated monthly.

10. The system for constructing a credibility assessment model based on a Bayesian network according to claim 7, characterized in that: In the network structure construction module, the visual interactive interface supports multi-language switching and multi-user collaborative operation. Multi-user collaboration supports multiple experts to build network structures simultaneously in different locations.

Images