Mining Audit Business Rules Based On Interpretable Machine Learning

With the continuous advancement of audit informatization,the refinement requirements of the audit business are constantly being improved.In order to improve the compliance of procedures,the authenticity of content,the sufficiency of reasons,and the efficiency of audit judgments,researchers need to conduct regular knowledge discovery from data,Extract audit business rules.Traditional audit business rules mostly use expert experience or statistical analysis methods,but this approach has certain subjective and one-sided deficiencies.Audit business rules mining based on multi-dimensional data and machine learning models can make up for the above deficiencies,but the current popular depth Ways such as learning also have the inexplicability of the black box,which will limit the credibility of the audit rules that are excavated.Therefore,it is necessary to build an interpretable machine learning model.This paper is based on the engineering change data of a certain municipal audit bureau and uses association rules,decision trees and Bayesian networks to extract rules with strong interpretable machine learning methods.In terms of structured data,use the constrained association rule method to discover the association between other features and the target feature,and use such association rules as auditing business rules.The decision tree is used as the classification model of audit business data,and the relationship between the root node and the leaf node is output as the audit business rule.The Bayesian network model is used to study the structure of audit business data to form a directed acyclic graph of the characteristics of the audit business data,and then the parameters are learned.Finally,the inference result of the Bayesian network is used as the audit business rule.In terms of unstructured data,these three methods are also used to mine the conditional dependencies between the subject words in the audit text as audit business rules,revealing the inherent attributes of the audit text.The extraction of audit business rules is a key link in the formation of the audit evidence chain from "engineering data ? audit doubts ? audit evidence ?audit judgment".As an unsupervised machine learning method,association rules are used for knowledge discovery.It has a good effect on the extraction of audit business rules.The rules are intuitive and have confidence as an evaluation of the effectiveness of the rules themselves.Decision tree and Bayesian network model have good classification effect on structured data of audit business.Among them,the accuracy rate,recall rate,F1 value,and AUC value of the decision tree reached82.3%,91.4%,86.6%,and 85.7% respectively.The classification effect of Bayesian network is89.4%,99.0%,94.0%,93.6% in accuracy rate,recall rate,F1 value,and AUC value respectively.Due to the imbalance of audit data samples,the AUC value is used as a more reference value classification evaluation index,and at the same time as the confidence of the output rule.Use interpretable machine learning methods to mine audit business rules,compare the differences and pros and cons of various machine learning models,explore the applicable scenarios of different methods,and expand the application of interpretable machine learning models.The audit business rules mined by interpretable machine learning are objective and credible,and can be used as an auxiliary basis for auditors when making audit decisions.