Intelligent expense reimbursement auditing system and method for financial sharing center

Abstract

The invention discloses an intelligent expense reimbursement auditing system and method for a financial sharing center, and relates to the technical field of intelligent financial risk control, and the method comprises the steps: collecting reimbursement voucher data, and carrying out the preprocessing of the reimbursement voucher data, and generating standardized data; the method comprises the following steps: constructing a dual-task learning architecture based on an XGBoost and LightGBM framework, performing enhancement through a weight loss function and a multi-dimensional dynamic feature cross mechanism to generate an enhanced dual-task learning architecture, obtaining a credit-risk two-dimensional evaluation model through training, and performing credit score regression prediction and risk level classification prediction to obtain a credit score and a risk level; performing integrity verification on the credit score and the risk level through an intelligent shunting strategy engine, and performing credit-risk weight fusion, load dynamic adjustment and risk conduction analysis through a preset shunting strategy matrix to obtain a work order shunting result; according to the invention, through the credit-risk two-dimensional evaluation model, the problem of misjudgment caused by credit and risk separation analysis is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent financial risk control, and particularly to an intelligent expense reimbursement audit system and method for a financial sharing center. Background Art

[0002] In the field of financial sharing centers, expense reimbursement audit technology has gradually evolved from traditional manual processing to an automated audit system based on a rule inference engine. Existing technologies adopt multi-level rule verification and a static risk assessment model, and judge the compliance of reimbursement vouchers through preset thresholds, significantly improving the audit efficiency. Such systems can achieve basic anomaly detection through structured data extraction and rule matching, and support standardized process management.

[0003] There are limitations in the collaborative optimization of credit assessment and risk identification in existing technologies. In particular, the dynamic correlation analysis of credit scores and risk levels is insufficient, resulting in limited flexibility and adaptability of audit strategies. For example, when the applicant has good credit but the voucher has complex risk characteristics, it is difficult for the system to achieve accurate classification through a single rule or static model, which may increase the workload of manual review; existing solutions lack a closed-loop mechanism for continuous self-optimization and cannot adjust audit strategies in real time according to new risk patterns, affecting the long-term risk control effect. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent expense reimbursement audit method for a financial sharing center to solve the problem of rigid audit strategies caused by insufficient credit-risk collaborative assessment.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides an intelligent expense reimbursement audit method for a financial shared service center, which includes collecting reimbursement voucher data and preprocessing it to generate normalized data; constructing a dual-task learning architecture based on the XGBoost and LightGBM frameworks, and enhancing it through a weighted loss function and a multi-dimensional dynamic feature crossing mechanism to generate an enhanced dual-task learning architecture, training a credit-risk two-dimensional evaluation model, and performing credit score regression prediction and risk level classification prediction to obtain a credit score and a risk level; performing integrity verification on the credit score and the risk level through an intelligent shunt strategy engine, and through a preset shunt strategy matrix, performing credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to obtain a work order shunt result; extracting abnormal features and performing compliance verification on the work order shunt result through a dual intelligent audit mechanism to generate an initial expense reimbursement audit decision package; based on the initial expense reimbursement audit decision package, through a rule mining engine and a feature weight optimization algorithm, performing incremental knowledge extraction and parameter adjustment of the dual intelligent audit mechanism to obtain an updated audit knowledge base, and through the Monte Carlo random sampling and adversarial strategy simulation method, performing multi-dimensional deduction on the evolution path of fiscal and tax policies and the evolution trend of fraud means to construct a risk evolution simulation sandbox to obtain a decay-resistant knowledge base; based on the decay-resistant knowledge base, performing final decision analysis on the work order shunt result through a dual intelligent audit mechanism to obtain a final expense reimbursement audit decision package.

[0007] As a preferred embodiment of the intelligent expense reimbursement audit method for the financial shared service center of the present invention, wherein: the steps of collecting reimbursement voucher data and preprocessing it to generate normalized data are as follows, Real-time collect electronic invoice data, itinerary data, payment voucher data, and approval process data to obtain reimbursement voucher data; Perform OCR recognition, data cleaning, format standardization, and association verification on the reimbursement voucher data to generate normalized data.

[0008] As a preferred embodiment of the intelligent expense reimbursement audit method for the financial shared service center of the present invention, wherein: constructing a dual-task learning architecture based on the XGBoost and LightGBM frameworks, and enhancing it through a weighted loss function and a multi-dimensional dynamic feature crossing mechanism to generate an enhanced dual-task learning architecture, the steps are as follows, Combine the XGBoost and LightGBM frameworks and share the feature extraction layer, retain the credit score output end of the XGBoost framework and the risk assessment output end of the LightGBM framework, and perform initialization and space mapping to construct a dual-task learning architecture; Enhance and optimize the performance of the dual-task learning architecture through a dynamic weighted loss function and a multi-dimensional dynamic feature crossing mechanism to obtain an enhanced dual-task learning architecture.

[0009] As a preferred solution of the intelligent expense reimbursement audit method of the financial sharing center described in the present invention, the following steps are included: A credit-risk two-dimensional evaluation model is obtained through training, and credit score regression prediction and risk level classification prediction are carried out to obtain the credit score and risk level. The specific steps are as follows Based on historical normalized data, the enhanced dual-task learning architecture is trained by the phased feature progressive training method combined with the time series cross-validation method to obtain the credit-risk two-dimensional evaluation model; The normalized data is input into the credit-risk two-dimensional evaluation model for credit score regression prediction and risk level classification prediction to obtain the credit score and risk level.

[0010] As a preferred solution of the intelligent expense reimbursement audit method of the financial sharing center described in the present invention, the following steps are included: The integrity of the credit score and risk level is verified by the intelligent shunt strategy engine, and through the preset shunt strategy matrix, credit-risk weight fusion, load dynamic adjustment and risk conduction analysis are carried out to obtain the work order shunt result. The specific steps are as follows Through the intelligent shunt strategy engine, the integrity of the credit score and risk level is verified to generate a standardized two-dimensional evaluation data packet; The standardized two-dimensional evaluation data packet is mapped to the cells of the preset shunt strategy matrix through the coordinate mapping mechanism, and credit-risk weight fusion, load dynamic adjustment and risk conduction analysis are carried out to generate the work order shunt result.

[0011] As a preferred solution of the intelligent expense reimbursement audit method of the financial sharing center described in the present invention, the following steps are included: Abnormal feature extraction and compliance verification are carried out on the work order shunt result through the dual intelligent audit mechanism to generate the initial expense reimbursement audit decision package. The specific steps are as follows Based on the work order shunt result, abnormal features are extracted through the deep learning model in the dual intelligent audit mechanism, and at the same time, the rule inference engine is used for compliance verification to obtain the preliminary audit conclusion and risk identification; Through the decision weighting algorithm, the confidence of the preliminary audit conclusion and risk identification is fused to generate the initial expense reimbursement audit decision package.

[0012] As a preferred solution of the intelligent expense reimbursement audit method of the financial sharing center described in the present invention, the following steps are included: Based on the initial expense reimbursement audit decision package, through the rule mining engine and the feature weight optimization algorithm, incremental knowledge extraction and parameter adjustment of the dual intelligent audit mechanism are carried out to obtain the updated audit knowledge base. The specific steps are as follows The initial expense reimbursement audit decision package is input into the rule mining engine, the FP-Growth algorithm is used to mine association rules, and the confidence of the association rules is calculated. The association rules higher than the association rule confidence threshold are screened to generate a new set of association rules; Based on the new association rule set, adjust the feature weight matrix and priority parameters in the dual intelligent review mechanism through the gradient descent method, and output the optimized parameter set of the dual intelligent review mechanism; Fuse and resolve conflicts for the new association rule set and the optimized parameter set of the dual intelligent review mechanism through the versioned incremental merge mechanism to generate an updated review knowledge base.

[0013] As a preferred solution of the intelligent expense reimbursement review method for the financial sharing center described in the present invention, wherein: through the Monte Carlo random sampling and adversarial strategy simulation method, perform multi-dimensional deduction on the evolution path of fiscal and tax policies and the evolution trend of fraud means, and construct a risk evolution simulation sandbox to obtain an anti-decay knowledge base. The specific steps are as follows. Based on the updated review knowledge base, perform simulation on the evolution path of fiscal and tax policies through the Monte Carlo random sampling method to obtain a simulation data set of the evolution path of fiscal and tax policies, and predict the evolution trend of fraud means to obtain the prediction result of the evolution trend of fraud means; Based on the simulation data set of the evolution path of fiscal and tax policies and the prediction result of the evolution trend of fraud means, construct a risk evolution simulation sandbox through the multi-dimensional deduction method to obtain an anti-decay fortified review knowledge base.

[0014] As a preferred solution of the intelligent expense reimbursement review method for the financial sharing center described in the present invention, wherein: based on the anti-decay knowledge base, perform final decision analysis on the work order diversion result through the dual intelligent review mechanism to obtain the final expense reimbursement review decision package. The specific steps are as follows. Load the anti-decay fortified review knowledge base through the dual intelligent review mechanism, perform anomaly feature detection and compliance verification on the work order diversion result, and obtain the preliminary review determination result and risk correction suggestions; Perform final decision confidence fusion calculation and encapsulation on the preliminary review determination result and risk correction suggestions to generate an expense reimbursement review decision package.

[0015] In a second aspect, the present invention provides an intelligent expense reimbursement audit system for a financial sharing center, including a data collection, evaluation module, shunt strategy module, audit evolution module, optimization module, and output module; the data collection is used to collect reimbursement voucher data and preprocess it to generate normalized data; the evaluation module is used to construct a dual-task learning architecture based on the XGBoost and LightGBM frameworks, and enhance it through a weighted loss function and a multi-dimensional dynamic feature crossing mechanism to generate an enhanced dual-task learning architecture. A credit-risk two-dimensional evaluation model is obtained through training, and credit score regression prediction and risk level classification prediction are performed to obtain a credit score and a risk level; the shunt strategy module is used to perform integrity verification on the credit score and risk level through an intelligent shunt strategy engine, and through a preset shunt strategy matrix, perform credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to obtain a work order shunt result; the audit evolution module is used to extract abnormal features and verify compliance for the work order shunt result through a dual intelligent audit mechanism to generate an initial expense reimbursement audit decision package; the optimization module is used to, based on the initial expense reimbursement audit decision package, perform incremental knowledge extraction and parameter adjustment of the dual intelligent audit mechanism through a rule mining engine and a feature weight optimization algorithm to obtain an updated audit knowledge base, and through the Monte Carlo random sampling and adversarial strategy simulation method, perform multi-dimensional deduction on the evolution path of fiscal and tax policies and the evolution trend of fraud means to construct a risk evolution simulation sandbox to obtain an anti-decay knowledge base; the output module is used to, based on the anti-decay knowledge base, perform final decision analysis on the work order shunt result through a dual intelligent audit mechanism to obtain a final expense reimbursement audit decision package.

[0016] In a third aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the intelligent expense reimbursement audit method for a financial sharing center as described in the first aspect of the present invention is implemented.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of the intelligent expense reimbursement audit method for a financial sharing center as described in the first aspect of the present invention is implemented.

[0018] The beneficial effects of the present invention are as follows: By constructing a credit-risk two-dimensional evaluation model, synergistically integrating the XGBoost and LightGBM frameworks and adopting a dynamic feature crossing mechanism, unified modeling of credit evaluation and risk identification is achieved, reducing the misjudgment problems caused by the separate analysis of credit and risk, and improving the accuracy of auditing and business adaptability; at the same time, through incremental knowledge extraction and dynamic parameter adjustment, continuously updating the audit knowledge base based on rule mining and gradient optimization, the self-evolution of audit rules and the dynamic optimization of model parameters are realized, solving the problem that the static audit system is difficult to adapt to business changes, and ultimately improving the intelligent level of the system and the timeliness of risk control, forming a closed-loop optimization from data input to decision output, making the entire audit process have accuracy, adaptability and continuous evolution ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is a flowchart of the intelligent expense reimbursement audit method for the financial shared service center.

[0021] Figure 2 It is a schematic diagram of the intelligent expense reimbursement audit system for the financial shared service center.

[0022] Figure 3 It is a flowchart of generating credit scores and risk levels.

[0023] Figure 4 It is a flowchart of generating the result of work order diversion. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings of the specification.

[0025] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0026] Second, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure or characteristic that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that are mutually exclusive of other embodiments.

[0027] Referring to Figures 1 to 4 , which is an embodiment of the present invention. This embodiment provides an intelligent expense reimbursement audit method for a financial sharing center, including the following steps: S1. Collect reimbursement voucher data and perform preprocessing to generate normalized data.

[0028] S1.1. Real-time collect electronic invoice data, itinerary data, payment voucher data and approval process data to obtain reimbursement voucher data.

[0029] Specifically, connect to the electronic invoice platform through an application programming interface to obtain electronic invoice data, extract itinerary data from the airline and railway system interfaces, capture payment voucher data by associating with the bank payment gateway, and synchronize the enterprise approval system interface to obtain approval process data; the collection of electronic invoice data includes fields such as invoice code, invoice number, invoice date, and verification code; the collection of itinerary data includes fields such as passenger name, flight number, departure and arrival locations, and ticket price; the collection of payment voucher data includes fields such as transaction time, transaction amount, payee name, and transaction serial number; the collection of approval process data includes fields such as approver position, approval opinion, and approval timestamp; a unique identifier is generated for all collected fields through the MD5 algorithm, and after verifying the data integrity, they are merged into reimbursement voucher data.

[0030] S1.2. Perform OCR recognition, data cleaning, format standardization and correlation verification on the reimbursement voucher data to generate normalized data.

[0031] Specifically, use the Tesseract OCR engine to perform optical character recognition on the electronic invoice data and itinerary data, and extract printed text information, such as the invoice code, invoice number, and invoice date fields in the electronic invoice data, and the passenger name and flight number fields in the itinerary data. The recognition result is output as UTF-8 encoded text data; remove non-standard characters in the itinerary data through regular expression matching, such as deleting the currency symbol in the ticket price field; perform outlier detection on the payment voucher data, such as marking the data as to be verified when the transaction amount is negative; verify the integrity of the required fields in the electronic invoice data, and trigger the completion process when key fields are missing; Uniformly convert the transaction time of payment voucher data into a standardized 24-hour date and time format; uniformly retain two decimal places for monetary amounts; reorganize the multi-level approval opinions in the approval process data in hierarchical order; perform the conversion from abbreviated name to full name for the purchaser name in the e-invoice data. Compare the verification code of the e-invoice data with the tax system filing record through the hash value; verify the consistency between the passenger ID number in the itinerary data and the applicant information in the approval process data; perform real-time verification of the transaction serial number of the payment voucher data with the bank system record, and generate standardized data including invoice identifiers, itinerary numbers, payment serial numbers, and approval form numbers.

[0032] S2. Construct a dual-task learning architecture based on the XGBoost and LightGBM frameworks, and enhance it through the weighted loss function and multi-dimensional dynamic feature crossing mechanism to generate an enhanced dual-task learning architecture.

[0033] S2.1. Combine the XGBoost (extreme gradient boosting framework) and LightGBM framework (lightweight gradient boosting framework) and share the feature extraction layer. Retain the credit score output end of the XGBoost framework and the risk assessment output end of the LightGBM framework, and perform initialization and space mapping to construct a dual-task learning architecture.

[0034] Specifically, the get_booster.get_score (obtain booster - obtain feature importance score) method of the XGBoost framework is used to extract the reimbursement feature importance ranking of the normalized data, and the split_feature_importance (split feature importance evaluation method) method of the LightGBM framework is used to obtain the reimbursement feature importance vector of the normalized data of the same dimension; align the top 20 common features of the reimbursement feature importance of the normalized data of the XGBoost framework and the LightGBM framework, such as invoice amount, invoicing frequency, merchant type field; input the normalized data corresponding to the common features into the layer for parameter sharing, and use the Glorot (Glorot uniform distribution initialization method) uniform distribution to initialize the weight matrix of the shared layer to complete the initialization of the network parameters; through the t-SNE dimensionality reduction algorithm, obtain the similarity probability distribution of the feature vectors in the high-dimensional space, and map the high-dimensional feature vectors extracted by the XGBoost and LightGBM architectures to a two-dimensional plane. Randomly distribute the positions of the feature points in the low-dimensional space during the initialization stage to generate the corresponding relationship between the high-dimensional feature similarity and the distance of the low-dimensional space points. In the iteration stage, adjust the positions of the low-dimensional points to maintain the consistency of the probability distributions in the high-dimensional space and the low-dimensional space, so that similar features in the original high-dimensional feature space are kept in close proximity after dimensionality reduction, and different features are kept far apart. Observe the change trend of the probability distribution difference in each iteration, and determine that the convergence is completed when the fluctuation of the distribution difference tends to be stable. In the finally obtained two-dimensional projection, the common features such as invoice amount and invoicing frequency extracted by the two XGBoost and LightGBM architectures form clear clustering regions, and the non-common features are scattered at the edge positions, realizing the visual alignment and comparison of the feature spaces of the two algorithms; the XGBoost framework retains the fully connected layer as the credit score output end, and the LightGBM framework retains the Sigmoid output layer as the risk assessment output end; integrate the shared feature extraction layer and the dual output ends into a unified dual-task learning architecture.

[0035] S2.2. Enhance and optimize the dual-task learning architecture through the dynamic weighted loss function and the multi-dimensional dynamic feature cross mechanism to obtain the enhanced dual-task learning architecture.

[0036] Specifically, it is obtained by randomly splitting the historical normalized data in a 7:3 ratio. 70% of the historical normalized data is used as the training set of the historical normalized data, and 30% is used as the validation set of the historical normalized data. During the splitting, the distribution consistency of the credit score and the risk level is maintained; a mean squared error loss function is configured for the credit score output end, and a cross-entropy loss function is configured for the risk assessment output end. The exemplary initial weight ratio is set to 0.7:0.3; during the training process, the loss weight is dynamically adjusted according to the change in the comprehensive index score of the performance on the historical normalized data validation set. For example, when the comprehensive index of the risk assessment performance drops by more than 5%, the loss weight is increased by 0.1; the invoice amount and the invoicing frequency in the structured reimbursement features are combined by Cartesian product to generate two-dimensional cross features, and the merchant type and the invoicing time interval are combined by one-hot encoding to generate three-dimensional cross features; after splicing the structured reimbursement features and the three-dimensional cross features, they are input into the shared layer of the dual-task learning architecture; after every 5 trainings, the R-squared value of the credit score output and the AUC value of the risk assessment output are calculated using the validation set. When the increase in the R-squared value is less than the exemplary 0.01 and the increase in the AUC value is less than 0.005, the training is terminated, and the current parameters are saved as the enhanced dual-task learning architecture; It should be noted that the expressions for calculating the R-squared value of the credit score output and the AUC value of the risk assessment output are as follows: ; Among them, is the exemplary value range (0 - 1) of the R-squared value of the credit score output, is the total number of samples in the validation set, is the sample index variable of the validation set, is the true credit score value of the th sample in the validation set, with an exemplary value range (300 - 850), is the predicted credit score value of the th sample by the dual-task learning architecture, with an exemplary value range (300 - 850), is the average value of all true credit scores in the validation set, with an exemplary value range (650 ± 50), is the exemplary value range (0.5 - 1) of the AUC value of the risk assessment output, is the true positive rate, is the false positive rate, is the arithmetic operator, is the exemplary value range (0 - 1) of the false positive rate FPR, is the infinitesimal increment of FPR.

[0037] S3. Obtain a credit-risk two-dimensional evaluation model through training, and perform credit score regression prediction and risk level classification prediction to obtain the credit score and the risk level.

[0038] S3.1. Based on the historical normalized data, the enhanced dual-task learning architecture is trained by the phased feature progressive training method combined with the time series cross-validation method to obtain a credit-risk two-dimensional evaluation model.

[0039] Specifically, the historical normalized data is sorted by timestamp and divided into 5 exemplary time windows, each window containing 6 consecutive months of data; in the first stage of training, only the basic feature fields of time windows 1-3 are used, such as invoice amount, invoicing frequency, merchant type, and input into the enhanced dual-task learning architecture for 10 rounds of iteration; in the second stage, the derived feature fields of time window 4 are added, such as the cross-feature of merchant type and invoicing frequency, and continue to train for 5 rounds; in the third stage, all the feature fields of time window 5 are introduced, including time series features such as payment interval duration, approval timeliness, etc., and complete the final 3 rounds of training; in each stage, the TimeSeriesSplit method is used for exemplary 3-fold time series cross-validation to ensure that the time of the historical normalized data test set is always later than that of the historical normalized data training set; when the R-squared value of the credit score output on the historical normalized data validation set fluctuates less than exemplary 0.01 for 3 consecutive rounds and the AUC value of the risk assessment output fluctuates less than 0.005, the training is terminated, and the parameters are saved to generate a credit-risk two-dimensional evaluation model.

[0040] S3.2. Input the normalized data into the credit-risk two-dimensional evaluation model for credit score regression prediction and risk level classification prediction to obtain the credit score and risk level.

[0041] Specifically, the invoice amount field in the normalized data is processed by Min-Max normalization and mapped to the [0,1] interval; the merchant type field is converted into a one-hot encoded vector to generate a sparse binary vector containing, for example, 20 dimensions; the sparse binary vector is input into the shared feature extraction layer of the credit-risk two-dimensional evaluation model, and a 128-dimensional feature representation is generated through 3 fully connected networks; the 128-dimensional feature representation is respectively input into the linear regression layer of the credit score output end and the Sigmoid classification layer of the risk assessment output end; the credit score output layer uses a linear activation function to generate a continuous numerical value in the range of, for example, [300,850]; the risk level output layer generates a risk level probability value in the [0,1] interval through the Sigmoid function, and an exemplary risk threshold of 0.65 is set. When the risk level probability value ≥ 0.65, it is determined as a high risk level, otherwise it is a low risk level; finally, the exemplary result is output: the credit score is 725 points, and the risk level is "high risk"; Furthermore, the process of setting the risk threshold is as follows: calculate the distribution of risk probability values output by the credit-risk two-dimensional evaluation model on the validation set, and plot the true positive rate (TPR) and false positive rate (FPR) curves under different risk thresholds; determine the optimal risk threshold through the principle of maximizing the Youden index (TPR - FPR).

[0042] S4. Integrity verification is performed on the credit score and risk level through the intelligent shunt strategy engine, and credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis are carried out through the preset shunt strategy matrix to obtain the work order shunt result.

[0043] S4.1. Through the intelligent shunt strategy engine, integrity verification is performed on the credit score and risk level to generate a standardized two-dimensional evaluation data packet.

[0044] Specifically, the intelligent shunt strategy engine receives the credit score and risk level data output by the credit-risk two-dimensional evaluation model, performs a numerical range check on the credit score to confirm that the credit score is within the exemplary range of [300, 850]; performs an enumerated value check on the risk level to confirm that the risk level is "high risk" or "low risk"; checks the association logic between the credit score and the risk level, and exemplarily verifies the logical consistency of the high-risk level corresponding to a credit score lower than 600 points; encapsulates the credit score and risk level data that pass the verification in JSON format, marks the credit score field as "creditScore", and the risk level field as "riskLevel"; adds a timestamp field to record the evaluation time; finally generates a standardized two-dimensional evaluation data packet containing the credit score, risk level, and timestamp.

[0045] S4.2. Map the standardized two-dimensional evaluation data packet to the cells of the preset shunt strategy matrix through the coordinate mapping mechanism, and perform credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to generate the work order shunt result.

[0046] Specifically, map the credit score of 725 points in the standardized two-dimensional evaluation data packet to the 3rd row of the preset shunt strategy matrix according to the preset interval [700 - 750]; map the risk level of "high risk" to the 5th column of the matrix; read the initial weight configuration of the cell (3, 5) of the preset shunt strategy matrix, and exemplarily obtain a credit weight of 0.6 and a risk weight of 0.4; dynamically adjust the risk weight according to the current load status of the review queue, and exemplarily reduce the risk weight by 0.1 when it is detected that the backlog of high-risk work orders exceeds 50; retrieve the risk conduction records of the same type of work orders in the past 3 hours, and exemplarily increase the risk weight by 0.2 if 3 associated merchant risk events are found; substitute the final risk weight into the linear weighting formula to calculate the priority score; match the preset shunt strategy matrix according to the priority score interval to generate a work order shunt result containing the channel identifier and weight details; It should be noted that the setting process of the preset diversion strategy matrix is as follows: based on the statistical analysis of historical audit data, the credit score is divided into 5 intervals (the exemplary interval boundaries are [300-500), [500-600), [600-700), [700-800), and [800-850]), and the risk level is divided into 5 levels (the exemplary levels 1 to 5 correspond to the risk probabilities [0-0.2), [0.2-0.4), [0.4-0.6), [0.6-0.8), and [0.8-1.0]); the diversion effects of different interval combinations are tested on the validation set by the grid search method, and it is determined that the audit efficiency is optimal when the credit interval width is 200 points and the risk level interval is 0.2; each matrix cell presets the initial weight, and the exemplary high-risk cell (1,5) is set with a credit weight of 0.3 and a risk weight of 0.7, and the low-risk cell (5,1) is set with a credit weight of 0.8 and a risk weight of 0.2; and finally a 5-row × 5-column diversion strategy matrix is generated.

[0047] S5. Through the dual intelligent review mechanism, abnormal features are extracted and compliance verification is performed on the work order diversion results to generate an initial expense reimbursement review decision package.

[0048] S5.1. Based on the work order diversion results, the deep learning model in the dual intelligent audit mechanism is used to extract abnormal features, and the rule reasoning engine is used to verify compliance to obtain preliminary audit conclusions and risk identification.

[0049] Specifically, a high-risk work order training set is constructed based on high-risk work orders in the historical work order diversion results, where the "transferred to manual review" work order is marked as 1 and the "automatically passed" work order is marked as 0; invoice image features are extracted, and the digital recognition in the invoice amount area is focused on as an example; the Transformer editor is used to analyze the time series features in the work order diversion results, and the time series pattern of "transferred to manual review" for three consecutive times is modeled as an example; the weighted cross entropy loss function is used in the training process, and the abnormal sample weight is set to 3.0 as an example; the model parameters are frozen after the weighted cross entropy loss function score on the high-risk work order verification set reaches an example of 0.88, and the trained deep learning model is obtained; Extract 200 compliance clauses from historical normalized data, including "the amount of a single invoice shall not exceed 10,000 yuan" as an example; use the Drools rule inference engine syntax to write a decision table, and set a hard rule that "travel expense reimbursement must be accompanied by an approval form" as an example; verify the rule coverage through the test set, and ensure that 95% of common violation situations can be covered by the rules as an example, and obtain the trained rule inference engine; Input the work order diversion result into the trained deep learning model to extract the invoice image features, and exemplarily obtain the digital features of the invoice amount area; input the historical sequence of work order diversion into the Transformer encoder to exemplarily analyze the timing pattern of "transfer to manual review" for 3 consecutive times; splice the extracted invoice image features and timing pattern and input them into the fully connected layer, and output the anomaly probability value between 0 and 1 through the Sigmoid activation function, and exemplarily calculate an anomaly score of 0.87; synchronously trigger the rule inference engine for verification, and exemplarily detect a violation item of "the invoice amount of 12,000 yuan exceeds the limit"; merge the anomaly probability value of the deep learning model with the violation entries of the rule inference engine, and exemplarily generate a preliminary review conclusion and risk identification including "high risk probability 0.87" and "amount exceeding the standard violation".

[0050] S5.2. Through the decision-weighted algorithm, fuse the confidence levels of the preliminary review conclusion and risk identification to generate the initial expense reimbursement review decision package.

[0051] Specifically, by analyzing the anomaly probability confidence level output by the deep learning model in the preliminary review conclusion; synchronously reading the detection results of the rule inference engine in the risk identification and analyzing the rule matching degree with the rule inference engine detection to obtain the risk identification compliance confidence level; according to the weight configuration parameters in the credit-risk two-dimensional evaluation model, obtain the weight allocation ratio of the deep learning model result and the rule inference engine detection result in the current business scenario; perform weighted synthesis on the anomaly probability confidence level and the risk identification compliance confidence level according to the weight ratio; retrieve the preset review standard library and match the review standard corresponding to the current business type; compare the fused anomaly probability confidence level and compliance confidence level with the review standard, and generate a "not passed" conclusion when the fused anomaly probability confidence level and compliance confidence level reach or exceed the review standard, and generate a "passed" conclusion when they do not reach, and finally generate the review conclusion; encapsulate the review conclusion, anomaly probability confidence level, and compliance confidence level into the initial expense reimbursement review decision package; It should be noted that the process of the preset review standard library: analyze the rule effectiveness in the historical review data, and count the actual accuracy performance of the rules in different business scenarios; combine the risk control requirements in the enterprise financial management system to determine the minimum review passing standard for each business type; finally, store the minimum review standard by business category as the review standard library.

[0052] S6. Based on the initial expense reimbursement review decision package, through the rule mining engine and feature weight optimization algorithm, perform incremental knowledge extraction and adjust the parameters of the dual intelligent review mechanism to obtain the updated review knowledge base.

[0053] S6.1. Input the initial expense reimbursement review decision package into the rule mining engine, use the FP-Growth algorithm to mine association rules, calculate the confidence of the association rules, filter out association rules that are higher than the confidence threshold of the association rules, and generate a new association rule set.

[0054] It should be noted that the expression for calculating the confidence of association rules is: ; in, is the confidence of the association rule, It satisfies the conditions at the same time And Conclusion The number of work orders, is the number of work orders that meet the conditions only, is the number of work orders that meet the conclusion; Specifically, parse the initial expense reimbursement review decision package, read the review conclusion fields one by one, and filter out the work order records with the conclusion of "not passed"; convert the feature data of each rejected work order into a transaction item set format; configure the minimum support of the FP-Growth algorithm to 5%, the minimum association rule confidence to 75%, execute the FP-Growth algorithm to mine frequent item sets, and exemplarily find that the support of the combination of ["amount exceeds limit", "hotel category"] is 20%; generate candidate association rules for each frequent item set, and exemplarily generate association rules with the antecedent of "amount exceeds limit" and the consequent of "not passed" "; calculate the confidence of the candidate association rules, count the number of work orders containing both "amount exceeded limit" and "not approved" in the initial expense reimbursement review decision package, 200, and the number of work orders containing only "amount exceeded limit" is 250, calculate the confidence of the association rules; compare the association rule confidence example 80% with the preset association rule confidence threshold 75%, 80%>75% meets the condition; store the association rules in a structured form; traverse all frequent item sets and repeat the above association rule generation and screening process, each association rule meets the conditions of support ≥5% and confidence ≥75%. Generate a new association rule set; It should be noted that the preset process of the preset association rule confidence threshold is: analyze the actual performance distribution of association rules in historical expense reimbursement review decisions, and use the ROC curve analysis method to evaluate the effects of different association rule confidence thresholds; observe the balance between the true positive rate and the false positive rate, and exemplify that when the confidence rate is 75%, an 85% high-risk transaction recognition rate and a 10% misjudgment rate are achieved; and 75% is used as the association rule confidence threshold.

[0055] S6.2. Based on the new association rule set, the feature weight matrix and priority parameters in the dual intelligent audit mechanism are adjusted by the gradient descent method, and the optimized dual intelligent audit mechanism parameter set is output.

[0056] Specifically, parse the rule entries in the new association rule set, and exemplarily extract high-confidence association rules such as "amount exceeding limit → not approved"; construct a training data set including historical expense reimbursement review decisions and new association rules, and exemplarily integrate 10,000 review records and 15 new association rules; initialize the parameters of the dual intelligent review mechanism, and exemplarily set the initial value of the feature weight matrix of the deep learning model to 0.5 and the initial value of the priority parameter of the rule inference engine to 1.0; configure the gradient descent optimizer, and exemplarily set the learning rate to 0.01 and the batch size to 32; perform parameter iterative update, obtain the matching error between the current expense reimbursement review decision and the new association rules, and exemplarily obtain an error value of 0.15; adjust the parameters in the direction of the error gradient, and exemplarily update the feature weight matrix of the deep learning model to 0.48 and the priority parameter of the rule inference engine to 1.05; repeat the iterative process until the exemplary maximum number of iterations of 100 times is reached; save the parameter values of the finally optimized dual intelligent review mechanism, and exemplarily generate a parameter set of the dual intelligent review mechanism including a feature weight matrix of 0.47 and a priority parameter of 1.08.

[0057] S6.3. Integrate and resolve conflicts between the new association rule set and the optimized parameter set of the dual intelligent review mechanism through a versioned incremental merge mechanism to generate an updated review knowledge base.

[0058] Specifically, compare each rule entry in the new association rule set with the rule inference engine in the existing dual intelligent review mechanism one by one. Exemplarily, it is found that there is an overlap in business logic between the newly added association rule "amount exceeding limit → not approved" and the old rule "large amount invoice → manual review"; execute the conflict resolution strategy, and preferentially retain the rule with a higher confidence level. Exemplarily, select to retain the new rule with a confidence level of 80% and discard the old rule with a confidence level of 75%; convert the optimized parameter set of the dual intelligent review mechanism into a configuration file format; finally, package and generate an updated review knowledge base including exemplarily 15 new association rules.

[0059] S7. Through the Monte Carlo random sampling and adversarial strategy simulation method, conduct multi-dimensional deduction on the evolution path of fiscal and tax policies and the evolution trend of fraud means, and construct a risk evolution simulation sandbox to obtain an anti-decay knowledge base.

[0060] S7.1. Based on the updated review knowledge base, simulate the evolution path of fiscal and tax policies through the Monte Carlo random sampling method to obtain a simulation data set of the evolution path of fiscal and tax policies, and predict the evolution trend of fraud means to obtain the prediction result of the evolution trend of fraud means.

[0061] Specifically, the historical change records of fiscal and tax policies in the updated audit knowledge base are structured, and feature dimensions such as policy type, effective time, and adjustment range are extracted to establish a policy feature vector library. Based on the probability distribution characteristics of historical change records, the Monte Carlo random sampling method is used to conduct multiple rounds of independent sampling in the fiscal and tax policy feature library. Each sampling randomly selects a fiscal and tax policy category from the fiscal and tax policy type distribution, generates fiscal and tax policy change parameters according to the historical adjustment range distribution, and generates simulated effective time points according to the statistical law of the fiscal and tax policy change time interval. For each combination of fiscal and tax policy features generated by sampling, the historical tax payment data of enterprises is associated to generate policy impact evaluation indicators such as tax rate change sensitivity and deductible item impact range. The fiscal and tax policy feature vectors, corresponding timestamps, and generated evaluation indicators of each sampling are structurally combined. After multiple rounds of iterative sampling, all sampling results are sorted and integrated according to the time series to form a simulated dataset of the fiscal and tax policy change path with complete time series characteristics and impact evaluation indicators. Based on the fiscal and tax policy type, adjustment range, and effective time characteristics in the simulated dataset of the fiscal and tax policy change path, feature matching is performed with the fraud patterns recorded in the historical fraud case library to form a fiscal and tax policy-fraud correspondence. The adversarial strategy simulation method is used to conduct multiple rounds of strategy simulations within the preset strategy space of three types of fraud behaviors: false invoicing, fictitious expenses, and forged bills. In each round of simulation, the corresponding initial fraud strategy is selected according to the fiscal and tax policy change characteristics, and the change in the fraud strategy intensity is dynamically generated according to the fiscal and tax policy adjustment range, such as the correlation ratio between the change in the value-added tax rate and the fictitious amount. The mutation characteristics of fraud means generated in each round of simulation are recorded, and the fraud feature combinations that frequently appear are integrated according to the fiscal and tax policy timeline to output the prediction results of the evolution trend of fraud means, including fraud type, evolution time node, and intensity change characteristics.

[0062] S7.2. Based on the simulated dataset of the fiscal and tax policy change path and the prediction results of the evolution trend of fraud means, a risk evolution simulation sandbox is constructed through a multi-dimensional deduction method to obtain an audit knowledge base with anti-decay reinforcement.

[0063] Specifically, based on the tax and fiscal policy types, adjustment amplitudes, and effective time characteristics in the tax and fiscal policy change path simulation dataset, a spatio-temporal alignment and matching is performed with the fraud types, evolution time nodes, and intensity change characteristics in the fraud means evolution trend prediction results to form a tax and fiscal policy-fraud association matrix. In the multi-dimensional deduction environment, three deduction axes of time dimension, tax and fiscal policy dimension, and fraud dimension are set. The time dimension is divided by exemplary quarters, the tax and fiscal policy dimension is divided into three categories of exemplary tax rate, deduction, and collection and management, and the fraud dimension is divided into three categories of exemplary bills, amounts, and subjects. When performing the deduction, the corresponding fraud evolution characteristics are inserted on the time axis at the tax and fiscal policy effective nodes. When the adjustment amplitude of the tax and fiscal policy exceeds the exemplary 5%, the associated fraud intensity adjustment is triggered. When the occurrence frequency of the fraud characteristics exceeds the exemplary 60%, new audit rules are generated. The rule changes, parameter adjustments, and characteristic weight changes generated during the deduction process are updated in real time to the audit knowledge base. After multiple rounds of deduction iteration of the complete tax and fiscal policy cycle, an anti-decay fortified audit knowledge base containing time-sensitive rules, dynamic parameters, and enhanced characteristics is output.

[0064] S8. Based on the anti-decay knowledge base, through a dual intelligent audit mechanism, a final decision analysis is performed on the work order diversion result to obtain a final expense reimbursement audit decision package.

[0065] S8.1. Through the dual intelligent audit mechanism, load the anti-decay fortified audit knowledge base to perform an anomaly feature detection compliance check on the work order diversion result, and obtain a preliminary audit determination result and risk correction suggestions.

[0066] Specifically, through the dual intelligent audit mechanism, load the time-sensitive rules, dynamic parameters, and enhanced characteristics in the anti-decay fortified audit knowledge base to perform a structured analysis on the reimbursement voucher data in the work order diversion result, and extract audit feature items such as bill type, amount range, and expense subject. Match the time-sensitive rules of the audit knowledge base in the rule inference engine, and exemplarily execute the "large bill review rule is triggered when the amount of the special VAT invoice exceeds the exemplary 100,000 yuan"; in the deep learning model, apply dynamic parameters to allocate the feature weights of the deep learning model, and exemplarily increase the weight of the newly added expense subject after the policy adjustment by the exemplary 30%. When performing anomaly feature detection, compare the bill authenticity verification result with the bill feature library of the audit knowledge base, and exemplarily detect abnormal bills with invoice codes that do not conform to the latest coding rules of the tax authorities; when performing compliance verification, check the expense subject against the deductible subject list after the policy adjustment, and exemplarily find that the subjects abolished by the policy are still in use. Output a preliminary audit determination result containing abnormal bill marks and non-compliant subject identifications, and at the same time generate risk correction suggestions such as exemplarily "replace with a compliant invoice" and "adjust the expense subject".

[0067] S8.2. Perform a final decision confidence fusion calculation on the preliminary review judgment result and the risk correction suggestion, and encapsulate them to generate an expense reimbursement review decision package.

[0068] It should be noted that the expression for performing a final decision confidence fusion calculation on the preliminary review judgment result and the risk correction suggestion is: ; Where, is an exemplary value of the final decision confidence (0 - 1), is an exemplary value of the deep learning model confidence (0 - 1), is an exemplary value of the confidence output by the rule inference engine (0 - 1), is an exemplary value of the risk correction suggestion compensation coefficient (0.2 - 0.8), is the weight of the deep learning model, is the weight of the rule inference engine, is an exemplary value of the risk compensation coefficient (0 - 0.1); Specifically, perform a weighted fusion on the deep learning model confidence and the confidence output by the rule inference engine in the preliminary review judgment result. The weight of the deep learning model is set to an exemplary 0.6, and the weight of the rule inference engine is set to an exemplary 0.4; evaluate the compensation coefficient for the risk correction suggestion. The compensation coefficient for exemplary major risk correction suggestions is taken as 0.8, and the compensation coefficient for general risk correction suggestions is taken as 0.5; perform a confidence fusion calculation, add the sum of the weighted model confidence and the rule confidence plus the risk compensation coefficient adjustment amount. The risk compensation coefficient is set to an exemplary 0.05; when the fusion calculation result exceeds an exemplary 0.85, it is determined as passed, and when it is lower than an exemplary 0.45, it is determined as rejected. An artificial review mark is generated in the intermediate interval; encapsulate the final decision result and the risk correction suggestion in JSON format, including three data segments: the review conclusion, the confidence score, and the risk prompt item, to generate an expense reimbursement review decision package for structured storage.

[0069] This embodiment also provides an intelligent expense reimbursement review system for a financial shared center, including: A data collection, evaluation module, diversion strategy module, review evolution module, optimization module, and output module; Data collection is used to collect reimbursement voucher data and preprocess it to generate normalized data; The evaluation module is used to construct a dual-task learning architecture based on the XGBoost and LightGBM frameworks, and enhance it through a weighted loss function and a multi-dimensional dynamic feature cross mechanism to generate an enhanced dual-task learning architecture. Obtain a credit-risk two-dimensional evaluation model through training, and perform credit score regression prediction and risk level classification prediction to obtain the credit score and the risk level; The shunt strategy module is used to perform integrity verification on the credit score and risk level through the intelligent shunt strategy engine, and through the preset shunt strategy matrix, perform credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to obtain the work order shunt result; The audit evolution module is used to extract abnormal features and perform compliance verification on the work order shunt result through a dual intelligent audit mechanism, and generate an initial expense reimbursement audit decision package; The optimization module is used to, based on the initial expense reimbursement audit decision package, through the rule mining engine and the feature weight optimization algorithm, perform incremental knowledge extraction and parameter adjustment of the dual intelligent audit mechanism to obtain an updated audit knowledge base, and through the Monte Carlo random sampling and adversarial strategy simulation method, perform multi-dimensional deduction on the change path of fiscal and tax policies and the evolution trend of fraud means, and construct a risk evolution simulation sandbox to obtain an anti-decay knowledge base; The output module is used to, based on the anti-decay knowledge base, perform final decision analysis on the work order shunt result through the dual intelligent audit mechanism to obtain the final expense reimbursement audit decision package.

[0070] This embodiment also provides a computer device applicable to the intelligent expense reimbursement audit method in the financial shared center, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent expense reimbursement audit method in the financial shared center as proposed in the above embodiment.

[0071] This computer device can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of this computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0072] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the intelligent expense reimbursement audit method for the financial sharing center proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, abbreviated as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, abbreviated as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, abbreviated as EPROM), programmable read-only memory (Programmable Red-Only Memory, abbreviated as PROM), read-only memory (Read-Only Memory, abbreviated as ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0073] In summary, the present invention achieves the following: by constructing a credit-risk two-dimensional evaluation model, synergistically integrating the XGBoost and LightGBM frameworks and adopting a dynamic feature crossing mechanism, it realizes the unified modeling of credit evaluation and risk identification, reduces the misjudgment problem caused by the split analysis of credit and risk, and improves the audit accuracy and business adaptability; at the same time, through incremental knowledge extraction and dynamic parameter adjustment, continuously updating the audit knowledge base based on rule mining and gradient optimization, it realizes the self-evolution of audit rules and the dynamic optimization of model parameters, solves the problem that the static audit system is difficult to adapt to business changes, and ultimately improves the system intelligence level and risk control timeliness. The synergistic effect of these two steps forms a closed-loop optimization from data input to decision output, making the entire audit process have accuracy, adaptability and continuous evolution ability.

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An intelligent expense reimbursement audit method for a financial shared service center, characterized in that: including, collecting reimbursement voucher data and preprocessing it to generate normalized data; constructing a dual-task learning architecture based on the XGBoost and LightGBM frameworks, enhancing it through a weighted loss function and a multi-dimensional dynamic feature crossing mechanism to generate an enhanced dual-task learning architecture, training to obtain a credit-risk two-dimensional evaluation model, and performing credit score regression prediction and risk level classification prediction to obtain a credit score and a risk level; performing integrity verification on the credit score and risk level through an intelligent shunting strategy engine, and through a preset shunting strategy matrix, performing credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to obtain a work order shunting result; extracting abnormal features and performing compliance verification on the work order shunting result through a dual intelligent audit mechanism to generate an initial expense reimbursement audit decision package; based on the initial expense reimbursement audit decision package, through a rule mining engine and a feature weight optimization algorithm, performing incremental knowledge extraction and parameter adjustment of the dual intelligent audit mechanism to obtain an updated audit knowledge base, and through the Monte Carlo random sampling and adversarial strategy simulation method, performing multi-dimensional deduction on the evolution path of fiscal and tax policies and the evolution trend of fraud means to construct a risk evolution simulation sandbox to obtain a decay-resistant knowledge base; based on the decay-resistant knowledge base, performing final decision analysis on the work order shunting result through a dual intelligent audit mechanism to obtain a final expense reimbursement audit decision package.

2. The intelligent expense reimbursement review method for a financial shared center according to claim 1, wherein: The specific steps for collecting reimbursement voucher data and preprocessing it to generate normalized data are as follows. Real-time collecting electronic invoice data, itinerary data, payment voucher data, and approval process data to obtain reimbursement voucher data, and performing OCR recognition, data cleaning, format standardization, and association verification to generate normalized data.

3. The intelligent expense reimbursement audit method for a financial shared center according to claim 1, wherein: The specific steps for generating the enhanced dual-task learning architecture are as follows. Combining the XGBoost and LightGBM frameworks and sharing the feature extraction layer, retaining the credit score output end of the XGBoost framework and the risk assessment output end of the LightGBM framework, and performing initialization and space mapping to construct a dual-task learning architecture; Enhancing and optimizing the performance of the dual-task learning architecture through a dynamic weighted loss function and a multi-dimensional dynamic feature crossing mechanism to obtain an enhanced dual-task learning architecture.

4. The intelligent expense reimbursement audit method for a financial shared center according to claim 1, characterized in that: The specific steps for obtaining the credit score and risk level are as follows. Based on historical normalized data, training the enhanced dual-task learning architecture through a phased feature progressive training method combined with a time series cross-validation method to obtain a credit-risk two-dimensional evaluation model; Inputting the normalized data into the credit-risk two-dimensional evaluation model to perform credit score regression prediction and risk level classification prediction to obtain a credit score and a risk level.

5. The intelligent expense reimbursement audit method for the financial sharing center according to claim 1, characterized in that: The specific steps for obtaining the work order shunting result are as follows. Performing integrity verification on the credit score and risk level through an intelligent shunting strategy engine to generate a standardized two-dimensional evaluation data packet; Mapping the standardized two-dimensional evaluation data packet to the cells of a preset shunting strategy matrix through a coordinate mapping mechanism, and performing credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to generate a work order shunting result.

6. The intelligent expense reimbursement review method of the financial shared center according to claim 1, wherein: The generation of the initial expense reimbursement review decision package is specifically carried out as follows: Based on the work order diversion result, extract abnormal features through the deep learning model in the dual intelligent review mechanism, and at the same time use the rule inference engine for compliance verification to obtain the preliminary review conclusion and risk identification; Through the decision weighting algorithm, fuse the confidence levels of the preliminary review conclusion and risk identification to generate the initial expense reimbursement review decision package.

7. The intelligent expense reimbursement audit method for a financial shared center according to claim 1, characterized in that: The obtaining of the updated review knowledge base is specifically carried out as follows: Input the initial expense reimbursement review decision package into the rule mining engine, use the FP-Growth algorithm to mine association rules, calculate the confidence levels of the association rules, screen the association rules with confidence levels higher than the association rule confidence level threshold, and generate a new set of association rules; Based on the new set of association rules, adjust the feature weight matrix and priority parameters in the dual intelligent review mechanism through the gradient descent method, and output the optimized parameter set of the dual intelligent review mechanism; Fuse and resolve conflicts for the new set of association rules and the optimized parameter set of the dual intelligent review mechanism through the versioned incremental merge mechanism to generate the updated review knowledge base.

8. The intelligent expense reimbursement audit method for a financial shared center according to claim 1, wherein: The obtaining of the anti-decay knowledge base is specifically carried out as follows: Based on the updated review knowledge base, perform simulation of the path of changes in fiscal and tax policies through the Monte Carlo random sampling method to obtain the simulation data set of the path of changes in fiscal and tax policies, and predict the evolution trend of fraud means to obtain the prediction result of the evolution trend of fraud means; Based on the simulation data set of the path of changes in fiscal and tax policies and the prediction result of the evolution trend of fraud means, construct a risk evolution simulation sandbox through the multi-dimensional deduction method to obtain the anti-decay fortified review knowledge base.

9. The intelligent expense reimbursement review method for a financial sharing center according to claim 1, wherein: The obtaining of the final expense reimbursement review decision package is specifically carried out as follows: Load the anti-decay fortified review knowledge base through the dual intelligent review mechanism to perform anomaly feature detection and compliance verification on the work order diversion result, and obtain the preliminary review determination result and risk correction suggestions; Perform final decision confidence level fusion calculation and encapsulation on the preliminary review determination result and risk correction suggestions to generate the expense reimbursement review decision package.

10. An intelligent expense reimbursement audit system for a financial shared service center, based on the intelligent expense reimbursement audit method for a financial shared service center according to any one of claims 1 to 7, characterized in that: It includes a data collection, evaluation module, diversion strategy module, review evolution module, optimization module, and output module; Data collection is used to collect reimbursement voucher data and preprocess it to generate normalized data; The evaluation module is used to construct a dual-task learning architecture based on the XGBoost and LightGBM frameworks, and enhance it through the weighted loss function and multi-dimensional dynamic feature cross mechanism to generate an enhanced dual-task learning architecture. Through training, a credit-risk two-dimensional evaluation model is obtained, and credit score regression prediction and risk level classification prediction are performed to obtain the credit score and risk level; The diversion strategy module is used to perform integrity verification on the credit score and risk level through the intelligent diversion strategy engine, and through the preset diversion strategy matrix, perform credit-risk weight fusion, load dynamic adjustment, and risk conduction analysis to obtain the work order diversion result; The review evolution module is used to extract abnormal features and perform compliance verification on the work order diversion result through the dual intelligent review mechanism to generate the initial expense reimbursement review decision package; Optimization module, which is used to perform incremental knowledge extraction and adjust the parameters of the dual intelligent audit mechanism based on the initial expense reimbursement audit decision package through a rule mining engine and a feature weight optimization algorithm, obtain an updated audit knowledge base, and conduct multi-dimensional deduction on the change path of fiscal and tax policies and the evolution trend of fraud means through the Monte Carlo random sampling and adversarial strategy simulation method, and construct a risk evolution simulation sandbox to obtain a decay-resistant knowledge base; Output module, which is used to perform final decision analysis on the work order diversion result through the dual intelligent audit mechanism based on the decay-resistant knowledge base to obtain the final expense reimbursement audit decision package.

Images