A method and system for identifying the evolution of enterprise tax risk based on reinforcement learning

Abstract

This invention discloses a method and system for evolutionary identification of corporate tax-related risks based on reinforcement learning, belonging to the field of corporate tax-related risk identification technology. It includes: collecting corporate financial data and extracting state vectors; establishing multiple verification actions and strategies, obtaining verification results and establishing a sample set through these actions; calculating the risk score of the verification results and merging them into the samples, identifying verification results with risk scores reaching a threshold as output results; optimizing the verification strategies; and verifying the output results, removing samples corresponding to verification results from the sample set if they do not meet the standards. This invention, by establishing multiple verification actions, divides the verification operation into various different actions, reducing the computational load of individual verification actions. By using a deep Q-network to determine the selectivity of verification actions and tracing the source of tax-related risks through individual verification operations, it can provide an accurate review scope for manual review and improve verification efficiency.

Description

Technical Field

[0001] This invention relates to the field of enterprise tax risk identification technology, specifically to an evolutionary identification method and system for enterprise tax risks based on reinforcement learning. Background Technology

[0002] Tax risks faced by enterprises refer to the potential for economic losses due to penalties incurred by enterprises in the course of their operations if they fail to comply with tax laws and regulations. The quantifiable indicators of enterprise tax risks are reflected in tax returns, balance sheets, profit and loss statements, and cash flow statements for various tax types. These indicators can reflect problems in the revenue, sales, and period expenses presented in tax returns, potentially leading to non-compliant tax practices and thus tax risks. To effectively reduce tax risks, enterprises should conduct self-checks before filing tax returns.

[0003] When identifying tax-related risks, it is necessary to use an automated identification system to assist manual verification. For example, the automatic identification of corporate tax-related risks based on neural networks, as described in patent publication number CN115204590A, includes: collecting datasets of tax-related risk indicators for enterprises at different risk levels; constructing a neural network model based on these indicators and initializing the weights of the indicators using a particle swarm optimization algorithm; inputting the dataset and the corresponding enterprise risk levels into the initialized neural network model for training; obtaining the target enterprise's tax-related risk indicators, inputting these indicators into the model, and using the neural network model to obtain the enterprise risk level corresponding to the target enterprise's tax-related risk indicators; and outputting the target enterprise's enterprise risk level. This approach can solve the problems of accurate, efficient, and timely tax-related risk identification for enterprises under the current massive amount of tax-related data, utilizing deep learning theory in neural network models to mine potential patterns in the data and promptly identify corporate tax-related risks.

[0004] The measurement indicators of corporate tax-related risks are reflected in tax returns, balance sheets, profit and loss statements, and cash flow statements for various tax types. Tracing the specific sources of tax-related risks requires a variety of detailed calculations and analyses, such as analyzing cash flow, verifying invoices, and comparing contracts with invoices. The amount of calculation required for tracing the sources is too large to be carried out. Therefore, the above methods and existing tax-related risk identification methods can only assess the risk level and cannot trace the sources of tax-related risks. Summary of the Invention

[0005] The purpose of this invention is to provide an evolutionary identification method for corporate tax-related risks based on reinforcement learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an evolutionary identification method for corporate tax-related risks based on reinforcement learning, comprising: collecting data: collecting corporate financial data, and extracting features from the corporate financial data using a feature extraction method to obtain a state vector;

[0007] Establish a dynamic decision-making model for risk identification: Set up multiple verification actions and establish verification strategies based on state vectors. Select verification actions to verify corporate financial data through verification strategies, obtain verification results, record the verification results as samples, and establish a sample set.

[0008] Preliminary determination of output results: Based on the verification results and through the risk score calculation method, determine the risk score of the verification results, merge the risk scores into the sample, and determine the verification results with risk scores reaching the threshold as the output results;

[0009] Optimize the dynamic decision-making model for risk identification: Establish a method for determining verification actions to optimize the verification strategy, and improve the method for determining verification actions through a sample set;

[0010] Output result verification: The output result is verified using a time-sequential method. If it meets the standard, the output result is output. If it does not meet the standard, the sample corresponding to the verification result is deleted from the sample set at the same time as the output result is output.

[0011] Preferably, the feature extraction method includes:

[0012] The state vector includes: tax burden deviation, gross profit margin, cash flow rate, related-party transaction ratio, and invoice anomaly rate;

[0013] The method for calculating the tax burden rate deviation is as follows: extract the actual tax burden rate of the enterprise from the enterprise's financial data and collect the industry average tax burden rate, calculate the deviation ratio between the enterprise's actual tax burden rate and the industry average tax burden rate, and obtain the tax burden rate deviation.

[0014] The method for calculating the cash return rate is as follows: extract the cash return amount and sales revenue from the company's financial data, calculate the ratio of cash return amount to sales revenue, and obtain the cash return rate.

[0015] The method for calculating the related-party transaction ratio is as follows: extract the related-party transaction amount and the total amount of similar transactions in the same period from the company's financial data, calculate the ratio of the related-party transaction amount to the total amount of similar transactions in the same period, and obtain the related-party transaction ratio.

[0016] The method for calculating the invoice anomaly rate is as follows: extract the total amount of abnormal invoices and the total amount of invoices in the same period from the company's financial data, calculate the ratio of the total amount of abnormal invoices to the total amount of invoices in the same period, and obtain the invoice anomaly rate;

[0017] The gross profit margin is calculated as follows: extract the sales revenue and cost of sales for the same period from the company's financial data, calculate the ratio of the difference between sales revenue and cost of sales to cost of sales, and obtain the gross profit margin.

[0018] Preferably, the establishment of the risk identification dynamic decision-making model specifically involves:

[0019] Set up verification actions: analyze fund flows, verify invoices, compare contracts and invoices, and check related-party transactions;

[0020] Establish a verification strategy: If the cash flow rate is >10% or the tax burden deviation is <-0.2, prioritize analyzing cash flow; otherwise, select randomly.

[0021] Perform the verification action: merge the number of risk points into the state vector, update the state vector, the initial number of risk points is 0, and execute the selected verification action to determine the number and type of risk points in this verification.

[0022] Update the state vector for the next time step: merge the number of risk points in the state vector with the number of risk points in the current time step, and record it as the number of risk points in the state vector at the next time step. Other parameters of the state vector at the next time step remain unchanged.

[0023] Establish a sample set: merge the updated state vector, the selected verification action, and the state vector at the next moment into a single sample, and merge multiple samples into a sample set.

[0024] Preferably, the risk score calculation method includes:

[0025] The point weights of different verification actions for each enterprise are obtained through an enterprise weight allocation method;

[0026] The verification results are weighted according to the integral weight of different verification actions, and the risk score is calculated according to the formula, as follows:

[0027]

[0028] in Indicates risk score, Indicates the number of risk points. This indicates the number of false alarms for risk points in the first 10 risk identifications. This represents the average time spent on the first 10 risk identification attempts. This indicates the time spent on this risk identification. This indicates the integral weight corresponding to the verification action used in this risk identification;

[0029] The risk scores are merged into the samples for sample updates;

[0030] When the risk score is higher than 0.4, the verification result will be marked as a result to be output.

[0031] Preferably, the enterprise weight allocation method includes:

[0032] Collect historical data, including historical audit records of tax-related risks of similar enterprises, and map the risk points that appear in the historical audit records to the audit actions;

[0033] The number of risk points for each verification action is counted, and the number of risk points for each verification action is normalized to obtain an integral weight that sums to 1.

[0034] Preferably, the method for determining the verification action includes:

[0035] Building a deep Q-network:

[0036] Input layer: 6 neurons, each corresponding to a different dimension in the state vector;

[0037] Hidden layers: 2 fully connected layers, 64 neurons per layer, with ReLU activation function;

[0038] Output layer: 4 neurons, each corresponding to the action Q-value of one of the four check actions, with no activation function;

[0039] Training sample set: capacity 1000, training starts after storing 500 samples, the parameters of the initial deep Q network are randomly generated;

[0040] Training process: Randomly select several samples and calculate the Q-value of the predicted action for each sample, specifically as follows:

[0041] ;

[0042] in This represents the predicted action Q-value for sample t. This represents the risk integral of sample t. This indicates that when the state vector of sample t+1 is input into a deep Q-network, and the maximum value of the four output action Q-values ​​is selected, the parameter set of the deep Q-network is: ;

[0043] Then, the parameter set of the deep Q-network is updated by minimizing the loss function:

[0044] ;

[0045] in Let n represent the loss function, and n represent the number of samples selected in one training iteration. This represents the target Q value of the i-th sample. This means substituting the state vector of the i-th sample into the parameter set. The parameter set is determined by minimizing the loss function, taking the four action Q-values ​​output by the deep Q-network that correspond to the check action of the i-th sample. Represents the updated parameter set;

[0046] After training and updating the parameter set of the deep Q network, the state vector to be identified is fed into the deep Q network. Based on the four output action Q values, the verification action corresponding to the action with the largest action Q value is selected to perform risk verification, replacing the original strategy.

[0047] Preferably, when selecting verification actions using a deep Q-network, the set of parameters used by the deep Q-network can be fixed, and after each risk identification, a new set of parameters can be trained based on the set of parameters used by the deep Q-network to obtain a secondary storage set of parameters. After several training iterations, the secondary storage set of parameters obtained from the training can be used as a new training basis for training and the secondary storage set of parameters can be updated again. After updating the training basis several times, the set of parameters used can be replaced with the secondary storage set of parameters.

[0048] Preferably, the time transition method includes:

[0049] Determine the number of risk points corresponding to each risk point type based on the output results, and count the number of risk points corresponding to each risk point type in the output results of the first 10 risk identifications.

[0050] Calculate the average rate of change of the number of risk points for each risk type in the previous 10 risk identifications, and calculate the rate of change of the number of risk points for each risk type in this risk identification compared to the previous risk identification.

[0051] If more than 60% of the risk points change rate exceeds the average change rate of the previous 10 times, then while outputting the results to be output, the risk identification in this case will be marked as unexpected data, and the sample corresponding to the verification result will be deleted from the sample set and will not participate in subsequent model training.

[0052] An evolutionary identification system for corporate tax-related risks based on reinforcement learning is provided, which utilizes the aforementioned method for evolutionary identification of corporate tax-related risks based on reinforcement learning.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] By establishing multiple verification actions, the verification process is divided into various identification operations such as analyzing fund flows, verifying invoices, comparing contracts and invoices, and checking related-party transactions. This reduces the computational load of a single verification action. By using a deep Q-network to determine the selectivity of verification actions and tracing the source of tax-related risks through a single verification action, the system can provide an accurate scope for manual review and improve verification efficiency.

[0055] Meanwhile, a reinforcement learning framework is introduced to capture the impact of corporate strategy adjustments on risk identification in real time through dynamic game between the agent and the environment, thereby achieving strategy evolution. In the continuous process of tax risk identification, the deep Q network can continuously learn from the samples, optimize the selection strategy for verification actions, improve the accuracy of tax risk identification, and enhance the ability of tax risk identification to judge new risks.

[0056] In addition, by using the enterprise weight allocation method, the focus of the verification actions can be changed according to the business scope and nature of each enterprise, further optimizing the selection strategy of verification actions, supporting risk analysis of multiple industries and multiple tax types, and having a wide range of applications. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the enterprise tax risk identification and evolution identification method of the present invention;

[0058] Figure 2 This is a flowchart illustrating the risk score calculation method in this invention;

[0059] Figure 3 This is a flowchart illustrating the method for determining the verification action in this invention;

[0060] Figure 4 This is a flowchart illustrating the time transition method in this invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] like Figures 1-4 As shown, the present invention provides a technical solution: a method and system for evolutionary identification of corporate tax-related risks based on reinforcement learning, comprising:

[0063] Data collection: Collect corporate financial data and extract features from the corporate financial data using feature extraction methods to obtain state vectors;

[0064] Establish a dynamic decision-making model for risk identification: Set up multiple verification actions and establish verification strategies based on state vectors. Select verification actions to verify corporate financial data through verification strategies, obtain verification results, record the verification results as samples, and establish a sample set.

[0065] Preliminary determination of output results: Based on the verification results and through the risk score calculation method, determine the risk score of the verification results, merge the risk scores into the sample, and determine the verification results with risk scores reaching the threshold as the output results;

[0066] Optimize the dynamic decision-making model for risk identification: Establish a method for determining verification actions to optimize the verification strategy, and improve the method for determining verification actions through a sample set;

[0067] Output result verification: The output result is verified using a time-sequential method. If it meets the standard, the output result is output. If it does not meet the standard, the sample corresponding to the verification result is deleted from the sample set at the same time as the output result is output.

[0068] It is important to note that multiple verification actions are performed by technical personnel. The purpose of establishing these actions is to categorize identification behaviors in tax-related identification. For example, verification actions can be divided into various identification operations such as analyzing cash flow, verifying invoices, comparing contracts and invoices, and checking related-party transactions. The more verification actions set up, the less computation and time each action takes. The rule base for verification actions can refer to regulations such as the "China Tax Audit Yearbook 2022" and the "Provisional Regulations of the People's Republic of China on Value-Added Tax," as well as high-frequency risk clauses in the company's historical contracts (such as payment terms and acceptance standards). Finally, the identification results are manually reviewed to correct misjudgments and update the rule base. Each verification action can identify risk points within the current verification scope. For example, the risk of concealed output tax can be discovered through cash flow comparison.

[0069] Feature extraction methods include:

[0070] The state vector includes: tax burden deviation, gross profit margin, cash flow rate, related-party transaction ratio, and invoice anomaly rate;

[0071] The method for calculating the tax burden rate deviation is as follows: extract the actual tax burden rate of the enterprise from the enterprise's financial data and collect the industry average tax burden rate, calculate the deviation ratio between the enterprise's actual tax burden rate and the industry average tax burden rate, and obtain the tax burden rate deviation.

[0072] The method for calculating the cash return rate is as follows: extract the cash return amount and sales revenue from the company's financial data, calculate the ratio of cash return amount to sales revenue, and obtain the cash return rate.

[0073] The method for calculating the related-party transaction ratio is as follows: extract the related-party transaction amount and the total amount of similar transactions in the same period from the company's financial data, calculate the ratio of the related-party transaction amount to the total amount of similar transactions in the same period, and obtain the related-party transaction ratio.

[0074] The method for calculating the invoice anomaly rate is as follows: extract the total amount of abnormal invoices and the total amount of invoices in the same period from the company's financial data, calculate the ratio of the total amount of abnormal invoices to the total amount of invoices in the same period, and obtain the invoice anomaly rate;

[0075] The gross profit margin is calculated as follows: extract the sales revenue and cost of sales for the same period from the company's financial data, calculate the ratio of the difference between sales revenue and cost of sales to cost of sales, and obtain the gross profit margin.

[0076] It should be noted that, for ease of understanding, the following simulated data is used:

[0077] The state vector is a multi-dimensional vector, which includes the tax burden deviation, gross profit margin, cash flow rate, related-party transaction ratio, invoice anomaly rate, and number of declaration anomalies.

[0078] Taking the tax burden rate deviation calculation method as an example:

[0079] Assuming that the actual tax burden rate of the enterprise and the industry average tax burden rate are extracted from the enterprise's financial data as 0.9% and 1.2% respectively, then the deviation of the tax burden rate is: (0.9%-1.2%) / 1.2%=-0.25.

[0080] Gross profit margin, cash flow rate, related-party transaction ratio, and invoice anomaly rate are all calculated using the above method, and will not be demonstrated in detail here.

[0081] like Figure 1 As shown, the specific steps for establishing a dynamic decision-making model for risk identification are as follows:

[0082] Set up verification actions: analyze fund flows, verify invoices, compare contracts and invoices, and check related-party transactions;

[0083] Establish a verification strategy: If the cash flow rate is >10% or the tax burden deviation is <-0.2, prioritize analyzing cash flow; otherwise, select randomly.

[0084] Perform the verification action: merge the number of risk points into the state vector, update the state vector, the initial number of risk points is 0, and execute the selected verification action to determine the number and type of risk points in this verification.

[0085] Update the state vector for the next time step: merge the number of risk points in the state vector with the number of risk points in the current time step, and record it as the number of risk points in the state vector at the next time step. Other parameters of the state vector at the next time step remain unchanged.

[0086] Establish a sample set: merge the updated state vector, the selected verification action, and the state vector at the next moment into a single sample, and merge multiple samples into a sample set.

[0087] It should be noted that, for the sake of easy understanding, the following simulated data is used:

[0088] The analysis of fund flows, verification of invoices, comparison of contracts and invoices, and inspection of related-party transactions are respectively labeled as a0, a1, a2, and a3;

[0089] In a risk identification process:

[0090] Company A's state vector:

[0091] s0 = [Tax burden rate deviation = −0.15, Gross profit margin = 18%, Capital return rate = 8%, Related transaction ratio = 25%, Invoice abnormality rate = 0.02, Number of risk points = 0];

[0092] Company A has a cash flow rate of 8% < 10% and a tax burden deviation of -0.15 > -0.2. It randomly selects action a1 (checking invoices). Assuming that after executing a1, it finds two invoices with incorrect codes and identifies one risk point (concealed income), then the state vector at the next moment is s1 = [tax burden deviation = −0.15, gross profit margin = 18%, cash flow rate = 8%, related party transaction ratio = 25%, invoice anomaly rate = 0.02, number of risk points = 1].

[0093] Then, (s0, a1, s1) is recorded as a sample, and multiple samples are calculated in the same way to build a sample set.

[0094] like Figure 2 As shown, the risk score calculation method includes:

[0095] The point weights of different verification actions for each enterprise are obtained through an enterprise weight allocation method;

[0096] The verification results are weighted according to the integral weight of different verification actions, and the risk score is calculated according to the formula, as follows:

[0097]

[0098] in Indicates risk score, Indicates the number of risk points. This indicates the number of false alarms for risk points in the first 10 risk identifications. This represents the average time spent on the first 10 risk identification attempts. This indicates the time spent on this risk identification. This indicates the integral weight corresponding to the verification action used in this risk identification;

[0099] The risk scores are merged into the samples for sample updates;

[0100] When the risk score is higher than 0.4, the verification result will be marked as a result to be output.

[0101] It should be noted that, for ease of calculation, the following simulated data is used:

[0102] Suppose that in a risk identification process, Company A assigns weights of 0.2, 0.6, 0.15, and 0.05 to different verification actions (analyzing cash flow a0, verifying invoices a1, comparing contracts and invoices a2, and checking related-party transactions a3), respectively. In the previous 10 risk identifications, after manual review (each risk identification requires manual verification and confirmation, and the rule base for the verification action is optimized based on the cause of the error), one risk identification error was found. The average time spent on the previous 10 risk identifications was 4 hours. For the sample (s0, a1, s1), the risk identification time was 2 hours. The calculated risk score is: 0.6×1 - 0.5×1 + 0.2×2 = 0.5. The updated sample is (s0, a1, 0.5, s1), and the score threshold is 0.4. Since the risk score is higher than the score threshold, the verification result is marked as pending output.

[0103] like Figure 2 As shown, the enterprise weight allocation methods include:

[0104] Collect historical data, including historical audit records of tax-related risks of similar enterprises, and map the risk points that appear in the historical audit records to the audit actions;

[0105] The number of risk points for each verification action is counted, and the number of risk points for each verification action is normalized to obtain an integral weight that sums to 1.

[0106] It should be noted that, for ease of understanding, the following simulated data is used:

[0107] Suppose that a total of 56 companies' historical tax risk verification records are collected, and the verification actions include four types: analyzing cash flow, verifying invoices, comparing contracts and invoices, and checking related-party transactions. The number of risk points corresponding to different verification actions are 20, 60, 15, and 5, respectively. After normalization, the integral weights of the four verification actions—analyzing cash flow, verifying invoices, comparing contracts and invoices, and checking related-party transactions—are 0.2, 0.6, 0.15, and 0.05, respectively.

[0108] like Figure 3 As shown, the methods for determining the verification actions include:

[0109] Building a deep Q-network:

[0110] Input layer: 6 neurons, each corresponding to a different dimension in the state vector;

[0111] Hidden layers: 2 fully connected layers, 64 neurons per layer, with ReLU activation function;

[0112] Output layer: 4 neurons, each corresponding to the action Q-value of one of the four check actions, with no activation function;

[0113] Training sample set: capacity 1000, training starts after storing 500 samples, the parameters of the initial deep Q network are randomly generated;

[0114] Training process: Randomly select several samples and calculate the Q-value of the predicted action for each sample, specifically as follows:

[0115] ;

[0116] in This represents the predicted action Q-value for sample t. This represents the risk integral of sample t. This indicates that when the state vector of sample t+1 is input into a deep Q-network, and the maximum value of the four output action Q-values ​​is selected, the parameter set of the deep Q-network is: ;

[0117] Then, the parameter set of the deep Q-network is updated by minimizing the loss function:

[0118] ;

[0119] in Let n represent the loss function, and n represent the number of samples selected in one training iteration. This represents the target Q value of the i-th sample. This means substituting the state vector of the i-th sample into the parameter set. The parameter set is determined by minimizing the loss function, taking the four action Q-values ​​output by the deep Q-network that correspond to the check action of the i-th sample. Represents the updated parameter set;

[0120] After training and updating the parameter set of the deep Q network, the state vector to be identified is fed into the deep Q network. Based on the four output action Q values, the verification action corresponding to the action with the largest action Q value is selected to perform risk verification, replacing the original strategy.

[0121] It should be noted that, for ease of calculation, the following simulated data is used:

[0122] Suppose that 32 samples are randomly sampled from the sample set during a single training session;

[0123] Calculate the target Q value (taking the sample (s0, a1, 0.5, s1) of company A as an example):

[0124] Suppose we input the sample (s0, a1, 0.5, s1) from company A, If the value is 0.3, then the target Q value is: =0.5 + 0.95 × 0.3 = 0.785.

[0125] Calculate the loss function:

[0126] ;

[0127] Update the parameter set of the deep Q network using the Adam optimizer. The learning rate is 0.001.

[0128] After obtaining a certain number of samples through the verification strategy, the deep Q-network can be trained. The initial parameter set is randomly generated. After a large number of training sessions (a certain number of training sessions or a training threshold can be set; when the number of training sessions is reached or several consecutive loss functions are below the training threshold), the original verification strategy can be canceled, and the deep Q-network can be used to determine the optimal verification action. After each new risk identification, the sample set can be updated, and the training of the deep Q-network can be restarted periodically to maintain the ability of risk identification to adapt to new risks.

[0129] like Figure 3 As shown, when selecting verification actions using a deep Q-network, the set of parameters used by the deep Q-network can be fixed. After each risk identification, a new set of parameters is trained based on the set of parameters used by the deep Q-network to obtain a secondary storage parameter set. After several training iterations, the secondary storage parameter set obtained from the training is used as a new training basis for training, and the secondary storage parameter set is updated again. After updating the training basis several times, the set of parameters used is replaced with the secondary storage parameter set.

[0130] It is important to note that when using a deep Q-network, fixing the parameter set ensures greater stability of each risk identification result, guaranteeing the accuracy and stability of risk identification. The parameter set update of the deep Q-network depends on the original parameter set. During risk identification, the parameter set is continuously updated and iterated in the binary space, and the iterated parameter set can be periodically put into the deep Q-network for use. This ensures both the stability of risk identification and the self-evolution of the deep Q-network, improving its adaptability to new risk points.

[0131] like Figure 4As shown, the time transition methods include:

[0132] Determine the number of risk points corresponding to each risk point type based on the output results, and count the number of risk points corresponding to each risk point type in the output results of the first 10 risk identifications.

[0133] Calculate the average rate of change of the number of risk points for each risk type in the previous 10 risk identifications, and calculate the rate of change of the number of risk points for each risk type in this risk identification compared to the previous risk identification.

[0134] If more than 60% of the risk points change rate exceeds the average change rate of the previous 10 times, then while outputting the results to be output, the risk identification in this case will be marked as unexpected data, and the sample corresponding to the verification result will be deleted from the sample set and will not participate in subsequent model training.

[0135] It should be noted that, for ease of understanding, the following simulated data is used:

[0136] Assume the types and number of risk points identified in this risk identification, as well as the rate of change compared to the previous risk identification, are as follows: False invoices (4, rate of change 0.33), abuse of tax incentives (12, rate of change 0.2), failure to withhold and pay individual income tax (26, rate of change 0.04), and non-compliant invoice issuance (6, this risk point was not identified in the previous risk identification, therefore there is no rate of change). The previous 10 risk identifications revealed three types of risk points: failure to withhold and pay individual income tax, false invoices, and abuse of tax incentives, with average rates of change of 0.08, 0.23, and 0.14 respectively. Regarding non-compliant invoice issuance... For withholding individual income tax, the change rate this time is lower than the average change rate of the previous 10 times. For false invoices, the change rate this time is higher than the average change rate of the previous 10 times. For the abuse of tax incentives, the change rate this time is higher than the average change rate of the previous 10 times. For non-compliant invoice issuance, it did not occur in the previous 10 times, indicating that the average change rate of the previous 10 times is 0. The change rate this time is relatively large. There are a total of 4 types of risk points. In this risk identification, 3 of them are relatively high, reaching 60%. Therefore, after outputting the results of this time, the risk identification is marked as unexpected data, and the sample corresponding to the verification result is deleted from the sample set to avoid contamination of the sample set.

[0137] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A reinforcement learning-based method for the evolutionary identification of corporate tax risks, characterized in that: include: Data collection: Collect corporate financial data and extract features from the corporate financial data using feature extraction methods to obtain state vectors; Establish a dynamic decision-making model for risk identification: Set up multiple verification actions and establish verification strategies based on state vectors. Select verification actions to verify corporate financial data through verification strategies, obtain verification results, record the verification results as samples, and establish a sample set. Preliminary determination of output results: Based on the verification results and through the risk score calculation method, determine the risk score of the verification results, merge the risk scores into the sample, and determine the verification results with risk scores reaching the threshold as the output results; Optimize the dynamic decision-making model for risk identification: Establish a method for determining verification actions to optimize the verification strategy, and improve the method for determining verification actions through a sample set; Output result verification: The output result is verified using a time-sequential method. If it meets the standard, the output result is output. If it does not meet the standard, the sample corresponding to the verification result is deleted from the sample set at the same time as the output result is output.

2. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 1, characterized in that: The feature extraction method includes: The state vector includes: tax burden deviation, gross profit margin, cash flow rate, related-party transaction ratio, and invoice anomaly rate; The method for calculating the tax burden rate deviation is as follows: extract the actual tax burden rate of the enterprise from the enterprise's financial data and collect the industry average tax burden rate, calculate the deviation ratio between the enterprise's actual tax burden rate and the industry average tax burden rate, and obtain the tax burden rate deviation. The method for calculating the cash return rate is as follows: extract the cash return amount and sales revenue from the company's financial data, calculate the ratio of cash return amount to sales revenue, and obtain the cash return rate. The method for calculating the related-party transaction ratio is as follows: extract the related-party transaction amount and the total amount of similar transactions in the same period from the company's financial data, calculate the ratio of the related-party transaction amount to the total amount of similar transactions in the same period, and obtain the related-party transaction ratio. The method for calculating the invoice anomaly rate is as follows: extract the total amount of abnormal invoices and the total amount of invoices in the same period from the company's financial data, calculate the ratio of the total amount of abnormal invoices to the total amount of invoices in the same period, and obtain the invoice anomaly rate; The gross profit margin is calculated as follows: extract the sales revenue and cost of sales for the same period from the company's financial data, calculate the ratio of the difference between sales revenue and cost of sales to cost of sales, and obtain the gross profit margin.

3. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 2, characterized in that: The establishment of the dynamic decision-making model for risk identification specifically involves: Set up verification actions: analyze fund flows, verify invoices, compare contracts and invoices, and check related-party transactions; Establish a verification strategy: If the cash flow rate is >10% or the tax burden deviation is <-0.2, prioritize analyzing cash flow; otherwise, select randomly. Perform the verification action: merge the number of risk points into the state vector, update the state vector, the initial number of risk points is 0, and execute the selected verification action to determine the number and type of risk points in this verification. Update the state vector for the next time step: merge the number of risk points in the state vector with the number of risk points in the current time step, and record it as the number of risk points in the state vector at the next time step. Other parameters of the state vector at the next time step remain unchanged. Establish a sample set: merge the updated state vector, the selected verification action, and the state vector at the next moment into a single sample, and merge multiple samples into a sample set.

4. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 3, characterized in that: The risk score calculation method includes: The point weights of different verification actions for each enterprise are obtained through an enterprise weight allocation method; The verification results are weighted according to the integral weight of different verification actions, and the risk score is calculated according to the formula, as follows: in Indicates risk score, Indicates the number of risk points. This indicates the number of false alarms for risk points in the first 10 risk identifications. This represents the average time spent on the first 10 risk identification attempts. This indicates the time spent on this risk identification. This indicates the integral weight corresponding to the verification action used in this risk identification; The risk scores are merged into the samples for sample updates; When the risk score is higher than 0.4, the verification result will be marked as a result to be output.

5. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 4, characterized in that: The enterprise weight allocation method includes: Collect historical data, including historical audit records of tax-related risks of similar enterprises, and map the risk points that appear in the historical audit records to the audit actions; The number of risk points for each verification action is counted, and the number of risk points for each verification action is normalized to obtain an integral weight that sums to 1.

6. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 4, characterized in that: The method for determining the verification action includes: Building a deep Q-network: Input layer: 6 neurons, each corresponding to a different dimension in the state vector; Hidden layers: 2 fully connected layers, 64 neurons per layer, with ReLU activation function; Output layer: 4 neurons, each corresponding to the action Q-value of one of the four check actions, with no activation function; Training sample set: capacity 1000, training starts after storing 500 samples, the parameters of the initial deep Q network are randomly generated; Training process: Randomly select several samples and calculate the Q-value of the predicted action for each sample, specifically as follows: ； in This represents the predicted action Q-value for sample t. This represents the risk integral of sample t. This indicates that when the state vector of sample t+1 is input into a deep Q-network, and the maximum value of the four output action Q-values ​​is selected, the parameter set of the deep Q-network is: ; Then, the parameter set of the deep Q-network is updated by minimizing the loss function: ； in Let n represent the loss function, and n represent the number of samples selected in one training iteration. This represents the target Q value of the i-th sample. This means substituting the state vector of the i-th sample into the parameter set. The parameter set is determined by minimizing the loss function, taking the four action Q-values ​​output by the deep Q-network that correspond to the check action of the i-th sample. Represents the updated parameter set; After training and updating the parameter set of the deep Q network, the state vector to be identified is fed into the deep Q network. Based on the four output action Q values, the verification action corresponding to the action with the largest action Q value is selected to perform risk verification, replacing the original strategy.

7. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 1, characterized in that: When selecting verification actions using a deep Q-network, the set of parameters used by the deep Q-network can be fixed. After each risk identification, a new set of parameters can be trained based on the set of parameters used by the deep Q-network to obtain a secondary storage set of parameters. After several training iterations, the secondary storage set of parameters obtained from the training can be used as a new training basis for training and the secondary storage set of parameters can be updated again. After updating the training basis several times, the set of parameters used can be replaced with the secondary storage set of parameters.

8. The method for evolutionary identification of enterprise tax-related risks based on reinforcement learning according to claim 1, characterized in that: The time transition method includes: Determine the number of risk points corresponding to each risk point type based on the output results, and count the number of risk points corresponding to each risk point type in the output results of the first 10 risk identifications. Calculate the average rate of change of the number of risk points for each risk type in the previous 10 risk identifications, and calculate the rate of change of the number of risk points for each risk type in this risk identification compared to the previous risk identification. If more than 60% of the risk points change rate exceeds the average change rate of the previous 10 times, then while outputting the results to be output, the risk identification in this case will be marked as unexpected data, and the sample corresponding to the verification result will be deleted from the sample set and will not participate in subsequent model training.

9. A corporate tax risk evolution identification system based on reinforcement learning, characterized in that: The method for evolutionary identification of corporate tax-related risks based on reinforcement learning, as described in any one of claims 1-8, was used.

Images