A method for verifying invoice types based on long short-term memory neural network

Abstract

The present invention discloses a method for invoice type approval based on a long short-term memory neural network, which includes the following steps: 1) establishing an invoice type approval prediction model based on a long short-term memory neural network (LSTM) and setting input indicators and output indicators; 2) for each set taxpayer U, collecting the corresponding data of the taxpayer U for the last N consecutive months as the time series sample data of the taxpayer U according to the set input indicators and output indicators; 3) using the sample data to train the invoice type approval prediction model to obtain the optimal historical data length when the model performance is optimal; 4) intercepting the time series sample data according to the optimal historical data length to train the invoice type approval prediction model; 5) obtaining input indicator data of the optimal historical data length of the target taxpayer, and inputting it into the invoice type approval prediction model to perform invoice type approval prediction on the target taxpayer, and obtaining the target taxpayer's expansion incremental value, the maximum invoice limit for a single invoice, and the maximum monthly purchase quantity of invoices.

Description

Technical Field

[0001] The present invention relates to a method for verifying invoice types, and in particular to a method for verifying invoice types using a time series prediction model based on a long short-term memory neural network (LSTM), belonging to the field of artificial intelligence. Background Art

[0002] After completing tax registration, taxpayers need to obtain invoices and apply to the competent tax authorities for invoice collection procedures. The competent tax authorities will confirm the type, quantity, invoice limit and other matters of the invoices to be collected based on the taxpayer's business scope and scale. The above process is the approval of invoice types. Taxpayers who have completed invoice type approval and whose current type, quantity or invoice limit of invoices cannot meet their business needs can apply to the competent tax authorities for adjustments. Currently, the approval of invoice types mostly uses manual review, which is not only inefficient but also has problems such as high subjectivity. In addition, traditional methods cannot provide timely reminders to taxpayers who have changed their invoice types.

[0003] At present, time series prediction research is generally divided into three categories: the first is the traditional probability statistical model, which mainly includes regression models, such as the Autoregressive Integrated Moving Average Model (ARIMA); the second is artificial intelligence related models, which mainly include decision trees, Bayesian networks, support vector machines, artificial neural networks, such as support vector machine regression SVM (Support Vector Regression, SVM), and the neural network-based prediction model LSTM; the third is the hybrid model, which combines multiple algorithms for the final prediction.

[0004] ARIMA, proposed in the 1970s, is currently the most commonly used parametric statistical model. It is built on the basis of stationary time series and is used for short-term forecasting of highway traffic data. Since the data related to ticket type approval is nonlinear and non-stationary, these traditional time series models have very significant limitations when applied to ticket type approval forecasting. These limitations are mainly manifested in the following aspects: ① The assumptions of traditional time series models are relatively strict, and therefore the model has strict requirements on data; ② Since tax-related data is affected by a variety of complex factors and is usually non-stationary, traditional time series models can only process stationary data. Although data can be stabilized through differencing, this also causes data loss. SVM uses support vector machines to fit curves and perform regression analysis. It is more focused on estimation and prediction of small sample statistics.

[0005] Recurrent Neural Networks (RNNs) are a type of neural network designed to process sequential data. They were initially used in language models due to their ability to memorize long-term dependencies. However, RNNs have several drawbacks, such as requiring numerous nonlinear transformations, excessive parameters, prone to overfitting, insufficient number of layers, incomplete feature extraction, and vanishing gradients. Consequently, several RNN variants, such as the LSTM, have been proposed. The LSTM, a temporal recurrent neural network first proposed for language modeling in 1997, is renowned for its ability to memorize long-term dependencies, making it suitable for processing and predicting important events in time series with relatively long intervals and delays. However, the LSTM algorithm also has limitations in prediction. For example, when the sample sequence being learned contains linear relationships or noise, the LSTM algorithm can overfit the network, affecting prediction accuracy.

[0006] Historical invoicing information, operational information, and basic corporate information can reflect trends in invoice type changes. By using appropriate models, accurate predictions can be made regarding the number of invoices issued and the invoice limit, assisting tax authorities in approving invoice types. These predictions, pushed through the electronic tax bureau, can be used to alert taxpayers of changes in invoice types, provide feedback on the content of these alerts, and record the results. This is crucial for achieving automated and intelligent business processing. Summary of the Invention

[0007] In response to the above technical problems, the present invention provides an improved LSTM invoice type verification method, which predicts the maximum invoice quantity and the maximum invoice limit of an invoice, has a good prediction effect, and provides timely reminders to taxpayers who have changes in the maximum invoice quantity and the maximum invoice limit. The specific process of this method is shown as follows: Figure 1 shown.

[0008] Step 1 Determine the input indicators: Select stamp tax amount, property tax amount, invoice type code, invoice amount, sales industry code, input tax amount, and output tax amount as model input indicators. Take 10 as the base, and take the logarithm of the product of the maximum invoice limit for the same taxpayer in the i-th month and the maximum number of invoices purchased in the i-th month as the output indicator for the corresponding taxpayer in the i-th month, that is, the expansion increment.

[0009] Step 2: Extract data based on the indicators: Extract the corresponding data based on the above indicators. For taxpayers nationwide whose maximum single invoice issuance limit and monthly maximum invoice purchase quantity have changed in the past 24 months, extract the monthly stamp tax, property tax, invoice amount, sales industry code, input tax, and output tax data. Process outliers and missing values. Delete any anomalies in the above data, and fill in missing values ​​with zeros. Take the monthly indicator data for the same taxpayer U as a sample data set. Collect data for the same taxpayer U for N months, and you will get N corresponding sample data sets. The sequence formed by these N sample data sets is the time sample sequence for taxpayer U.

[0010] Step 3 Data stationarity test: Determine the stationarity of the data by observing the shape of the time series graph and the autocorrelation and partial autocorrelation function graphs of the sequence.

[0011] Step 4 proposes an improved LSTM time series prediction algorithm: starting from the time series samples, first use the variational auto-encoder (VAE) neural network idea to repair the unstable samples, and then use LSTM to establish the final ticket type approval prediction model.

[0012] Step 5: Design an optimal historical data length search method: Analyze the main factors affecting the prediction performance of the invoice type approval model, propose an optimal historical data length search method, and calculate the optimal historical data length, that is, the historical data length when the model performance is optimal (the model prediction accuracy is highest).

[0013] Step 6: Model training and testing: Based on the calculated optimal historical data length, extract the corresponding data and normalize it. Then, use 10-fold cross validation to split the dataset into ten parts, alternately using nine parts for training and one part for testing. The mean of the 10 results is used as an estimate of model accuracy. The root mean squared error (RMSE) is used as the evaluation metric to verify the model's generalization ability.

[0014] Step 7: Model evaluation and validation: Compare the difference between the model predictions and the actual number of tickets collected. If the results are not satisfactory, adjust the model parameters and retrain.

[0015] Step 8: Ticket type verification and prediction: Use the trained model to conduct ticket type verification and prediction, obtain the incremental value of each taxpayer, and further determine the maximum invoice limit for a single invoice and the maximum number of tickets purchased per month.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] (1) The present invention proposes an improved LSTM time series prediction algorithm. Compared with the traditional Arima time series model, it solves the problems of instability and nonlinearity. Compared with the traditional RNN recurrent neural network, it makes up for the defects such as too many parameters, easy overfitting, insufficient number of layers, insufficient feature information extraction, and gradient disappearance. The traditional LSTM algorithm directly sends the time series composed of samples as the input sequence into the input layer of the network, and fails to eliminate the interference caused by abnormal samples. If this interference is not eliminated, the prediction result may be inaccurate. The VAE algorithm can continuously "eliminate" or "decompose" the "unqualified data" in the sample through the encoding, decoding and reconstruction process of the sample, so the present invention combines the two types of algorithms to design an improved LSTM time series prediction algorithm.

[0018] (2) The present invention proposes a method for searching for the optimal historical data length, which selects the optimal ticket type to determine the relevant historical data length, which is conducive to the model fully mining the amount of information between the data, reducing errors, and improving the prediction performance of the model.

[0019] (3) This paper uses a model evaluation method that combines 10-fold cross-validation with actual ticket collection results. 10-fold cross-validation can obtain as much effective information as possible from limited learning data and start learning samples from multiple directions, which can effectively avoid falling into local minima and, to a certain extent, avoid overfitting. At the same time, comparison with actual ticket collection results can effectively verify the actual application effect of the model.

[0020] (4) The present invention can automatically calculate the approved value, including the recommended invoice types and the corresponding maximum invoice limit and number of invoices to be collected. Taxpayers can directly use the system recommended value or adjust it manually. If the adjusted value is less than or equal to the system recommended value, it will be automatically approved. If the adjusted value is greater than the system recommended value, it will need to be pushed to the tax terminal for review by tax personnel, thereby improving the efficiency of task processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the method flow of the present invention;

[0022] Figure 2 This is the flow chart of the improved LSTM time series prediction algorithm;

[0023] Figure 3 This is a time series diagram of stamp duty amount;

[0024] Figure 4 This is a time series diagram of property tax amount;

[0025] Figure 5 This is a time series diagram of the invoice amount;

[0026] Figure 6 It is the time series diagram of input tax;

[0027] Figure 7 This is the time series diagram of output tax amount;

[0028] Figure 8 This is the autocorrelation diagram of stamp duty amount;

[0029] Figure 9 This is the autocorrelation diagram of property tax amount;

[0030] Figure 10 This is the autocorrelation graph of the invoice amount;

[0031] Figure 11 is the autocorrelation graph of input tax;

[0032] Figure 12 is the output tax autocorrelation graph;

[0033] Figure 13 This is the partial correlation diagram of stamp duty amount;

[0034] Figure 14 is the partial correlation diagram of property tax amount;

[0035] Figure 15 This is the partial correlation diagram of the invoice amount;

[0036] Figure 16 is the partial correlation diagram of input tax;

[0037] Figure 17 is the partial correlation diagram of output tax;

[0038] Figure 18 This is a box plot of stamp duty amount;

[0039] Figure 19 This is a box plot of property tax amount;

[0040] Figure 20 It is a box plot of the invoiced amount;

[0041] Figure 21 is the box plot of input tax;

[0042] Figure 22 is the box plot of output tax;

[0043] Figure 23 This is the LSTM network structure diagram;

[0044] Figure 24 Flowchart of the method for searching for the optimal history data length;

[0045] Figure 25 This is a graph of loss values ​​for different lengths of historical data;

[0046] Figure 26 This is the LSTM training result diagram;

[0047] Figure 27 The following is a chart showing the distribution of the number of taxpayers. DETAILED DESCRIPTION

[0048] The present invention is described in further detail below with reference to the accompanying drawings.

[0049] 1. Improved LSTM time series prediction algorithm

[0050] Although the traditional LSTM neural network improves the network prediction accuracy through the gating device, it ignores the linear relationship between samples and the consideration of "noise", and directly sends the sequence samples to the input layer of the network. If this interference is not eliminated, the prediction results may be inaccurate, thereby reducing the model prediction performance. Therefore, the present invention repairs the samples with the help of the VAE algorithm to "eliminate" or "decompose" the "unqualified data" in the samples. In summary, the present invention combines the two types of algorithms to design an improved LSTM time series prediction algorithm. The specific process of the algorithm is as follows Figure 2 shown.

[0051] ① For the training samples, manually observe the moving average and moving mean square error over time to see if the data shows a clear upward or downward trend. Generally, the invoice type approval data for normal taxpayers is independent of each other. If the data shows a clear upward or downward trend over time, it is considered unreasonable, and the sample is resampled and repeated. When resampling, try to avoid taxpayers who frequently change invoice types (for example, taxpayers whose number of invoice type changes exceeds the set threshold N) and taxpayers who may be at risk of false invoices.

[0052] ② Perform sample stationarity analysis by observing the shape of the time series graph and performing a unit root test. If the sample is stationary, proceed to step ④. Otherwise, proceed to step ③.

[0053] ③ Repair the sample to obtain the reconstructed sample sequence. The repair of the sample by VAE can be regarded as the original sample X={X1,X2,...X m}, encoded as Z={Z1,Z2,...Z m}, and then decoded to process.

[0054] ④ Build an LSTM model and use the "qualified sample" sequence as the input data of the prediction neural network to train the prediction model until the loss function requirements are met.

[0055] ⑤ Input the test sample into the model and obtain the predicted output sequence. Compare the results and if they are within the acceptable RMS error range, proceed to the next step; otherwise, adjust the parameters and proceed to ④.

[0056] ⑥ Output the improved LSTM time series prediction algorithm model.

[0057] 2. Stability analysis

[0058] Stationarity can be categorized as strong or weak. Strong stationarity is very demanding, requiring that any statistical properties between two sets of data remain constant over time. This stringent requirement is difficult to prove theoretically and verify in practice, resulting in limited application.

[0059] Weak stationarity relaxes the stationarity condition and only requires that the low-order moments are stationary, that is, the mathematical expectation (mean) and variance do not change with time and location. The conditions for a weakly stationary process are:

[0060] ①The mean function is always constant over the sampling time;

[0061] ② There is a second-order moment;

[0062] ③ The autocovariance is the same for all times t and time lags k.

[0063] The second-order moment, also known as variance, reflects the fluctuation of a random variable around its mean. The larger the variance, the greater the volatility. The autocovariance describes the second-order moment between the values ​​of a random variable at any two different times, t1 and t2. It describes the correlation between the fluctuations in the values ​​of a random variable at two different times (relative to the mean).

[0064] Therefore, we can judge whether the time series is stationary by graphical observation of the time series diagram and unit root test, as follows:

[0065] ①Observe the shape of the time series graph to preliminarily judge its stationarity

[0066] According to the definition of weak stationarity, the mean and variance of a time series are constant, so its time series graph should fluctuate around a horizontal line with roughly the same amplitude. If the time series graph has obvious increasing, decreasing, or cyclical fluctuations, the time series is likely non-stationary.

[0067] ②Observe the autocorrelation and partial autocorrelation function graphs of the sequence. For a stationary time series, its autocorrelation or partial autocorrelation coefficient generally decreases rapidly to near 0 or becomes 0 after a certain order, while the autocorrelation coefficient of a non-stationary time series generally decreases slowly rather than decreasing rapidly.

[0068] ③ Unit root test. By observing the time series graph, autocorrelation and partial autocorrelation graph to judge the stationarity of the time series, different conclusions may be drawn due to different judgments of the observer on the graph. In order to examine the stationarity of the time series more objectively, a statistical test method, namely the unit root test, is introduced. The present invention adopts the ADF test (Augmented Dickey-Fuller Test) method. The unit root test refers to the test of whether there is a unit root in the sequence. If there is a unit root, the sequence is non-stationary. Therefore, the H0 hypothesis of the ADF test is the existence of a unit root. If the significance test statistic obtained is less than three confidence levels (10%, 5%, 1%), then there is a corresponding (90%, 95, 99%) degree of confidence to reject the null hypothesis.

[0069] The present invention uses the statsmodels module in Python to perform an ADF test and calculate the t-statistic value, p-value value, and icbest value. Among them, t-statistic, T test, is the hypothesis test value; p-value is the hypothesis test result; icbest value is the maximum information criterion value.

[0070] The time series diagrams of stamp tax amount, property tax amount, invoice amount, input tax amount and output tax amount are as follows: Figure 3 、 4 , 5, 6, and 7. The autocorrelation diagrams are shown in Figure 8 、 9 , 10, 11, and 12, the partial correlation diagrams are shown as Figure 13 、 14 , 15, 16, and 17.

[0071] The ADF test results are as follows:

[0072] t-statistic: -0.452

[0073] p-value: 0.901

[0074] icbest: {'5%':-2.876,'1%':-3.464,'10%':-2.575}

[0075] From the above time series diagram, autocorrelation diagram, partial correlation diagram and ADF test results, it can be seen that the taxpayers' stamp duty, property tax, invoice amount, input tax and output tax data are all unstable.

[0076] 3. Outlier processing

[0077] Boxplots are used to analyze the data for outliers and remove them. Boxplots are a method of describing data using five statistical quantities: minimum, first quartile, median, third quartile, and maximum. They can also roughly show whether the data is symmetrical and the degree of dispersion of the distribution, and are particularly useful for comparing several samples.

[0078] The box plots of stamp tax amount, property tax amount, invoice amount, input tax amount and output tax amount are as follows: Figure 18 、 19 , 20, 21, and 22, it can be seen from the figures that there are certain outliers. The data smaller than the first quartile and larger than the third quartile are deleted.

[0079] 4. Model building

[0080] The present invention builds an LSTM model through Keras. Keras is an advanced neural network API written in Python, which can run with TensorFlow, CNTK or Theano as the backend. The present invention builds an LSTM multi-feature input ticket type approval prediction model, wherein the model input is stamp tax amount, property tax amount, invoice type code, invoice amount, input tax amount, output tax amount, industry code, and the output is the logarithm of the product value of the maximum invoice limit for a single invoice and the maximum number of tickets purchased per month with 10 as the base, that is, the expansion increment. The model uses two LSTM hidden layers and a Dense output layer with a linear rectified function (Rectified Linear Unit, ReLU) as the activation function, selects the Adaptive Moment Estimation (Addam) algorithm as the optimizer, the mean square error function (MSE) as the loss function of the model, and RMSE as the evaluation index of the model. The model network structure is as follows: Figure 23 shown.

[0081] ReLU, also known as the rectified linear unit, is an activation function commonly used in artificial neural networks. Compared with the Sigmoid and tanh functions, it converges faster.

[0082] Adam combines the advantages of two extended stochastic gradient descent algorithms, AdaGrad and RMSProp. Like RMSProp, Adam not only calculates the adaptive parameter learning rate based on the mean of the first-order moments, but also fully utilizes the mean of the second-order moments of the gradient (i.e., the uncentered variance).

[0083] MSE is the sum of the squares of the differences between the true value and the predicted value. It is often used as a criterion for evaluating the prediction effect of a model and a loss function of the model. The smaller the MSE value, the better the model fitting effect. Its calculation formula is:

[0084]

[0085] 5. Optimal historical data length search method

[0086] The length of historical data directly affects the model prediction performance. This paper analyzes the impact of the key parameter of historical data length on model performance and establishes a method for searching for the optimal historical data length. The specific process is as follows: Figure 24 As shown. This paper adopts the single-step prediction method. If a test sample is selected, the prediction error value has a certain degree of randomness, and the optimal historical data length searched is not universal. Therefore, this paper selects a certain amount of samples as the test set, and determines the historical data length when the average error of the test set is the lowest as the optimal historical data length. First, select part of the original data for testing and normalize it. When normalizing the sample data, min-max normalization is used to linearly transform the original data so that the result value is mapped above [0,1]. Secondly, the model prediction output data length f=1 is fixed, the window moving step s=1 is fixed, the historical data length is changed, and the coordinate delay method is used to construct the input and output data sets. Then, the training set and test set ratio are set, and the input and output data sets are divided into training set and test set at 8:2. The training set is input into the LSTM network training model, and the test set is input into the trained LSTM network to realize prediction. Finally, the average error loss between the sample estimate and the true value in the test set is calculated, and the historical data length when the average error loss is the smallest is selected as the optimal historical data length. The loss values ​​of different historical data lengths are as follows Figure 25 As shown in the figure, it can be seen that the loss value is the smallest when w=24, so 24 months of historical data are selected as the input data of the LSTM model.

[0087] 6. Model training and evaluation

[0088] After building the LSTM model, select an appropriate batchsize and train the model. Batchsize is the number of samples selected for one training. First, do not use Dropout and regularization terms, use a smaller data set (take a small part from the original data set) to train and fit this data set. Then, in one epoch, print out the input and output to check the correctness of the data. Next, observe the initial loss value and estimate the loss. Visualize the training process, calculate the loss and RMSE on the validation set after each epoch training, and record the loss and RMSE of the training set and validation set after each epoch. After many experiments, the present invention selects a batchsize of 128 and sets the learning rate attenuation index to 0.01.

[0089] The present invention uses ten-fold cross validation to evaluate the model. The data set is divided into ten parts, and nine of them are used as training data and one as test data in turn. Each test will produce a corresponding RMSE. The average of the 10 results is used as an estimate of the model effect. The training results are as follows: Figure 26 As shown in Figure 2, the average RMSE of the LSTM model is 0.54.

[0090] This paper compares the maximum monthly purchase volume and actual receipt volume of invoices for taxpayers who have changed their invoice types nationwide. The six invoice types are: 2008 Edition VAT Ordinary Invoice (two-copy unlimited amount version), 2008 Edition VAT Ordinary Invoice (five-copy unlimited amount version), 2016 Edition VAT Ordinary Invoice (two-copy folded version), 2016 Edition VAT Ordinary Invoice (five-copy folded version), Special VAT Invoice (three-copy Chinese unlimited amount version), and Special VAT Invoice (six-copy Chinese unlimited amount version).

[0091] The difference between (number of approved invoices - actual number of invoices) is selected as the analysis indicator. The minimum value of the difference is 0, the maximum value is 600, 400, 188, 100, 199, 499, and the standard deviation is 28, 67, 24, 21, 32, and 26. Figure 27 As shown in the figure, the difference between the assessed and actual amounts of most taxpayers is between 10 and 100, while the difference for a small number of taxpayers is greater than 1000, indicating a good prediction effect.

[0092] 7. Ticket type approval, prediction and result processing

[0093] To predict taxpayers nationwide, we collected 24 months of stamp duty, property tax, invoice amounts, input tax, and output tax records for the taxpayers under test, sorting them by time. We fed the trained LSTM model into the model output, which we then converted into predictions for the invoice type as follows.

[0094] Ticket type approval forecast result = 10 模型输出结果 (4)

[0095] The maximum invoice limit for a single invoice is 100, 1000, 10,000, 100,000, 1,000,000, 10,000,000, and 100,000,000. Divide each of these maximum invoice limits by the predicted expansion increments, rounding to the nearest integer to obtain the monthly maximum purchase quantity. The value with the smallest difference from the previous monthly maximum purchase quantity is selected as the final predicted monthly maximum purchase quantity, thereby obtaining the final predictions for the maximum invoice limit for a single invoice and the monthly maximum purchase quantity.

Claims

1. A method for verifying invoice types based on a long short-term memory neural network, comprising the following steps: 1) Establish a ticket type approval prediction model based on the long short-term memory neural network (LSTM) and set input and output indicators; 2) For each taxpayer U, the corresponding data for the last N consecutive months is collected based on the set input and output indicators as the time series sample data of taxpayer U. The ticket type approval prediction model is a two-layer LSTM recurrent neural network model, including two LSTM hidden layers and a Dense output layer with a linear rectifier function as the activation function. The adaptive moment estimation algorithm is selected as the optimizer for training the ticket type approval prediction model, and the mean square error function is used as the loss function for training the ticket type approval prediction model. 3) Using the variational autoencoder neural network to repair the time series sample data to obtain stationary sample data; then using the stationary sample data to train the ticket type approval prediction model to obtain the optimal historical data length that makes the ticket type approval prediction model perform optimally; The method of using variational autoencoder neural network to repair time series sample data and obtain stable sample data is as follows: 31) For the time series sample data of the assumed taxpayer U, determine whether the time series sample data is reasonable based on the moving average change graph and the moving mean square deviation change graph of the time series sample data. If it is unreasonable, resample the assumed taxpayer U; if it is reasonable, proceed to step 32); 32) Performing a sample stationarity analysis on the time series sample data. If the set parameter requirements are met, the data is considered to be stationary sample data; otherwise, proceeding to step 33); the set parameter requirements include: a) the mean function of the time series sample data is constant at all times, b) the time series sample data has a second-order moment, and c) the autocovariance of the time series sample data is the same for all times t and time lags k; 33) Use the variational autoencoder neural network to encode the time series sample data to obtain sequence Z, and then decode the sequence Z to obtain the reconstructed sample sequence As the stationary sample data corresponding to the time series sample data; The resulting sequence Distance from sequence X Minimum, X is the time series sample data; The method for obtaining the optimal historical data length is as follows: selecting multiple time series sample data; then setting the model prediction output data length f=1, the fixed window moving step s=1, changing the length of the time series sample data, and constructing the input and output data sets using the coordinate delay method; then dividing the input and output data sets into a training set and a test set, using the training set to train the ticket type approval prediction model, obtaining an estimated value for each sample in the training set, statistically calculating the average error loss between the sample estimated value and the corresponding true value at different lengths, and selecting the time series sample data length corresponding to the minimum average error loss as the optimal historical data length; 4) Based on the optimal historical data length obtained, the time series sample data is intercepted to train the ticket type approval prediction model; 5) Obtain input indicator data of the optimal historical data length of the target taxpayer and input it into the ticket type approval prediction model trained in step 4) to perform ticket type approval prediction for the target taxpayer, and obtain the target taxpayer's expanded version increment value, the maximum invoice limit for a single invoice, and the maximum number of tickets purchased per month.

2. The method according to claim 1, wherein The set input indicators include stamp tax amount, property tax amount, invoice type code, invoice amount, input tax amount, output tax amount, and industry code; the set output indicators are the logarithm of the product of the maximum invoice limit for a single invoice of the same taxpayer in the i-th month and the maximum number of invoices purchased in the i-th month, based on 10, that is, the expansion increment value.

Images