Judicial data time sequence rule type knowledge mining method

Abstract

The invention provides a judicial data time sequence rule type knowledge mining method, which comprises the following steps of: randomly acquiring exponential data of various different case types from a judicial data table as original data for training a time sequence mining model, including various different case quantity data, and taking a set number value as a training set; preprocessing data, and dividing a training set and a test set; correlation analysis is carried out through correlation coefficients and chi-square verification; performing ADF test on the training set to obtain a test result, and performing parameter selection after the data is stable; inputting the training set into various models to obtain a training set prediction result; applying the trained model to a test set to obtain a test set prediction result; and comparing the prediction precision of the multiple models, endowing prediction results of the multiple models with weights based on different characteristics of data captured by different models, and combining the prediction results to obtain a new combined prediction model. According to the method, the data quality is improved, the model prediction effect is improved, and the mining performance of the time sequence mining model is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data mining, and in particular to a method for mining judicial data temporal regularity knowledge. Background Art

[0002] The judicial knowledge system includes four categories: descriptive knowledge, relational knowledge, rule-based knowledge and regularity-based knowledge.

[0003] like Figure 1 As shown in the figure, descriptive knowledge explains the essence of judicial-related things and helps information systems understand their core content, such as litigation process, trial process, etc. Relational knowledge describes the relationship between case procedures, parties, judges, lawyers and other subjects, such as case-case relationship, person-case relationship, etc. Rule-based knowledge is a summary of judicial business experience, providing operating guidelines for information systems, such as simplified procedure judgment rules and element-based judgment rules. Regularity-based knowledge forms trend summaries through long-term data analysis of trial trends and system operations, such as the first-instance acceptance of judgments and the annual case closing rules.

[0004] Law-based knowledge is further subdivided into judicial situation laws, judicial index laws, and economic and social index laws. Judicial situation laws reflect the development trend and internal logic of judicial activities, such as the quality and efficiency laws of trial, execution, management, and litigation services. Judicial index laws monitor judicial openness, fairness, efficiency, and credibility through evaluation indicators, comprehensively reflect the status of judicial operations, and help discover and improve problems. Economic and social index laws analyze the relationship between justice and the economy and society through quantitative indicators, revealing the impact of economic development, social structure, and population changes on justice.

[0005] However, at present, due to the problems of excessive data volume and data instability in judicial data, and the inability to automatically mine regular knowledge, the mining performance and quality of judicial data are relatively low. Summary of the invention

[0006] The present invention proposes a method for mining the temporal regularity knowledge of judicial data. The regional distribution law of the trial situation is mined through statistical methods such as mean, median, and percentile, and the temporal law is mined by clustering, classification, and pattern discovery. In order to solve the interpretability and computational burden of high-dimensional data, methods such as Pearson correlation coefficient, chi-square test, and feature selection are used for data dimensionality reduction. For the knowledge of causal laws, mining is carried out through causal graph analysis, and multiple models (such as ARIMA, LSTM, etc.) are used for training and comparison to improve the prediction effect of the model and data quality.

[0007] The present invention provides a method for mining judicial data temporal regularity knowledge, comprising the following steps (eg Fig.37 shown):

[0008] Randomly obtain multiple categories (preferably 4 categories) of exponential data of different case types from judicial data as raw data for time series mining model training, the data set contains multiple categories (preferably 4 categories) of exponential data, preferably, each category of exponential data contains 60 values; the raw data contains multiple categories (preferably 6 categories) of different case volume data, and a set number (preferably 60) of case volume data of each category is taken as training set data;

[0009] Performing data preprocessing on the raw data, and dividing the preprocessed data into a training set and a test set at a ratio of 5 to 1;

[0010] Due to the excessive noise in the sample data and the instability of the original sequence, the classic time series prediction models such as ARIMA, LSTM, and INAR (1) all show poor results. Preferably, VMD (variational mode decomposition) is used to preprocess the data, decomposing the complex original sequence into several simple subsequences that contain the structure and periodicity of the original sequence, and then the subsequences (most of the subsequences obtained by VMD have relatively good properties such as stability) are modeled separately to achieve the purpose of improving the prediction accuracy. The experimental results show that the judicial case volume data processed by VMD is stable and more suitable for the LSTM model. The three evaluation indicators obtained on the test set are greatly reduced.

[0011] The data were analyzed for correlation using correlation coefficient and chi-square verification method. The training set data was tested using ADF (Augmented Dickey-Fuller Testing) to obtain test results, and the model parameters were selected after the data became stable.

[0012] The training set data is input into the ARIMA model, LSTM model, LR-ARIMA-LSTM model, INAR (1) model, VMD-LSTM model, etc. to obtain the training set prediction results, and the parameters are slightly adjusted according to the results.

[0013] Apply the trained model to the test set, obtain the test set prediction results, and draw a visual image;

[0014] The prediction accuracy of the test set prediction results of multiple models is compared. Based on the different features of data that different models can capture, the prediction results of multiple models are weighted and combined to obtain a new combined prediction model, and the combined prediction model is used to predict the data. This can more accurately capture data features and obtain more accurate prediction results.

[0015] Furthermore, the data preprocessing method includes: filtering out outliers from the original data and interpolating using the mean; the exponential data preprocessing method is to use the box plot method to filter out outliers from each variable in the data set, taking the first variable as the force, replacing all outliers in the exponential data set with the mean of the remaining data, and subtracting the mean from each value and dividing it by the standard deviation, thereby converting it into standard data with a mean of 0 and a variance of 1, using the first fifty data as training data to train model parameters; and using the last ten data as test data to test the model and make a model evaluation;

[0016] The preprocessing method of the case volume data is as follows: standardize the data, subtract the mean from each value and divide it by the standard deviation, thereby converting it into standard data with a mean of 0 and a variance of 1; use the first fifty data as training data to train model parameters; and use the last ten data as test data to test the model and make a model evaluation.

[0017] Furthermore, the model evaluation method includes: using accuracy Acc, mean square error MSE, mean absolute error MAE, root mean square error RMSE, mean absolute error rate MAPE, and symmetric mean absolute error rate SMAPE indicators to evaluate model performance, wherein MSE and MAE measure the size of the error between the predicted value and the true value, MAPE and SMAPE measure the relative size of the error value to the true value, and RMSE measures the size of the error between the predicted value and the true value. The closer these five indicators are to 0, the better the model effect.

[0018] Furthermore, the correlation analysis method includes: respectively calculating the Pearson correlation coefficient and the Kendall tau rank correlation coefficient between two columns of data, the Pearson correlation coefficient captures linear correlation, and the Kendall tau rank correlation coefficient captures nonlinear correlation, and the correlation coefficient is used to measure the correlation between different columns, with a value of -1 to 1. The closer the correlation coefficient is to 0, the weaker the correlation between the two columns of indicators, and the closer it is to plus or minus 1, the stronger the correlation. The correlation between variables is intuitively represented by a heat map.

[0019] Furthermore, the new combined prediction model obtained by weighting and combining the prediction results of multiple models includes:

[0020] Mix-ARIMA-LSTM model, assuming that the LSTM model weight is β and the ARIMA-Pro model weight is (1-β), then the prediction result of the Mix-ARIMA-LSTM model is:

[0021] β(lstm_predict)+(1-β)(arima-pro_predict)(1)

[0022] The LR-ARIMA-LSTM model uses the prediction results of the ARIMA-Pro and LSTM training sets as covariates to perform linear regression on the true values. The coefficients and intercept terms of the ARIMA-Pro and LSTM models are obtained by performing linear regression on the two model results of the training set and the true values.

[0023] Assuming the LSTM model coefficient is α, the ARIMA-Pro model coefficient is β, and the intercept term is b, the prediction result of the LR-ARIMA-LSTM model is: α(lstm_predict)+β(arima-pro_predict)+b(2)

[0024] In formula (1) and (2), lstm_predict and arima-pro_predict are the prediction results of LSTM and ARIMA-Pro models respectively.

[0025] Furthermore, the parameter selection of the model includes: the ARIMA model selects the order according to the ACF and PACF diagrams of the data, and the ARIMA-Pro model selects the order according to the AIC, BIC criteria and the cyclic structure.

[0026] The correlation analysis method includes: for two types of case volumes x and y that are strongly linearly correlated, further estimating the marginal distribution function of the two types of case volumes x and y to obtain F x (x), F y (y) and determine the optimal marginal distribution through the KS test.

[0027] Furthermore, the correlation analysis method also includes: using a Copula function (such as Frank, Clayton, Gumbel, Gaussian) to fit the data of two types of case volumes that are strongly linearly correlated, and estimating the fitting parameter values ​​to identify the dependency structure of the two types of case volumes that are strongly linearly correlated.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The judicial data temporal regularity knowledge mining method provided by the present invention utilizes descriptive statistical methods such as mean, median, percentile, variance, range, skewness, kurtosis, etc. to conduct macro-scale mining of the knowledge of the regional distribution regularity of the trial situation, and adopts clustering, classification, repeated pattern mining, rule mining, temporal retrieval, anomaly discovery, etc. to mine the temporal domain regularity of the trial situation; performs feature selection by methods such as Pearson correlation coefficient and chi-square verification to achieve data space dimension reduction; constructs training and test comparisons through multiple models such as ARIMA, ARIMA-Pro, LSTM, Mix-ARIMA-LSTM, L-ARIMA-LSTM, etc., and finally adopts the L-ARIMA-LSTM model to mine the temporal regularity of exponential judicial data, thereby realizing automatic mining of regularity knowledge; in view of the problem that the case volume data is too large and unstable, and the ARIMA, LSTM and INAR (1) models cannot achieve good prediction accuracy, VMD is further adopted to preprocess the case volume data, thereby improving the data quality, thereby improving the model prediction effect, and significantly improving the mining performance of the temporal mining model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0031] In the attached picture: Figure 1 Design and classify schematic diagrams for the diversified knowledge system of judicial knowledge system;

[0032] Figure 2 A mechanism diagram for regular knowledge mining of the present invention;

[0033] Figure 3 This is a schematic diagram of the LSTM model principle of an embodiment of the present invention;

[0034] Figure 4 This is a long short-term memory unit structure diagram of the LSTM model of an embodiment of the present invention;

[0035] Figure 5 It is a visualization diagram of an exponential raw data set according to an embodiment of the present invention;

[0036] Figure 6 This is a visualization diagram of the raw data of the case volume of Category 1 and Category 3 of the embodiment of the present invention;

[0037] Figure 7 This is a visualization diagram of the original data of the case volume of categories 2, 4, and 5 of the embodiment of the present invention;

[0038] Figure 8This is a visualization diagram of the original data of the case volume of Category 3 of the embodiment of the present invention;

[0039] Fig. 9 An exponential data box plot according to an embodiment of the present invention;

[0040] Fig.10 It is a heat map of the Pearson correlation coefficient of exponential data in an embodiment of the present invention;

[0041] Fig.11 It is a heat map of Kendall tau rank correlation coefficient of exponential data in an embodiment of the present invention;

[0042] Fig.12 The Pearson correlation coefficient heat map of the case volume data of the embodiment of the present invention;

[0043] Fig.13 The Kendall tau correlation coefficient heat map of the case volume data of the embodiment of the present invention;

[0044] Fig.14 It is a structural density function diagram of the dependent case volume of the first category and the third category of the embodiment of the present invention;

[0045] Fig.15 A Gaussian Copula function fitting QQ graph of the case volume dependency structure of the first category and the third category of the embodiment of the present invention;

[0046] Fig.16 Result diagram of variable regression analysis of cases of category 1 and category 3 of the embodiment of the present invention;

[0047] Fig.17 is the ACF diagram of variable 1 of the embodiment of the present invention;

[0048] Fig.18 is a PACF diagram of variable 1 of an embodiment of the present invention;

[0049] Fig.19 A visualization diagram of the time series prediction results of the ARIMA model according to an embodiment of the present invention;

[0050] Fig. 20 A visualization diagram of the time series prediction results of the ARIMA-Pro model according to an embodiment of the present invention;

[0051] Fig.21 The first-order difference PACF diagram of the first category case volume data of the embodiment of the present invention;

[0052] Fig. 22 The first-order difference ACF diagram of the case volume data of Category 1 according to the embodiment of the present invention;

[0053] Fig.23A visualization diagram of the ARIMA-Pro model of case volume data according to an embodiment of the present invention;

[0054] Fig.24 This is a visualization diagram of the time series prediction results of the LSTM model of an embodiment of the present invention;

[0055] Fig.25 This is a visualization diagram of the case volume data LSTM model of an embodiment of the present invention;

[0056] Fig.26 This is a visualization diagram of the time series prediction results of the Mix-ARIMA-LSTM model according to an embodiment of the present invention;

[0057] Fig. 27 This is a visualization diagram of the time series prediction results of the LRC-ARIMA-LSTM model according to an embodiment of the present invention;

[0058] Fig.28 This is a diagram of the prediction results of the INAR(1)-YW model according to an embodiment of the present invention;

[0059] Fig.29 This is a diagram of the prediction results of the INAR(1)-CLS model according to an embodiment of the present invention;

[0060] Fig.30 The original data of the first category of cases and four modal function diagrams of VMD decomposition according to the embodiment of the present invention;

[0061] Fig.31 This is a diagram of the prediction results of the VMD-LSTM model of an embodiment of the present invention;

[0062] Fig.32 This is a time series prediction result diagram of variable 1 of an embodiment of the present invention;

[0063] Fig.33 This is a time series prediction result diagram of variable 2 of an embodiment of the present invention;

[0064] Fig.34 This is a time series prediction result diagram of variable 3 of an embodiment of the present invention;

[0065] Fig.35 This is a time series prediction result diagram of variable 4 of an embodiment of the present invention;

[0066] Fig.36 A comparative analysis diagram of different models in an embodiment of the present invention;

[0067] Fig.37 It is a flow chart of the judicial data temporal regularity knowledge mining method of the present invention;

[0068] Fig.38 The figure is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings.

[0070] The embodiment of the present invention provides a method for mining judicial data time series regularity knowledge, which is specifically implemented as follows:

[0071] (1) ARIMA-Pro model: The ARIMA model is an integration of the autoregressive AR model and the moving average MA model, and a difference process is added on this basis. Specifically:

[0072] AR: represents the relationship between the next period response value and the previous period response value.

[0073] I: represents the difference process performed by the model to establish a stationary series;

[0074] MA: represents the relationship between the next period response value and the previous disturbance term.

[0075] Since the original data has time series characteristics, the ARIMA model can be used to predict a single column of future data. (p, d, q) are the model parameters in the model. The mathematical description of the model is as follows:

[0076]

[0077] in, express After d times of difference transformation, the sequence is is the random error term at time t, are the AR model fitting parameters, are the fitting parameters of the MA model, p and q are the orders of the moving average model (AR) and the autoregressive model (MA). How to accurately select parameters to achieve the best prediction effect while avoiding overfitting is the key to this model.

[0078] The ARIMA model and ARIMA-Pro model processes are as follows:

[0079] Data preprocessing: filter out outliers from the original data and use mean interpolation;

[0080] 1) Dataset division: Divide the preprocessed data into training set and test set at a ratio of 5 to 1;

[0081] 2) Stationarity test. Use ADF test on the training set data to get the result. If the data is stationary, proceed to the next step.

[0082] 3) Parameter selection. The ARIMA model selects the order based on the ACF and PACF plots of the data, while ARIMA-Pro selects the order based on the AIC, BIC criteria and cyclic structure;

[0083] 4) Training model. Input the training set data into the model of a certain order, obtain the training set prediction results and slightly adjust the parameters according to the results;

[0084] 5) Test the model. Apply the trained model to the test set, obtain the test set results and draw a visual image.

[0085] (2) LSTM model:

[0086] The information of each column of the original data is related to time, and thus it is predicted by the neural network time series model. The LSTM (Long Short Term Memory) model [2] is a special recurrent neural network (RNN) model. This model improves the internal structure of the neurons in the hidden layer of the RNN. Its working principle is to selectively forget some historical information, remember some new information at present, and then merge the past memory with the present and output it. Figure 3 As shown:

[0087] The long short-term memory network slightly improves the RNN structure. In the original hidden state h t Based on this, add a new state unit c t (memory cell) to store long-term states, so that it can preserve the key information in the long-term input sequence for a long time. The structure of a single LSTM neuron is as follows Figure 4 shown.

[0088] The activation function sigmoid of the forget gate changes the input to a value between (0,1). The closer the output value is to 0, the more information should be deleted. Otherwise, the information is retained. Which information:

[0089]

[0090] The input gate also uses the sigmoid function to filter the information of the current unit input and determine the What information is stored in:

[0091] Candidate Status The calculation of tanh function is equivalent to , The information contained is summarized as follows:

[0092]

[0093] Current unit status Contains all input information of the history: .

[0094] The output gate is responsible for controlling how much information from the current input is used for the output state: .

[0095] The final output of LSTM is determined by the output gate and the cell state: .

[0096] In the above calculation formula They are the corresponding unit inputs at the current moment. , The weight matrix and bias term of Denotes the Hadamard product.

[0097] Long short-term memory network through unit state It can store key information in historical inputs, which effectively solves the problem of long-term dependence of recurrent neural networks. The long short-term memory model has shown good performance in processing sequence problems. Here, different numbers of LSTM layers and fully connected layers will be stacked according to the data characteristics of different columns to achieve the best prediction effect.

[0098] (3) INAR(1) model: The INAR model is a useful non-negative dependent variable series model. The first-order INAR (INAR(1)) model is widely used, partly because of its explanatory power. The INAR(1) model assumes that the current number of individuals is the sum of the number of individuals that remained (or survived) in the previous period and the number of individuals that entered (or were born) in the intermediate period.

[0099] The number of cases at time t is ,INAR(1) model can be expressed as: ,

[0100] in represents the binomial thinning operator, which is defined as: , represents a list of independent and identically distributed 0-1 random variables with success probability , is expected to be A non-negative integer-valued random variable with parameter Poisson distribution.

[0101] (4) Copula function: The Copula function describes the correlation between random variables. It connects the marginal distribution function of each random variable with its joint distribution function, which is conducive to the overall study of the dependency mechanism between random variables. It is defined as:

[0102] ,in are the joint distribution function and its marginal distribution function respectively.

[0103] The expression of Gaussian Copula function is: ,

[0104] In the formula is the inverse function of the standard normal distribution function.

[0105] (5) Variational Mode Decomposition (VMD)

[0106] VMD is a classic data-driven signal processing algorithm. For time series, its essence is to convert a complex time series into Decompose into multiple intrinsic mode functions (IMFs) with different time scales and residual components The problem can be expressed as the solution of the following constraint problem:

[0107] min u k ,{ ω k } ∑ k=1 K ∂ t [(δ t + j πt * u k (t))] e -j ω k t 2 2 st ∑ k=1 K u k t =f(t) .

[0108] This constraint problem can be solved by using the Lagrange multiplier method and the variational method.

[0109] Dataset:

[0110] (1) Index-based judicial data set

[0111] This study randomly obtained four types of exponential data of different case types from the judicial data as the original data for model training. The data set contains four types of exponential data, each type of exponential data contains 60 values. For data security reasons, this study first anonymized all the data. The first five rows of data after anonymization are shown in the following table:

[0112] Table 1: Exponential data set (first 5 rows)

[0113]

[0114] The original data set of exponential data is visualized as follows Figure 5 shown.

[0115] (2) Caseload Judicial Dataset

[0116] The original judicial case data contains 6 different types of case volumes, and 60 values ​​are taken from each type as the training data set. The first 5 rows of data are shown in the following table:

[0117] Table 2: Caseload original data set (first 5 rows)

[0118]

[0119] Since the number of cases in different categories varies significantly, the raw data can be visualized as Figure 6 , Figure 7 , Figure 8 shown.

[0120] Preprocessing and data set division: Since the time series analysis model has high requirements for data stability, it is necessary to filter out abnormal data and preprocess it before training the model.

[0121] (1) Exponential data preprocessing

[0122] The outliers of the data can be obtained by using the box plot method. Fig. 9 As shown, based on the graphical results, the outliers of each variable in the data set are screened out. Taking the first variable as the force, its outliers are as follows:

[0123] Table 3: Outliers for variable 1

[0124]

[0125] For all outliers in the exponential data set, replace them with the mean of the remaining data, subtract the mean from each value and divide it by the standard deviation to convert it into standard data with a mean of 0 and a variance of 1.

[0126] On this basis, the first fifty data are used as training data to train the model parameters; the last ten data are used as test data to test and evaluate the model; among them, the test data set is as follows, with time points from 51 to 60.

[0127] Table 4: Exponential data test set

[0128]

[0129] (2) Caseload data preprocessing

[0130] Before using the LSTM model to model case volume data, the data needs to be standardized. The specific operation is to subtract the mean from each value and then divide it by the standard deviation to convert it into standard data with a mean of 0 and a variance of 1.

[0131] The first fifty data are used as training data to train the model parameters; the last ten data are used as test data to test and evaluate the model. The test data is as follows, from time 51 to 60.

[0132] Table 5: Caseload data test set

[0133]

[0134] Model performance evaluation indicators: Assume that the true value on the test set is: , the model's predicted value is: The following indicators are used to evaluate the model performance.

[0135] 1) Accuracy (Acc), whose mathematical definition is as follows: , (r is the number of errors within the range of positive and negative c),

[0136] 2) Mean square error (MSE), which is mathematically defined as follows:

[0137] 3) Mean absolute error (MAE), which is mathematically defined as follows:

[0138] 4) Root mean square error (RMSE), which is mathematically defined as follows:

[0139] 5) Mean absolute error rate (MAPE), which is mathematically defined as follows:

[0140]

[0141] 6) Symmetric mean absolute error (SMAPE), which is mathematically defined as follows:

[0142]

[0143] Among them, MSE and MAE can measure the size of the error between the predicted value and the true value, MAPE and SMAPE can measure the relative size of the error value to the true value, and RMSE can measure the size of the error between the predicted value and the true value. The closer these five indicators are to 0, the better the model effect.

[0144] Correlation analysis: The correlation coefficient can measure the correlation between different columns, with a value of -1 to 1. The closer the correlation coefficient is to 0, the weaker the correlation between the two columns of indicators, and the closer it is to positive or negative 1, the stronger the correlation. Calculate the Pearson correlation coefficient and Kendall tau rank correlation coefficient between the two columns of data respectively, and use a heat map to intuitively represent the correlation between variables.

[0145] (1) Index-based judicial dataset

[0146] The Pearson correlation coefficient can capture linear correlation. Fig.10 Display: The correlation coefficient between variable 1 and variable 2 is 0.396, indicating that there is a weak positive linear correlation; the correlation coefficient between any two columns is too small, so it can be considered that there is no linear relationship.

[0147] Kendall tau rank correlation coefficient can capture nonlinear correlation. Fig.11The results are almost the same. Therefore, it can be considered that there is a weak correlation between the first two variables, and there is no correlation between any other two variables.

[0148] (2) Original case data set

[0149] Calculate the Pearson correlation coefficient and Kendall tau rank correlation coefficient between the two columns of data respectively, and use the heat map as shown in Fig.12 shown.

[0150] The Pearson correlation coefficient can capture linear correlation. Fig.12 It can be seen that the correlation coefficient between the case volume of Category 1 and Category 3 is 0.889, indicating that there is a strong positive linear correlation; the correlation coefficients between the case volume of Category 6 and the case volumes of Category 2 and Category 3 are 0.687 and 0.572 respectively, indicating that there is a weak positive linear correlation between Category 6 and both Category 2 and Category 3 cases; the correlation coefficient between the case volume of Category 4 and Category 5 is 0.569, indicating that there is a weak positive linear correlation; and the correlation coefficients between any two other categories of cases are too small, and it can be considered that there is no linear relationship.

[0151] Kendall tau rank correlation coefficient can capture nonlinear correlation. Fig.13 Shows Fig.12 Almost the same results.

[0152] Therefore, for the strongly correlated case volumes of category 1 and category 3, we further estimate their marginal distribution functions to obtain F ̂_x (x), F ̂_y (y), and determine the optimal marginal distribution through the KS test.

[0153] In order to characterize the dependency structure of the two types of case variables, the Copula function (such as Frank, Clayton, Gumbel, Gaussian) is used to fit the data and estimate its parameter values. Finally, it is identified that the dependency structure of the first and third columns can be characterized by the Gaussian Copula function.

[0154] Based on the training set data, the unknown parameter ρ in the Gaussian Copula function is estimated to be (ρ ) ̂=0.8874. Substitute (ρ ) ̂= 0.8874 into the Gaussian Copula function and plot its density function as shown below: Fig.14 , Fig.15 shown.

[0155] Depend on Fig.15 It can be seen that the Gaussian copula can well fit the joint distribution of the number of cases in category 1 and category 3.

[0156] Due to the strong linear correlation between the number of cases in category 1 and category 3, the number of cases in category 3 was further used as the covariate and the number of cases in category 1 as the response variable to obtain the linear regression results in the training set as follows: Fig.16 shown.

[0157] Depend on Fig.16 It can be seen that the sample points are scattered around the fitting line and the calculated R2 is 0.61, which is small. Therefore, the linear regression model fitting effect of the first and third category cases is average. The prediction effect of the test set obtained by linear regression is as follows:

[0158] Table 6: LR value results based on linear regression

[0159]

[0160] The calculated MAPE of the case volume test set is 9.97%. Although the linear regression model has a general fitting effect on the data, due to the strong correlation of the data, the effect of predicting the case volume of category 1 through the case volume of category 3 is much higher than the single variable prediction result.

[0161] Stationarity test:

[0162] (1) When conducting time series modeling for exponential judicial data sets, we must first conduct a stationarity test. Taking the first column of variables as an example, we conduct an ADF test (Augmented Dickey-Fuller Testing) on ​​it. The results are:

[0163] Table 7: ADF test results for exponential data variable 1

[0164]

[0165] From the test results in Table 7, we can see that the Test Statistic Value is -4.9019, which is less than -3.5714, and the p-value is 0.00003. Therefore, the time domain sequence stationarity of variable 1 is 99%, which means that the conditions for time series modeling of variable 1 are met.

[0166] (2) Case volume original data set Taking the case volume data of category 1 as an example, the ADF test (Augmented Dickey-Fuller Testing) was performed on the original data of “category 1”. The results are as follows.

[0167] Table 8: ADF test results of the case volume variable in category 1

[0168]

[0169] As shown in Table 8, the test result of the first category case volume is -1.175215, which is greater than -2.59933, and the p-value is 0.684166. Therefore, the time domain series of this variable is not stable. In order to model the time series, it is necessary to further perform differential processing on it. The ADF test results of the first category case volume data after processing are as follows:

[0170] Table 9: ADF test results after first-order difference of case volume in category 1

[0171]

[0172] As shown in Table 9, the test result of the first-order difference processing of the case volume of Category 1 is -3.005977, which is less than -2.92262, and the p-value is 0.03434, which means that there is a 95% confidence that the series is stationary.

[0173] ARIMA / ARIMA-Pro Models:

[0174] (1) Index-type judicial data set For index-type judicial data, before ARIMA modeling, it is necessary to further perform ACF and PACF analysis on the data, such as Fig.17 and 18 shown.

[0175] The ARIMA model parameters are determined by the ACF and PACF plots. For variable 1 data, the parameters are selected as: p=5, d=0, q=9.

[0176] The prediction results of the trained ARIMA model on the test set are shown in the 'arima' column of the following table:

[0177] Table 10: ARIMA model prediction results

[0178]

[0179] The results of the ARIMA model on the training set and test set are visualized, such as Fig.19 shown.

[0180] Fig.19 The blue and green lines are the true values, the orange line is the prediction result of the training set, and the red line is the prediction result of the test set.

[0181] Different from the ARIMA model parameter determination method of observing ACF and PACF graphs, the ARIMA-Pro model parameters are determined by AIC, BIC criteria and a three-layer loop structure. For variable 1 data, the final parameters are: p=6, d=0, q=3.

[0182] The prediction results of ARIMA-Pro on the test set are shown in the 'arima-pro' column of the following table:

[0183] Table 11: ARIMA-Pro model prediction results for exponential data

[0184]

[0185] Visualize the results of the model on the training set and test set. Fig. 20 . Fig. 20 The blue and green lines are true values, the orange line is the prediction result of the training set, and the red line is the prediction result of the test set. By comparing the results of the two models, it is not difficult to find that the ARIMA model works better in the training set, and has extremely accurate prediction results at the seven time nodes of 51-56 and 60, but because the other three data are extreme values, the prediction effect is average. The ARIMA-Pro model has a smaller error in the test data because it can more accurately capture the changing trend of the original data. Overall, the ARIMA-Pro model avoids overfitting in the training set and achieves higher prediction accuracy.

[0186] (2) Case volume original data set For the case volume data, further ACF and PACF analysis results are as follows Fig.21 and 22 For the case volume data of category 1, the ARIMA-Pro model parameters can be determined as p = 4, d = 1, q = 4 based on the AIC and BIC criteria. The prediction results of the ARIMA-Pro model in the test set are shown in the 'arima-pro' column of the following table:

[0187] Table 12: Case volume data ARIMA-Pro model prediction results

[0188]

[0189] The calculated test set MAPE = 21.8%. The results of the model on the training set and test set are visualized, as shown in Fig.23 shown.

[0190] Fig.23 The blue and green lines are the true values, the orange line is the prediction result of the training set, and the red line is the prediction result of the test set. Fig.23 It can be seen that the ARIMA-Pro model has a smaller error in the training data and can more accurately capture the changing trend of the original data. The predictions at the first four time points of the test data are very accurate, but the errors become larger afterwards. This is mainly due to the following two reasons:

[0191] 1) The original data has large noise and the sequence is not stable. Even if the first-order difference is used, there is still a large error;

[0192] 2) The original data is too large.

[0193] LSTM Model:

[0194] The LSTM model used is composed of a stack of LSTM layers and fully connected layers. The main parameters of the LSTM layer are as follows:

[0195] input_size=(1,6,1),batch_size=1,timestep=6,input_dim=1,

[0196] hidden_size=12,num_layers=3,dropout=0.25.

[0197] Fully connected layer parameters: input_size=12, output_size=1.

[0198] Other model parameters:

[0199] learning rate=0.002

[0200] epoch=200

[0201] optimizer=Adam

[0202] loss function = MSELoss

[0203] (1) Exponential judicial data set For variable 1 data, the values ​​of the past 6 time points are used to predict the value at the next moment. The prediction results of the LSTM model on the exponential data test set are shown in the 'lstm' column of the following table:

[0204] Table 13: Exponential data LSTM model prediction results (variable 1)

[0205]

[0206] Visualize the results of the model on the test set as follows Fig.24 shown.

[0207] Fig.24 The blue line in the middle is the true value of the training set, the orange line is the true value of the test set, and the green line is the predicted result of the test set.

[0208] (2) Original case data set

[0209] The prediction results of the LSTM model on the case test set are shown in the 'lstm' column of the following table:

[0210] Table 14: LSTM model prediction results for case volume data (Category 1 cases)

[0211]

[0212] The calculated test set MAPE = 17.14%.

[0213] The results of the LSTM model on the test set are visualized as follows Fig.25 shown. Fig.25 The blue line is the true value of the training set, the orange line is the true value of the test set, and the green line is the predicted result of the test set. Fig.25 It can be seen that the LSTM model is also more accurate in the first four steps, but the error becomes larger after four steps. The reason is that the original data is too large. Even if the LSTM model's prediction error for normalized data is only 0.1, the error can reach 100982 after denormalization.

[0214] Ensemble model: Since different models can capture different features of data, two other combined prediction models are given here. The prediction results of the two models are weighted and combined to obtain a new model, so as to more accurately capture the data features and obtain more accurate prediction results.

[0215] Mix-ARIMA-LSTM model: The prediction results of the ARIMA-Pro and LSTM training sets are assigned different weights. The weights are selected by minimizing the mean square error (MSE) between the combined model results on the training set and the real data.

[0216] Assuming the LSTM model weight is β and the ARIMA-Pro model weight is (1-β), the prediction result of the Mix-ARIMA-LSTM model is: β(lstm_predict)+(1-β)(arima-pro_predict).

[0217] The second combined prediction model is the LR-ARIMA-LSTM model. The prediction results of the ARIMA-Pro and LSTM training sets are used as covariates to perform linear regression on the true values. The coefficients and intercept terms of the two models are obtained by performing linear regression on the results of the two models in the training set and the true values.

[0218] Assume that the LSTM model coefficient is α, the ARIMA-Pro model coefficient is β, and the intercept term is b. The prediction result of the LR-ARIMA-LSTM model is: α(lstm_predict)+β(arima-pro_predict)+b.

[0219] In the above two formulas, lstm_predict and arima-pro_predict are the prediction results of LSTM and ARIMA-Pro models respectively.

[0220] Table 15: Mix-ARIMA-LSTM model prediction results

[0221]

[0222] Visualize the results of the Mix-ARIMA-LSTM model on the test set and get Fig.26 .

[0223] Fig.26 The blue line in the middle is the true value of the training set, the orange line is the true value of the test set, and the green line is the predicted result of the test set.

[0224] LR-ARIMA-LSTM Model:

[0225] The LRR-ARIMA-LSTM parameters are obtained by linear regression method: α=0.7053, β=0.2863, b=0.6793.

[0226] The prediction results of LR-ARIMA-LSTM on the test set are shown in the 'lr_arima_lstm' column of the following table:

[0227] Table 16: LR-ARIMA-LSTM prediction results

[0228]

[0229] Visualize the results of the model on the test set as follows Fig. 27 shown.

[0230] Fig. 27 The blue line in the middle is the true value of the training set, the orange line is the true value of the test set, and the green line is the predicted result of the test set.

[0231] INAR (1)-YW model: INAR (1) model refinement operator parameters as well as The expectation is , usually estimated using the sample autocovariance function, the equation is as follows:

[0232]

[0233]

[0234] The parameter estimates are calculated using the Yule-Walker equation. The parameters are then fed into the INAR(1) model to obtain the INAR(1)-YW model, and the prediction results are as follows:

[0235] Table 17: INAR(1)-YW model prediction results

[0236]

[0237] The calculation shows that the MAPE index of the INAR(1)-YW model in the case test set is 20%. The INAR(1)-YW model is visualized as follows: Fig.28 shown.

[0238] INAR1(1)-CLS model: Another way to estimate the parameters of the INAR(1) model is to consider it as right The coefficients of the conditional least squares regression are solved. The parameters estimated by the conditional least squares regression are applied to the INAR(1) model to obtain the INAR(1)-CLS model, and its prediction results are shown in the following table.

[0239] Table 18: INAR(1)-CLS model prediction results

[0240]

[0241] The calculation shows that the MAPE index of the INAR(1)-CLS model in the case test set is 20.75%. The INAR(1)-CLS model is visualized as follows: Fig.29 shown.

[0242] Depend on Fig.28 and Fig.29 By comparison, the results obtained by the two parameter estimation methods are basically the same. Although the INAR (1) model has a greater advantage in capturing data trends, the prediction is not accurate. This is because the data is too dispersed (the variance is much larger than the mean), which leads to inaccurate parameter estimation and thus affects the accuracy of the model.

[0243] YND-LSTM model: For time domain regularity mining, the research object is the root opportunity judicial time series data, and a time series prediction model that conforms to its data characteristics is constructed. However, due to the excessive noise in the sample data and the instability of the original sequence, the classic time series prediction models such as ARIMA, LSTM, and INAR (1) all show poor results. Therefore, in order to solve the instability of the original data, VMD is further used to preprocess the data and decompose the complex original sequence into several simple subsequences that contain the structure and periodicity of the original sequence. Then, the subsequences (generally speaking, the subsequences obtained by VMD mostly have relatively good properties such as stability) are modeled separately to achieve the purpose of improving the prediction accuracy.

[0244] First, perform modal decomposition on the original data and decompose the original sequence into k subsequences , and then for each sequence The LSTM model is applied to obtain the subsequence prediction results, and finally the prediction results of each subsequence are added together to obtain the test set prediction results.

[0245] For the first category of cases, after parameter selection, the number of modes is finally determined to be 4, that is, the original data (sequence) of the first category of cases is decomposed into 4 modes .like Fig.30 shown.

[0246] Fig.30 Category 1 represents the original data of the first category of cases, 'mode1' represents the first mode function obtained by VMD decomposition of the original data, and the rest are the second, third, and fourth mode functions. Then, LSTM is used to model and predict the four modes separately and then add them together to obtain the test set prediction results of the VMD-LSTM model, as shown in the following table:

[0247] Table 19: Prediction results of VMD-LSTM model

[0248]

[0249] The calculated MAPE index of the VMD-LSTM model is 10.56%. The results are visualized as follows Fig.31 shown.

[0250] The blue curve 'y1_true' represents the true data of category 1, and the orange curve 'y1_predict' represents the prediction result of VMD-LSTM. Fig.31 It can be seen that the case volume data after VMD decomposition and then modeling and predicting can accurately capture the changing trend of the original data and greatly improve the prediction effect. However, the model's prediction at the extreme point is slightly worse. This is because the sample size is insufficient, resulting in the high-frequency mode being too rough, which in turn affects the prediction performance of LSTM.

[0251] Model Evaluation:

[0252] (1) Index-based judicial data set

[0253] Assume the true value on the test set is: y=(y_1,y_2,⋯,y_10);

[0254] The predicted value of the model is: y'=(y_1',y_2',⋯,y_10').

[0255] There are five evaluation indicators, namely accuracy, mean square error (MSE), mean absolute error (MAE), mean absolute error rate (MAPE), and symmetric mean absolute error rate (SMAPE). The definitions are as follows:

[0256] , (r is the number of errors within the range of positive and negative c),

[0257] ,

[0258] ,

[0259] ,

[0260] .

[0261] Among them, MSE and MAE can measure the size of the error between the predicted value and the true value, MAPE and SMAPE can measure the relative size of the error value to the true value. The closer the four indicators are to 0, the better the model effect.

[0262] For the data of variable 1, the prediction results of the five models are visualized as follows Fig.32 shown.

[0263] The indicator results of variable 1 data are shown in the following table (c=1):

[0264] Table 20: Evaluation index results of variable 1 data

[0265]

[0266] As can be seen from the table above, for the prediction of variable 1 data, the Mix-ARIMA-LSTM model achieves the best score and the highest accuracy. Therefore, the Mix-ARIMA-LSTM model can be selected to predict variable 1 data.

[0267] For the data of variable 2, the prediction results of the five models are visualized as follows Fig.33 shown.

[0268] The indicator results of variable 2 data are shown in the following table (c=1.5):

[0269] Table 21: Evaluation index results of variable 2 data

[0270]

[0271] As can be seen from the table, for the prediction of variable 2 data, the LSTM model achieves the best score except for MAE, and its MAE is only 0.004 different from the other minimum MAE, which can be ignored. It can be considered that the result obtained by LSTM is the most accurate. Therefore, the LSTM model can be selected to predict the variable 2 data.

[0272] For the data of variable 3, the prediction results of the five models are visualized as follows Fig.34 shown.

[0273] The indicator results of variable 3 data are shown in the following table (c=1.5):

[0274] Table 22: Evaluation index results of variable 3 data

[0275]

[0276] As can be seen from the table, for the prediction of variable 3 data, the ARIMA model achieves the best score and the result has the highest accuracy. Therefore, the ARIMA model can be selected to predict variable 3 data.

[0277] For the data of variable 4, the prediction results of the five models are visualized as follows Fig.35 shown.

[0278] The indicator results of variable 4 data are shown in the following table (c=1.5):

[0279] Table 23: Evaluation index results of variable 4 data

[0280]

[0281] As can be seen from the table, for the prediction of variable 4 data, the LSTM model achieves the best score except MAE, but its model accuracy is higher and the other indicators are not much different from LSTM. It can be considered that the result obtained by L-ARIMA-LSTM is the most accurate. Therefore, the L-ARIMA-LSTM model can be selected to predict the data of variable 4.

[0282] (2) The original case data set compares and analyzes five models, including ARIMA-Pro, LSTM, INAR(1)-YW, INAR(1)-CLS, and VMD-LSTM. Fig.36 shown.

[0283] Among them, the data results for the number of cases in Category 1 are shown in the following table:

[0284] Table 24: Performance of the time series regularity model for the first category of cases

[0285]

[0286] It can be seen from Table 24 that for the time series regularity of case volume data, the three evaluation indicators of the VMD-LSTM model on the test set are reduced by nearly half, and MAPE and SMAPE both reach about 10%, showing the best results. Therefore, the optimal model for mining the time series regularity of case volume data is VMD-LSTM.

[0287] An embodiment of the present invention further provides a computer device, Fig.38 is a schematic diagram of the structure of a computer device provided by an embodiment of the present invention; see the accompanying drawings Fig.38 As shown, the computer device includes: an input system 23, an output system 24, a memory 22 and a processor 21; the memory 22 is used to store one or more programs; when the one or more programs are executed by the one or more processors 21, the one or more processors 21 implement the judicial data time series regularity knowledge mining method provided in the above embodiment; wherein the input system 23, the output system 24, the memory 22 and the processor 21 can be connected by a bus or other means, Fig.38 The example of connecting through bus is taken in the following.

[0288] The memory 22 is a readable and writable storage medium of a computing device, which can be used to store software programs and computer executable programs, such as the program instructions corresponding to the judicial data time series regularity knowledge mining method described in the embodiment of the present invention; the memory 22 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function; the data storage area can store data created according to the use of the device, etc.; in addition, the memory 22 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device; in some instances, the memory 22 can further include a memory remotely arranged relative to the processor 21, and these remote memories can be connected to the device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0289] The input system 23 may be used to receive input digital or character information, and to generate key signal input related to user settings and function control of the device; the output system 24 may include display devices such as display screens.

[0290] The processor 21 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 22, that is, realizes the above-mentioned judicial data time series regularity knowledge mining method.

[0291] The computer device provided above can be used to execute the judicial data time series regularity knowledge mining method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0292] The embodiment of the present invention also provides a storage medium containing computer executable instructions, which are used to execute the judicial data time series regularity knowledge mining method provided in the above embodiment when executed by a computer processor. The storage medium is any of various types of memory devices or storage devices, and the storage medium includes: installation media, such as CD-ROM, floppy disk or tape system; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media (such as hard disk or optical storage); registers or other similar types of memory elements, etc.; the storage medium may also include other types of memory or combinations thereof; in addition, the storage medium may be located in a first computer system in which the program is executed, or may be located in a different second computer system, which is connected to the first computer system via a network (such as the Internet); the second computer system may provide program instructions to the first computer for execution. The storage medium includes two or more storage media that may reside in different locations (for example, in different computer systems connected via a network). The storage medium may store program instructions (for example, specifically implemented as a computer program) that can be executed by one or more processors.

[0293] Of course, the storage medium containing computer executable instructions provided in an embodiment of the present invention is not limited to the judicial data time series regularity knowledge mining method described in the above embodiment, and can also execute related operations in the judicial data time series regularity knowledge mining method provided in any embodiment of the present invention.

[0294] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for mining judicial data temporal regularity knowledge, characterized by: The following steps are involved: Randomly obtain multiple types of exponential data of different case types from judicial data as raw data for time series mining model training, the data set includes multiple types of exponential data; the raw data includes multiple types of case volume data, and a set number of case volume data of each type is taken as training set data; Performing data preprocessing on the raw data, and dividing the preprocessed data into a training set and a test set at a ratio of 5 to 1; The data were analyzed for correlation using correlation coefficient and chi-square verification method; The training set data is tested using the ADF test to obtain the test results, and the model parameters are selected after the data is stable; The training set data is input into the ARIMA model, LSTM model, LR-ARIMA-LSTM model, INAR (1) model, VMD-LSTM model, etc. to obtain the training set prediction results, and the parameters are slightly adjusted according to the results. Apply the trained model to the test set, obtain the test set prediction results, and draw a visual image; The prediction accuracy of the test set prediction results of multiple models is compared. Based on the fact that different models can capture different characteristics of the data, the prediction results of multiple models are weighted and combined to obtain a new combined prediction model, which is then used to predict the data.

2. The judicial data temporal regularity knowledge mining method according to claim 1 is characterized in that: The data preprocessing method includes: filtering out outliers from the original data and interpolating using the mean; the exponential data preprocessing method is to use the box plot method to filter out outliers for each variable in the data set, taking the first variable as the force, replacing all outliers in the exponential data set with the mean of the remaining data, and subtracting the mean from each value and dividing it by the standard deviation, thereby converting it into standard data with a mean of 0 and a variance of 1, using the first fifty data as training data to train model parameters; and using the last ten data as test data to test the model and make a model evaluation; The preprocessing method of the case volume data is as follows: standardize the data, subtract the mean from each value and divide it by the standard deviation, thereby converting it into standard data with a mean of 0 and a variance of 1; use the first fifty data as training data to train model parameters; and use the last ten data as test data to test the model and make a model evaluation.

3. The judicial data temporal regularity knowledge mining method according to claim 2 is characterized in that: The model evaluation method includes: The accuracy Acc, mean square error MSE, mean absolute error MAE, root mean square error RMSE, mean absolute error rate MAPE, and symmetric mean absolute error rate SMAPE are used to evaluate the model performance. Among them, MSE and MAE measure the size of the error between the predicted value and the true value, MAPE and SMAPE measure the relative size of the error value to the true value, and RMSE measures the size of the error between the predicted value and the true value. The closer these five indicators are to 0, the better the model effect.

4. The judicial data temporal regularity knowledge mining method according to claim 1 is characterized in that: The method of correlation analysis comprises: The Pearson correlation coefficient and Kendall tau rank correlation coefficient between the two columns of data are calculated respectively. The Pearson correlation coefficient captures linear correlation, and the Kendall tau rank correlation coefficient captures nonlinear correlation. The correlation coefficient is used to measure the correlation between different columns, with a value ranging from -1 to 1. The closer the correlation coefficient is to 0, the weaker the correlation between the two columns of indicators, and the closer it is to plus or minus 1, the stronger the correlation.

5. The judicial data temporal regularity knowledge mining method according to claim 1 is characterized in that: The new combined prediction model obtained by weighting and combining the prediction results of multiple models includes: Mix-ARIMA-LSTM model, assuming that the LSTM model weight is β and the ARIMA-Pro model weight is (1-β), then the prediction result of the Mix-ARIMA-LSTM model is: β(lstm_predict)+(1-β)(arima-pro_predict)(1) The LR-ARIMA-LSTM model uses the prediction results of the ARIMA-Pro and LSTM training sets as covariates to perform linear regression on the true values. The coefficients and intercept terms of the ARIMA-Pro and LSTM models are obtained by performing linear regression on the two model results of the training set and the true values. Assuming the LSTM model coefficient is α, the ARIMA-Pro model coefficient is β, and the intercept term is b, the prediction result of the LR-ARIMA-LSTM model is: α(lstm_predict)+β(arima-pro_predict)+b(2) In formula (1) and (2), lstm_predict and arima-pro_predict are the prediction results of LSTM and ARIMA-Pro models respectively.

6. The judicial data temporal regularity knowledge mining method according to claim 1 is characterized in that: The parameter selection of the model includes: The ARIMA model selects the order based on the ACF and PACF plots of the data, and the ARIMA-Pro model selects the order based on the AIC, BIC criteria and cyclic structure.

7. The judicial data temporal regularity knowledge mining method according to claim 1, characterized in that: The correlation analysis method includes: for two types of case volumes x and y that are strongly linearly correlated, further estimating the marginal distribution function of the two types of case volumes x and y to obtain F x (x), F y (y) and determine the optimal marginal distribution through the KS test.

8. The method for mining judicial data temporal regularity knowledge according to claim 7, characterized in that: The correlation analysis method also includes: using a Copula function to fit the data of two types of case volumes that are strongly linearly correlated, and estimating fitting parameter values ​​to identify the dependency structure of the two types of case volumes that are strongly linearly correlated.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the judicial data temporal regularity knowledge mining method described in any one of claims 1-8 are implemented.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the judicial data time series regularity knowledge mining method as described in any one of claims 1-8 are implemented.

Images