An enterprise risk intelligent identification method based on financial big data

Abstract

A kind of enterprise risk intelligent identification method based on financial big data, steps as follows:1) judge whether training cluster center and decision tree is needed;2) data acquisition, obtain the business data, legal data, category data of enterprise, generate high-risk data set, low-risk data set;3) data preprocessing, the data in two data sets are split into high-risk business data, legal data, category data matrix, low-risk business data, legal data matrix, low-risk matrix, category data matrix;4) data clustering, constructs K-Means model, respectively trains cluster center to high-risk business data matrix, high-risk legal data matrix;5) extract nonlinear feature: obtain high-risk feature matrix and low-risk feature matrix;6) construct decision tree, based on high-risk feature matrix and low-risk feature matrix constructs decision tree model;7) identify enterprise risk, load decision tree model, obtain security risk analysis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of enterprise risk identification, in particular to an enterprise risk intelligent identification method based on financial big data. BACKGROUND

[0002] For the convenience of management and supervision, the business data such as annual report, registered capital and the legal data such as execution, litigation of the registered enterprise will be disclosed on the relevant website. With the development of big data analysis technology, the modeling can be carried out by using the disclosed data of the registered enterprise and the normal enterprise, and the hidden rules in the data can be analyzed to further realize the risk warning of the enterprise.

[0003] The judgment of enterprise security risk is essentially to classify the enterprise by using the obtained business and legal data, and the enterprise is divided into high risk, low risk and other categories according to the risk degree. The solution to the classification problem generally includes two steps of feature extraction and modeling classification.

[0004] Among them, the common methods of feature extraction include: 1. Principal component analysis, which maps the n-dimensional features of the data to fewer k-dimensional features and ensures the orthogonality of these k-dimensional features. This method is divided into two kinds of feature decomposition which tries to retain the features and singular value decomposition which tries to ensure the sparsity of data points; 2. Linear discriminant analysis, which projects the data to low dimension, so that the data of the same category is as close as possible, and the distance between different categories is as large as possible; 3. Multidimensional scaling analysis, which simplifies the data samples in the multidimensional space to the low-dimensional space for positioning, analysis and classification, while retaining the original relationship between objects. The principal component analysis and linear discriminant analysis are linear dimension reduction methods, which are only suitable for linearly separable data samples; the multidimensional scaling analysis is a nonlinear dimension reduction method, but it needs to manually determine the dimension of the low-dimensional space and the actual meaning of each dimension, which has certain limitations.

[0005] Common classification models include: 1. Bayesian discriminant method based on probability, which is only suitable for linearly separable samples and needs to obtain the prior probability of each class; 2. Decision tree algorithm, which is prone to overfitting; 3. Support vector machine method, which is sensitive to data and difficult to use non-continuous data such as class attributes for classification; 4. Artificial neural network, which has poor interpretability and high machine learning cost. SUMMARY

[0006] The present application aims at the deficiencies of the prior art, and provides an enterprise risk intelligent identification method based on financial big data, which can continuously train the model using enterprise data and effectively identify enterprise risks.

[0007] The technical scheme of the present application is: an enterprise risk intelligent identification method based on financial big data, the steps are as follows:

[0008] 1) Determine whether the clustering center and decision tree need to be trained, if not, directly collect enterprise data for enterprise risk identification, if yes, go to the next step;

[0009] 2) Data collection, obtain the business data, legal data and category data of enterprises from the website, and store the data of the 1-2 years before the cancellation of the cancelled enterprises in the high-risk data set dfAcc, and store the data of the enterprises not cancelled for more than 5 years in the low-risk data set dfNorm;

[0010] 3) Data preprocessing, delete all 0 or invalid data with incomplete data in the high-risk data set dfAcc and the low-risk data set dfNorm, and split the data in the two data sets into high-risk business data matrix, high-risk legal data matrix, high-risk category data matrix, low-risk business data matrix, low-risk legal data matrix, and low-risk category data matrix by range standardization of business data and legal data in the two data sets respectively;

[0011]

[0012] 4) Data clustering, build a K-Means model, train the clustering center using the high-risk business data matrix and the high-risk legal data matrix, and determine the number of clusters according to the dissimilarity of the high-risk business data matrix, the high-risk legal data matrix, the low-risk business data matrix, and the low-risk legal data matrix in the clustering result;

[0013] 5) Extract nonlinear features:

[0014] 5-1) According to the two sets of clustering centers obtained in step 3), select 4 clustering centers with performance difference exceeding the threshold or maximum difference from the high-risk business data matrix and the high-risk legal data matrix respectively, and select 4 clustering centers with performance difference exceeding the threshold or maximum difference from the low-risk business data matrix and the low-risk legal data matrix respectively;

[0015] 5-2) Calculate the modified cosine similarity of each group of data distance to the corresponding clustering center, and obtain the high-risk business data distance matrix, the legal distance matrix, and the low-risk business data distance matrix, the legal distance matrix;

[0016] 5-3) Horizontally merge the high-risk business data distance matrix, the legal distance matrix, and the high-risk category data matrix to obtain the high-risk feature matrix chAcc and the low-risk feature matrix chNorm;

[0017] 6) Build a decision tree, build a decision tree model based on the high-risk feature matrix chAcc and the low-risk feature matrix chNorm;

[0018] ​7) Identify enterprise risks, load decision tree model, classify the obtained nonlinear features, and obtain safety risk analysis results.

[0019] Further, the business data includes registered capital, main income, total liabilities, main business income, net profit, total tax, and owner's equity, the legal data includes external guarantee, administrative punishment, external investment, plaintiff case, defendant case, defendant court, and judgment loss, and the category data includes region, industry, and enterprise nature.

[0020] Further, in step 3), the specific steps of data preprocessing are as follows,

[0021] 3-1) Delete the data in the high-risk data set dfAcc and the low-risk data set dfNorm that are repeated, incomplete, all 0, and registered capital 0;

[0022] 3-2) Split the high-risk data set dfAcc horizontally into a high-risk non-category data matrix dataAcc without registered capital and a category data matrix TypeAcc, and split the low-risk data set dfNorm horizontally into a low-risk non-category data matrix dataNorm without registered capital and a category data matrix TypeNorm;

[0023] 3-3) Divide each element in the high-risk non-category data matrix dataAcc without registered capital and the low-risk non-category data matrix dataNorm without registered capital by the registered capital item in the same row, respectively, to obtain two groups of high-risk and low-risk data matrices;

[0024] 3-4) According to the range d and the minimum value m of each item of data in the high-risk non-category data matrix dataAcc without registered capital, the two groups of data matrices obtained in step 3-3) are subjected to range standardization, and the formula is as follows:

[0025]

[0026] In the formula, data ij represents the range standardization of the high-risk or low-risk data matrix, m j represents the minimum value of the jth item, d j represents the range of the jth item.

[0027] 3-5) Split the data matrix of step 3-4) by data item category to obtain high-risk business data matrix dataCorpAcc, high-risk legal data matrix dataLawAcc, high-risk category data matrix dataTypeAcc, low-risk business data matrix dataCorpNorm, low-risk legal data matrix dataLawNorm, and low-risk category data matrix dataTypeNorm.

[0028] Further, step 4), the specific method of data clustering is as follows,

[0029] 4-1) Set the range of the number of cluster centers to be enumerated k1, k2, enumerate the number of cluster centers k, so that k1 < k < k2, and randomly select k data points in the high-risk business data matrix and the high-risk legal data matrix as the initial cluster centers nowCentre k*m = (c1, c2, …, c k ) T ;

[0030] 4-2) Based on the modified cosine similarity as the distance measurement method, the distance is uniformly mapped to (0, 1) to obtain the high-risk business data distance matrix and the high-risk legal data distance matrix, respectively. The calculation formula of the high-risk business data distance matrix is as follows,

[0031] distAcc = (x ij ),

[0032]

[0033] In the formula, distAcc represents the high-risk business data distance matrix, x ij represents the element of the i-th row and the j-th column of distAcc, data il represents the element of the i-th row and the l-th column of distAcc, represents the average value of the l-th column of dataCorpAcc, centre jl represents the element of the i-th row and the j-th column of nowCentre;;

[0034] The calculation formula of the high-risk legal data distance matrix has the same principle as the high-risk business data distance matrix.

[0035] 4-3) Each point in each distance matrix is classified into the class represented by the nearest cluster center, and the cluster center is updated with the average value of each class, and steps 3-2) and 3-3) are repeated.

[0036] 4-4) When the clustering situation no longer changes or the number of loop iterations reaches the upper limit, save the current clustering center nowCentre as the local optimal clustering center cen;

[0037] 4-5) Obtain the high-risk enterprise dataset and the low-risk enterprise dataset respectively according to the classification situation of the local optimal clustering center cen: count = (c1, c2,..., c k )), c i represents the number of data points belonging to the i-th class, and save them to the number of data points count1 belonging to each class in the high-risk dataset and the number of data points count2 belonging to each class in the low-risk dataset respectively;

[0038] 4-6) Calculate the Pearson correlation coefficient nowRela according to count1 and count2. The calculation formula is as follows,

[0039]

[0040] If nowRela < rela (the initial value of rela is 1), then save the local optimal clustering center cen as the optimal clustering center centre, and save rela as the current nowRale;

[0041] If nowRela > rela, keep the clustering center unchanged.

[0042] Further, in step 1), the method for judging whether to train the clustering center and the decision tree is as follows,

[0043] 1-1) Set the mode mode = 1 to represent predicting the enterprise risk, distName ='refinedCosine' to represent training the clustering center based on the modified cosine similarity, trnCentre = 0 to represent not retraining the clustering center, pickCentre = 1 to represent selecting fewer centers for feature extraction, and trnTree = 1 to represent retraining the decision tree;

[0044] 1-2) When trnCentre = 0, refineCentre = 1, and distName ='refinedCentre', check the folder where the clustering center is saved. If there is no trained clustering center, then it is necessary to retrain the clustering center and the decision tree that depends on the clustering center.

[0045] Further, in step 5-1), after traversing k, the centers are saved to the corresponding text files. Since refinedCentre = 1, the 8 centers with the largest difference between count1 and count2 from the optimal cluster centers are selected as pickedCentres and saved to the corresponding text files.

[0046] Furthermore, in step 6), when constructing the decision tree model, the Gini coefficient method is used to calculate the information gain. When selecting node parameters, the set of features with the strongest information gain is selected for branching, and the post-pruning strategy with minimum support is used for pruning.

[0047] The beneficial effects of adopting the above technical solution are as follows:

[0048] 1. This method can achieve intelligent identification of enterprise risks;

[0049] 2. This method is suitable for nonlinear feature extraction from financial big data;

[0050] 3. The risk identification model constructed using this method has been tested with test data and its accuracy rate reaches 99%.

[0051] 4. The risk identification model constructed by this method is adaptive: as the financial environment changes, new data or new data items can be continuously used to train the model, and it can solve the problem of overfitting in decision tree models.

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0053] Fig. 1 This is a logic flowchart of the present invention;

[0054] Fig. 2 This is a flowchart illustrating the logic of data preprocessing in this invention.

[0055] Fig. 3 This is a flowchart illustrating the logic of training cluster centers in this invention. Detailed Implementation

[0056] See Figs. 1 to 3 A method for intelligent identification of enterprise risk based on financial big data, the specific steps of which are as follows:

[0057] 1) Determine whether cluster centers and decision trees need to be trained. If not, directly collect enterprise data for enterprise risk identification. If so, proceed to the next step. The specific determination method is as follows:

[0058] 1-1) Setting mode=1 indicates predicting enterprise risk, distName='refinedCosine' indicates training cluster centers based on modified cosine similarity, trnCentre=0 indicates not retraining cluster centers, pickCentre=1 indicates selecting fewer centers for feature extraction, and trnTree=1 indicates retraining the decision tree.

[0059] 1-2) When trnCentre=0, refineCentre=1, distName='refinedCentre', check the folder where the cluster centers are saved. If there are no trained cluster centers, then the cluster centers and the decision trees that depend on the cluster centers need to be retrained.

[0060] 2) Data collection: Obtain enterprise business registration data, legal data, and category data from the website. The business registration data includes registered capital, main operating income, total liabilities, main business revenue, net profit, total tax payment, and owner's equity. The legal data includes external guarantees, administrative penalties, external investments, plaintiff's case filing, defendant's case filing, defendant's court appearance, and judgment loss. The category data includes region, industry, and enterprise nature. Data of deregistered enterprises for 1-2 years before deregistration is stored in the high-risk dataset dfAcc, and data of enterprises that have not been deregistered for more than 5 years is stored in the low-risk dataset dfNorm.

[0061] 3) Data preprocessing, such as Fig. 2 As shown, invalid data with all zeros or incomplete data in the high-risk dataset dfAcc and the low-risk dataset dfNorm are deleted. Then, by performing range standardization on the business data and legal data in both datasets respectively, the data in the two datasets are split into a high-risk business data matrix, a high-risk legal data matrix, a high-risk category data matrix, a low-risk business data matrix, a low-risk legal data matrix, and a low-risk category data matrix. The specific method is as follows:

[0062] 3-1) In this embodiment, duplicate, incomplete, all zero, and registered capital zero data are deleted from the high-risk dataset dfAcc and the low-risk dataset dfNorm. 150,000 data entries are randomly selected from the high-risk dataset dfAcc and 450,000 data entries are randomly selected from the low-risk dataset dfNorm.

[0063] 3-2) Horizontally split the high-risk dataset dfAcc into a high-risk non-categorical data matrix dataAcc (excluding registered capital) and a categorical data matrix TypeAcc, and horizontally split the low-risk dataset dfNorm into a low-risk non-categorical data matrix dataNorm (excluding registered capital) and a categorical data matrix TypeNorm;

[0064] 3 - 3) Divide each element in the high - risk non - category data matrix dataAcc without registered capital and the low - risk non - category data matrix dataNorm without registered capital by the registered capital item in the same row to obtain two groups of high - risk and low - risk data matrices;

[0065] 3 - 4) According to the range d and the minimum value m of the data in the high - risk non - category data matrix dataAcc without registered capital, perform range normalization on the data in the two groups of data matrices obtained in step 3 - 3). The formula is as follows:

[0066]

[0067] In the formula, data ij represents the range normalization of the high - risk or low - risk data matrix, m j represents the minimum value of the j - th item, and d j represents the range of the j - th item;

[0068] 3 - 5) Split the data matrix normalized by range in step 3 - 4) according to the data item category to obtain the high - risk industrial and commercial data matrix dataCorpAcc 150000×6 , the high - risk legal affairs data matrix dataLawAcc 150000×7 , the high - risk category data matrix dataTypeAcc 150000×3 , the low - risk industrial and commercial data matrix dataCorpNorm 450000×6 , the low - risk legal affairs data matrix dataLawNorm 450000×7 , and the low - risk category data matrix dataTypeNorm 450000×3 .

[0069] 4) Data clustering. As shown in Fig. 3 , construct a K - Means model, respectively perform clustering center training on the high - risk industrial and commercial data matrix and the high - risk legal affairs data matrix, and determine the number of clusters according to the dissimilarity of the high - risk industrial and commercial data matrix, the high - risk legal affairs data matrix, the low - risk industrial and commercial data matrix, and the low - risk legal affairs data matrix in the clustering results. The specific method is as follows:

[0070] 4 - 1) Set the range k1, k2 of the number of clustering centers to be enumerated, enumerate the number of clustering centers k, so that k1 < k < k2, and randomly select k data points in the high - risk industrial and commercial data matrix and the high - risk legal affairs data matrix respectively as the initial clustering centers nowCentre k*m =(c1, c2,..., c k ) T ;

[0071] 4-2) Based on modified cosine similarity as the distance metric, the distances are uniformly mapped to (0,1) to obtain the distance matrices for high-risk business data and high-risk legal data. Taking the calculation of the distance matrix for high-risk business data as an example, the calculation formula is as follows:

[0072] distAcc 150000×k =(x ij )

[0073]

[0074] In the formula, distAcc 150000×k distAcc represents the distance matrix for high-risk business data. The matrix has 150,000 rows and k columns, x ij This represents the element in the i-th row and j-th column of distAcc, data il This represents the element in the i-th row and l-th column of distAcc. This represents the average value of the l-th column of dataCorpAcc, where centre jl This represents the element in the i-th row and j-th column of nowCentre;

[0075] The calculation principle of the distance matrix for high-risk business data is the same as that for the distance matrix for high-risk business data.

[0076] 4-3) Assign each point in the distance matrix of high-risk business and legal data to the class represented by the nearest cluster center, update the cluster center with the average value of each class, and repeat steps 3-2) and 3-3).

[0077] 4-4) When the clustering situation no longer changes or the number of iterations reaches the upper limit, the current cluster center nowCentre is saved as the local optimal cluster center cen. The upper limit of the number of iterations can be set according to actual needs. In this embodiment, the upper limit of the number of iterations is set to 80,000.

[0078] 4-5) Obtain the classification results of the high-risk dataset dfAcc and the low-risk dataset dfNorm according to the local optimal cluster center cen: count = (c1, c2, ..., c k ), c i This represents the number of data points belonging to the i-th class, and is stored in the high-risk dataset as count1, and in the low-risk dataset as count2, respectively.

[0079] 4-6) Calculate the Pearson correlation coefficient nowRela based on count1 and count2. The calculation formula is as follows:

[0080]

[0081] If nowRela < rela (the initial value of rela is 1), then save the local optimal clustering center cen as the optimal clustering center centre, and save rela as the current Pearson correlation coefficient nowRale;

[0082] If nowRela > rela, keep the clustering center unchanged.

[0083] 5) Extract non - linear features:

[0084] 5 - 1) According to the two sets of clustering centers obtained in step 3), respectively select 4 clustering centers with a performance difference exceeding the threshold or the largest difference from the high - risk industrial and commercial data matrix and the high - risk legal affairs data matrix, and respectively select 4 clustering centers with a performance difference exceeding the threshold or the largest difference from the low - risk industrial and commercial data matrix and the low - risk legal affairs data matrix, as follows:

[0085] After traversing k, save the optimal clustering center centre to the corresponding text file. / centres / refinedCosine / corp / all / centres.csv. Since refinedCentre = 1, select 8 clustering centers with the largest difference between count1 and count2 from the optimal clustering center centre as pickedCentre, and save them to the corresponding text file. / centres / refinedCosine / corp / pick / centres.csv;

[0086] 5 - 2) Calculate the modified cosine similarity between each group of data and the corresponding clustering center. The principle of the modified cosine similarity formula is the same as that in step 4 - 2), and obtain the high - risk industrial and commercial data distance matrix distCorpAcc, the high - risk legal affairs distance matrix distLawAcc, the low - risk industrial and commercial data distance matrix distCorpNorm, and the low - risk legal affairs data distance matrix distLawNorm;

[0087] 5 - 3) Horizontally merge the high - risk industrial and commercial data distance matrix distCorpAcc and the legal affairs distance matrix distLawAcc with the high - risk category data matrix dataTypeAcc, and horizontally merge the low - risk industrial and commercial data distance matrix distCorpNorm and the legal affairs data distance matrix distLawNorm with the low - risk category data matrix dataTypeNorm, respectively obtaining the high - risk feature matrix chAcc and the low - risk feature matrix chNorm.

[0088] 6) Construct a decision tree. Train a C5.0 decision tree model based on the high-risk feature matrix chAcc and the low-risk feature matrix chNorm. When training the decision tree model, use the Gini coefficient method to calculate the information gain. When selecting node parameters, select the set of features with the strongest information gain for branching, and use the minimum support post-pruning strategy to prune the tree to avoid overfitting. Save the trained decision tree to the corresponding text file . / model / tree.pkl.

[0089] 7) Identify enterprise risks, read the enterprise data to be predicted for risk, process the data using the method in step 3), then use the cluster centers obtained in step 4) and the method in step 5) to extract nonlinear features to obtain the corresponding feature matrix, then load the decision tree obtained in step 6), substitute it into the feature matrix, classify the obtained nonlinear features, and obtain the safety risk analysis results. The analysis results are shown in the table below:

[0090]

[0091] The test results show that, compared with commonly used methods (directly building a decision tree model), the accuracy and recall of the model built using this method, after data preprocessing, nonlinear feature extraction, and construction of a decision tree risk prediction model, are significantly improved.

[0092] This method first uses a decision tree model to analyze the initial data and identify 17 data items with the highest predictive importance. Then, it uses the K-Means algorithm with modified cosine similarity as the distance metric to cluster the preprocessed data. Next, it extracts nonlinear features based on the distance to the cluster centers, and finally performs classification based on the decision tree model. This method enables intelligent identification of enterprise risks, significantly improving the accuracy of risk identification. Furthermore, the risk identification model constructed by this method is highly adaptable, allowing for continuous training with new data or data items, and it solves the overfitting problem that occurs when directly using decision tree models.

Claims

1. A method for intelligent identification of enterprise risk based on financial big data, characterized in that... The steps are as follows: 1) Determine whether it is necessary to train cluster centers and decision trees. If not, directly collect enterprise data for enterprise risk identification. If so, proceed to the next step. The method for determining whether cluster centers and decision trees need to be trained is as follows. 1-1) Setting mode=1 indicates predicting enterprise risk, distName = 'refinedCosine' indicates training cluster centers based on modified cosine similarity, trnCentre=0 indicates not retraining cluster centers, pickCentre = 1 indicates selecting fewer centers for feature extraction, and trnTree = 1 indicates retraining the decision tree. 1-2) When trnCentre = 0, refineCentre = 1, distName = 'refinedCentre', check the folder where the cluster centers are saved. If there are no trained cluster centers, then the cluster centers and the decision trees that depend on the cluster centers need to be retrained. 2) Data collection: Obtain business registration data, legal data, and category data of enterprises from the website, and store the data of deregistered enterprises for 1-2 years before deregistration in the high-risk dataset dfAcc, and store the data of enterprises that have not been deregistered for more than 5 years in the low-risk dataset dfNorm; 3) Data preprocessing: Remove invalid data that is all 0 or incomplete in the high-risk dataset dfAcc and the low-risk dataset dfNorm. Then, by performing range standardization on the business data and legal data in the two datasets respectively, the data in the two datasets are split into a high-risk business data matrix, a high-risk legal data matrix, a high-risk category data matrix, a low-risk business data matrix, a low-risk legal data matrix, and a low-risk category data matrix. The specific steps of the data preprocessing are as follows. 3-1) Delete duplicate, incomplete, all-zero, and registered capital zero data from the high-risk dataset dfAcc and the low-risk dataset dfNorm; 3-2) Horizontally split the high-risk dataset dfAcc into a high-risk non-categorical data matrix dataAcc (excluding registered capital) and a categorical data matrix TypeAcc, and horizontally split the low-risk dataset dfNorm into a low-risk non-categorical data matrix dataNorm (excluding registered capital) and a categorical data matrix TypeNorm; 3-3) Divide each element in the high-risk non-categorical data matrix dataAcc (excluding registered capital) and the low-risk non-categorical data matrix dataNorm (excluding registered capital) by the registered capital item in the same row to obtain two sets of high-risk and low-risk data matrices. 3-4) Based on the range d and minimum value m of each data item in the high-risk, non-categorical data matrix dataAcc (excluding registered capital), perform range standardization on the two sets of data matrices obtained in step 3-3), using the following formula: ； In the formula, The range standardization represents the range of a high-risk or low-risk data matrix. express j The minimum value of each data item. express j The range of each data item; 3-5) Split the range-standardized data matrix from step 3-4 according to data item categories to obtain the high-risk business data matrix. High-risk legal data matrix High-risk category data matrix Low-risk business data matrix Low-risk legal data matrix Low-risk category data matrix ; 4) Data clustering: Construct a K-Means model, train cluster centers using a high-risk business data matrix and a high-risk legal data matrix respectively, and determine the number of clusters based on the dissimilarity of the high-risk business data matrix, the high-risk legal data matrix, the low-risk business data matrix, and the low-risk legal data matrix in the clustering results. The specific methods for data clustering are as follows: 4-1) Define the range of cluster centers to be enumerated as k1 and k2, and enumerate the number of cluster centers k such that it satisfies Furthermore, k data points were randomly selected from both the high-risk business data matrix and the high-risk legal data matrix as initial cluster centers. ; 4-2) Based on modified cosine similarity as the distance metric, the distance is nonlinearly mapped to (0,1) to obtain the distance matrices for high-risk business data and high-risk legal data, respectively. The calculation formula for the distance matrix of high-risk business data is as follows: ； In the formula, distAcc This represents the distance matrix of high-risk business data. express distAcc No. i Line number j Column elements, express distAc c No. i Line number l Column elements, express dataCorpAcc No. l The average of the column, express No. j Line number l Column elements; The calculation formula for the distance matrix of high-risk legal data is the same as that for the distance matrix of high-risk business data. 4-3) Assign each point in each distance matrix to the class represented by the nearest cluster center, update the cluster center with the average value of each class, and repeat steps 3-2) and 3-3). 4-4) When the clustering situation no longer changes or the maximum number of iterations is reached, the current cluster centers are changed. Save as local optimal cluster centers ; 4-5) Obtain the classification results of the high-risk enterprise dataset and the low-risk enterprise dataset according to the local optimal cluster center cen: , This represents the number of data points belonging to the i-th class, and is stored in the high-risk dataset for the number of data points belonging to each class respectively. The number of data points belonging to each category in the low-risk dataset ; 4-6) According to , Calculate the Pearson correlation coefficient The calculation formula is as follows: ； like , The initial value is 1, which will determine the local optimal cluster centers. Save as optimal cluster center And save it as the current ; like If so, the cluster centers remain unchanged; 5) Extracting nonlinear features: 5-1) Based on the two sets of high-risk non-class datasets obtained in step 3), select four cluster centers with performance differences exceeding the threshold or the largest differences from the high-risk business data matrix and the high-risk legal data matrix, respectively. 5-2) Calculate the modified cosine similarity between each group of data and the corresponding cluster center to obtain the high-risk business data distance matrix, legal data distance matrix, and low-risk business data distance matrix, legal data distance matrix; 5-3) The distance matrix of high-risk business data, the distance matrix of legal data, and the high-risk category data matrix are horizontally merged to obtain the high-risk feature matrix chAcc and the low-risk feature matrix chNorm; 6) Construct a decision tree model based on the high-risk feature matrix chAcc and the low-risk feature matrix chNorm; When constructing the decision tree model, the Gini coefficient method is used to calculate the information gain. When selecting node parameters, the set of features with the strongest information gain is selected for branching, and the post-pruning strategy with minimum support is used for pruning. 7) Identify enterprise risks, load a decision tree model, classify the acquired nonlinear features, and obtain the results of the safety risk analysis.

2. The enterprise risk intelligent identification method based on financial big data according to claim 1, characterized in that: The business registration data includes registered capital, main operating income, total liabilities, main business revenue, net profit, total tax payment, and owner's equity. The legal data includes external guarantees, administrative penalties, external investments, plaintiff's case filing, defendant's case filing, defendant's court appearance, and judgment loss. The category data includes region, industry, and enterprise nature.

3. The enterprise risk intelligent identification method based on financial big data according to claim 1, characterized in that: In step 5-1), after traversing k, the center is saved to the corresponding text file. Since refinedCentre=1, the 8 centers with the largest difference between count1 and count2 from the optimal cluster center center are selected as pickedCentre and saved to the corresponding text file.

Images