Enterprise portrait big data analysis method

Abstract

This invention specifically relates to a big data analysis method for enterprise profiling, belonging to the field of big data analysis technology. It includes: calculating the association strength score between each associated enterprise in a set of associated enterprises and the target enterprise; the association strength score is obtained by weighted summation of three factors: personnel overlap factor, transaction closeness factor, and industry similarity factor. In this invention, by constructing a set of associated enterprises and utilizing cross-domain profiling information to complete missing dimensions, it solves the technical problem that small and micro enterprises and newly registered enterprises suffer from a large number of missing profiling dimensions due to data sparsity, and that traditional filling methods cannot effectively distinguish individual enterprise differences. By calculating the association strength score based on the weighted summation of personnel overlap, transaction closeness, and industry similarity factors, enterprises with higher association levels receive greater weight in density estimation, thus improving the rationality of cross-domain information compensation.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a big data analytics method for enterprise profiling. Background Technology

[0002] With the development of big data technology, enterprise profiling is widely used in scenarios such as financial risk control, supply chain management, and market analysis. Existing enterprise profiling big data analysis technologies typically extract features from direct data such as the enterprise's own business registration information, financial statements, and tax records, and use statistical analysis or machine learning methods to calculate profile label values ​​for each dimension; when data for certain dimensions is missing, mean imputation or imputation based on the statistical distribution of enterprises in the same industry is used.

[0003] However, the available data for micro and small enterprises and newly registered enterprises is extremely sparse, and many profile dimensions lack direct data support. The method of filling in the data based on industry averages ignores the individual differences between enterprises, leading to a homogenization of many micro and small enterprise profiles that fail to reflect their true operating characteristics. At the same time, existing methods do not effectively utilize the cross-domain relationship information contained in related entities such as legal representatives, core shareholders, and upstream and downstream partners, resulting in a significant lack of profile dimensions for micro and small enterprises and newly registered enterprises, and an inability to effectively distinguish individual enterprise differences. Summary of the Invention

[0004] The purpose of this invention is to propose a big data analysis method for enterprise profiling in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A big data analysis method for enterprise profiling includes: Obtain the basic business registration data of the target company, extract the legal representative identifier, shareholder identifier list, registered address and business scope text of the target company; traverse the tag values ​​of each dimension in the known profile tag set of the target company, identify dimensions with empty tag values ​​or marked as invalid, and generate a list of dimensions to be completed; Based on the list of legal representative and shareholder identifiers, other companies where these individuals serve as legal representatives, shareholders, or senior executives are retrieved from the enterprise association database to form a group of related enterprises. Based on the target enterprise's supply chain transaction records, upstream suppliers and downstream buyers that have transacted with the target enterprise are extracted to form a group of related enterprises. The group of related enterprises and the group of related enterprises are merged, deduplicated, and the target enterprise itself is excluded to generate a set of related enterprises. For each associated enterprise in the associated enterprise set, calculate its association strength score with the target enterprise. The association strength score is obtained by weighted summation of three factors: personnel overlap factor, transaction closeness factor, and industry similarity factor, with the sum of the three weight coefficients being one. Among them, the personnel overlap factor is the ratio of the number of common natural persons to the total number of core personnel of the target enterprise, the transaction closeness factor is the ratio of the transaction amount between the two parties to the total transaction amount of the target enterprise, and the industry similarity factor is the cosine similarity of the business scope text of the two parties after TF-IDF vectorization. For each missing dimension in the dimension completion list, related companies with valid label values ​​in the dimension are selected from the set of related companies to form a subset of related companies for the dimension. The label values ​​of each related company in the dimension are used as sample points, and the corresponding association strength scores are used as sample weights. The weighted kernel density estimation algorithm is then used to estimate the probability distribution of the label values ​​of the target company in the dimension. The mode of the probability distribution is taken as the candidate completion value, and the information entropy of the probability distribution is calculated as the completion confidence index. The candidate completion values ​​and the completion confidence index are written into the target company profile.

[0006] Preferably, the method further includes: Word embedding is performed on the target company's business scope text to obtain semantic vectors. In the full enterprise profile database, the cosine similarity between the semantic vectors and the semantic vectors of each company's business scope is calculated. The companies are sorted from high to low according to the cosine similarity, and the top K companies whose profile label values ​​in all dimensions are valid are selected to form a semantic nearest neighbor company set. The label values ​​of each dimension corresponding to the missing dimension list of each company in the semantic nearest neighbor company set are statistically analyzed to generate the industry semantic benchmark distribution of each missing dimension. Here, K is a preset retrieval quantity parameter.

[0007] Preferably, the method further includes: For each missing dimension, the industry semantic benchmark distribution is used as the prior distribution, and the probability distribution of the dimension candidate completion values ​​obtained based on the set of related enterprises is used as the likelihood function. The posterior distribution is calculated according to Bayes' theorem. The expected value of the posterior distribution is taken as the final completion label value of the dimension, and the variance of the posterior distribution is calculated as the uncertainty index. The final completion label value of each missing dimension and the corresponding uncertainty index are written into the target enterprise profile, replacing the candidate completion value written in claim 1.

[0008] Preferably, in the generated list of dimensions to be completed, an empty label value means that the dimension did not obtain any valid data source during the data collection phase; an invalid label value means that although the dimension has data records, it is marked as unusable because the data timeliness exceeds the preset period or the credibility of the data source is lower than the preset threshold.

[0009] Preferably, the extraction scope of the transaction-related enterprise group is: from the target enterprise's supply chain transaction records, select enterprises that have transaction records with the target enterprise within a preset time window, where the transaction status is completed and the transaction amount is greater than zero.

[0010] Preferably, when a related enterprise is only associated through transaction relationships and there is no personnel overlap, the personnel overlap factor of the related enterprise is zero; when a related enterprise is only associated through personnel relationships and there are no transaction records, the transaction closeness factor of the related enterprise is zero.

[0011] Preferably, the total number of core personnel is the total number of natural persons after deduplication of the target company's legal representative, all shareholders, and all senior executives.

[0012] Preferably, when the subset of associated enterprises corresponding to a missing dimension is an empty set, the dimension is not completed temporarily, and the dimension is retained in the list of dimensions to be completed and marked as pending further processing.

[0013] Preferably, the generation method of the industry semantic benchmark distribution is as follows: for the continuous label value dimension, the kernel density estimation algorithm is used to estimate the label value of each enterprise in the semantic nearest enterprise set in the dimension to obtain the probability density function; for the discrete label value dimension, the frequency of each label value appearing in the semantic nearest enterprise set is used as the probability distribution.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention solves the technical problem that small and micro enterprises and newly registered enterprises have a large number of missing profile dimensions due to their own data sparsity, and that traditional filling methods cannot effectively distinguish the differences between individual enterprises. It constructs a set of related enterprises and uses cross-domain profile information to fill in missing dimensions. Based on the three factors of personnel overlap, transaction closeness and industry similarity, the association strength score is calculated, so that enterprises with higher association degree get greater weight in density estimation, thereby improving the rationality of cross-domain information compensation.

[0015] 2. This invention integrates the candidate completion distribution of related enterprises with the industry semantic benchmark distribution using Bayesian methods. When the information of related enterprises is sufficient, personalized related information takes the lead, while when the information of related enterprises is insufficient, the industry semantic benchmark distribution is automatically relied upon. This ensures that the profile completion results remain reasonable under different data conditions. The completion results are accompanied by confidence and uncertainty indicators for downstream applications to refer to the reliability of each dimension. Attached Figure Description

[0016] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1This is a flowchart of the method of the present invention. Detailed Implementation

[0017] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0019] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0020] In this embodiment, it includes: Step 1: Obtain basic data of the target company and generate a list of dimensions to be completed. Obtain the basic business registration data of the target company, extract its legal representative identifier, shareholder identifier list, registered address, and business scope text; at the same time, obtain the known profile tag set of the target company, traverse the tag values ​​of each dimension in the known profile tag set, identify the dimensions with empty tag values ​​or marked as invalid, and compile these dimensions to generate a list of dimensions to be completed.

[0021] It should be noted that the aforementioned set of known profile tags refers to the collection of all dimension tags of the target company that have been assigned or have not yet been assigned in the multi-dimensional profile system. Profile dimensions include, but are not limited to, technological innovation capabilities, supply chain stability, actual operating scale, financial health, and market activity. An empty tag value means that no valid data source was obtained for that dimension during the data collection phase; a tag value marked as invalid means that although there are data records for that dimension, they are marked as unusable because the data quality verification failed (such as the data timeliness exceeding the preset period, the data source credibility being lower than the preset threshold, etc.).

[0022] Step 2: Construct a set of related enterprises based on personnel and transaction relationships. Based on the lists of legal representative and shareholder identifiers, other companies in which these natural persons or legal persons serve as legal representatives, shareholders, or senior executives are retrieved from the enterprise association database. The retrieved companies are then grouped into a personnel-related enterprise group. Based on the target company's supply chain transaction records, upstream suppliers and downstream buyers that have transacted with the target company are extracted. The extracted companies are then grouped into a transaction-related enterprise group. The personnel-related enterprise group and the transaction-related enterprise group are merged, duplicate companies are removed, and the target company itself is excluded to generate a set of related enterprises.

[0023] It should be noted that the search scope of the aforementioned personnel-related enterprise group refers to matching enterprise records that appear in the legal representative, shareholder, or senior management fields of any enterprise in the enterprise association database, using the individual or legal person identifiers in the legal representative and shareholder identifier lists as search criteria. When the same individual holds different roles in multiple enterprises, all of these enterprises are included in the personnel-related enterprise group.

[0024] It should be noted that the scope of the aforementioned group of related companies refers to selecting companies from the target company's supply chain transaction records that have valid transaction records with the target company within a preset time window. Valid transaction records are those with a completed transaction status and a transaction amount greater than zero.

[0025] Step 3: Calculate the association strength score between each related company and the target company, and extract the profile label values ​​of the related companies. For each associated company in the associated company set, extract its existing profile tag value corresponding to the list of dimensions to be completed; at the same time, calculate the association strength score between the associated company and the target company.

[0026] The correlation strength score is obtained by weighted summation of three factors: personnel overlap factor, transaction closeness factor, and industry similarity factor. The calculation formula is as follows: in, For the first The correlation strength score between the related companies and the target company This refers to the index of the related enterprise within the set of related enterprises. The personnel overlap factor is defined as the first... The ratio of the number of natural persons jointly owned by the related companies and the target company to the total number of core personnel of the target company, including the legal representative, shareholders and senior executives; The transaction tightness factor is defined as the first... The ratio of the transaction amount between a related company and the target company to the total transaction amount of the target company; The industry similarity factor is defined as the first... The cosine similarity between the business scope texts of the related companies and the business scope texts of the target company after TF-IDF vectorization is calculated. , , The preset weighting coefficients satisfy... .

[0027] It should be noted that the total number of core personnel mentioned above refers to the total number of natural persons after deduplication of the target company's legal representative, all shareholders, and all senior executives. When a related company is only linked through transactions and there is no overlap in personnel, the personnel overlap factor for that related company... The value is zero; when a related enterprise is only linked through personal relationships and there are no transaction records, the transaction closeness factor of that related enterprise is zero. The value is zero.

[0028] It should be noted that the above TF-IDF vectorization process refers to the process of segmenting the business scope text, calculating the term frequency-inverse document frequency value of each term, and converting the text into a high-dimensional vector with the term TF-IDF value as a component. The cosine similarity algorithm is used to calculate the similarity score, which is calculated from the two TF-IDF vectors mentioned above. The input is the similarity score, which is between zero and one.

[0029] Step 4: Based on weighted kernel density estimation, generate candidate completion values ​​for each missing dimension and write them into the target enterprise profile. For each missing dimension in the dimension completion list, related companies with valid label values ​​for that dimension are selected from the set of related companies, forming a subset of related companies for that dimension. The label values ​​of each related company in this subset are used as sample points, and the correlation strength score for each related company is calculated. As sample weights, they are input into the weighted kernel density estimation algorithm to estimate the probability distribution of the label values ​​of the target company in this dimension; the mode of this probability distribution is taken as the candidate completion value of this dimension, and the information entropy of this probability distribution is calculated as the completion confidence index; the candidate completion values ​​of each missing dimension and the corresponding completion confidence index are written into the target company profile.

[0030] It should be noted that the weighted kernel density estimation algorithm described above is a non-parametric density estimation method. Its input consists of the label value sample points of each associated enterprise in the associated enterprise subset and the corresponding association strength score in that dimension. As sample weights, the output is a probability density function over the domain of label values. Association strength score. The higher the degree of association between the companies, the greater the weight of their label values ​​in the density estimation, which makes the information of companies with a higher degree of association with the target company have a greater impact on the estimation results.

[0031] It should be noted that the above completion confidence index uses information entropy as a measure. When the label values ​​of each related enterprise in the related enterprise subset are relatively concentrated in this dimension, the information entropy of the probability distribution is low, and the corresponding completion confidence is high; when the label values ​​are relatively dispersed, the information entropy is high, and the corresponding completion confidence is low. The completion confidence index is used to identify the reliability of the completion results in each dimension.

[0032] It should be noted that when the subset of associated companies corresponding to a missing dimension is empty (i.e., no company in the set of associated companies has a valid label value for that dimension), the dimension will not be completed for the time being. Instead, the dimension will be retained in the list of dimensions to be completed and marked as pending further processing.

[0033] In this embodiment of the application, in order to obtain a reasonable completion result even when the information of the associated enterprise is insufficient, the following steps are also included in step 4: Step 5: Retrieve the semantic nearest neighbor set and generate the industry semantic benchmark distribution. Word embedding processing is performed on the business scope text of the target enterprise to convert it into a semantic vector. In the full enterprise profile database, the cosine similarity between the semantic vector and the semantic vectors of each enterprise's business scope is calculated, and the top-ranked vectors are sorted from highest to lowest cosine similarity. The complete profiles of enterprises are used to form a semantic neighbor enterprise set. The label values ​​of each enterprise in the semantic neighbor enterprise set for each dimension in the missing dimension list are statistically analyzed, and the distribution characteristics of each dimension label value are extracted to generate the industry semantic benchmark distribution for each missing dimension. Among these, This is the preset search quantity parameter.

[0034] It should be noted that the word embedding process mentioned above refers to using pre-trained word vectors to map each word in the business scope text to vectors, and then obtaining the semantic vector representation of the entire text by averaging or weighted averaging the vectors of each word.

[0035] It should be noted that the aforementioned companies with complete profiles refer to those whose profile tag values ​​for all dimensions are valid in the full enterprise profile database.

[0036] It should be noted that the aforementioned industry semantic benchmark distribution refers to the empirical distribution of label values ​​for each enterprise in a certain dimension within the set of semantically nearest enterprises. For continuous label value dimensions, the industry semantic benchmark distribution uses a kernel density estimation algorithm to obtain the probability density function. Its input is the sample points of label values ​​for each enterprise in the set of semantically nearest enterprises in that dimension, and its output is the probability density function over the domain of label values ​​in that dimension. For discrete label value dimensions, the industry semantic benchmark distribution uses the frequency of each label value as the probability distribution.

[0037] Step 6: Perform Bayesian fusion of the probability distribution of candidate completion values ​​and the industry semantic benchmark distribution to generate the final completion label value. For each missing dimension, the industry semantic benchmark distribution is used as the prior distribution, and the probability distribution of the candidate completion values ​​for that dimension obtained based on the set of related enterprises is used as the likelihood function. The posterior distribution is calculated according to Bayes' theorem. The expected value of the posterior distribution is taken as the final completion label value for that dimension, and the variance of the posterior distribution is calculated as the uncertainty index. The final completion label value and the corresponding uncertainty index of each missing dimension are written into the target enterprise profile, replacing the candidate completion values ​​written in step 4.

[0038] It should be noted that the purpose of the Bayesian fusion mentioned above is to address situations where the number of related enterprise subsets corresponding to a missing dimension is large and the association strength score is high. When the likelihood function is high, its influence on the posterior distribution is dominant, and the final completed label value is mainly determined by the information of the associated companies; when the sample size of the subset of associated companies is small or the association strength score is low... When the values ​​are generally low, the prior distribution has a dominant influence on the posterior distribution, and the final completed label values ​​tend to be closer to the industry semantic benchmark values.

[0039] It should be noted that the uncertainty index mentioned above represents the variance of the posterior distribution. A smaller variance indicates a more certain final result after fusion; a larger variance indicates a higher degree of uncertainty in the final result. The uncertainty index is written into the target company profile along with the final completed label value, providing downstream analysis with a reference for the reliability of each dimension when using the profile data.

[0040] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0041] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0042] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0043] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0044] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0045] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0046] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0047] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0048] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An enterprise portrait big data analysis method, characterized in that, include: Obtain the target company's basic business registration data, and extract the target company's legal representative identification, shareholder identification list, registered address, and business scope text; Iterate through the known profile tag set of the target company and the tag values ​​of each dimension, identify dimensions with empty tag values ​​or marked as invalid, and generate a list of dimensions to be completed; Based on the lists of legal representative and shareholder identifiers, other companies where these individuals serve as legal representatives, shareholders, or senior executives are retrieved from the enterprise association database to form a group of enterprises associated with these individuals; based on the target enterprise's supply chain transaction records, upstream suppliers and downstream buyers that have transacted with the target enterprise are extracted to form a group of enterprises associated with transactions. Merge the personnel-related enterprise group and the transaction-related enterprise group, remove duplicates, and exclude the target enterprise itself to generate a set of related enterprises; For each associated company in the associated company set, calculate its association strength score with the target company; The association strength score is obtained by weighted summation of three factors: personnel overlap factor, transaction closeness factor, and industry similarity factor, with the sum of the three weight coefficients being one. Among them, the personnel overlap factor is the ratio of the number of common natural persons to the total number of core personnel of the target company, the transaction closeness factor is the ratio of the transaction amount between the two parties to the total transaction amount of the target company, and the industry similarity factor is the cosine similarity of the business scope text of the two parties after TF-IDF vectorization. For each missing dimension in the dimension completion list, related companies with valid label values ​​in the dimension are selected from the set of related companies to form a subset of related companies for the dimension. The label values ​​of each related company in the dimension are used as sample points, and the corresponding association strength scores are used as sample weights. The weighted kernel density estimation algorithm is then used to estimate the probability distribution of the label values ​​of the target company in the dimension. The mode of the probability distribution is taken as the candidate completion value, and the information entropy of the probability distribution is calculated as the completion confidence index. The candidate completion values ​​and the completion confidence index are written into the target company profile.

2. The enterprise profiling big data analysis method according to claim 1, characterized in that, Also includes: Word embedding is performed on the business scope text of the target enterprise to obtain semantic vectors; in the full enterprise profile database, the cosine similarity between the semantic vector and the semantic vector of each enterprise's business scope is calculated, and the enterprises are sorted from high to low according to the cosine similarity. The top K enterprises whose profile label values ​​in all dimensions are valid are selected to form a set of semantically nearest enterprises. The label values ​​of each enterprise in the semantic nearest neighbor enterprise set corresponding to the dimension list to be completed are statistically analyzed to generate the industry semantic benchmark distribution of each missing dimension; where K is a preset retrieval quantity parameter.

3. The enterprise profiling big data analysis method according to claim 2, characterized in that, Also includes: For each missing dimension, the industry semantic benchmark distribution is used as the prior distribution, and the probability distribution of the dimension candidate completion values ​​obtained based on the set of related enterprises is used as the likelihood function. The posterior distribution is calculated according to Bayes' theorem. The expected value of the posterior distribution is taken as the final completion label value of the dimension, and the variance of the posterior distribution is calculated as the uncertainty index. The final completion label value of each missing dimension and the corresponding uncertainty index are written into the target enterprise profile, replacing the candidate completion value written in claim 1.

4. The enterprise profiling big data analysis method according to claim 1, characterized in that, In the list of dimensions to be completed, an empty label value means that the dimension did not obtain any valid data source during the data collection phase; an invalid label value means that although the dimension has data records, it is marked as unusable because the data timeliness exceeds the preset period or the credibility of the data source is lower than the preset threshold.

5. The enterprise profiling big data analysis method according to claim 1, characterized in that, The scope of the transaction-related enterprise group is: from the target enterprise's supply chain transaction records, select enterprises that have transaction records with the target enterprise within a preset time window, where the transaction status is "completed" and the transaction amount is greater than zero.

6. The enterprise profiling big data analysis method according to claim 1, characterized in that, When a related enterprise is only connected through transactions and there is no overlap in personnel, the personnel overlap factor of the related enterprise is zero; when a related enterprise is only connected through personnel and there are no transaction records, the transaction closeness factor of the related enterprise is zero.

7. The enterprise profiling big data analysis method according to claim 1, characterized in that, The total number of core personnel is the total number of natural persons after deduplication of the target company's legal representative, all shareholders, and all senior executives.

8. The enterprise profiling big data analysis method according to claim 1, characterized in that, When the subset of related enterprises corresponding to a missing dimension is empty, the dimension will not be completed temporarily. Instead, the dimension will be retained in the list of dimensions to be completed and marked as pending further processing.

9. The enterprise profiling big data analysis method according to claim 1, characterized in that, The industry semantic benchmark distribution is generated as follows: For the continuous label value dimension, the kernel density estimation algorithm is used to estimate the label value of each enterprise in the semantic nearest enterprise set in the dimension to obtain the probability density function; for the discrete label value dimension, the frequency of each label value in the semantic nearest enterprise set is used as the probability distribution.

Images