Enterprise qualification identification method and device based on multiple models

Abstract

The invention discloses a multi-model-based enterprise qualification identification method and device, and relates to the technical field of data processing, and the method comprises the steps: carrying out the type division of enterprises, and enabling the types of the enterprises to comprise science and technology enterprises, enterprises with the standard scale, and general enterprises; performing qualification identification on the science and technology enterprises by adopting the science and technology enterprise initial score model to obtain adjusted final qualification scores of the science and technology enterprises; a scale standard-reaching enterprise model is adopted to carry out qualification identification on the scale standard-reaching enterprises to obtain adjusted final qualification scores of the scale standard-reaching enterprises; a score card optimization model is adopted to carry out qualification identification on the general enterprise, and a final qualification score of the general enterprise is obtained; and evaluating the science and technology enterprise initial score model, the scale up-to-standard enterprise model and the general enterprise model respectively, and when evaluation results are unqualified, retraining the corresponding models. According to the method, different models are adopted for qualification judgment for different enterprises, so that the judgment accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and more specifically, to a method and device for enterprise qualification identification based on multiple models. Background Art

[0002] Currently, when banks market enterprise customers, especially small and micro enterprises, they will obtain a large number of enterprise lists from various channels. In the traditional marketing model, marketing activities are often large-scale, but lack personalization and customization, and are often restricted by factors such as information asymmetry. In the marketing stages such as customer telemarketing and door-to-door visits, a large amount of time is wasted in finding contact information and other reach links, and it is found that the customer qualifications are poor during door-to-door visit and research, resulting in a large consumption of time and human resources, a low success rate of implementation, and low marketing efficiency.

[0003] To solve this problem, some prediction models are also used to identify enterprise qualifications. However, since these models were not created specifically for the problem of enterprise qualification identification in the pre-research stage, and there are many types of enterprises, using the same set of models is likely to cause a large deviation in the identification effect for different types of enterprises. For example, with the development of technology, technology enterprises have received increasing attention. Different from traditional enterprises, the typical characteristics of technology enterprises are high-tech content, high investment, high growth, high risk, and light assets. At present, for the way of enterprise identification, the judgment criteria of traditional enterprises are still adopted, which is likely to lead to high-quality technology enterprises being identified as ordinary enterprises, thus missing high-quality customers.

[0004] Therefore, how to make appropriate qualification identification for different types of enterprises with the change of the enterprise industrial structure has become an urgent problem to be solved at present. Summary of the Invention

[0005] In view of the problems in the related art, the present invention provides a method and device for enterprise qualification identification based on multiple models to overcome the above-mentioned technical problems existing in the related art.

[0006] For this purpose, the specific technical solutions adopted by the present invention are as follows:

[0007] According to one aspect of the present invention, there is provided a method for enterprise qualification identification based on multiple models, including:

[0008] S1. Classify enterprises, and the enterprise types include technology enterprises, scale-compliant enterprises, and general enterprises;

[0009] S2. Use the initial score model for technology enterprises to identify the qualifications of technology enterprises and obtain the adjusted final qualification score of technology enterprises;

[0010] S3. Use the scale-compliant enterprise model to identify the qualifications of scale-compliant enterprises and obtain the adjusted final qualification score of scale-compliant enterprises;

[0011] S4. Use a scoring card optimization model to identify the qualifications of general enterprises and obtain the final qualification scores of general enterprises;

[0012] Using a scoring card optimization model to identify the qualifications of general enterprises includes:

[0013] Train a scoring card model based on general enterprise data with preset labels, and combine a score conversion mechanism to obtain a scoring card optimization model;

[0014] By substituting the filtered features of general enterprises into the scoring card optimization model, output the final qualification scores of the corresponding general enterprises.

[0015] Furthermore, the classification of enterprises includes:

[0016] When an enterprise obtains a science and technology qualification certification, the corresponding enterprise is classified as a science and technology enterprise; otherwise, by comparing the enterprise's scale index with the index threshold, obtain the type of the corresponding enterprise, including enterprises that meet the scale standard and general enterprises.

[0017] Furthermore, use the initial score model for science and technology enterprises to identify the qualifications of science and technology enterprises, and obtain the adjusted final qualification scores of science and technology enterprises, including:

[0018] Train a tree model based on science and technology enterprise data with preset labels to obtain an initial score model for science and technology enterprises;

[0019] By substituting the filtered features of science and technology enterprises into the initial score model for science and technology enterprises, output the initial qualification scores of the corresponding science and technology enterprises;

[0020] Use a preset risk rule to convert the initial qualification scores of science and technology enterprises to obtain the final qualification scores of science and technology enterprises;

[0021] Based on the experience database of science and technology enterprises, adjust the final qualification scores of science and technology enterprises to obtain the adjusted final qualification scores of science and technology enterprises.

[0022] Furthermore, by substituting the filtered features of science and technology enterprises into the initial score model for science and technology enterprises, output the initial qualification scores of the corresponding science and technology enterprises, including:

[0023] Obtain the importance scores of the corresponding science and technology enterprise features through the importance of science and technology enterprise features in the tree model;

[0024] Rank the science and technology enterprise features according to the importance scores, and select the science and technology enterprise features ranked within the preset ranking.

[0025] Further, by adopting the model of enterprises meeting the scale standard, the qualification identification of enterprises meeting the scale standard is carried out, and the adjusted final qualification score of enterprises meeting the scale standard includes:

[0026] Based on the data of enterprises meeting the scale standard with preset labels, the extreme gradient boosting algorithm model is trained to obtain the model of enterprises meeting the scale standard;

[0027] By substituting the screened features of enterprises meeting the scale standard into the model of enterprises meeting the scale standard, the qualification score of the corresponding enterprises meeting the scale standard is output;

[0028] Based on the experience database of enterprises meeting the scale standard, the qualification score of enterprises meeting the scale standard is adjusted to obtain the adjusted final qualification score of enterprises meeting the scale standard.

[0029] Further, the screening of the features of enterprises meeting the scale standard includes:

[0030] Based on the importance calculation formula, the importance score of the features of enterprises meeting the scale standard is calculated. The importance calculation formula is:

[0031]

[0032] In the formula, V(t) represents the importance score of the features of enterprises meeting the scale standard, t represents the node, and M represents the number of all trees;

[0033] N(m) represents the number of non-leaf nodes of the m-th tree, and β(m,i) represents the splitting feature of the i-th non-leaf node of the m-th tree;

[0034] G γ(m,i) and H γ(m,i) respectively represent the first-order derivative and the second-order derivative of all samples falling on the i-th non-leaf node of the m-th tree;

[0035] L represents the left node, R represents the right node, and λ represents the hyperparameter of the regularization term;

[0036] I represents the indicator function, G γ(m,i,L) and G γ(m,i,R) respectively represent the sum of the first-order derivatives of the left and right nodes falling on the i-th non-leaf node of the m-th tree, and H γ(m,i,L) and H γ(m,i,R) respectively represent the sum of the second-order derivatives of the left and right nodes falling on the i-th non-leaf node of the t-th tree;

[0037] Rank the features of enterprises meeting the scale standard according to the importance score, and select the features of enterprises meeting the scale standard ranked within the preset ranking.

[0038] Further, the adoption of the scoring card optimization model for the qualification identification of general enterprises also includes:

[0039] Based on the experience database of general enterprises, adjust the final qualification score of general enterprises to obtain the adjusted final qualification score of general enterprises.

[0040] Furthermore, the score conversion mechanism includes:

[0041] Calculate the score of the traditional scorecard of general enterprises, and take the average of the scores of the traditional scorecard to obtain the basic score; calculate the logistic regression results of each dimension of general enterprises, and multiply them by the weights of the corresponding dimensions to obtain the dimension scores;

[0042] Add the basic score of general enterprises and all dimension scores to obtain the final qualification score of general enterprises.

[0043] Furthermore, after identifying the qualifications of technology-based enterprises, scale-compliant enterprises and general enterprises, it also includes:

[0044] When it is necessary to adjust the output result of the model to be evaluated, increase the number of trigger mechanism enterprises of the corresponding model to be evaluated, where the models to be evaluated include the initial score model of technology-based enterprises, the scale-compliant enterprise model and the general enterprise model;

[0045] When the ratio of the number of trigger mechanism enterprises of the model to be evaluated to the number of enterprises of the corresponding enterprise type is greater than the preset value, the evaluation result of the model to be evaluated is unqualified; when the evaluation result is unqualified, retrain the corresponding model.

[0046] According to another aspect of the present invention, there is also provided an enterprise qualification identification device based on multiple models, which includes: a classification module, a technology-based enterprise identification module, a scale-compliant enterprise identification module and a general enterprise identification module.

[0047] Among them, the classification module is used to divide the types of enterprises, and the enterprise types include technology-based enterprises, scale-compliant enterprises and general enterprises. The technology-based enterprise identification module is used to adopt the initial score model of technology-based enterprises to identify the qualifications of technology-based enterprises and obtain the adjusted final qualification score of technology-based enterprises. The scale-compliant enterprise identification module is used to adopt the scale-compliant enterprise model to identify the qualifications of scale-compliant enterprises and obtain the adjusted final qualification score of scale-compliant enterprises. The general enterprise identification module is used to adopt the optimized scorecard model to identify the qualifications of general enterprises and obtain the final qualification score of general enterprises.

[0048] The beneficial effects of the present invention are:

[0049] (1) Starting from the qualification situation of enterprises, the present invention divides the evaluation of enterprises into three sub-models based on different categories of enterprises, highlighting the characteristics of different customer groups and improving the credibility of the models.

[0050] (2) In terms of the score adjustment of the experience database and the model iteration of the present invention, the algorithm model is integrated with expert experience, and the non-batch and non-public data of enterprises are integrated into the model results, further improving the accuracy of model evaluation.

[0051] (3) As a customer screening model in the marketing stage, the qualification score model implements the concept of letting data run more and letting customer managers run less. Especially in the marketing stage of small and micro batch customer groups, in the face of various massive enterprise lists, it helps customer managers to stratify and classify customer groups, screen out target enterprise customer groups that match their own marketing qualifications, give marketing priorities, and improve marketing efficiency and implementation rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0053] Figure 1 It is a flowchart of the enterprise qualification recognition method based on multiple models according to an embodiment of the present invention;

[0054] Figure 2 It is a block diagram of the modules of the enterprise qualification recognition device based on multiple models according to an embodiment of the present invention;

[0055] Figure 3 It is a model classification diagram according to an embodiment of the present invention;

[0056] Figure 4 It is a flowchart in specific applications according to an embodiment of the present invention.

[0057] In the figure:

[0058] 1. Classification module; 2. Technology-based enterprise identification module; 3. Scale-compliant enterprise identification module; 4. General enterprise identification module; 5. Model evaluation module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] To further illustrate the embodiments, the present invention provides drawings. These drawings are part of the disclosure of the present invention, mainly used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, those of ordinary skill in the art should be able to understand other possible implementation manners and the advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are usually used to represent similar components.

[0060] According to an embodiment of the present invention, there are provided an enterprise qualification recognition method and device based on multiple models.

[0061] The present invention will be further described in conjunction with the accompanying drawings and specific embodiments. As Figure 1 shown, according to an embodiment of the present invention, a method for identifying enterprise qualifications based on multiple models is provided, including:

[0062] S1. Classify the enterprises, and the enterprise types include technology-based enterprises, scale-compliant enterprises, and general enterprises.

[0063] S2. Adopt the initial score model for technology-based enterprises to identify the qualifications of technology-based enterprises and obtain the adjusted final qualification scores of technology-based enterprises.

[0064] S3. Adopt the scale-compliant enterprise model to identify the qualifications of scale-compliant enterprises and obtain the adjusted final qualification scores of scale-compliant enterprises.

[0065] S4. Adopt the scoring card optimization model to identify the qualifications of general enterprises and obtain the final qualification scores of general enterprises.

[0066] Adopting the scoring card optimization model to identify the qualifications of general enterprises includes:

[0067] Based on the general enterprise data with preset labels, train the scoring card model, and combine the score conversion mechanism to obtain the scoring card optimization model.

[0068] By substituting the screened features of general enterprises into the scoring card optimization model, output the final qualification scores of the corresponding general enterprises.

[0069] In one embodiment, classifying the enterprises includes:

[0070] When an enterprise obtains a technology-based qualification certification, classify the corresponding enterprise as a technology-based enterprise; otherwise, compare the scale indicators of the enterprise with the indicator thresholds to obtain the types of the corresponding enterprises, including scale-compliant enterprises and general enterprises.

[0071] In one embodiment, adopting the initial score model for technology-based enterprises to identify the qualifications of technology-based enterprises and obtain the adjusted final qualification scores of technology-based enterprises includes:

[0072] Based on the technology-based enterprise data with preset labels, train the tree model to obtain the initial score model for technology-based enterprises.

[0073] By substituting the screened features of technology-based enterprises into the initial score model for technology-based enterprises, output the initial qualification scores of the corresponding technology-based enterprises.

[0074] Utilize the preset risk rules to convert the initial qualification scores of technology-based enterprises to obtain the final qualification scores of technology-based enterprises.

[0075] Based on the experience database of technology-based enterprises, adjust the final qualification score of technology-based enterprises to obtain the adjusted final qualification score of technology-based enterprises.

[0076] In one embodiment, by substituting the filtered features of technology-based enterprises into the initial score model of technology-based enterprises, the output initial qualification scores of corresponding technology-based enterprises include:

[0077] Obtain the importance score of the corresponding technology-based enterprise features through the importance of technology-based enterprise features in the tree model.

[0078] Rank the technology-based enterprise features according to the importance score, and select the technology-based enterprise features ranked within the preset ranking.

[0079] In one embodiment, adopt the scale-compliant enterprise model to identify the qualifications of scale-compliant enterprises, and the adjusted final qualification scores of scale-compliant enterprises obtained include:

[0080] Based on the scale-compliant enterprise data with preset labels, train the extreme gradient boosting algorithm model to obtain the scale-compliant enterprise model.

[0081] By substituting the filtered features of scale-compliant enterprises into the scale-compliant enterprise model, output the qualification scores of corresponding scale-compliant enterprises.

[0082] Based on the experience database of scale-compliant enterprises, adjust the qualification scores of scale-compliant enterprises to obtain the adjusted final qualification scores of scale-compliant enterprises.

[0083] In one embodiment, the screening of the features of scale-compliant enterprises includes:

[0084] Based on the importance calculation formula, calculate the importance scores of the features of scale-compliant enterprises. The importance calculation formula is:

[0085]

[0086] In the formula, V(t) represents the importance score of the features of scale-compliant enterprises, t represents the node, M represents the number of all trees; N(m) represents the number of non-leaf nodes of the m-th tree, and β(m,i) represents the splitting feature of the i-th non-leaf node of the m-th tree; G γ(m,i) and H γ(m,i) respectively represent the first-order derivative and the second-order derivative of all samples falling on the i-th non-leaf node of the m-th tree; L represents the left node, R represents the right node, and λ represents the hyperparameter of the regularization term; I represents the indicator function, G γ(m,i,L) and G γ(m,i,R) respectively represent the sum of the first-order derivatives of the left and right nodes falling on the i-th non-leaf node of the m-th tree, and H γ(m,i,L) and Hγ(m,i,R) respectively represent the sum of the second-order derivatives of the left and right nodes of the i-th non-leaf node on the t-th tree.

[0087] Rank the characteristics of enterprises meeting the scale standard according to the importance score, and select the characteristics of enterprises meeting the scale standard within the preset ranking.

[0088] In one embodiment, using the scoring card optimization model, the qualification identification of general enterprises further includes:

[0089] Based on the experience database of general enterprises, adjust the final qualification score of general enterprises to obtain the adjusted final qualification score of general enterprises.

[0090] In one embodiment, the score conversion mechanism includes:

[0091] Calculate the traditional scoring card score of general enterprises, and take the average of the traditional scoring card scores to obtain the basic score; calculate the logistic regression results of each dimension of general enterprises, and multiply them by the weights of the corresponding dimensions to obtain the dimension scores;

[0092] Add the basic score of general enterprises to all dimension scores to obtain the final qualification score of general enterprises.

[0093] In one embodiment, after the qualification identification of technology-based enterprises, enterprises meeting the scale standard and general enterprises, it further includes:

[0094] When it is necessary to adjust the output result of the model to be evaluated, increase the number of trigger mechanism enterprises of the corresponding model to be evaluated, where the models to be evaluated include the initial score model of technology-based enterprises, the model of enterprises meeting the scale standard, and the model of general enterprises.

[0095] When the ratio of the number of trigger mechanism enterprises of the model to be evaluated to the number of enterprises of the corresponding enterprise type is greater than the preset value, the evaluation result of the model to be evaluated is unqualified; when the evaluation result is unqualified, retrain the corresponding model.

[0096] As Figure 2 shown, according to another embodiment of the present invention, there is also provided an enterprise qualification identification device based on multiple models, and the device includes: a classification module 1, a technology-based enterprise identification module 2, an enterprise identification module 3 meeting the scale standard, and a general enterprise identification module 4.

[0097] The classification module 1 is used to divide the types of enterprises, and the enterprise types include technology-based enterprises, enterprises meeting the scale standard, and general enterprises.

[0098] The technology-based enterprise identification module 2 is used to adopt the initial score model of technology-based enterprises to identify the qualifications of technology-based enterprises and obtain the adjusted final qualification score of technology-based enterprises.

[0099] The scale-compliant enterprise identification module 3 is used to adopt the scale-compliant enterprise model to perform qualification identification 4 on the scale-compliant enterprises and obtain the adjusted final qualification score of the scale-compliant enterprises.

[0100] The general enterprise identification module is used to identify the qualifications of general enterprises using a scoring card optimization model to obtain the final qualification score of the general enterprises.

[0101] In order to facilitate understanding of the above technical solutions of the present invention, the working principle of the present invention in the actual process is described in detail below. Figures 3-4 shown.

[0102] Step 1: Based on the acquired enterprise data, enterprises are divided into technology enterprises, non-small enterprises and general enterprises.

[0103] The qualification model can be divided into three sub-models according to the type of predicted enterprise, namely the technology enterprise model, the non-small enterprise model and the general enterprise model. The latter two are collectively referred to as the non-technology enterprise model. Since different types of enterprises have different group characteristics, the three models differ in input indicators and identification. If the enterprise meets the technology qualification certification, the technology enterprise model is applicable. If the enterprise lacks technology qualification certification, it continues to determine whether the scale of the enterprise is a non-small and micro enterprise. If the enterprise is a non-small and micro enterprise, the non-small enterprise model, that is, the scale-compliant enterprise model, is applicable. Otherwise, the general enterprise model is applicable.

[0104] The standards for identifying small and micro enterprises are as follows:

[0105] If the enterprise is an industrial enterprise, the standards are: annual taxable income does not exceed 300,000 yuan; the number of employees does not exceed 100; the total assets do not exceed 30 million yuan. If the enterprise is other enterprises, the standards are: annual taxable income does not exceed 300,000 yuan; the number of employees does not exceed 80; the total assets do not exceed 10 million yuan.

[0106] Step 2: Identify the enterprise qualifications under the set type.

[0107] Step 21: For technology-related enterprises, a pre-trained technology-related enterprise tree model is used to identify their qualifications; the technology-related enterprise tree model is determined through training iterations based on a gradient-boosted decision tree algorithm.

[0108] If the enterprise is a technology enterprise, the enterprise qualification assessment is carried out through the preset technology enterprise tree model to generate the enterprise qualification score. Among them, the preset technology enterprise tree model is based on external public data and is obtained by training and iterating using the decision tree algorithm based on gradient boosting. The specific method is:

[0109] Step 211: Clean, preprocess, and perform feature engineering on the original enterprise data. Among them, feature engineering is the logical processing of indicators based on experience. Traditional models for evaluating enterprise qualifications usually cover dimensions such as enterprise scale, business stability, risk performance, and finance, which are more inclined to the fundamental evaluation of mature enterprises. However, the core competitiveness of technology-based enterprises lies in their technological innovation capabilities. Therefore, the innovation dimension and growth dimension are added on the basis of the original evaluation dimensions, and in-depth mining of indicators is carried out: at the indicator level, it includes the number of patents, the proportion of invention patents, the number of intellectual property rights under application, which reflect the quantity and quality of the enterprise's intellectual property rights; the proportion of R & D personnel, which reflects the construction of the talent team; the growth rate of the number of enterprise insured persons and the proportion of registered capital changes, which reflect the growth of the enterprise.

[0110] The feature dimensions involved in the identification of technology-based enterprises include the enterprise's finance and taxation, innovation, stability, growth, qualifications, and performance. Among them, the selected indicators are shown in Table 1.

[0111] Table 1 Selected indicators of technology-based enterprises

[0112]

[0113]

[0114] Step 212: Select the preset labels of the training model for technology-based enterprises.

[0115] The preset label is the good sample of the training sample, defined as the enterprise that obtains equity financing: an enterprise invested by professional primary market investment and financing institutions such as venture capital institutions and risk investment institutions, or a listed company.

[0116] Step 213: Determine the feature importance of the indicators, sort them according to the indicator importance, and select the features ranked within the preset ranking based on the ranking.

[0117] The importance of the technology-based enterprise tree model is determined by the following method:

[0118] The global importance of feature j is measured by the average of the importance of feature j in a single tree:

[0119]

[0120] Among them, M represents the number of trees, represents the importance of feature j in the m-th tree; the importance of feature j in a single tree is as follows:

[0121]

[0122] Among them, L represents the number of leaf nodes of the tree, v t represents the feature associated with node t, Represents the reduction value of the squared loss after node t splits.

[0123] The greater the importance, the greater the ability of this feature to reduce loss, that is, the better the performance of the indicator in the model.

[0124] Step 214: Train the tree model for technology-based enterprises and solidify it into the initial score model for technology-based enterprises.

[0125] Step 215: Use the indicators of the enterprise as features, i.e., the input items of the tree model, and input them into the initial score model to obtain the initial qualification score of the technology-based enterprise.

[0126] Step 216: Based on preset risk and other rules, adjust and transform the initial qualification score of the technology-based enterprise to obtain the final qualification score Score_0 of the technology-based enterprise 1 , and the score range is 0 - 100 points.

[0127] Step 217: Use the experience database to adjust the final qualification score of the technology-based enterprise. Since the data source of the initial model is pure external data, and external data is difficult to fully reflect the characteristics of the enterprise. According to experience, for example, during on-site visits by account managers, internal information can be obtained through communication with enterprise leaders or financial personnel, obtaining the current actual market competition position and future development potential of the enterprise. Therefore, an experience database is established. The adjustment items include too high e1, too low e2, accurate e3. If the score is considered too high, then select too high; if the score is considered too low, then select too low. Assume that the number of trigger mechanisms for considering a certain enterprise's score as too high (i.e., clicking e1) is a1, and the number of trigger mechanisms for considering it as too low (i.e., clicking e2) is a2. When the number of a1 or a2 reaches the threshold A0, then adjust the enterprise qualification score of the technology-based enterprise:

[0128] Score_adj 1 = Score_0 1 - a1, where a1 ≤ A0, if a1 > A0 then

[0129] Score_adj 1 = Score_0 1 - A0; if Score_adj 1 < 0, then take 0.

[0130] Score_adj 1 = Score_0 1 + a2, where a2 <= A0, if a2 > A0 then

[0131] Score_adj 1 = Score_0 1 + A0; if Score_adj 1If it is > 100, then take 100.

[0132] The experience database includes the number of each enterprise obtaining three types of labels. Taking the number of high - value labels as an example: Suppose the number of enterprises in the experience database is 100, and the number of enterprises obtaining high - value labels is 20. Among them, the number of high - value labels obtained by each enterprise can be 1, 2, 3, 4, etc. Take the 75th percentile of the number of high - value labels of these 20 enterprises as the threshold of the number of this label.

[0133] Step 218: Set the model accuracy monitoring index.

[0134] M = (the number of enterprises triggered by the trigger mechanism) / (the number of enterprises visited by the customer manager) * 100%.

[0135] When M > M0, the model is re - iteratively trained to form a new model. M0 is the monitoring index threshold for model trigger iteration in the experience database, that is, the 75th percentile threshold of the number of high - value / low - value enterprises / total number of enterprises mentioned above.

[0136] Step 22: For non - small enterprises, conduct enterprise qualification determination through a preset non - small enterprise model; the preset non - small enterprise model is determined by using the XGBoost algorithm (extreme gradient boosting algorithm).

[0137] If the enterprise is a non - technology enterprise and its scale is not a small and micro enterprise, then conduct enterprise qualification assessment through a preset non - small enterprise model to generate an enterprise qualification score. Among them, the preset non - small enterprise model is obtained by using the XGBoost algorithm based on public external data, and the data involved covers the enterprise's industry and commerce, intellectual property, judiciary, operation, public opinion and other data, and judges the enterprise qualification from four dimensions: scale, stability, innovation, and growth.

[0138] The specific method is as follows:

[0139] Step 221: Conduct data cleaning, data pre - processing and feature engineering processing on the original data of non - small enterprises, and the screened indicators are shown in Table 2.

[0140] Table 2 Screened indicators of non - small enterprises

[0141]

[0142]

[0143] Step 222: Select the preset labels of the training model for non - technology non - small enterprises.

[0144] The preset tags are training samples, including listed companies excluding ST and financial industry enterprises; among them, the Shanghai and Shenzhen Stock Exchanges conduct special treatment (Special treatment) on the stock trading of listed companies with abnormal financial conditions or other conditions, and prefix ST before the abbreviation, so such stocks are called ST stocks.

[0145] Step 223: Determine the feature importance of non-small enterprise indicators, sort them according to importance, and select features based on the ranking.

[0146] The importance is determined according to the following method:

[0147]

[0148] Among them, V(t) represents the importance score of the characteristics of enterprises that meet the scale standard, t represents the node, M represents the number of all trees; N(m) represents the number of non-leaf nodes of the m-th tree, and β(m,i) represents the division feature of the i-th non-leaf node of the m-th tree; G γ(m,i) and H γ(m,i) respectively represent the first-order derivative and the second-order derivative of all samples falling on the i-th non-leaf node of the m-th tree; L represents the left node, R represents the right node, and λ represents the hyperparameter of the regularization term; I represents the indicator function, G γ(m,i,L) and G γ(m,i,R) respectively represent the sum of the first-order derivatives of the left and right nodes falling on the i-th non-leaf node of the m-th tree, and H γ(m,i,L) and H γ(m,i,R) respectively represent the sum of the second-order derivatives of the left and right nodes falling on the i-th non-leaf node of the t-th tree. The greater the importance, the better the performance of this indicator.

[0149] Step 224: Based on the selected features and the corresponding training data, train the non-small enterprise tree model and solidify it into the initial non-small enterprise model.

[0150] Step 225: Through the non-small enterprise model, use the characteristics of non-small enterprises as input items to obtain the qualification score Score_0 2 , and the score range is 0-100 points.

[0151] Step 226: Use the experience database to adjust the final qualification score of non-small enterprises. The adjustment items include too high e1, too low e2, and accurate e3. If it is considered that the score is too high, select too high; if the score is too low, select too low.

[0152] Suppose the number of trigger mechanisms for which a certain enterprise's score is considered too high (i.e., click e1) is a1, and the number of trigger mechanisms for which the score is considered too low (i.e., click e2) is a2. When the number of a1 or a2 reaches the threshold A0, then adjust the enterprise qualification score of non-small enterprises:

[0153] Score_adj 2 =Score_0 2 -a1, where a1<=A0, if a1>A0 then

[0154] Score_adj 2 =Score_0 2 -A0; if Score_adj 2 <0, then take 0.

[0155] Score_adj 2 =Score_0 2 +a2, where a2<=A0, if a2>A0 then

[0156] Score_adj 2 =Score_0 2 +A0; if Score_adj 2 >100, then take 100.

[0157] Step 227: Set the model accuracy monitoring indicator:

[0158] M = number of enterprises that trigger the mechanism / number of enterprises that the account manager visits*100%.

[0159] When M>M0, the model is re-trained iteratively to form a new model.

[0160] Step 23: For other enterprises, enterprise qualifications are identified through the preset general enterprise model; wherein, the enterprise model is optimized by using a scoring card.

[0161] If the enterprise is neither a technology-qualified enterprise nor a non-small enterprise, the general enterprise model is applicable. At this point, the forecast portfolio has covered all external industrial and commercial enterprises. Since general enterprises are aimed at non-technology-type small and micro enterprises, the nature of enterprises is wide, and the scorecard model is more conducive to exploring the characteristics of different enterprise groups and generating enterprise qualification scores.

[0162] The traditional scoring card model is a method of determining the score based on the coefficient of logistic regression and WOE (Weight of Evidence), which converts the logistic regression model into a standard score. However, the application optimizes the algorithm in the score conversion stage of the score card. The score conversion optimization algorithm can highlight the score of each dimension of the quantitative enterprise, so as to provide different strategic solutions for enterprises with different characteristics.

[0163] The preset scorecard optimization model is a scorecard sub-model obtained by iterative training based on data with preset labels, and then a scorecard optimization model is generated through a score conversion formula.

[0164] The specific qualification score evaluation method is as follows:

[0165] Step 231: Clean, preprocess and perform feature engineering on the original enterprise data, evaluate the enterprise qualification from six dimensions of finance and taxation, growth, innovation, scale, performance and stability, perform index processing, and the screened indexes are shown in Table 3.

[0166] Table 3 Screened indexes of general enterprises

[0167]

[0168]

[0169] Step 232: Select the preset labels of the training model for non-tech small and micro enterprises; the preset label is enterprises above designated size. Enterprises above designated size refer to those enterprises whose main business income, sales volume or operating income and other indicators reach a certain scale standard within a certain period (usually one year). The recognition standards for enterprises above designated size in different industries are different, as follows:

[0170] Industry: Industrial legal persons with an annual main business income of 20 million yuan or more.

[0171] Construction: Legal persons with an annual operating income of 300 million yuan or more. Qualified construction legal persons are enterprises above designated size.

[0172] Wholesale and retail: Legal persons with an annual sales volume exceeding 100 million yuan or an operating income exceeding 5 million yuan. Wholesale enterprises with an annual commodity sales volume of 20 million yuan or more and retail enterprises with an annual commodity sales volume of 5 million yuan or more are enterprises above designated size.

[0173] Accommodation and catering: Legal persons with an annual turnover exceeding 2 million yuan. In addition, legal persons of accommodation and catering enterprises above the quota also belong to enterprises above designated size.

[0174] Transportation, warehousing and postal services: Legal persons with an annual operating income exceeding 20 million yuan.

[0175] Service industry: Service enterprises with an annual main business income of 10 million yuan or more. Legal persons of service enterprises above designated size are enterprises above designated size.

[0176] Step 233: Determine the regression coefficient ranking of the indexes and select features.

[0177] Step 234: Based on the selected features and the corresponding training data, train the scoring card model to obtain the scoring card model result.

[0178] Step 235: Using the general model, take the characteristics of the enterprise as input items to obtain the general initial qualification score Score_0 of the enterprise, with a score range of 0 - 100 points.

[0179] Step 236: Based on the preset risk and other rules, adjust and optimize the initial qualification score of the enterprise to obtain the final qualification score of the enterprise.

[0180] The specific score conversion and optimization algorithm is as follows:

[0181] First, calculate the traditional scorecard score of the enterprise, denoted as score0.

[0182] Classify the model indicators into six dimensions: finance and taxation, growth, innovation, scale, performance, and stability. The indicators for each dimension are The corresponding logistic regression coefficients are Then the logistic regression result for each dimension is:

[0183]

[0184] Use the CRITIC weighting method to give the weight w for each dimension i , i = 1, …, 6.

[0185] Among them, C j = S j × R j ,

[0186] p represents the number of indicators within the i-th dimension; S j represents the standard deviation of the j-th indicator, reflecting the volatility of the indicator; r ij represents the correlation coefficient between the i-th indicator and the j-th indicator; R j represents the sum of (1 - r ij ) for all indicators within the i-th dimension, reflecting the conflict between indicators; C j represents the information content of the j-th indicator, expressed as the product of volatility and conflict.

[0187] Determine the final qualification score of the enterprise = basic score + dimension score:

[0188]

[0189] Among them, is the mean of the standard scores obtained from the traditional logistic regression scorecard of the sample customer group, serving as the basic score for optimizing the model score.

[0190] Step 237: Adjust the final qualification score using the experience database. The adjustment items include being on the high side e1, being on the low side e2, and being accurate e3. If it is considered that the score is on the high side, then select being on the high side; if the score is on the low side, then select being on the low side.

[0191] Suppose the number of trigger mechanisms for which it is considered that a certain enterprise's score is on the high side (i.e., clicking e1) is a1, and the number of trigger mechanisms for which it is considered that the score is on the low side (i.e., clicking e2) is a2. When the number of a1 or a2 reaches the threshold A0, then adjust the enterprise qualification score of the general enterprise:

[0192] Score_adj 3 = Score - a1, where a1 <= A0. If a1 > A0, then Score_adj 3 = Score - A0;

[0193] If Score_adj 3 < 0, then take 0.

[0194] Score_adj 3 = Score + a2, where a2 <= A0. If a2 > A0, then Score_adj 3 = Score + A0;

[0195] If Score_adj 3 > 100, then take 100.

[0196] Step 238: Set the model accuracy monitoring index.

[0197] M = (the number of enterprises with trigger mechanisms) / (the number of enterprises visited by customer managers) * 100%.

[0198] When M > M0, the model is re-iteratively trained to form a new model.

[0199] Step 3: Transmit the final qualification score to the downstream application end and perform model iteration according to the application effect.

[0200] The present invention also provides an enterprise qualification recognition device based on multiple models. The device includes:

[0201] A classification module, configured to classify enterprises into technology-based enterprises, non-small enterprises, and general enterprises according to the acquired enterprise data.

[0202] An identification module, configured to respectively perform enterprise qualification identification on enterprise qualifications under the set types; the identification module includes a technology-based enterprise identification module, a non-small enterprise identification module, and a general enterprise identification module.

[0203] A technology enterprise identification module is used to identify the qualifications of technology enterprises by using a pre-trained technology enterprise tree model; the technology enterprise tree model is determined through training and iteration based on the gradient-boosted decision tree algorithm.

[0204] A non-small enterprise identification module is used to identify the enterprise qualifications of non-small enterprises through a preset non-small enterprise model; the preset non-small enterprise model is determined by using the XGBoost algorithm.

[0205] A general enterprise identification module is used to identify the enterprise qualifications of other enterprises through a preset general enterprise model; among them, the general enterprise model adopts the method of scorecard optimization.

[0206] It can be seen from the above technical solutions that:

[0207] The qualification score model can be classified into three sub-models according to different enterprise types, namely the technology enterprise model, the non-small enterprise model and the general enterprise model. The latter two are collectively referred to as the non-technology enterprise model. If an enterprise has certain technology-related qualification identifications, the technology enterprise model is applicable. If an enterprise lacks technology-related qualification identifications, it continues to judge whether the scale of the enterprise is a non-small and micro enterprise. If the enterprise is a non-small and micro enterprise, the non-small enterprise model is applicable, otherwise the general enterprise model is applicable. The specific implementation of the present invention includes:

[0208] 1. Obtain enterprise data.

[0209] 2. Judge the enterprise type.

[0210] Judge whether the enterprise has technology-related qualification identifications and whether it is a small and micro enterprise.

[0211] 3. Based on the enterprise data and the enterprise type, evaluate the enterprise qualifications through a preset sub-model and output the qualification score.

[0212] 1) Technology enterprise model

[0213] If the enterprise is a technology enterprise, the enterprise qualifications are evaluated through a preset technology enterprise tree model to generate the enterprise qualification score. Among them, the preset technology enterprise tree model is obtained through training and iteration based on data with preset labels, and the data involved covers the enterprise's industry and commerce, qualifications, intellectual property rights, judicature, operations, investment and financing, etc.

[0214] The specific method is as follows:

[0215] Perform data cleaning, data preprocessing and feature engineering on the enterprise original data.

[0216] Select the preset labels of the training model for technology enterprises.

[0217] Determine the feature importance of data metrics, sort them by importance, and select several features based on the ranking.

[0218] Based on the selected several features and the corresponding training data, train a tree model and solidify it into an initial scoring model.

[0219] Through the above model, use the features of the enterprise as input items to obtain the initial qualification score of the enterprise.

[0220] Based on preset risk and other rules, adjust and transform the initial qualification score of the enterprise to obtain the final qualification score of the enterprise.

[0221] Transmit the model results to the downstream application side, and the customer manager evaluates the customer during the actual visit to the customer to form feedback on the accuracy of the front-end score

[0222] When the model feedback volume and the inaccurate rate of the feedback results reach a certain threshold, the back-end model will adjust the preset label customer group range and then perform iterative training on the model to form a new model.

[0223] 2) Non-small enterprise model

[0224] If the enterprise is a non-technology enterprise and the enterprise scale is a non-small micro enterprise, then use a preset non-small enterprise tree model to evaluate the enterprise qualification and generate the enterprise qualification score. The preset non-small enterprise tree model is obtained through training and iteration based on data with preset labels, and the data involved covers the enterprise's industry and commerce, qualifications, intellectual property, judiciary, operations, investment and financing, etc.

[0225] The specific method is as follows:

[0226] Perform data cleaning, data preprocessing, and feature engineering on the enterprise's original data.

[0227] Select the training model preset labels for non-technology non-small enterprises.

[0228] Determine the feature importance of data metrics, sort them by importance, and select several features based on the ranking.

[0229] Based on the selected several features and the corresponding training data, train a tree model and solidify it into a model.

[0230] Through the above model, use the features of the enterprise as input items to obtain the qualification score of the enterprise.

[0231] Transmit the model results to the downstream application side, and the customer manager evaluates the customer during the actual visit to the customer to form feedback on the accuracy of the front-end score

[0232] When the model feedback volume and the feedback result inaccuracy rate reach a certain threshold, the backend model will adjust the preset label customer group range and then perform iterative training on the model to form a new model.

[0233] 3) General enterprise model

[0234] The enterprise qualification is evaluated by optimizing the model through a preset technology enterprise scoring card to generate an enterprise qualification score. The preset scoring card optimization model is a scoring card sub-model obtained through training and iteration based on data with preset labels, and then the scoring card optimization model generated through a score conversion formula. The data involved covers the industry and commerce, qualifications, intellectual property rights, judicature, operations, investment and financing, etc. of the enterprise.

[0235] The specific method is as follows:

[0236] Perform data cleaning, data preprocessing, and feature engineering on the original enterprise data.

[0237] Select the preset labels of the training model for general enterprises.

[0238] Determine the feature importance of the data indicators, sort them according to the importance, and select several features based on the ranking.

[0239] Based on the selected several features and the corresponding training data, train the scoring card model, and convert the result of the scoring card model through a preset score conversion formula to obtain the general enterprise qualification score and solidify it into an initial score model.

[0240] Using the above model, take the features of the enterprise as input items to obtain the initial qualification score of the enterprise.

[0241] Based on the preset risk rules, adjust and convert the initial qualification score of the enterprise to obtain the final qualification score of the enterprise.

[0242] The model result is transmitted to the downstream application side, and the customer manager evaluates the customer during the actual visit to the customer to form a feedback on the accuracy of the front-end score.

[0243] When the model feedback volume and the feedback result inaccuracy rate reach a certain threshold, the backend model will adjust the preset label customer group range and then perform iterative training on the model to form a new model.

[0244] IV. Output the enterprise qualification score and transmit it to the downstream application side.

[0245] In addition, the present invention further includes:

[0246] 1) Technology-based enterprises cover a wide range of industries with numerous and detailed field divisions, such as new materials, new energy, biomedicine, new-generation information technology, artificial intelligence, etc. At the same time, technology innovation enterprises often exhibit characteristics such as intellectual property intensity, technology specialization, and research frontierization. Especially for start-up enterprises, there may be uncertainties in the transformation of scientific and technological achievements. Simply using a scoring model where experts evaluate each enterprise one by one is not only inefficient but also difficult to control the consistency of the rules measured by manual judgment, resulting in high costs. The model focuses on examining the technology innovation attributes of enterprises. For enterprises that trigger preset risk matters of technology-based enterprises, through additional adjustment and conversion of the qualification scores of the enterprises, risk warnings are given to users.

[0247] 2) The scale, enterprise qualification, and risk tolerance of non-small enterprises are quite different from those of small and micro enterprises. The present invention separately models and evaluates this customer group to ensure the credibility of the enterprise qualification assessment metrics.

[0248] 3) The general enterprise model is the underlying model of the model. Compared with the technology-based enterprise model, it examines more of the basic business operations and competitive qualifications of enterprises rather than technology innovation attributes, etc.

[0249] Due to considering the interpretability of the model, the current enterprise evaluation models in the banking industry will preferentially choose the widely used classic scoring card model. The qualification score model selects the tree model in the technology-based enterprise model and the non-small model. The tree model is data-driven. By learning a large amount of data, it discovers the hidden laws and characteristics behind the data, without the need for manual analysis and rule construction like expert rules. Moreover, the machine learning method has a high degree of automation and can automatically perform processes such as feature extraction, model selection, and parameter optimization, reducing manual intervention and costs. The model of the present invention is a daily batch model, ensuring the timeliness and accuracy of enterprise qualification evaluation.

[0250] The main pain point solved by the qualification score model is that before the customer manager conducts batch enterprise marketing, the list of enterprises in hand can be pre-screened according to the qualification scores of the enterprises, and the target marketing customer groups can be preferentially selected to achieve hierarchical marketing, efficient marketing, and precise marketing.

[0251] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An enterprise qualification recognition method based on multiple models, characterized in that Including: S1. Classify enterprises into different types, and the enterprise types include technology-based enterprises, enterprises meeting the scale standard, and general enterprises; S2. Use the initial score model for technology-based enterprises to identify the qualifications of technology-based enterprises and obtain the adjusted final qualification score of technology-based enterprises; S3. Use the model for enterprises meeting the scale standard to identify the qualifications of enterprises meeting the scale standard and obtain the adjusted final qualification score of enterprises meeting the scale standard; S4. Use the optimized scoring card model to identify the qualifications of general enterprises and obtain the final qualification score of general enterprises; Using the optimized scoring card model to identify the qualifications of general enterprises includes: Train the scoring card model based on the general enterprise data with preset labels and combine the score conversion mechanism to obtain the optimized scoring card model; Substitute the screened features of general enterprises into the optimized scoring card model and output the final qualification score corresponding to the general enterprises.

2. The enterprise qualification recognition method based on multiple models according to claim 1, wherein The classification of enterprises includes: When an enterprise obtains a technology-based qualification certification, the corresponding enterprise is classified as a technology-based enterprise; otherwise, compare the scale index of the enterprise with the index threshold to obtain the type of the corresponding enterprise, including enterprises meeting the scale standard and general enterprises.

3. The enterprise qualification recognition method based on multiple models according to claim 1, wherein The use of the initial score model for technology-based enterprises to identify the qualifications of technology-based enterprises and obtain the adjusted final qualification score of technology-based enterprises includes: Train the tree model based on the technology-based enterprise data with preset labels to obtain the initial score model for technology-based enterprises; Substitute the screened features of technology-based enterprises into the initial score model for technology-based enterprises and output the initial qualification score corresponding to the technology-based enterprises; Use the preset risk rules to convert the initial qualification score of technology-based enterprises to obtain the final qualification score of technology-based enterprises; Based on the experience database of technology-based enterprises, adjust the final qualification score of technology-based enterprises to obtain the adjusted final qualification score of technology-based enterprises.

4. The enterprise qualification recognition method based on multiple models according to claim 3, wherein The substitution of the screened features of technology-based enterprises into the initial score model for technology-based enterprises and the output of the initial qualification score corresponding to the technology-based enterprises includes: Obtain the importance score of the corresponding technology-based enterprise features through the importance of technology-based enterprise features in the tree model; Rank the technology-based enterprise features according to the importance score and select the technology-based enterprise features ranked within the preset ranking.

5. The enterprise qualification recognition method based on multiple models according to claim 1, wherein, The use of the model for enterprises meeting the scale standard to identify the qualifications of enterprises meeting the scale standard and obtain the adjusted final qualification score of enterprises meeting the scale standard includes: Train the extreme gradient boosting algorithm model based on the data of enterprises meeting the scale standard with preset labels to obtain the model for enterprises meeting the scale standard; Substitute the screened features of enterprises meeting the scale standard into the model for enterprises meeting the scale standard and output the qualification score corresponding to the enterprises meeting the scale standard; Based on the experience database of enterprises meeting the scale standard, adjust the qualification score of enterprises meeting the scale standard to obtain the adjusted final qualification score of enterprises meeting the scale standard.

6. The method for identifying enterprise qualifications based on multiple models according to claim 5, wherein, The screening of the features of enterprises meeting the scale standard includes: Calculate the importance score of the features of enterprises meeting the scale standard based on the importance calculation formula. The importance calculation formula is: In the formula, V(t) represents the importance score of the features of enterprises meeting the scale standard, t represents the node, and M represents the number of all trees; Let \(N(m)\) denote the number of non - leaf nodes of the \(m\) - th tree, and \(\beta(m,i)\) denote the partitioning feature of the \(i\) - th non - leaf node of the \(m\) - th tree. G γ(m,i) and H γ(m,i) respectively represent the first-order derivative and the second-order derivative of all samples that fall on the i-th non-leaf node of the m-th tree; Let \(L\) denote the left node, \(R\) denote the right node, and \(\lambda\) denote the hyper - parameter of the regularization term. I represents the indicator function, G γ(m,i,L) and G γ(m,i,R) respectively represent the sum of the first-order derivatives on the left and right nodes of the i-th non-leaf node on the m-th tree, H γ(m,i,L) and H γ(m,i,R) respectively represent the sum of the second-order derivatives on the left and right nodes of the i-th non-leaf node on the t-th tree; Rank the features of enterprises that meet the scale standard according to the importance score, and select the features of enterprises that meet the scale standard within the preset ranking.

7. The method for identifying enterprise qualifications based on multiple models according to claim 1, wherein The use of the optimized scoring - card model for qualification identification of general enterprises also includes: Based on the experience database of general enterprises, adjust the final qualification score of general enterprises to obtain the adjusted final qualification score of general enterprises.

8. The method for identifying enterprise qualifications based on multiple models according to claim 7, wherein The score conversion mechanism includes: Calculate the traditional scoring - card score of general enterprises, and take the mean of the traditional scoring - card scores to obtain the basic score; calculate the logistic regression results of each dimension of general enterprises, and multiply them by the weights of the corresponding dimensions to obtain the dimension scores. Add the basic score of general enterprises and all dimension scores to obtain the final qualification score of general enterprises.

9. The enterprise qualification recognition method based on multiple models according to claim 1, characterized in that It also includes: When it is necessary to adjust the output result of the model to be evaluated, increase the number of trigger - mechanism enterprises of the corresponding model to be evaluated, where the models to be evaluated include the initial score model of technology - based enterprises, the model of enterprises that meet the scale standard, and the model of general enterprises. When the ratio of the number of trigger - mechanism enterprises of the model to be evaluated to the number of enterprises of the corresponding enterprise type is greater than the preset value, the evaluation result of the model to be evaluated is unqualified; when the evaluation result is unqualified, retrain the corresponding model.

10. An enterprise qualification recognition device based on multiple models, which is used to implement the enterprise qualification recognition method based on multiple models according to any one of claims 1-9, and is characterized in that, The device includes: a classification module, a technology - based enterprise identification module, a scale - standard - meeting enterprise identification module, and a general - enterprise identification module. Among them, the classification module is used to classify the types of enterprises, and the enterprise types include technology - based enterprises, scale - standard - meeting enterprises, and general enterprises. The technology - based enterprise identification module is used to adopt the initial score model of technology - based enterprises to identify the qualifications of technology - based enterprises and obtain the adjusted final qualification score of technology - based enterprises. The scale - standard - meeting enterprise identification module is used to adopt the model of scale - standard - meeting enterprises to identify the qualifications of scale - standard - meeting enterprises and obtain the adjusted final qualification score of scale - standard - meeting enterprises. The general - enterprise identification module is used to adopt the optimized scoring - card model to identify the qualifications of general enterprises and obtain the final qualification score of general enterprises.

Images