Agricultural subject credit dynamic evaluation method and device

Abstract

The invention provides an agricultural subject credit dynamic evaluation method, which comprises the following steps: a data acquisition and management step: acquiring source data related to agricultural subject credit, and taking the source data as sample data; a prediction model management step: performing feature extraction and screening on the sample data to obtain screened features, and performing integrated improvement training on the basic model by using the screened features to obtain an optimal model; a credit evaluation index processing step of screening out important features and calculating weights of the important features in combination with the feature weight in the optimal model, the feature weight in the SHAP visual model and the feature weight of expert evaluation, and setting a credit evaluation rule based on a feature set and the weights of the important features; and a credit evaluation step of calculating an evaluation result of the credit of the agricultural subject from the feature data of the to-be-evaluated agricultural subject by using the current credit evaluation rule. The agricultural subject credit dynamic evaluation method has objectivity and transparency, and can dynamically evaluate the agricultural subject credit in time.

Description

Technical Field

[0001] The present invention relates to the field of data processing and intelligent decision-making, and in particular to a method and device for dynamic evaluation of agricultural subject credit. Background Art

[0002] For functional departments and financial institutions, the credit rating of agricultural entities is the key decision-making basis when conducting industry supervision and credit decisions. Agricultural entities have both high risks and high growth, but the lack of collateral, incomplete management and information disclosure have caused problems in the credit evaluation of agricultural entities. At present, the credit evaluation of agricultural entities mainly relies on the personal experience of experts and limited financial information, which makes the evaluation process susceptible to subjective factors and reduces the objectivity and accuracy of credit evaluation.

[0003] Although some agricultural entity credit evaluation methods have gathered multi-source information of agricultural entities and conducted agricultural entity credit evaluation based on machine learning algorithms, which have improved the accuracy of credit evaluation to a certain extent compared with traditional methods such as hierarchical analysis and logistic regression, the black box properties of machine learning algorithms can easily lead to the credit evaluation process and evaluation results being unable to be objectively and reasonably explained.

[0004] Moreover, in the existing credit evaluation process of agricultural entities, model training and credit evaluation are completely separated and managed offline, resulting in slow model upgrades, untimely credit data updates, and even inconsistent evaluation rules for new and old data. With the increase in the dimensions and volume of credit-related data of agricultural entities, the accuracy requirements for credit evaluation are getting higher and higher, and the prediction models and credit scoring and rating rules need to be updated in a timely manner in line with the needs of data, technology, and supervision. Summary of the invention

[0005] 1. Technical issues to be resolved The present invention is proposed in view of the above-mentioned technical problems, and aims to provide a method and device for dynamic evaluation of agricultural entity credit which is not affected by subjective factors, has a transparent and controllable process and can be dynamically evaluated.

[0006] (II) Technical solution In view of the above technical problems to be solved, the present invention provides a method for dynamic evaluation of agricultural subject credit, which comprises: Data collection and management steps: collecting source data related to agricultural subject credit and using it as sample data; Prediction model management step: extract and filter the features of sample data to obtain the filter features, and use the filter features to integrate and improve the basic model to obtain the optimal model; The credit evaluation processing step combines the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts to screen out important features and calculate their weights, and set credit evaluation rules based on the feature set and weights of important features; The credit evaluation step uses the current credit evaluation rules to calculate the credit evaluation results of the agricultural entity based on the characteristic data of the agricultural entity to be evaluated.

[0007] Among them, feature extraction and screening are performed on sample data to obtain screening features, including: Extract samples from the sample data, extract features, obtain preliminary screening features, set feature and result labels at the same time, use the Pearson coefficient and the Gini coefficient of random forest to perform preliminary screening on the features, and obtain screening features.

[0008] In addition, the basic model is integrated and improved by using the screening features, including: From the existing basic models, select two or more models as base models, use the training set to train the base models, get their respective prediction results, use them as new features to form a new feature set, and then select a new model from the existing basic models and integrated models as a meta-model integration to learn these new features and result labels, get prediction results and prediction accuracy, until the model converges or reaches satisfaction, thus obtaining a process model; using the same validation set, compare the accuracy of the current process model with the existing basic models and integrated models, and the model with the highest accuracy is the optimal model.

[0009] In addition, the basic model includes at least one of SMOTE, logistic regression, BP neural network, SVM, random forest, AdaBoost, GBDT, XGBoost, and LightGBM, and the integrated model includes at least one of Stacking, voting, boosting, and bagging.

[0010] In addition, the set feature is at least one of the agricultural entity's asset data, production and operation data, agricultural insurance data, and subsidy data from the competent authority, and the result label is whether there is a default or a credit score.

[0011] In addition, the method combines the feature weights in the optimal model, the feature weights in the SHAP visualization model and the feature weights evaluated by experts to screen out key features and calculate their weights, including: arranging all the features of the optimal model in descending order of importance, the importance of each feature is the corresponding weight of the feature, the SHAP method visualizes the influence of each feature in the optimal model on the prediction result and defines each feature weight with a SHAP value, and at the same time inputs the feature weights evaluated by experts, represents the set of these features and weights as a set of triples, and then sets the influencing factors of the importance of each feature set of the optimal model, the SHAP method and the expert evaluation, and performs weighted summation of the three sets to obtain the comprehensive weight of each feature, thereby obtaining a feature set of important features.

[0012] The present invention also provides a device for dynamic evaluation of agricultural subject credit, which comprises: The data collection and management module collects source data related to the credit of agricultural entities and uses it as sample data; The prediction model management module extracts and filters the features of the sample data to obtain the screening features, and uses the screening features to integrate and improve the basic model to obtain the optimal model; The credit evaluation index processing module combines the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts to screen out important features and calculate their weights, and sets credit evaluation rules based on the feature set and weights of important features; The credit evaluation module uses the current credit evaluation rules to calculate the credit evaluation results of the agricultural entity based on the characteristic data of the agricultural entity to be evaluated.

[0013] (III) Beneficial effects The dynamic credit evaluation method for agricultural entities according to the present invention can centrally and dynamically manage the multi-source data, prediction models, credit evaluation rules, etc. of agricultural entities, so that the credit evaluation of agricultural entities is transparent and controllable in the whole process of data processing, feature extraction, model training, credit evaluation and rating. The base model can be flexibly selected to build an integrated model with higher accuracy, and the key factors affecting the credit of agricultural entities can be visually screened in combination with the SHAP method, and the credit evaluation (including scoring and rating) can be calculated according to the current credit evaluation rules. The present invention overcomes the interference of subjective factors and the black box defects of machine learning algorithms in the credit evaluation process. The prediction model and credit evaluation rules can be dynamically customized and updated in time according to changes in the data of agricultural entities. The credit evaluation results of agricultural entities are also updated accordingly, which improves the flexibility and interpretability of the credit evaluation process and the objectivity, accuracy and real-time nature of the evaluation results, which is conducive to the functional departments and financial institutions to dynamically evaluate the credit situation of agricultural entities and carry out real-time supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1It is a flow chart of a method for dynamic evaluation of agricultural entity credit provided by an embodiment of the present invention.

[0015] Figure 2 It is a schematic diagram of feature screening and optimal model construction.

[0016] Figure 3 It is a flowchart that combines machine learning, SHAP, and expert evaluation to screen important features and set weights.

[0017] Figure 4 It is a dynamic credit evaluation device for agricultural entities. DETAILED DESCRIPTION

[0018] The present invention is described below in conjunction with the accompanying drawings. The embodiments are only for illustration and are not intended to limit the present invention. Solutions that have substantially the same composition and exert the same effects as the technical concept described in the claims of the present invention are all included in the technical scope of the present invention.

[0019] Figure 1 Flowchart of the method for dynamic credit evaluation of agricultural entities provided by the embodiment of the present invention. Figure 1 As shown, the agricultural subject credit dynamic evaluation method of the present invention includes: Data collection and management: Collect source data related to the credit of agricultural entities and use it as sample data; Prediction model management: extract and filter sample data features to obtain filter features, use the filter features to integrate and improve the basic model and obtain the optimal model; Credit evaluation index processing: Combine the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts to screen out important features and calculate their weights, and set credit evaluation rules based on the feature set and weights of important features; Credit evaluation: The credit evaluation result of the agricultural entity is calculated based on the characteristic data of the agricultural entity to be evaluated using the current credit evaluation rules.

[0020] Among them, first of all, source data related to the credit of agricultural entities are collected, and the source data include but are not limited to: collecting basic information, asset and credit information of agricultural entities from the agricultural and rural big data platform, and the asset information includes at least one of land, facilities, factory buildings, agricultural products, agricultural machinery, subsidies and insurance; collecting educational information of the person in charge of the agricultural entity from the China Higher Education Student Information and Career Center; collecting liability information, loan records, credit ratings and other information from financial institutions; collecting credit information from credit reporting agencies; collecting credit information of corporate agricultural entities from the National Enterprise Credit Publicity System; collecting judicial information of agricultural entities from the China Law Enforcement Information Disclosure Network; collecting industrial and commercial information of corporate agricultural entities from the enterprise information query platform; collecting news information related to agricultural entities from the Internet and other platforms.

[0021] These source data are stored in the system database for preprocessing, which includes but is not limited to: removing outliers, filling missing values, cleaning, association, numerical standardization, data consistency verification, etc. After the source data is preprocessed, it is stored in the database as sample data.

[0022] Next, extract and filter the features of the sample data to obtain the filtering features. Specifically, it includes: extracting samples from the sample data, extracting features, obtaining preliminary screening features, setting feature and result labels at the same time, using the Pearson coefficient and the Gini coefficient of the random forest to perform preliminary screening on the features to obtain the screening features. For example, a sample database is established, and for a large number of sample data, screening conditions such as the time window or subject type of the sample are set, the corresponding samples are filtered, and saved as a classification sample library. Figure 2 As shown in , this process can be called index selection. More specifically, for example, the credit data of family farms in the past three years are used as samples, and then the sample data are processed with unbalanced samples using SMOTE oversampling technology, stratified sampling technology, etc., and named as the three-year sample database of family farms. Then, feature extraction is performed on all sample data in the sample database, such as Figure 2 As shown in Figure 1, samples are extracted from the agricultural subject credit database, and training set, test set, and validation set are set. The ratio of training set, test set, and validation set is set to 6:2:2. Then, the features are further screened. Figure 2 As shown, in data processing and feature engineering, including but not limited to using the Pearson coefficient and the gini coefficient of random forest to screen features, preliminarily eliminating duplicate and invalid features, so as to obtain screening features. Among them, including but not limited to: the features are set as the asset data, production and operation data, agricultural insurance data, and subsidy data of the competent department of the agricultural subject, and the result label is set as whether it is a default or a credit score.

[0023] Next, the basic model is integrated and improved by using the screening features, which includes: selecting two or more models from the existing basic models as the base models, using the training set to train the base models, obtaining their respective prediction results, using them as new features to form a new feature set, and then selecting a new model from the existing basic models and integrated models as a meta-model integration to learn these new features and result labels, obtain prediction results and prediction accuracy, until the model converges or reaches satisfaction, thereby obtaining a process model; using the same validation set to compare the accuracy of the current process model with the existing basic model and integrated model, and the model with the highest accuracy is the optimal model.

[0024] Among them, the basic models include but are not limited to SMOTE, logistic regression, BP neural network, SVM, random forest, AdaBoost, GBDT, XGBoost, LightGBM, but random forest, XGBoost, and LightGBM are preferred. In addition, the integration methods include but are not limited to Stacking, voting, boosting, and bagging, among which Stacking is preferred, which can make full use of the advantages and characteristics of different models to improve the accuracy and generalization ability of the overall model. For example: D is the initial training set, is the sample set obtained after training with random forest, XGBoost, and LightGBM. is the logistic regression prediction function, This is the predicted value of the model after Stacking integration.

[0025] In addition, if the process model is the optimal model, it will be saved as an integrated model for model upgrade and iteration, otherwise it will be deleted. The accuracy, AUC value, and F1-Score can be used to comprehensively evaluate the prediction effect of the model. The higher the value, the better the prediction effect of the model. In addition, the test set is used to evaluate the generalization ability of the model.

[0026] Next, combining the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts, screen out important features and calculate their weights, and set credit evaluation rules based on the feature set and weights of important features. The method of combining the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts, screen out key features and calculate their comprehensive weights, including: arranging all the features of the optimal model in descending order of importance, the importance of each feature is the corresponding weight of the feature, the SHAP method visualizes the degree of influence of each feature in the optimal model on the prediction result and defines each feature weight with the SHAP value, and inputs each feature weight assessed by experts at the same time, represents the set of these features and weights as a set of triples, and then sets the influencing factors of the importance of each feature set assessed by the optimal model, the SHAP method, and the expert, and weights the three sets to obtain the comprehensive weight of each feature, thereby obtaining the feature set of important features.

[0027] For example, all the features of the optimal model are arranged in descending order of importance, and the importance of each feature is the corresponding weight of the feature; the SHAP method visualizes the influence of each feature in the optimal model on the prediction result, and uses the SHAP value to define the weight of each feature, and inputs the weight of each feature evaluated by experts. This set of features and weights is represented as a set of triples, {(f1,w f11 ,w f12,w f13 ),(f2,w f21 ,w f22 ,w f23 ),…,(f n ,w fn1 ,w fn2 ,w fn3 )} Among them, f i For a certain feature, w fi1 、w fi2 、w fi3 The feature f i The weights within the optimal model, the weights in the SHAP visualization model, and the weights of expert evaluation.

[0028] in, .

[0029] Further set the optimal model, SHAP method, and the influencing factors of the feature set importance evaluated by experts, such as w s1 :w s2 :w s3 =3:5:2, and the weighted sum of the three sets is taken to obtain each feature and weight. For another example, the impact factor of the optimal model and expert evaluation is set to 0, that is, w s1 and w s3 If it is 0, then the credit evaluation index of the agricultural entity is completely set by the SHAP method.

[0030] Calculation formula: Feature f i The comprehensive weight is expressed as For feature f i The comprehensive weight W i Sort in descending order and save, and select important feature indicators, and remove invalid features or features with very weak influence. For example, set the first m features, or features with weight ranking in the top t%, or features with weight above s, as the feature set of the prediction model. Figure 3As shown, important features can be screened and weights can be set by combining machine learning, SHAP, and expert evaluation. The feature screening of important features and the setting of feature weights avoid the "black box" defect of machine learning, making the decision process of the prediction model transparent and controllable. In addition, multiple credit scoring rules can be managed in advance, and each credit scoring rule corresponds to an integrated model, a feature set, and a weight. Credit evaluation rules are further set based on the feature set and weights of the important features obtained above. After a rule is set as a formal credit evaluation rule, it is used to calculate the credit scores of each subject in the system. That is, the credit scoring rule corresponds to the integrated model and the feature set, and the integrated model or feature set can be used as the input of the credit evaluation, the prediction result of the integrated model can be used as the calculation basis of the credit evaluation, or the features and weights of the feature set can be used as the calculation factors of the credit evaluation. For the convenience of calculation, the credit evaluation can be implemented in percentage.

[0031] For example, the formula for calculating a credit score: Normalize the feature data and feature weights to be predicted. The normalization includes but is not limited to minimum-maximum normalization and Z-score normalization.

[0032] The normalized feature data is weighted and summed to convert it into a percentage score: Among them, S min and S max They are respectively the minimum and maximum values ​​of all possible credit score values, which can be obtained from historical data statistics or data distribution.

[0033] In addition, after obtaining the credit score, the score can be used as input to calculate the credit rating according to the credit rating rules. For example, the credit rating of agricultural entities can be divided according to the four grades and ten levels of excellent, good, medium and poor, AAA, AA, A, BBB, BB, B, CCC, CC, C, and D, and the credit rating range can be compared. Among them, excellent corresponds to AAA, AA, and A, good corresponds to BBB, BB, and B, medium corresponds to CCC, CC, and C, and poor corresponds to D.

[0034] In order to meet actual needs, the agricultural subject credit dynamic evaluation method of the present invention can update subject data, replace prediction models, credit evaluation (including scoring and rating) indicator processing methods or credit evaluation (including scoring and rating) rules at any time. The system will re-evaluate all subject data according to the new model and new rules to maintain the timeliness of credit evaluation and the consistency of evaluation results with current evaluation standards.

[0035] In addition, the present invention also provides a dynamic credit evaluation device for agricultural entities, such as Figure 4 As shown, it includes: The data collection and management module collects source data related to the credit of agricultural entities and uses it as sample data; The prediction model management module extracts and screens the sample data to obtain the screening features, and uses the screening features to integrate and improve the basic model to obtain the optimal model; The credit evaluation index processing module combines the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts to screen out important features and calculate their weights, and sets credit evaluation rules based on the feature set and weights of important features; The credit evaluation module uses the current credit evaluation rules to calculate the credit evaluation results of the agricultural entity based on the characteristic data of the agricultural entity to be evaluated.

[0036] In addition, the agricultural subject credit dynamic evaluation device of the present invention can also have a user management module, a log management module, a credit evaluation report generation module, and a data sharing module. The data sharing module can encrypt and share the agricultural subject credit score and rating with banks or other competent functional departments.

[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic credit evaluation of agricultural entities, characterized in that: It includes: Data collection and management steps: collecting source data related to agricultural subject credit and using it as sample data; Prediction model management step: extract and filter the features of sample data to obtain the filter features, and use the filter features to integrate and improve the basic model to obtain the optimal model; The credit evaluation index processing step combines the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts to screen out important features and calculate their weights, and set credit evaluation rules based on the feature set and weights of important features; The credit evaluation step uses the current credit evaluation rules to calculate the credit evaluation results of the agricultural entity based on the characteristic data of the agricultural entity to be evaluated.

2. The agricultural subject credit dynamic evaluation method according to claim 1 is characterized in that: Extract and filter the sample data to obtain the filtering features, including: Extract samples from the sample data, extract features, obtain preliminary screening features, set feature and result labels at the same time, use the Pearson coefficient and the Gini coefficient of random forest to perform preliminary screening on the features, and obtain screening features.

3. The agricultural subject credit dynamic evaluation method according to claim 2 is characterized in that: The set feature is at least one of the agricultural entity's asset data, production and operation data, agricultural insurance data, and subsidy data from the competent department, and the result label is whether it is in default or a credit score.

4. The agricultural subject credit dynamic evaluation method according to claim 1 is characterized in that: Use screening features to improve the integrated training of the basic model, including: From the existing basic models, select two or more models as base models, use the training set to train the base models, get their respective prediction results, use them as new features to form a new feature set, and then select a new model from the existing basic models and integrated models as a meta-model integration to learn these new features and result labels, get prediction results and prediction accuracy, until the model converges or reaches satisfaction, thus obtaining a process model; using the same validation set, compare the accuracy of the current process model with the existing basic models and integrated models, and the model with the highest accuracy is the optimal model.

5. The agricultural subject credit dynamic evaluation method according to claim 4 is characterized in that: The basic model includes at least one of SMOTE, logistic regression, BP neural network, SVM, random forest, AdaBoost, GBDT, XGBoost, and LightGBM, and the integrated model includes at least one of Stacking, voting, boosting, and bagging.

6. The agricultural subject credit dynamic evaluation method according to claim 1 is characterized in that: The method combines the feature weights in the optimal model, the feature weights in the SHAP visualization model and the feature weights evaluated by experts to screen out key features and calculate their weights, including: arranging all the features of the optimal model in descending order of importance, the importance of each feature is the corresponding weight of the feature, visualizing the influence of each feature in the optimal model on the prediction result by the SHAP method and defining each feature weight with a SHAP value, inputting each feature weight evaluated by experts at the same time, expressing the set of these features and weights as a set of triples, setting the influencing factors of the importance of each feature set of the optimal model, the SHAP method and the expert evaluation, and weighted summing the three sets to obtain the comprehensive weight of each feature, thereby obtaining a feature set of important features.

7. A dynamic credit evaluation device for agricultural entities, characterized in that: It includes: The data collection and management module collects source data related to the credit of agricultural entities and uses it as sample data; The prediction model management module extracts and filters the features of the sample data to obtain the screening features, and uses the screening features to integrate and improve the basic model to obtain the optimal model; The credit evaluation index processing module combines the feature weights in the optimal model, the feature weights in the SHAP visualization model, and the feature weights assessed by experts to screen out important features and calculate their weights, and sets credit evaluation rules based on the feature set and weights of important features; The credit evaluation module uses the current credit evaluation rules to calculate the credit evaluation results of the agricultural entity based on the characteristic data of the agricultural entity to be evaluated.

Images