Strategy rule generation method and system based on decision tree

Abstract

The invention provides a strategy rule generation method and system based on a decision tree, and the method comprises the steps: obtaining the historical credit performance data of a credit institution, and carrying out the processing and sampling of the historical credit performance data according to a business scene, and obtaining an original data set; adjusting an original data set, establishing a decision tree model, and dividing the data set according to variable features; based on the decision tree model, mining rules of all nodes of the decision tree, and selecting calculation indexes in combination with a credit business scene to obtain the rules and index information corresponding to the rules; and screening and sorting the rules output by the decision tree model in combination with credit business indexes to obtain a strategy rule combination. The system comprises a model construction module, an information acquisition module and a rule output module. Through the steps of data processing, model establishment, rule mining and screening and the like, a strategy rule which is high in interpretability and combined with a business index can be provided, and more accurate and efficient risk control and decision support are provided for a credit institution.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and in particular to a decision tree-based strategy rule generation method and system. Background Art

[0002] Mining strategy rules play a vital role in the field of credit risk control; they are based on a large amount of data and statistical analysis to determine the credit risk and repayment ability of borrowers. Through mining strategy rules, multiple factors such as borrowers' personal information, financial status and employment status can be analyzed to predict their repayment ability and default risk. These rules can help banks and financial institutions evaluate borrowers' credit ratings, decide whether to approve loan applications, and set appropriate loan interest rates and amounts. Mining strategy rules usually use machine learning and data mining techniques to discover the laws and patterns hidden in the data through the analysis and modeling of a large amount of historical data; these rules can be updated and adjusted according to real-time market changes and borrowers' specific circumstances to improve the accuracy and efficiency of risk assessment. In the field of credit risk control, the design and optimization of mining strategy rules is an ongoing process that requires continuous collection and analysis of new data, and adjustment and improvement based on actual conditions; this can help financial institutions better manage risks, improve the accuracy and reliability of loans, and protect their own interests.

[0003] At present, mining strategy rules often requires combining personal business experience and several features to form a risk control strategy. For example, when formulating a bank credit strategy based on historical data, a strategy rule is passed for companies that have no overdue behavior within six months, have a registered capital of more than 1 million, and have a staff size of more than 50 people. However, manual analysis is highly dependent on experience and is more likely to consume a lot of time. The decision tree is a tree model based on if-then rules, which can be regarded as a collection of rules. Therefore, it has become an important tool for mining risk control rules. In summary, how to use the decision tree model to efficiently mine risk control strategy rules and assist or replace manual efforts to improve strategy search efficiency is an important issue in the field of risk control.

[0004] Prior art 1, Chinese patent, application number: CN202310550673.3 discloses a method for mining risk control strategies for cash flow loans based on paired sample T test, including: using the cash flow data based on the company's history and the company's default data, finding effective rules through the paired sample T test method and visualizing their data, and calculating the default rate and pass rate corresponding to the rules. Although the rules with lower default rates and higher pass rates are selected and discussed with experts to form the final rule base; the rule base is updated as the data is updated to ensure the timeliness of the rule parameters. However, the strategy generation efficiency is low, which affects the accuracy of the cash flow loan risk control strategy mining.

[0005] Prior art 2, Chinese patent, application number: CN202310854965.6 discloses a method, system, medium and device for generating risk control strategies based on automatic rules, including: obtaining user data and order information; labeling user samples based on user data to obtain a multivariate labeled data set; preprocessing order information to obtain different feature sets; using different gradient tree models to train different feature sets and multivariate labeled data sets to obtain output decision tree groups; automatically screening and disassembling output decision trees that meet preset screening conditions to obtain automatic rules; judging whether a user is a risky user according to automatic rules. Although it has important application value in solving problems such as poor interpretability of risk control models and difficulty in persuading users, it can provide a more efficient and interpretable risk control solution for the new retail industry. However, a large amount of user data and order information is required as input, and the quality and accuracy of the data have a great impact on the effect of the model; if there are errors, omissions or deviations in the data, the generated automatic rules may be inaccurate or invalid.

[0006] Prior art three, Chinese patent, application number: CN202311732289.1 discloses an intelligent risk control based on an AI large model, including: collecting sample data, collecting business data and user behavior data that need to be connected to risk control; processing data, marking sensitive data that needs risk control; constructing a multi-dimensional risk factor system, establishing a reasonable user portrait, improving the identification effect of black market users, statistical factors and advanced factors; automatically generating risk control models and rules through machine learning, and automatically generating a strategy rule engine algorithm library; the business system accesses the risk control model and calls the generated strategy rule library. Although by building an AI model to privatize risk control gpt, through risk control data training, generate strategies and rules for different scenarios, and then through program offline testing, the test is automatically generated through the strategy to the rule engine, improve the risk control construction process and the efficiency of risk control strategy generation, shorten the risk control construction cycle, form an independent risk control product, and practice risk control strategies in real time. However, the model requirements are high, which is directly related to the accuracy of the risk control strategy results, and the implementation requirements are high.

[0007] At present, the existing technologies 1, 2 and 3 have the problem of low efficiency in mining risk control strategy rules, which reduces the data processing efficiency, affects the degree of automation of strategy generation, and increases the cost of strategy development. Therefore, the present invention provides a decision tree-based strategy rule generation method and system, which realizes efficient strategy mining and rule integration based on an innovative decision tree method. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides a decision tree-based policy rule generation method, comprising the following steps:

[0009] Obtain historical credit performance data from credit institutions, process and sample the historical credit performance data according to business scenarios to obtain the original data set; establish a decision tree model;

[0010] Based on the decision tree model, the rules of all nodes of the decision tree are mined, and the calculation indicators are selected in combination with the credit business scenario to obtain the rules and their corresponding indicator information;

[0011] Combined with credit business indicators, the rules output by the decision tree model are screened and sorted to obtain a strategy rule combination.

[0012] Optionally, the process of obtaining the original data set includes the following steps:

[0013] Obtain historical customer credit performance data from credit institutions and adjust variable data types based on actual business scenarios;

[0014] According to the proportion of each type of sample in the credit performance data, the data samples are stratified and sampled, and the weight variables of the sampled data are calculated;

[0015] The sampled data is divided into training set and test set to obtain a usable original data set.

[0016] Optionally, the prediction variable is the behavior information of the credit user, and its value type is numeric, character, or Boolean.

[0017] Optionally, the target variable is obtained based on whether the credit customer is overdue, which is a binary categorical variable represented by good or bad. Good means that the user has no overdue behavior, and bad means that the user has overdue behavior.

[0018] Optionally, the training set is a data set used to train and generate a decision tree model, and the test set is a data set used to test and measure the effectiveness of the model.

[0019] Optionally, adjust the original data set, build a decision tree model, and divide the data set according to variable characteristics.

[0020] Optionally, the process of obtaining a policy rule combination includes the following steps:

[0021] For each rule, apply it to the sample set and calculate the effect of applying the rule, which includes the improvement of prediction accuracy and the reduction of risk;

[0022] For the calculated rule effects, the rules whose effects are greater than a specified threshold are screened out; the screened out rules are used as candidate sets of policy rules;

[0023] For the candidate rules, select evaluation indicators according to the credit business indicators, and calculate the score or value of each rule on the specified indicator; sort the rules according to the indicator score or value, from high to low.

[0024] The decision tree-based policy rule generation system provided by the present invention comprises:

[0025] The model building module is responsible for obtaining the historical credit performance data of credit institutions, processing and sampling the historical credit performance data according to business scenarios to obtain the original data set; adjusting the original data set, establishing a decision tree model, and dividing the data set according to variable characteristics;

[0026] The information acquisition module is responsible for mining the rules of all nodes of the decision tree based on the decision tree model, selecting calculation indicators based on the credit business scenario, and obtaining the rules and their corresponding indicator information;

[0027] The rule output module is responsible for screening and sorting the rules output by the decision tree model in combination with credit business indicators to obtain a strategy rule combination.

[0028] Optional, model building modules, including:

[0029] The data acquisition submodule is responsible for obtaining historical customer credit performance data from credit institutions and adjusting variable data types according to actual business scenarios;

[0030] The stratified sampling submodule is responsible for stratifying the data samples according to the proportion of each type of samples in the credit performance data and calculating the weight variables of the sampled data;

[0031] The data partitioning submodule is responsible for partitioning the sampled data into training sets and test sets based on a specified method to obtain a usable original data set.

[0032] Optional, rule output module, including:

[0033] The effect calculation submodule is responsible for applying each rule to the sample set and calculating the effect after applying the rule, which includes the improvement of prediction accuracy and the reduction of risk.

[0034] The rule screening submodule is responsible for screening out the rules whose effects are greater than a specified threshold from the calculated rule effects; the screened rules are used as candidate sets of policy rules;

[0035] The indicator evaluation submodule is responsible for selecting evaluation indicators for candidate rules according to the credit business indicator lift, calculating the score or value of each rule on the specified indicator, and sorting the rules from high to low according to the indicator score or value.

[0036] The present invention first obtains the historical credit performance data of the credit institution, processes and samples the historical credit performance data according to the business scenario, and obtains the original data set; adjusts the original data set, establishes a decision tree model, and divides the data set according to the variable characteristics; secondly, based on the decision tree model, mines the rules of all nodes of the decision tree, selects the calculation index in combination with the credit business scenario, and obtains the rules and the corresponding index information; finally, in combination with the credit business index, the rules output by the decision tree model are screened and sorted to obtain the strategy rule combination. A representative data set is obtained for subsequent modeling and analysis; it is ensured that the training and evaluation of the decision tree model are based on real and reliable data, and the accuracy and credibility of the decision tree model are improved. The rules of each node in the decision tree model and the index information related thereto are obtained; the rules with strong interpretability are extracted from the decision tree model to help understand and explain the decision-making process of the model. At the same time, the calculation index is selected according to the credit business scenario, so that the rules can be better mined and optimized according to the business needs. The rules are screened and sorted according to the credit business indicators to obtain the final strategy rule combination; the rules output by the decision tree model are combined with the business indicators to ensure that the generated rules meet the actual business needs and risk control requirements. At the same time, the generation of strategy rule combinations can help credit institutions formulate effective risk control strategies and improve the accuracy of loan decisions and risk control capabilities.

[0037] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0038] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0040] Figure 1 This is a flow chart of a decision tree-based strategy rule generation method in Embodiment 1 of the present invention;

[0041] Figure 2 This is a diagram of the process of obtaining the original data set in Example 2 of the present invention;

[0042] Figure 3 Graph showing the calculation process of the weight variables of the sampling data in Embodiment 3 of the present invention;

[0043] Figure 4 A process diagram of dividing a data set according to variable characteristics in Example 4 of the present invention;

[0044] Figure 5 This is a process diagram of discretization of continuous numerical variables in Embodiment 5 of the present invention;

[0045] Figure 6 A process diagram for obtaining rules and their corresponding indicator information in Example 6 of the present invention;

[0046] Figure 7 A process diagram for obtaining rules and corresponding indicator information in Example 7 of the present invention;

[0047] Figure 8 This is a process diagram for obtaining a strategy rule combination in Embodiment 8 of the present invention;

[0048] Fig. 9 A block diagram of a policy rule generation system based on a decision tree in Embodiment 9 of the present invention;

[0049] Fig.10 This is a block diagram of a model construction module in Embodiment 10 of the present invention;

[0050] Fig.11 This is a block diagram of the information acquisition module in Example 11 of the present invention;

[0051] Fig.12 This is a block diagram of the rule output module in Example 12 of the present invention. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0053] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0054] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as described in detail. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0055] Example 1: Figure 1 As shown, an embodiment of the present invention provides a decision tree-based policy rule generation method, comprising the following steps:

[0056] S100: Obtain historical credit performance data of credit institutions, process and sample the historical credit performance data according to business scenarios to obtain original data sets; adjust the original data sets, establish a decision tree model, and divide the data sets according to variable characteristics;

[0057] S200: Based on the decision tree model, the rules of all nodes of the decision tree are mined, and calculation indicators are selected in combination with the credit business scenario to obtain the rules and their corresponding indicator information;

[0058] S300: In combination with the credit business indicators, the rules output by the decision tree model are screened and sorted to obtain a strategy rule combination.

[0059] The working principle and beneficial effects of the above technical solution are:

[0060] This embodiment first obtains the historical credit performance data of the credit institution, processes and samples the historical credit performance data according to the business scenario, and obtains the original data set; adjusts the original data set, establishes a decision tree model, and divides the data set according to the variable characteristics; secondly, based on the decision tree model, mines the rules of all nodes of the decision tree, selects calculation indicators in combination with the credit business scenario, and obtains the rules and their corresponding indicator information; finally, in combination with the credit business indicators, the rules output by the decision tree model are screened and sorted to obtain the strategy rule combination. Step S100 of the above scheme obtains a representative data set for subsequent modeling and analysis; ensures that the training and evaluation of the decision tree model are based on real and reliable data, and improves the accuracy and credibility of the decision tree model. Step S200 obtains the rules of each node in the decision tree model and the indicator information related thereto; extracts highly interpretable rules from the decision tree model to help understand and explain the decision-making process of the model. At the same time, by selecting calculation indicators according to the credit business scenario, the rules can be better mined and optimized according to business needs. Step S300 screens and sorts the rules according to the credit business indicators to obtain the final strategy rule combination; the rules output by the decision tree model are combined with the business indicators to ensure that the generated rules meet the actual business needs and risk control requirements. At the same time, the generation of strategy rule combinations can help credit institutions formulate effective risk control strategies and improve the accuracy of loan decisions and risk control capabilities.

[0061] In summary, the decision tree-based policy rule generation method of this embodiment can provide policy rules that are highly interpretable and combined with business indicators through steps such as data processing, model building, rule mining and screening, thereby providing credit institutions with more accurate and efficient risk control and decision support.

[0062] This embodiment is a method for efficiently utilizing a decision tree model to efficiently mine risk control policy rules. The innovative decision tree method also improves the efficiency and effectiveness of rule set modeling. Policy mining can assist or replace manual combination of policy rules, improve the degree of automation and search efficiency of the policy, and reduce the dependence of policy development on developer experience and business background.

[0063] Example 2: Figure 2 As shown, based on Example 1, the process of obtaining the original data set provided by the embodiment of the present invention includes the following steps:

[0064] S101: Obtain historical customer credit performance data from a credit institution, and adjust variable data types according to actual business scenarios; wherein the credit performance data includes prediction variables, target variables, and weight variables;

[0065] The predictor variable is the behavior information of the credit user, and its value type can be numeric, character, or Boolean;

[0066] The target variable is based on whether the credit customer is overdue. It is a binary categorical variable, represented by good or bad. Good means that the user has no overdue behavior, and bad means that the user has overdue behavior.

[0067] The weight variable is obtained after sampling the sample. In the business scenario, if the institution has a long history of lending, it has accumulated a large amount of data and the number of good samples usually far exceeds the number of bad samples. Therefore, the sample is sampled to improve the calculation efficiency, so the data set will contain weight variables;

[0068] S102: Perform stratified sampling on the data samples according to the proportion of each type of samples in the credit performance data, and calculate the weight variables of the sampled data;

[0069] S103: Divide the sampled data into a training set and a test set based on a specified method to obtain a usable original data set;

[0070] The training set is the data set used to train and generate the decision tree model, and the test set is the data set used to test and measure the model effect. The specified method is the method of dividing the training set and the test set selected according to the actual situation, such as division by fixed ratio (holdout method), K-fold cross-validation method, and custom test set.

[0071] The working principle and beneficial effects of the above technical solution are:

[0072] This embodiment first obtains the credit performance data of historical customers from a credit institution, and adjusts the variable data type according to the actual business scenario; secondly, stratified sampling is performed on the data samples according to the proportion of each type of sample in the credit performance data, and the weight variables of the sampled data are calculated; finally, the sampled data is divided into a training set and a test set based on a specified method to obtain a usable original data set. S101 of the above scheme obtains representative historical credit performance data, and adjusts the variable data type to meet the needs of subsequent analysis and modeling; ensures the accuracy and consistency of the data, and provides a reliable basis for subsequent data processing and analysis. Step S102 performs stratified sampling according to the sample proportion to obtain a representative sample set and generates weight variables; through stratified sampling, it is ensured that the sample set can accurately reflect the overall data distribution. At the same time, the introduction of weight variables can weight the samples in the subsequent modeling process, so as to model the sample distribution closer to the overall distribution, so that the model effect can be optimized in this business scenario. Step S103 divides the sampled data in a specified manner to obtain data sets for training and testing; the training set is used to build a decision tree model, and the test set is used to evaluate the performance of the model; by dividing the data set, the model can be trained and verified during the model building process, thereby evaluating the accuracy and generalization ability of the model.

[0073] In summary, this embodiment obtains available data sets that meet business needs, and performs data processing, sampling and division to provide a basis for the subsequent establishment and evaluation of decision tree models; ensures the quality and representativeness of data, improves the reliability and generalization ability of the model, and provides accuracy and reliability support for credit risk control decisions. This embodiment clarifies that the data set must include predictor variables, target variables, weight variables, and the meaning of each part; and standardizes the data format by adjusting the data format in combination with business scenarios and actual conditions.

[0074] Example 3: Figure 3 As shown, based on Example 2, the calculation process of the weight variable of the sampling data provided by the embodiment of the present invention includes the following steps:

[0075] S1021: Classify samples into different categories based on the target variable in the credit performance data; calculate the proportion of samples in each category;

[0076] S1022: Determine the sampling ratio of each category according to the ratio of samples of each category and the expected total sample size; and extract a corresponding number of samples from each category according to the determined sampling ratio;

[0077] S1023: Calculate the weight of each sample according to the sampling ratio and the overall sample size; for example, if the overall sample size is 10,000, the weight of the "good" sample is the overall sample size divided by the sampling sample size, that is, 10,000 / 800=12.5, and similarly, the weight of the "bad" sample is 10,000 / 200=50.

[0078] The working principle and beneficial effects of the above technical solution are:

[0079] This embodiment first divides the samples into different categories according to the target variable in the credit performance data; calculates the proportion of samples in each category; secondly, determines the sampling proportion of each category according to the proportion of samples in each category and the expected total sample size; extracts a corresponding number of samples from each category according to the determined sampling proportion; finally, calculates the weight of each sampled sample according to the sampling proportion and the overall sample size; for example, if the overall sample size is 10,000, the weight of the "good" sample is the overall sample size divided by the sampling sample size, that is, 10,000 / 800=12.5, and similarly, the weight of the "bad" sample is 10,000 / 200=50. Step S1021 of the above scheme divides the samples into different categories according to the target variable and calculates the proportion of samples in each category; understands the distribution of different categories in the sample, and provides a basis for subsequent stratified sampling and weight calculation. Step S1022 determines the sampling proportion of each category according to the sample proportion, and extracts a corresponding number of samples from each category; ensures that the sampled sample can represent the proportional relationship of the overall sample, thereby ensuring the accuracy and representativeness of the stratified sampling. Step S1023 calculates the weight of each sample according to the sampling ratio and the total sample size; provides sample weight information for subsequent model training and evaluation, ensures that different categories of samples are properly weighted during the modeling process, and more accurately reflects the contribution of different categories of samples.

[0080] In summary, this embodiment performs stratified sampling on samples and calculates the weights of samples, thereby ensuring the representativeness of samples and the accuracy of weights; ensuring the quality and representativeness of samples, improving the reliability and generalization ability of models, and providing support for accuracy and reliability of credit risk control and decision-making. Through the above steps, stratified sampling can be achieved according to the proportion of samples in each category, and corresponding weight variables can be generated; ensuring the representativeness of samples in the modeling and analysis process, and at the same time, the samples can be weighted in subsequent model training and evaluation to more accurately reflect the contribution of samples of different categories. In the actual implementation process, the above embodiment 3 also has certain technical disadvantages, that is, if the distribution of the target variable is very uneven, sampling according to the proportion may lead to insufficient information in certain categories. In view of this situation, the present invention adopts stratified sampling. In the process of stratified sampling, it is assumed that each layer is homogeneous. If the heterogeneity within the layer is high, stratified sampling may not be effective.

[0081] Therefore, this solution also adopts the following steps to solve the problem of high intra-layer heterogeneity when there is high intra-layer heterogeneity:

[0082] Acquire credit performance data, pre-process the credit performance data, and map the credit quality of the credit performance data;

[0083] Transform the credit performance data into a multi-dimensional matrix, which includes a credit score matrix, a credit covariance matrix, and a factor loading matrix;

[0084] Perform marginal contribution decomposition on the credit score matrix, credit covariance matrix and factor loading matrix to obtain marginal contribution coefficients;

[0085] Conduct correlation analysis on marginal contribution coefficients, screen out marginal contribution coefficients with low correlation, and obtain the optimal combination of marginal contribution coefficients;

[0086] According to the marginal contribution coefficient, the target variable in the credit performance data is selected to divide the samples into different categories, and the different categories are divided into test sets and training sets;

[0087] Statistically analyze the numerical range of marginal contribution coefficients of different categories, conduct intra-layer analysis of marginal contribution decomposition, determine the segmentation thresholds of marginal contribution coefficients of different categories, and analyze and compare the importance of different marginal contribution coefficients relative to different strata;

[0088] The credit performance data is classified by a gradient boosting machine to obtain a classification result of the credit performance data; wherein the gradient boosting machine includes a vector gradient boosting machine, an extreme gradient boosting machine, a light gradient boosting machine and a stochastic gradient descent boosting machine.

[0089] The principle of the above scheme is:

[0090] The present invention converts credit performance into a quantifiable credit quality indicator by mapping credit quality of credit data. By constructing a credit scoring matrix, a credit covariance matrix and a factor loading matrix, the credit data can be analyzed from multiple dimensions to reveal the intrinsic relationship between the data. The credit scoring matrix is ​​used to evaluate the degree of influence of different characteristics (such as income, debt ratio, credit history, etc.) on individual credit scores. Each characteristic will get a score according to its contribution to credit risk. Through the credit scoring matrix, the credit behavior of the customer can be converted into a numerical value, which can be used as a quantitative indicator of credit risk. The credit covariance matrix is ​​used to measure the correlation between different credit characteristics. This helps to identify which characteristics may have multicollinearity, that is, they may provide redundant information. Through the covariance matrix, the risk dispersion between different characteristics can be understood, which is crucial for building investment portfolios and risk management strategies. The factor loading matrix is ​​used for factor analysis, which reveals the relationship between the original variables and the potential factors. These potential factors usually represent some unobserved credit risk dimensions. Through the factor loading matrix, multiple related original variables can be summarized into a few unrelated factors, thereby achieving data dimensionality reduction.

[0091] By decomposing the matrix by marginal contribution, we can identify the contribution of each dimension to the overall credit risk. By analyzing the correlation between marginal contribution coefficients, we can filter out independent information, reduce multicollinearity problems, and improve the stability and predictive ability of the model.

[0092] Gradient boosting machine is a powerful ensemble learning algorithm that can effectively handle complex nonlinear relationships and provide high prediction performance by iteratively training decision trees to minimize the loss function. By performing intra-layer analysis on the marginal contribution coefficients of different categories, the segmentation threshold can be determined to better understand the importance of different categories in credit risk.

[0093] Through the above technical solution, if the heterogeneity within the layer is very high, the user's credit risk and some marginal characteristics of the user's credit data can be calculated, so that when stratifying, the data division result is clearer, and the data with large heterogeneity differences can be determined in the form of marginal contribution coefficients to achieve data differentiation.

[0094] Example 4: Figure 4 As shown, based on Example 1, the process of dividing the data set according to variable characteristics provided in the embodiment of the present invention includes the following steps:

[0095] S104: Based on the original data set, the original data is adjusted by discretizing the continuous numerical variables; based on the adjusted data set, the numerical variables are kept unchanged, the character and Boolean prediction variables are processed, the variable values ​​are sorted according to the proportion of customers with good credit performance in the sample, and the original values ​​are replaced by serial numbers to obtain the processed data set;

[0096] S105: Using the processed data set, establish the root node of the decision tree model; for each unsplit node in the decision tree model, determine whether the node data can be split, which satisfies the requirement that the data set of the current node is not empty and the value of the target variable of the current node is not unique;

[0097] Among them, for the splittable nodes, based on all the variables in the current data set, all the splitting points of all the variables are searched, and the optimal variables and splitting points are selected to split the nodes, so as to divide the credit population into two categories and generate two lower-level nodes;

[0098] S106: Repeat the above process of determining whether it can be split, searching for split points and splitting the lower layer for each newly generated node until all nodes cannot be split. Each node (except the root node) in the final decision tree is a rule.

[0099] In this embodiment, the discretization of continuous numerical variables refers to adjusting the sample data by binning, digit truncation and other operations on the continuous variables to reduce the number of unique values ​​of the sample values; the optimal variable and split point is: find the split that can obtain the largest Gini gain in the potential split of each variable, and compare all variables to select the variable with the largest Gini gain; the node cannot be split means: the label of the node data is unique, or the value of each attribute in the node is unique, or other restrictions, such as the number of node samples is lower than the user limit.

[0100] The working principle and beneficial effects of the above technical solution are:

[0101] This embodiment first adjusts the original data by discretizing continuous numerical variables based on the original data set; based on the adjusted data set, the numerical variables are kept unchanged, the character and Boolean prediction variables are processed, the variable values ​​are sorted according to the proportion of customers with good credit performance in the sample, and the original values ​​are replaced by serial numbers to obtain a processed data set; secondly, the root node of the decision tree model is established using the processed data set; for each unsplit node in the decision tree model, it is determined whether the node data can be split, and splitting means that the data set of the current node is not empty and the value of the target variable of the current node is not unique; finally, the above process of determining whether it can be split, searching for split points and splitting the lower layer is repeated for each newly generated node until all nodes cannot be split, and each node (except the root node) in the final decision tree is a rule. Step S104 of the above scheme discretizes the original data set and converts character and Boolean predictor variables; discretizes continuous variables to make them more suitable for establishing a decision tree model; at the same time, by sorting the variable values ​​according to the proportion of customers with good credit performance, the model's ability to identify important features can be improved. Step S105 establishes the root node of the decision tree based on the processed data set, and determines whether each unsplit node can be split; establishes the initial structure of the decision tree, and determines the nodes that can be split, providing a basis for the subsequent splitting process. Step S106 repeats the process of judging, searching for split points and splitting the lower layer for each splittable node based on the initial root node until all nodes cannot be split, thereby obtaining the final decision tree model; based on the division and segmentation of the data, a series of nodes and rules are generated to explain and predict the results of credit decisions.

[0102] In summary, this embodiment processes, divides and segments data, and finally generates the structure and rules of a decision tree model; through the decision tree model, decisions and predictions can be made based on the credit performance and other characteristics of the customer, providing accurate and reliable support for credit risk control and decision-making. This embodiment is a method for efficiently establishing a decision tree, which incorporates prior knowledge by sorting discrete attributes in combination with business scenarios, thereby increasing the reliability of segmentation; by discretizing continuous attributes, the complexity of searching for segmentation points is reduced, thereby improving the efficiency of the model.

[0103] Example 5: Figure 5 As shown, based on Example 4, the process of discretizing continuous numerical variables provided by the embodiment of the present invention includes the following steps:

[0104] S1041: For a continuous numerical variable, perform a binning operation, determine the number of bins or the binning method, divide the value range of the continuous numerical variable into multiple intervals, and each interval is used as a discrete category; according to the binning method, allocate sample data to each interval;

[0105] Among them, for equal-width binning, the value range of each interval is determined by calculating the minimum and maximum values ​​of the variable and the number of bins; for equal-frequency binning, the samples can be evenly distributed to each interval according to the number of samples;

[0106] S1042: within each binning interval, perform a digit truncation operation on the continuous numerical variable to retain the specified number of decimal places or digit range; and assign the sample data to the corresponding discrete category according to the truncation value range;

[0107] S1043: For character and Boolean predictor variables, sort them according to the proportion of customers with good credit performance in the sample and replace the original values ​​with serial numbers.

[0108] The working principle and beneficial effects of the above technical solution are:

[0109] This embodiment first performs binning operation on continuous numerical variables, determines the number of bins or the binning method, divides the value range of continuous numerical variables into multiple intervals, and each interval is a discrete category; according to the binning method, the sample data is assigned to each interval; secondly, within each binning interval, the continuous numerical variable is subjected to digit truncation operation, retaining the specified decimal places or digit range; according to the truncation value range, the sample data is assigned to the corresponding discrete category; finally, the character type and Boolean type prediction variables are sorted according to the proportion of customers with good credit performance in the sample, and the original values ​​are replaced by serial numbers. Step S1041 of the above scheme converts the continuous numerical variable into a discrete variable; reduces the uniqueness of the value of the continuous numerical variable, making it more suitable for the establishment of discrete models such as decision trees; the binning operation can also help discover the nonlinear relationship of the variable and improve the model's ability to recognize features. Step S1042 reduces the precision of the continuous numerical variable and converts it into a rougher discrete category; further reduces the uniqueness and details of the variable value, so that the model pays more attention to the overall trend of the variable and reduces the risk of overfitting. Step S1043 converts character and Boolean variables into ordered discrete variables, providing the model with sorting information on variable values, so that the model can better understand and explain the impact of the variables and the relationship between the variables.

[0110] In summary, this embodiment performs binning, digit truncation, and sorting on continuous numerical variables, and converts them into discrete variables to adapt to the establishment of discrete models; reduces the uniqueness of variable values, reduces the complexity of the model and the risk of overfitting, and provides better interpretability and explainability.

[0111] Example 6: Figure 6 As shown, based on Example 1, the process of obtaining a rule and its corresponding indicator information provided by the embodiment of the present invention includes the following steps:

[0112] S201: Initialize the root node and set the root node rule to be empty; starting from the root node, for each node except the root node, the node rule is the rule of its parent node plus the variable segmentation condition from its parent node to the current node;

[0113] S202: merging and organizing the obtained rules, and taking the intersection of different segmentation conditions of the same variable contained in the rules; sorting each segmentation condition contained in the organized rules in alphabetical order according to the name of the segmentation variable; accessing each node in the trained decision tree model based on the breadth-first order, and obtaining the rules of all nodes except the root node;

[0114] S203: Based on the data of each node, select calculation metrics in combination with the credit business scenario, calculate relevant metric information, and obtain rules and their corresponding metric information.

[0115] In this embodiment, the empty root node rule means that the root node contains all data sets without any conditional filtering; breadth-first means traversing the nodes of the tree along the breadth of the tree starting from the root node; if all nodes have been visited, the algorithm terminates; business-related metric information includes but is not limited to the number of samples hit by the rule, the good sample rate, the bad sample rate, and the proportion of lift samples in the rule samples, etc.

[0116] In this embodiment, the alphabetical order refers to sorting according to the ASCII code value; take the intersection of different segmentation conditions of the same variable included in the rule. For example, for continuous attributes, if the original rule is "2 < age < 5" and the current segmentation condition is "age > 3", it is sorted as "3 < age < 5"; for discrete attributes, if the original rule is "category in ["A", "B", "C"]" and the current segmentation condition is "category in ["A", "C"]", it is sorted as "category in ["A", "C"]".

[0117] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, first initialize the root node and make the root node rule empty; starting from the root node, for each node except the root node, the rule of the node is the rule of its parent node plus the variable segmentation condition from its parent node to the current node; secondly, merge and sort the obtained rules, and take the intersection of different segmentation conditions of the same variable included in the rules; for the sorted rules, sort each segmentation condition included therein according to the alphabetical order of the segmentation variable names; access each node in the trained decision tree model in breadth-first order to obtain the rules of all nodes except the root node; finally, based on the data of each node, select calculation metrics in combination with the credit business scenario, calculate relevant metric information, and obtain rules and their corresponding metric information. Step S201 of the above solution establishes the rules of each node in the decision tree model; determines the segmentation rules of each node for subsequent metric calculation and rule sorting. Step S202 sorts and arranges the rules for subsequent metric calculation and use; simplifies the expression form of the rules, removes duplicate segmentation conditions, and ensures the consistency and readability of the rules. Step S203 calculates metric information based on the data of the nodes, such as the number of samples of the node, the distribution of the target variable, the accuracy rate, etc.; evaluates the importance and quality of each node, helps understand the decision-making process of the model, and provides a basis for subsequent model evaluation and optimization.

[0118] In summary, this embodiment establishes the rules of each node in the decision tree model and calculates the relevant index information; the rules and index information help to understand the decision logic of the model, explain the results of the model, evaluate the performance of the model, and provide guidance for the optimization and improvement of the model; through these steps, the decision tree model can be transformed into a series of rules that are easy to understand and explain, and provide guidance for credit business decisions; all nodes in the rule set can be effectively and comprehensively traversed, and the index performance of the rules can be calculated in combination with business needs. The rule set obtained by the decision tree model is integrated, which reduces the complexity of the rules, increases orderliness, and is more convenient for observation and comparison between rules.

[0119] Example 7: Figure 7 As shown, based on Example 6, the process of obtaining a rule and its corresponding indicator information provided by the embodiment of the present invention includes the following steps:

[0120] S2031: Based on the data and credit business scenarios of each node, select indicators for calculation; indicators include sample size, distribution of target variables, accuracy, recall, precision, and F1 value, etc.;

[0121] S2032: For classification problems, calculate the number and proportion of samples in different categories and the degree of imbalance between categories; calculate the accuracy, recall and precision of nodes and other binary classification indicators;

[0122] S2033: For regression problems, calculate statistical indicators such as the mean, median and variance of the node to describe the distribution of the node data; based on the sample size of the node and the distribution of the target variable, calculate the purity or impurity indicators of the node, such as the Gini coefficient, information gain and entropy, to measure the purity of the node and the effect of the split.

[0123] The working principle and beneficial effects of the above technical solution are as follows: First, this embodiment selects the indicator calculation based on the data and credit business scenario of each node; the indicator includes the number of samples, the distribution of the target variable, accuracy, recall rate, precision rate and F1 value, etc.; secondly, for classification problems, the number of samples of different categories, the proportion and the imbalance degree between categories are calculated; the accuracy, recall rate and precision rate of the node are calculated; finally, for regression problems, the statistical indicators such as the mean, median and variance of the node are calculated to describe the distribution of the node data; based on the number of samples of the node and the distribution of the target variable, the purity or impurity indicators of the node, such as the Gini coefficient, information gain and entropy, are calculated to measure the purity of the node and the effect of splitting. Step S2031 of the above solution selects appropriate indicators to describe the data situation of the node and the performance of the model according to business needs; provides a comprehensive description of the node data characteristics and model effects to help understand the importance and quality of the node. Step S2032 calculates the number and proportion of samples of different categories to understand the distribution of node data and the balance between categories; measures the classification effect of the node and the prediction ability of the model for different categories. Step S2033 reveals the central trend, degree of dispersion and purity of node data by calculating statistical indicators and purity indicators; understands the distribution and purity of node data, provides a basis for model establishment, and optimizes the node splitting strategy.

[0124] In summary, through these steps, this embodiment can calculate relevant indicator information based on the data of each node, help understand the decision logic of the model, evaluate the performance of the model, and provide guidance for the optimization and improvement of the model; these indicator information helps to explain the results of the model, evaluate the accuracy and stability of the model, and provide reference and decision-making basis for credit business decisions. Through these steps, relevant indicator information can be calculated based on the data of each node, and these indicator information can be associated with rules; these indicator information helps to understand the decision logic of the model, evaluate the performance of the model, and provide a basis for subsequent model optimization and improvement.

[0125] Example 8: Figure 8 As shown, based on Example 1, the process of obtaining a policy rule combination provided by the embodiment of the present invention includes the following steps:

[0126] S301: For each rule, apply it to the sample set and calculate the effect after applying the rule, which includes the improvement of prediction accuracy and the reduction of risk;

[0127] S302: For the calculated rule effects, filter out the rules whose effects are greater than a specified threshold; and use the filtered out rules as a candidate set of policy rules;

[0128] S303: For the candidate rules, select evaluation indicators according to the credit business indicators, and for each rule, calculate its score or value on the specified indicator; sort the rules according to the indicator score or value, from high to low.

[0129] The working principle and beneficial effects of the above technical solution are as follows: First, for each rule, this embodiment applies it to the sample set and calculates the effect after applying the rule, which includes the improvement of prediction accuracy and risk reduction, etc.; secondly, for the calculated rule effect, the rules whose effect is greater than a specified threshold are screened out; the screened rules are used as candidate sets of strategy rules; finally, for the candidate rules, evaluation indicators are selected according to the credit business indicators, and for each rule, its score or value on the specified indicator is calculated; the rules are sorted from high to low according to the indicator score or value. Step S301 of the above scheme evaluates the improvement of prediction accuracy and risk reduction of the rules on the sample set by applying the rules; understands the degree of influence of each rule on the decision result, and provides a basis for rule screening and sorting. Step S302 determines the rules with obvious effects by screening out the rules whose effects are greater than a specified threshold, and uses them as candidate sets for subsequent strategy rules; improves the quality and effectiveness of strategy rules, reduces the redundancy and noise of rules, and improves the accuracy and effect of decision-making. Step S303 sorts the rules and determines the order of merit of the rules by selecting evaluation indicators and calculating scores; an ordered list of rules is provided for decision-making, which facilitates decision-makers to select the optimal combination of rules and improves the accuracy and efficiency of decision-making.

[0130] In summary, through the above steps, this embodiment can obtain a set of high-quality strategy rules by calculating the rule effects, screening rules and sorting rules; these rules can be used for credit risk control decisions, help improve the accuracy and effectiveness of decisions, and provide reference and basis for business decisions. At the same time, high-quality strategy rules can also reduce redundancy and noise in the decision-making process, and improve the efficiency and explainability of decisions. Through these steps, the rules can be screened and sorted according to the effects of the rules and credit business indicators to obtain a combination of a set of strategy rules; these strategy rules can be used for credit risk control decisions, help improve the accuracy and effectiveness of decisions, and provide reference and basis for business decisions.

[0131] Example 9: Fig. 9 As shown, based on Embodiment 1 to Embodiment 8, the decision tree-based policy rule generation system provided by the embodiment of the present invention includes:

[0132] The model building module is responsible for obtaining the historical credit performance data of credit institutions, processing and sampling the historical credit performance data according to business scenarios to obtain the original data set; adjusting the original data set, establishing a decision tree model, and dividing the data set according to variable characteristics;

[0133] The information acquisition module is responsible for mining the rules of all nodes of the decision tree based on the decision tree model, selecting calculation indicators based on the credit business scenario, and obtaining the rules and their corresponding indicator information;

[0134] The rule output module is responsible for screening and sorting the rules output by the decision tree model in combination with credit business indicators to obtain a strategy rule combination.

[0135] The working principle and beneficial effects of the above technical solution are as follows: the model building module of this embodiment obtains the historical credit performance data of the credit institution, processes and samples the historical credit performance data according to the business scenario, and obtains the original data set; adjusts the original data set, establishes a decision tree model, and divides the data set according to the variable characteristics; the information acquisition module mines the rules of all nodes of the decision tree based on the decision tree model, selects calculation indicators in combination with the credit business scenario, and obtains the rules and their corresponding indicator information; the rule output module screens and sorts the rules output by the decision tree model in combination with the credit business indicators to obtain the strategy rule combination. The above scheme obtains a representative data set for subsequent modeling and analysis; ensures that the training and evaluation of the decision tree model are based on real and reliable data, and improves the accuracy and credibility of the decision tree model. Obtain the rules of each node in the decision tree model and the indicator information related to it; extracts highly interpretable rules from the decision tree model to help understand and explain the decision-making process of the model. At the same time, by selecting calculation indicators according to the credit business scenario, the rules can be better mined and optimized according to business needs. The rules are screened and sorted according to the credit business indicators to obtain the final strategy rule combination; the rules output by the decision tree model are combined with the business indicators to ensure that the generated rules meet the actual business needs and risk control requirements. At the same time, the generation of strategy rule combinations can help credit institutions formulate effective risk control strategies and improve the accuracy of loan decisions and risk control capabilities.

[0136] In summary, the decision tree-based policy rule generation method of this embodiment can provide policy rules that are highly interpretable and combined with business indicators through steps such as data processing, model building, rule mining and screening, thereby providing credit institutions with more accurate and efficient risk control and decision support.

[0137] This embodiment is a method for efficiently utilizing a decision tree model to efficiently mine risk control policy rules. The innovative decision tree method also improves the efficiency and effectiveness of rule set modeling. Policy mining can assist or replace manual combination of policy rules, improve the degree of automation and search efficiency of the policy, and reduce the dependence of policy development on developer experience and business background.

[0138] Example 10: Fig.10As shown, based on Example 9, the model building module provided in this embodiment of the present invention includes:

[0139] The data acquisition submodule is responsible for obtaining historical customer credit performance data from credit institutions and adjusting variable data types according to actual business scenarios;

[0140] The stratified sampling submodule is responsible for stratifying the data samples according to the proportion of each type of samples in the credit performance data and calculating the weight variables of the sampled data;

[0141] The data partitioning submodule is responsible for partitioning the sampled data into training sets and test sets based on the specified method to obtain a usable original data set;

[0142] The discrete processing submodule is responsible for adjusting the original data by discretizing the continuous numerical variables based on the original data set; based on the adjusted data set, the numerical variables are kept unchanged, the character and Boolean prediction variables are processed, the variable values ​​are sorted according to the proportion of customers with good credit performance in the sample, and the original values ​​are replaced by serial numbers to obtain the processed data set;

[0143] The node splitting submodule is responsible for using the processed data set to establish the root node of the decision tree model; for each unsplit node in the decision tree model, it is determined whether the node data can be split. Splitting means that the data set of the current node is not empty and the value of the target variable of the current node is not unique;

[0144] The node judgment submodule is responsible for repeating the above process of judging whether it can be divided again, searching for the division point and dividing the lower layer for each newly generated node until all nodes cannot be divided. Each node in the final decision tree (except the root node) is a rule.

[0145] The working principle and beneficial effects of the above technical solution are as follows: the data acquisition submodule of this embodiment obtains the credit performance data of historical customers from the credit institution, and adjusts the variable data type according to the actual business scenario; the stratified sampling submodule performs stratified sampling on the data samples according to the proportion of each type of samples in the credit performance data, and calculates the weight variables of the sampled data; the data partitioning submodule divides the sampled data into training sets and test sets based on the specified method to obtain a usable original data set; the discrete processing submodule adjusts the original data by discretizing the continuous numerical variables based on the original data set; based on the adjusted data set, the numerical variables are kept unchanged, and the character and Boolean prediction variables are processed. Processing is performed, and the variable values ​​are sorted according to the proportion of customers with good credit performance in the sample, and the original values ​​are replaced by serial numbers to obtain the processed data set; the node splitting submodule uses the processed data set to establish the root node of the decision tree model; for each node in the decision tree model that has not been split, it is determined whether the node data can be split. Splitting means that the data set of the current node is not empty and the value of the target variable of the current node is not unique; the node judgment submodule repeats the above process of judging whether it can be split, searching for split points and splitting the lower layer for each newly generated node until all nodes cannot be split, and each node in the final decision tree (except the root node) is a rule. The above scheme obtains representative historical credit performance data and adjusts the variable data type to meet the needs of subsequent analysis and modeling; ensures the accuracy and consistency of the data, and provides a reliable basis for subsequent data processing and analysis. Stratified sampling is performed according to the sample ratio to obtain a representative sample set and generate weight variables; through stratified sampling, it is ensured that the sample set can accurately reflect the overall data distribution, and the introduction of weight variables can be used to weight the samples in the subsequent modeling process. The sampled data is divided in a specified way to obtain a data set for training and testing; the training set is used to build a decision tree model, and the test set is used to evaluate the performance of the model; by dividing the data set, the model can be trained and verified during the model building process, thereby evaluating the accuracy and generalization ability of the model. Discretize the original data set and convert character and Boolean predictor variables; discretize continuous variables to make them more suitable for building a decision tree model; at the same time, by sorting the variable values ​​according to the proportion of customers with good credit performance, the model's ability to identify important features can be improved. The root node of the decision tree is established based on the processed data set, and it is determined whether each unsplit node can be split; the initial structure of the decision tree is established, and the nodes that can be split are determined to provide a basis for the subsequent splitting process. Based on the initial root node, the process of repeatedly judging, searching for split points, and splitting the lower layer is repeated for each splittable node until all nodes are unsplittable, thereby obtaining the final decision tree model; based on the division and splitting of the data, a series of nodes and rules are generated to explain and predict the results of credit decisions.

[0146] In summary, this embodiment obtains available data sets that meet business needs, and performs data processing, sampling and division to provide a basis for the subsequent establishment and evaluation of decision tree models; ensures the quality and representativeness of data, improves the reliability and generalization ability of the model, and provides accuracy and reliability support for credit risk control decisions. This embodiment clarifies that the data set must include predictor variables, target variables, weight variables, and the meaning of each part; and standardizes the data format by adjusting the data format in combination with business scenarios and actual conditions.

[0147] This embodiment processes, divides and segments the data, and finally generates the structure and rules of the decision tree model; through the decision tree model, decisions and predictions can be made based on the customer's credit performance and other characteristics, providing accuracy and reliability support for credit risk control and decision-making. This embodiment is a method for efficiently establishing a decision tree. By sorting discrete attributes in combination with business scenarios, prior knowledge is incorporated to increase the reliability of segmentation; by discretizing continuous attributes, the range of data values ​​is reduced, the complexity of searching for segmentation points is reduced, and the efficiency of the model is improved.

[0148] Example 11: Fig.11 As shown, based on Example 9, the information acquisition module provided by the embodiment of the present invention includes:

[0149] The node initialization submodule is responsible for initializing the root node and setting the root node rule to empty. Starting from the root node, for each node except the root node, the node rule is the rule of its parent node plus the variable segmentation condition from its parent node to the current node.

[0150] The rule processing submodule is responsible for merging and organizing the obtained rules, taking the intersection of different segmentation conditions of the same variable contained in the rules; sorting each segmentation condition contained in the organized rules in alphabetical order according to the name of the segmentation variable; accessing each node in the trained decision tree model based on the breadth-first order, and obtaining the rules of all nodes except the root node;

[0151] The indicator calculation submodule is responsible for selecting calculation indicators based on the data of each node and combining the credit business scenario, calculating relevant indicator information, and obtaining rules and their corresponding indicator information.

[0152] The working principle and beneficial effects of the above technical solution are as follows: the node initialization submodule of this embodiment initializes the root node and sets the root node rule to be empty; starting from the root node, for each node except the root node, the node rule is the rule of its parent node plus the variable segmentation condition from its parent node to the current node; the rule processing submodule merges and organizes the obtained rules, and takes the intersection of different segmentation conditions of the same variable contained in the rules; for the sorted rules, each segmentation condition contained therein is sorted in alphabetical order according to the name of the segmentation variable; each node in the trained decision tree model is accessed in a breadth-first order to obtain the rules of all nodes except the root node; the indicator calculation submodule selects the calculation indicator based on the data of each node, combines the credit business scenario, calculates the relevant indicator information, and obtains the rule and its corresponding indicator information. The above scheme establishes the rules of each node in the decision tree model; determines the segmentation rule of each node for subsequent calculation of indicators and sorting of rules. The rules are sorted and sorted for subsequent calculation of indicators and use; the expression form of the rules is simplified, repeated segmentation conditions are removed, and the consistency and readability of the rules are ensured. Calculate indicator information based on node data, such as the number of node samples, the distribution of target variables, accuracy, etc.; evaluate the importance and quality of each node to help understand the decision-making process of the model and provide a basis for subsequent model evaluation and optimization.

[0153] In summary, this embodiment establishes the rules of each node in the decision tree model and calculates the relevant index information; the rules and index information help to understand the decision logic of the model, explain the results of the model, evaluate the performance of the model, and provide guidance for the optimization and improvement of the model; through these steps, the decision tree model can be transformed into a series of rules that are easy to understand and explain, and provide guidance for credit business decisions; all nodes in the rule set can be effectively and comprehensively traversed, and the index performance of the rules can be calculated in combination with business needs. The rule set obtained by the decision tree model is integrated, which reduces the complexity of the rules, increases orderliness, and is more convenient for observation and comparison between rules.

[0154] Example 12: Fig.12 As shown, based on Example 9, the rule output module provided in this embodiment of the present invention includes:

[0155] The effect calculation submodule is responsible for applying each rule to the sample set and calculating the effect after applying the rule, which includes the improvement of prediction accuracy and the reduction of risk.

[0156] The rule screening submodule is responsible for screening out the rules whose effects are greater than a specified threshold from the calculated rule effects; the screened rules are used as candidate sets of policy rules;

[0157] The indicator evaluation submodule is responsible for selecting evaluation indicators for candidate rules based on credit business indicators, calculating the score or value of each rule on the specified indicator, and sorting the rules from high to low based on the indicator score or value.

[0158] The working principle and beneficial effects of the above technical solution are as follows: the effect calculation submodule of this embodiment applies each rule to the sample set and calculates the effect after applying the rule, which includes the improvement of prediction accuracy and risk reduction, etc.; the rule screening submodule screens out the rules whose effects are greater than a specified threshold for the calculated rule effects; the screened rules are used as candidate sets of strategy rules; the indicator evaluation submodule selects evaluation indicators for the candidate rules according to the credit business indicators, and calculates the score or value of each rule on the specified indicator; the rules are sorted from high to low according to the indicator score or value. The above scheme evaluates the improvement of prediction accuracy and risk reduction of the rules on the sample set by applying the rules; understands the degree of influence of each rule on the decision result, and provides a basis for rule screening and sorting. Screen out the rules whose effects are greater than a specified threshold, determine the rules with obvious effects, and use them as candidate sets for subsequent strategy rules; improve the quality and effectiveness of strategy rules, reduce the redundancy and noise of rules, and improve the accuracy and effect of decision-making. By selecting evaluation indicators and calculating scores, the rules are sorted and the order of the advantages and disadvantages of the rules is determined; an ordered list of rules is provided for decision-making, which facilitates decision-makers to choose the optimal combination of rules and improves the accuracy and efficiency of decision-making.

[0159] In summary, through the above steps, this embodiment can obtain a set of high-quality strategy rules by calculating the rule effects, screening rules and sorting rules; these rules can be used for credit risk control decisions, help improve the accuracy and effectiveness of decisions, and provide reference and basis for business decisions. At the same time, high-quality strategy rules can also reduce redundancy and noise in the decision-making process, and improve the efficiency and explainability of decisions. Through these steps, the rules can be screened and sorted according to the effects of the rules and credit business indicators to obtain a combination of a set of strategy rules; these strategy rules can be used for credit risk control decisions, help improve the accuracy and effectiveness of decisions, and provide reference and basis for business decisions.

[0160] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of the present invention and its equivalent technologies, the present invention is also intended to include these changes and variations.

Claims

1. A decision tree-based strategy rule generation method, characterized in that: The following steps are involved: Obtain historical credit performance data from credit institutions, process and sample the data based on business scenarios to obtain the original data set; and establish a decision tree model. Based on the decision tree model, the rules of all nodes of the decision tree are mined, and calculation indicators are selected based on the credit business scenario to obtain the rules and their corresponding indicator information; Combined with credit business indicators, the rules output by the decision tree model are screened and sorted to obtain a strategy rule combination.

2. The decision tree-based policy rule generation method according to claim 1, wherein: The process of obtaining the original dataset includes the following steps: Obtain historical customer credit performance data from credit institutions and adjust variable data types based on actual business scenarios; Conduct stratified sampling of data samples according to the proportion of each type of sample in the credit performance data, and calculate the weight variables of the sampled data; The sampled data is divided into training set and test set to obtain the available original data set.

3. The decision tree-based policy rule generation method according to claim 2, wherein: The predictor variable is the behavior information of the credit user, and its value type is numeric, character, or Boolean.

4. The decision tree-based policy rule generation method according to claim 2, wherein: The target variable is a binary categorical variable based on whether the credit customer is overdue. It is expressed as good or bad. Good means that the user has no overdue behavior, and bad means that the user has overdue behavior.

5. The decision tree-based policy rule generation method according to claim 2, wherein: in, The training set is the data set used to train and generate the decision tree model, and the test set is the data set used to test and measure the effectiveness of the model.

6. The decision tree-based policy rule generation method according to claim 1, wherein: The original data set is adjusted, a decision tree model is established, and the data set is divided according to variable characteristics.

7. The decision tree-based policy rule generation method according to claim 1, wherein: The process of obtaining a policy rule combination includes the following steps: For each rule, apply it to the sample set and calculate the effect of applying the rule, which includes the improvement of prediction accuracy and the reduction of risk; For the calculated rule effects, the rules whose effects are greater than a specified threshold are filtered out; the filtered rules are used as candidate sets of policy rules; For the candidate rules, select evaluation indicators based on the credit business indicators, and calculate the score or value of each rule on the specified indicator; sort the rules according to the indicator score or value, from high to low.

8. A decision tree-based policy rule generation system, characterized in that: Include: The model building module is responsible for obtaining historical credit performance data from credit institutions, processing and sampling the historical credit performance data based on business scenarios to obtain the original data set; adjusting the original data set, building a decision tree model, and dividing the data set according to variable characteristics; The information acquisition module is responsible for mining the rules of all nodes in the decision tree based on the decision tree model, selecting calculation indicators based on the credit business scenario, and obtaining the rules and their corresponding indicator information; The rule output module is responsible for filtering and sorting the rules output by the decision tree model in combination with credit business indicators to obtain a strategy rule combination.

9. The decision tree-based policy rule generation system according to claim 8, characterized in that: Model building modules, including: The data acquisition submodule is responsible for obtaining historical customer credit performance data from credit institutions and adjusting variable data types based on actual business scenarios; The stratified sampling submodule is responsible for stratifying the data samples according to the proportion of each type of samples in the credit performance data and calculating the weight variables of the sampled data; The data partitioning submodule is responsible for dividing the sampled data into training sets and test sets based on the specified method to obtain a usable original data set.

10. The decision tree-based policy rule generation system according to claim 8, characterized in that: Rule output module, including: The effect calculation submodule is responsible for applying each rule to the sample set and calculating the effect of the rule application, which includes the improvement of prediction accuracy and the reduction of risk. The rule screening submodule is responsible for screening out the rules whose effects are greater than a specified threshold from the calculated rule effects, and using the screened rules as the candidate set of policy rules; The indicator evaluation submodule is responsible for selecting evaluation indicators for candidate rules based on credit business indicators, calculating the score or value of each rule on the specified indicator, and sorting the rules from high to low based on the indicator score or value.

Images