Method, device and equipment for identifying untrustworthy customers and storage medium

Abstract

The application discloses a method, device and equipment for identifying a dishonest customer and a storage medium. The method comprises the following steps: obtaining a plurality of pieces of to-be-identified data from a database when an identification instruction is acquired; selecting field information corresponding to each piece of to-be-identified data according to a preset template from the data; generating feature information according to the field information; inputting the feature information into a pre-trained target fusion model to obtain an output bad customer concentration mining result, and determining a corresponding dishonest customer according to the result, wherein the target fusion model is formed by superimposed fusion of a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model and a preset gradient boosting model. In the foregoing manner, the bad customer concentration is mined based on the model formed by superimposed fusion of the decision tree model, the Prim algorithm model, the logistic regression model and the gradient boosting model, so that the problem that the bad customer concentration identified when a single model is used to mine bad customers is low and the identification precision is not high is solved.

Description

Technical Field

[0001] This invention relates to the field of risk control management technology, and in particular to a method, apparatus, equipment and storage medium for identifying dishonest customers. Background Technology

[0002] Traditional risk control technology relies heavily on the experience and manual operation of risk control teams within various institutions. However, with the continuous development of technologies such as the internet and big data, traditional risk control methods are gradually becoming unable to support the business expansion of institutions. Big data, with its intelligent processing of multi-dimensional and massive amounts of data and standardized batch execution processes, is better suited to the development requirements of risk control in the information age. In the process of loan applications from micro and small enterprises, to ensure that all applicants are good customers (i.e., those who can repay loans), there is an urgent need for financial platforms to design a model that can identify all defaulting customers (also known as bad customers, i.e., those who do not repay loans). Generally, a single model is used for information mining, but this method results in a low concentration of bad customers, meaning low identification accuracy. Bad customer concentration refers to the ratio of identified bad customers to the total number of bad customers when faced with a large amount of data. The above content is only for assisting in understanding the technical solution of this invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0003] The main objective of this invention is to provide a method, apparatus, device, and storage medium for identifying dishonest customers, aiming to solve the technical problem that the concentration of identified bad customers is low and the identification accuracy is not high when using a single model to mine bad customers.

[0004] To achieve the above objectives, the present invention provides a method for identifying dishonest customers, the method comprising the following steps:

[0005] Upon receiving an instruction to identify a dishonest customer, several records of data to be identified are retrieved from the database;

[0006] Select the field information corresponding to each piece of data to be identified from the plurality of data to be identified according to the preset field selection template;

[0007] Generate feature information based on the field information;

[0008] The feature information is input into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model. The corresponding defaulting customers are determined based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model.

[0009] Optionally, before retrieving several pieces of data to be identified from the database upon receiving the instruction to identify a dishonest customer, the method further includes:

[0010] Retrieve all field information for each of the several customers from the dataset;

[0011] Perform dummy variable transformation on all the field information to obtain all feature information;

[0012] Based on all the feature information, determine the prediction capability information value corresponding to each field item;

[0013] Based on the predicted ability information value, each field item is filtered to obtain the remaining field items;

[0014] Generate a preset field selection template based on the remaining field items.

[0015] Optionally, the step of filtering each field item based on the predictive ability information value to obtain the remaining field items includes:

[0016] The prediction capability information value corresponding to each field item is compared with a preset numerical range to obtain the comparison result;

[0017] Based on the comparison results, each field item is filtered to obtain the remaining field items.

[0018] Optionally, after performing dummy variable transformation on all field information to obtain all feature information, the method further includes:

[0019] Write the information of each field and the corresponding feature information into the preset feature transformation mapping table;

[0020] The step of generating feature information based on the field information includes:

[0021] The corresponding feature information is determined by searching the preset feature transformation mapping table based on the field information.

[0022] Optionally, before retrieving several pieces of data to be identified from the database upon receiving the instruction to identify a dishonest customer, the method further includes:

[0023] Obtain the preset initial sample set;

[0024] Based on the preset field selection template, select from the preset initial sample set to obtain several sample data;

[0025] The initial fusion model is trained based on the aforementioned sample data to obtain the trained target fusion model.

[0026] Optionally, training the initial fusion model based on the plurality of sample data to obtain the trained target fusion model includes:

[0027] Based on the aforementioned sample data, a training set and a test set are randomly generated according to a preset ratio.

[0028] Based on the pre-trained target decision tree model, target Prim algorithm model, target logistic regression model and target gradient boosting model, predictions are made on the training set and the test set respectively to obtain a first prediction result and a second prediction result.

[0029] The training set is adjusted based on the first prediction result to generate a target training set;

[0030] The test set is adjusted based on the second prediction result to generate a target validation set;

[0031] The initial fusion model is trained based on the target training set and validated based on the target validation set. When the validation is successful, the trained target fusion model is obtained.

[0032] Optionally, before predicting the training set and the test set respectively based on the pre-trained target decision tree model, target Prim algorithm model, target logistic regression model, and target gradient boosting model to obtain the first prediction result and the second prediction result, the method further includes:

[0033] Obtain the configuration parameters corresponding to each algorithm model;

[0034] Configure the training conditions for each algorithm model according to the configuration parameters;

[0035] Under the training conditions, the initial decision tree model, the initial Prim algorithm model, the initial logistic regression model, and the initial gradient boosting model are trained according to the training set to obtain the target decision tree model, the target Prim algorithm model, the target logistic regression model, and the target gradient boosting model.

[0036] Furthermore, to achieve the above objectives, the present invention also proposes a device for identifying dishonest customers, the device comprising:

[0037] The acquisition module is used to retrieve several pieces of data to be identified from the database when a command to identify dishonest customers is received.

[0038] The selection module is used to select the field information corresponding to each piece of data to be identified from the plurality of data to be identified according to a preset field selection template.

[0039] The feature generation module is used to generate feature information based on the field information;

[0040] The identification module is used to input the feature information into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model, and to determine the corresponding defaulting customers based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model.

[0041] Furthermore, to achieve the above objectives, the present invention also proposes a device for identifying dishonest customers, the device comprising: a memory, a processor, and a dishonest customer identification program stored in the memory and executable on the processor, the dishonest customer identification program being configured to implement the dishonest customer identification method described above.

[0042] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a dishonest customer identification program, which, when executed by a processor, implements the dishonest customer identification method as described above.

[0043] This invention, upon receiving a command to identify dishonest customers, retrieves several data entries from a database; selects corresponding field information from each data entry according to a preset field selection template; generates feature information based on the field information; and inputs the feature information into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model. The corresponding dishonest customers are then identified based on this result. The target fusion model is composed of a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model, all fused together. This method, by selecting appropriate field information based on the preset field selection template and mining the bad customer concentration using the fused model (decision tree model, Prim algorithm model, logistic regression model, and gradient boosting model), identifies dishonest customers based on their corresponding data. This solves the problem of low bad customer concentration and low identification accuracy when using a single model to mine bad customers. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the structure of a dishonest customer identification device in the hardware operating environment involved in the embodiments of the present invention;

[0045] Figure 2 This is a flowchart illustrating the first embodiment of the method for identifying dishonest customers according to the present invention.

[0046] Figure 3 This is a flowchart illustrating the second embodiment of the method for identifying dishonest customers according to the present invention.

[0047] Figure 4 This is a flowchart illustrating the third embodiment of the method for identifying dishonest customers according to the present invention.

[0048] Figure 5 This is a decision tree diagram of an embodiment of the method for identifying dishonest customers according to the present invention;

[0049] Figure 6 This is a structural block diagram of the first embodiment of the dishonest customer identification device of the present invention.

[0050] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0051] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0052] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a defaulting customer identification device in the hardware operating environment involved in an embodiment of the present invention. Figure 1 As shown, the device for identifying dishonest customers may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to establish communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001. Those skilled in the art will understand that… Figure 1 The structure shown does not constitute a limitation on the identification device for dishonest customers and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0053] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a program for identifying untrustworthy customers.

[0054] exist Figure 1In the shown dishonest customer identification device, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the dishonest customer identification device of the present invention can be set in the dishonest customer identification device, and the dishonest customer identification device calls the dishonest customer identification program stored in the memory 1005 through the processor 1001 and executes the dishonest customer identification method provided in the embodiment of the present invention.

[0055] This invention provides a method for identifying dishonest customers, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the method for identifying dishonest customers according to the present invention.

[0056] In this embodiment, the method for identifying dishonest customers includes the following steps:

[0057] Step S10: Upon receiving the instruction to identify dishonest customers, retrieve several pieces of data to be identified from the database.

[0058] It is understood that the execution subject of this embodiment is a dishonest customer identification device. The dishonest customer identification device can be a computer, a server, or other devices with reasoning and computing capabilities. This embodiment does not limit the scope of such devices.

[0059] It should be noted that the bad customer identification instruction is an identification instruction entered by the administrator in the system. It is used to mine the concentration of bad customers and analyze the distribution of bad customers in the existing customer data. The database is the local big data platform database. Specifically, the bad customer identification device is equipped with a financial platform. After a micro-enterprise applies on the front-end page of the financial platform, it starts a query system based on the financial platform to obtain the relevant field information of the micro-enterprise. The field information is then stored in a structured manner in the local big data platform data warehouse. When a bad customer identification instruction is obtained, several pieces of data to be identified are retrieved from the local big data platform database. These several pieces of data to be identified can be all the customer data stored in the database, or they can be customer data that submitted applications within a certain time period, such as customer data that submitted applications within 5 days. Each piece of data to be identified represents all the field information corresponding to a customer.

[0060] Step S20: Select the field information corresponding to each piece of data to be identified from the plurality of data to be identified according to the preset field selection template.

[0061] Understandably, when identifying dishonest customers, all field information corresponding to all customers or several customers is obtained from the local big data platform database, i.e., several pieces of data to be identified. All field information is filtered according to the preset field selection template to obtain the field information corresponding to each piece of data to be identified. The field information is the indicator value corresponding to each indicator in each piece of data to be identified. For example, if the indicator name is "business owner's education level", the field information corresponding to the customer to be identified is "bachelor's degree".

[0062] In the specific implementation, multiple fields that meet the requirements are selected in advance based on the predictive ability information values ​​corresponding to each field. A preset field selection template is generated based on the multiple fields that meet the requirements, so as to select field information with effective information from all field information corresponding to each data to be identified.

[0063] Step S30: Generate feature information based on the field information.

[0064] It is understandable that feature information is information that the model can calculate. Optionally, dummy variable transformation is performed on field information with character type variables to obtain feature information. For example, if the indicator is named "Education of Business Owners", and the corresponding variables are "High School and Below", "Associate Degree", "Bachelor's Degree" and "Graduate Degree", after dummy variable transformation, the variable mapping value corresponding to the field information "Bachelor's Degree" of the customer to be identified is [0,1,0].

[0065] Step S40: Input the feature information into the pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model, and determine the corresponding defaulting customers based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model.

[0066] Understandably, the target fusion model is composed of a pre-set decision tree model, a pre-set Prim algorithm model, a pre-set logistic regression model, and a pre-set XGBoost model. The target fusion model is trained in advance, and based on the trained target fusion model, the feature information corresponding to each piece of data to be identified is identified to obtain the bad customer concentration mining result. Specifically, the bad customer concentration mining result can be the ratio between the number of currently identified bad customers and the total number of customers to be identified. Because the existing single model has a low bad customer concentration when mining bad customers, for example, if the same batch of data to be identified represents 20 customers, but there are actually 10 bad customers, the single model can only identify 6 bad customers. However, the target fusion model in this embodiment can identify 8 bad customers from 20 customers, improving the identification accuracy. This determines whether the customer corresponding to each piece of data to be identified is a customer who can repay the loan after the application. If not, the corresponding emergency strategy is selected to respond to the customer's loan application on the front-end page of the financial platform, for example, prompting "Application failed".

[0067] This embodiment retrieves several data entries to be identified from a database upon receiving a command to identify dishonest customers. From these entries, field information corresponding to each entry is selected according to a preset field selection template. Feature information is generated based on this field information. This feature information is then input into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model. The corresponding dishonest customers are then identified based on this result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model. By using this method, appropriate field information is selected based on the preset field selection template. The bad customer concentration is mined based on the model obtained by superimposing and fusing the decision tree model, Prim algorithm model, logistic regression model, and gradient boosting model. Dishonest customers are identified based on the data corresponding to each customer, thus solving the problem of low bad customer concentration and low identification accuracy when using a single model to mine bad customers.

[0068] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the method for identifying dishonest customers according to the present invention.

[0069] Based on the first embodiment described above, the method for identifying dishonest customers in this embodiment further includes the following before step S10:

[0070] Step S101: Obtain all field information corresponding to several customers from the dataset.

[0071] It is understood that all field information are the variable values ​​corresponding to each indicator in the enterprise modeling fields. Further, before step S101, all structured fields are randomly obtained from the local big data platform database, and several fields are randomly selected as enterprise modeling fields. Table 1 is an example of the enterprise modeling field table of an embodiment of the present invention. A large amount of sample data is obtained from the local big data platform database, and 16 indicators are randomly selected. Each indicator contains different variable values ​​to construct the enterprise modeling field table.

[0072] Suppose that in the dataset, the education level of business owner A is "Bachelor's Degree" and the location of the business is "Central China"..., while the education level of business owner B is "Master's Degree" and the location of the business is "North China". Retrieve the corresponding field information for business owners A and B from the dataset to obtain the corresponding fields for business owner A ("Bachelor's Degree", "Central China"...) and business owner B ("Master's Degree", "North China"...).

[0073] Table 1:

[0074]

[0075] Step S102: Perform dummy variable transformation on all the field information to obtain all feature information.

[0076] It should be noted that this embodiment performs dummy variable conversion on all character variables in the field information, thereby converting all character variables into variables that the model can calculate. Dummy variables, also called virtual variables, are used to reflect different attributes of a variable. Specifically, refer to Table 2, which is a preset feature conversion mapping table of an example of the present invention, containing variable mapping values ​​corresponding to 5 indicators.

[0077] Table 2:

[0078]

[0079] Specifically, based on all feature information (i.e., the values ​​of each variable), all categories of variables corresponding to each indicator are determined, and dummy variable transformation is performed based on all categories to determine the feature information corresponding to each category of variables.

[0080] Furthermore, after step S102, the method further includes: writing each field information and the corresponding feature information into a preset feature conversion mapping table;

[0081] Step S30 includes: searching the preset feature transformation mapping table based on the field information to determine the corresponding feature information.

[0082] Understandably, referring to Table 2, the field information of each type and the corresponding feature information are stored in a preset feature conversion mapping table. That is, the variables of each type and the corresponding variable mapping values ​​are stored in the preset feature conversion mapping table. When the variable values ​​of various indicators of the customer to be identified are obtained, the corresponding variable mapping values ​​are found in the preset feature conversion mapping table. For example, the field information corresponding to the education level indicator of the business owner of the customer to be identified is "Bachelor's Degree". The preset feature conversion mapping table is searched to determine the corresponding feature information as "0, 1, 0".

[0083] Step S103: Determine the prediction capability information value corresponding to each field item based on all the feature information.

[0084] It should be noted that the predictive ability information value is the IV value (Information Value), which is used to represent the predictive strength of the corresponding feature information. In the specific implementation, the IV value of each feature information is determined based on the WOE calculation. WOE stands for "weight of evidence". The WOE calculation can be understood as encoding a certain feature information using WOE. First, the feature information is grouped. For the i-th group, the corresponding WOE is calculated according to formula (1):

[0085]

[0086] Among them, #B i This is the number of bad customers in the i-th box, #G i This is the number of good customers in the i-th box, #B T It is the total number of bad customers in the total sample of all groups, #G T This represents the total number of good customers across all groups. WOE essentially indicates the difference in the proportion of good and bad customers in each group relative to the total number of good and bad customers. A larger absolute value of WOE indicates a more significant difference, while a smaller absolute value indicates a less significant difference. Assume variable x has n groups, and the WOE value for the i-th group is WOE. i Calculate the IV value corresponding to variable x according to formula (2):

[0087]

[0088] The percentage of bad customers is the ratio between the number of bad customers in the previous group and the total number of bad customers in the overall sample, while the percentage of good customers is the ratio between the number of good customers in the current group and the total number of good customers in the overall sample.

[0089] Step S104: Filter each field item according to the prediction ability information value to obtain the remaining field items.

[0090] It should be understood that, based on the predictive ability information value, suitable field items are selected from each field item to obtain the remaining field items. Specifically, step S104 includes: comparing the predictive ability information value corresponding to each field item with a preset numerical range to obtain a comparison result; and filtering each field item based on the comparison result to obtain the remaining field items.

[0091] It should be noted that, referring to Table 3 for example, Table 3 is an IV value distribution table for an example of the present invention. When the IV value of a field item is less than 0.02, it indicates that the field item does not have a significant contribution and should be removed from the dataset. When the IV value of a field item is between 0.02 and 0.5, the feature has a significant contribution, and the larger the IV value, the better the model performance after the corresponding feature information is input into the model. When the IV value of a field item is greater than 0.5, the feature is so significant that it will break through the model's performance after being input into the model, causing the model to overfit, and therefore the field item should be removed. In this embodiment, the preset numerical range is set to 0.02 to 0.5.

[0092] Table 3:

[0093] Variable name IV value Average monthly contribution of enterprise housing provident fund in the past two years 0.48 Monthly income 0.47 Details of corporate housing provident fund in the past year 0.42 Marital status of business owners 0.28 Current ratio 0.28 Debt-to-asset ratio 0.26 Enterprise size 0.21 The region where the company is located 0.17 Business owner's education 0.15 Daily average balance 0.04 Years of operation 0.03 Enterprise Type 0.01 Monthly billing amount 0.01 Average daily interest 0.01 Quick ratio 0.01 Cash Ratio 0.01

[0094] Referring to Table 3 for examples, the remaining fields include: average monthly contribution per employee to the enterprise's housing provident fund over the past two years, monthly income amount, details of whether the enterprise has a housing provident fund in the past year, marital status of the business owner, current ratio, debt-to-equity ratio, enterprise size, region where the enterprise is located, education level of the business owner, daily average balance, and years of operation.

[0095] Step S105: Generate a preset field selection template based on the remaining field items.

[0096] Understandably, using Table 3 as an example, a preset field selection template is generated based on the remaining fields: average monthly contribution to enterprise housing provident fund per person in the past two years, monthly income amount, details of whether or not there is enterprise housing provident fund in the past year, marital status of the business owner, current ratio, debt-to-asset ratio, enterprise size, enterprise location, business owner's education level, average daily balance, and years of operation. When selecting the field information corresponding to the customer to be identified from the database according to the preset field selection target, the field information corresponding to the average monthly contribution to enterprise housing provident fund per person in the past two years, the field information corresponding to the monthly income amount, etc., are selected from all fields, thereby obtaining feature information with significant contribution, which improves the model's recognition accuracy.

[0097] This embodiment obtains all field information corresponding to several customers from a dataset; performs dummy variable transformation on all field information to obtain all feature information; determines the predictive ability information value corresponding to each field item based on all feature information; filters each field item based on the predictive ability information value to obtain the remaining field items; generates a preset field selection template based on the remaining field items; when a bad customer identification instruction is obtained, several pieces of data to be identified are obtained from the database; field information corresponding to each piece of data to be identified is selected from the several pieces of data to be identified according to the preset field selection template; feature information is generated based on the field information; the feature information is input into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model; and the corresponding bad customer is determined based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model. The above method calculates the predictive ability information value of each field, generates a preset field selection template based on the fields with strong predictive ability, and selects the fields with strong predictive ability from all fields of each data to be identified based on the preset field selection template, providing more contributing feature information for model identification. The model is based on the superposition and fusion of decision tree model, Prim algorithm model, logistic regression model and gradient boosting model to mine bad customer concentration. Based on the data to be identified for each customer, the dishonest customers are identified, which solves the problem of low bad customer concentration and low identification accuracy when using a single model to mine bad customers.

[0098] refer to Figure 4 , Figure 4 This is a flowchart illustrating the third embodiment of the method for identifying dishonest customers according to the present invention.

[0099] Based on the first embodiment described above, the method for identifying dishonest customers in this embodiment further includes the following before step S10:

[0100] Step S01: Obtain the preset initial sample set.

[0101] Understandably, the initial sample set contains a large amount of sample data, including field information corresponding to all field items and annotation information, which is the workers' marking of good or bad customers.

[0102] Step S02: Select the preset initial sample set according to the preset field selection template to obtain several sample data.

[0103] It should be noted that, based on the preset field selection template, the more contributing field items are selected from all the field items corresponding to each sample to obtain several sample data. Each sample data contains multiple field information and annotation information. Referring to Table 3, after the preset field selection template, each sample data includes the field information corresponding to the average monthly contribution of enterprise housing provident fund in the past two years, but does not include the field information corresponding to the cash ratio.

[0104] Step S03: Train the initial fusion model based on the aforementioned sample data to obtain the trained target fusion model.

[0105] It should be noted that before training, the field information of character-type variables in each sample data is determined by looking up the pre-defined feature transformation mapping table to determine the corresponding feature information. The feature information corresponding to several sample data is input into the initial fusion model to obtain the model prediction results. The loss value is determined based on the annotation information and the prediction results. The internal parameters of the initial fusion model are adjusted based on the loss value to train the initial fusion model.

[0106] Further, step S03 includes: randomly generating a training set and a test set according to a preset ratio based on the plurality of sample data; predicting the training set and the test set respectively based on a pre-trained target decision tree model, a target Prim algorithm model, a target logistic regression model, and a target gradient boosting model to obtain a first prediction result and a second prediction result; adjusting the training set based on the first prediction result to generate a target training set; adjusting the test set based on the second prediction result to generate a target validation set; training the initial fusion model based on the target training set and validating it based on the target validation set; and obtaining the trained target fusion model when the validation passes.

[0107] Understandably, several sample data points are randomly split into a training set and a test set according to a preset ratio, such as 7:3. The trained target decision tree model, target prim model, target logistic regression model, and target XGBoost model predict the Training Data and Test Data respectively, obtaining the prediction results for each model. Multiple first prediction results from each model on the training set are combined to form new features, which serve as the new training set. The average of multiple second prediction results from each model on the test set is taken to obtain a new validation set. The initial fusion model is trained using the new target training set, and validated using the target validation set, resulting in the trained target fusion model.

[0108] Furthermore, before predicting the training set and the test set respectively based on the pre-trained target decision tree model, target Prim algorithm model, target logistic regression model, and target gradient boosting model to obtain the first prediction result and the second prediction result, the method further includes: obtaining the configuration parameters corresponding to each algorithm model; configuring the training conditions corresponding to each algorithm model according to the configuration parameters; and training the initial decision tree model, initial Prim algorithm model, initial logistic regression model, and initial gradient boosting model respectively based on the training set under the training conditions to obtain the target decision tree model, the target Prim algorithm model, the target logistic regression model, and the target gradient boosting model.

[0109] It should be noted that, referring to Table 4, which is a table of remaining field items and English variable representations for an example of the present invention, in this embodiment, Chinese variables are named as English variables according to the full or abbreviation of the corresponding Chinese names.

[0110] Table 4:

[0111] Variable name English name IV value Average monthly contribution of enterprise housing provident fund in the past two years bus_accfund_pay_amount_2y 0.48 Monthly income Mon_rece_amount 0.47 Details of corporate housing provident fund in the past year bus_accfund_detai_1y 0.42 Marital status of business owners bus_owners_mar 0.28 Current ratio current_ratio 0.28 Debt-to-asset ratio Asset_liability_ratio 0.26 Enterprise size bus_scale 0.21 The region where the company is located bus_location 0.17 Business owner's education bus_owners_edu 0.15 Daily average balance balance_day_ave 0.04 Years of operation management_years 0.03

[0112] In this embodiment, the parameters of the decision tree model are specifically set as follows: `max_depth` represents the maximum number of layers in the decision tree; to prevent overfitting, it is set to 4 layers. `min_samples_leaf` represents the number of samples on each leaf of the decision tree; in this embodiment, it is set to 50, meaning that a leaf node will be pruned if it has fewer than 50 samples. `min_samples_split` represents the minimum number of samples required for the decision tree to split and form a new branch; in this embodiment, it is set to 50, meaning that no new tree will be split if it has fewer than 50 samples. (Refer to...) Figure 5 , Figure 5 This is a decision tree diagram of an embodiment of the method for identifying dishonest customers of the present invention. It is generated by training a decision tree model using 11,257 specific sample instances. Figure 5 The bad customer concentration corresponding to the trained target decision tree model is calculated as: (4945*0.074) / (11257*0.046)=70.8%.

[0113] In this embodiment, the parameters of the Prim model are specifically set as follows: `peel_alpha` represents the initial hyperparameters, which are generally initialized in the order of 0.01, 0.05, and 0.1. In this embodiment, `peel_alpha` is set to 0.05; `mass_min` represents the target hit rate, which is set to 0.05 in this embodiment, meaning that regions with a bad customer ratio of less than 5% are automatically removed. Based on a specific sample instance of 11257 rows, the bad customer concentration mined by Prim in this embodiment is 75.2% (Coverage value).

[0114] In this embodiment, the parameters of the logistic regression model are specifically set as follows: the default value of C is 0. The smaller the C value, the stronger the regularization. To prevent overfitting, C is set to 0.1 in this embodiment. The class_weight is set to balanced, meaning that the logistic regression model will calculate the weights based on the training sample size. The more samples of the same type, the lower the weight; the fewer samples, the higher the weight. Referring to Table 5, which is a customer group model report of a logistic regression model for an example, combined with 11,257 rows of sample instances, the customer group model report shown in Table 5 is generated after training the logistic regression model. The first 6 boxes of samples account for 30% of the total number of people, and the bad customers captured account for 69.5% of all bad customers. This means that if the 30% of people with the lowest rejection scores are rejected, a 69.50% concentration of bad customers can be captured.

[0115] Table 5:

[0116]

[0117] In this embodiment, the parameters of the Xgboost model are specifically set as follows: boosting_type: GBDT gradient boosting algorithm is selected in this embodiment; max_depth: represents the tree depth. To prevent overfitting, the tree depth is selected as 2 in this embodiment; to maintain sample balance and prevent overfitting during model training, data sampling subsample and feature sampling colsample_bytree are both set to 0.7; initial learning rate learning_rate: in this embodiment, a relatively high learning rate of 0.1 is initially set, which will gradually flatten out during model iteration; AUC is used as the evaluation metric eval_metric, and the model is stopped after 100 training iterations (early_stopping_rounds). Referring to Table 6, which is an example of the Xgboost model customer group model report, combined with 11,257 specific sample instances, the Xgboost model training generated the customer group model report shown in Table 6. The first 6 boxes of samples account for 30% of the total number of people, and the bad customers captured account for 72.6% of all bad customers. This means that if the 30% of people with the lowest rejection scores are rejected, a 72.6% concentration of bad customers can be captured.

[0118] Table 6:

[0119]

[0120] Compared to single-model decision trees, Prim algorithm, logistic regression, and Xgboost model, the target fusion model in this embodiment predicts a higher concentration of bad customers, reaching 82.8%. The calculated concentration of bad customers is [(70.8% + 75.2% + 69.5% + 72.6%) * (1 + 0.15)] / 4 = 82.8%.

[0121] This embodiment obtains a preset initial sample set; selects several sample data points from the preset initial sample set according to a preset field selection template; trains the initial fusion model based on the sample data to obtain a trained target fusion model; when a bad customer identification instruction is received, several data points to be identified are obtained from the database; field information corresponding to each data point to be identified is selected from the data points to be identified according to the preset field selection template; feature information is generated based on the field information; the feature information is input into the pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model; and the corresponding bad customer is determined based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model. The above method selects contributing fields according to a preset field selection template based on the sample set to obtain sample data. The model based on the superposition and fusion of decision tree model, Prim algorithm model, logistic regression model and gradient boosting model is trained using the sample data to obtain a trained target fusion model. The concentration of bad customers is mined based on the trained target fusion model, and the dishonest customers are identified based on the data to be identified for each customer. This solves the problem of low bad customer concentration and low identification accuracy when using a single model to mine bad customers.

[0122] Furthermore, this embodiment of the invention also proposes a storage medium storing a dishonest customer identification program, which, when executed by a processor, implements the dishonest customer identification method as described above.

[0123] Since this storage medium adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be repeated here.

[0124] Reference Figure 6 , Figure 6 This is a structural block diagram of the first embodiment of the dishonest customer identification device of the present invention.

[0125] like Figure 6 As shown, the defaulter identification device proposed in this embodiment of the invention includes:

[0126] The acquisition module 10 is used to retrieve several pieces of data to be identified from the database when a bad customer identification instruction is received.

[0127] The selection module 20 is used to select the field information corresponding to each piece of data to be identified from the plurality of data to be identified according to a preset field selection template.

[0128] The feature generation module 30 is used to generate feature information based on the field information.

[0129] The identification module 40 is used to input the feature information into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model, and to determine the corresponding dishonest customers based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model.

[0130] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0131] This embodiment retrieves several data entries to be identified from a database upon receiving a command to identify dishonest customers. From these entries, field information corresponding to each entry is selected according to a preset field selection template. Feature information is generated based on this field information. This feature information is then input into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model. The corresponding dishonest customers are then identified based on this result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model. By using this method, appropriate field information is selected based on the preset field selection template. The bad customer concentration is mined based on the model obtained by superimposing and fusing the decision tree model, Prim algorithm model, logistic regression model, and gradient boosting model. Dishonest customers are identified based on the data corresponding to each customer, thus solving the problem of low bad customer concentration and low identification accuracy when using a single model to mine bad customers.

[0132] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0133] In addition, for technical details not described in detail in this embodiment, please refer to the method for identifying dishonest customers provided in any embodiment of the present invention, which will not be repeated here.

[0134] In one embodiment, the device for identifying dishonest customers further includes a template generation module;

[0135] The selection template generation module is used to obtain all field information corresponding to several customers from the dataset, perform dummy variable transformation on the all field information to obtain all feature information, determine the prediction ability information value corresponding to each field item based on the all feature information, filter each field item based on the prediction ability information value to obtain the remaining field items, and generate a preset field selection template based on the remaining field items.

[0136] In one embodiment, the selection template generation module is further configured to compare the prediction capability information value corresponding to each field item with a preset numerical range to obtain a comparison result, and filter each field item according to the comparison result to obtain the remaining field items.

[0137] In one embodiment, the feature generation module 30 is further configured to write each field information and the corresponding feature information into a preset feature conversion mapping table, and search the preset feature conversion mapping table according to the field information to determine the corresponding feature information.

[0138] In one embodiment, the device for identifying dishonest customers further includes a training module;

[0139] The training module is used to obtain a preset initial sample set, select a template based on the preset field to obtain several sample data, and train the initial fusion model based on the several sample data to obtain the trained target fusion model.

[0140] In one embodiment, the training module is further configured to: randomly generate a training set and a test set according to a preset ratio based on the plurality of sample data; predict the training set and the test set respectively based on a pre-trained target decision tree model, a target Prim algorithm model, a target logistic regression model, and a target gradient boosting model to obtain a first prediction result and a second prediction result; adjust the training set according to the first prediction result to generate a target training set; adjust the test set according to the second prediction result to generate a target validation set; train the initial fusion model according to the target training set and validate it according to the target validation set; and obtain the trained target fusion model when the validation passes.

[0141] In one embodiment, the training module is further configured to obtain configuration parameters corresponding to each algorithm model; configure training conditions corresponding to each algorithm model according to the configuration parameters; and train the initial decision tree model, the initial Prim algorithm model, the initial logistic regression model, and the initial gradient boosting model respectively according to the training set under the training conditions to obtain the target decision tree model, the target Prim algorithm model, the target logistic regression model, and the target gradient boosting model.

[0142] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0143] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0145] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for identifying dishonest customers, characterized in that, The method for identifying dishonest customers includes: Upon receiving an instruction to identify a dishonest customer, several records of data to be identified are retrieved from the database; Select the field information corresponding to each piece of data to be identified from the plurality of data to be identified according to the preset field selection template; Generate feature information based on the field information; The feature information is input into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model. The corresponding dishonest customers are determined based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model. The bad customer concentration is the ratio of the number of identified bad customers to the total number of bad customers. Before retrieving several pieces of data to be identified from the database upon receiving the instruction to identify a dishonest customer, the method further includes: Retrieve all field information for each of the several customers from the dataset; Perform dummy variable transformation on all the field information to obtain all feature information; Based on all the feature information, determine the prediction capability information value corresponding to each field item; Based on the predicted ability information value, each field item is filtered to obtain the remaining field items; Generate a preset field selection template based on the remaining field items.

2. The method for identifying dishonest customers as described in claim 1, characterized in that, The process of filtering each field item based on the predictive ability information value to obtain the remaining field items includes: The prediction capability information value corresponding to each field item is compared with a preset numerical range to obtain the comparison result; Based on the comparison results, each field item is filtered to obtain the remaining field items.

3. The method for identifying dishonest customers as described in claim 1, characterized in that, After performing dummy variable transformation on all field information to obtain all feature information, the method further includes: Write the information of each field and the corresponding feature information into the preset feature transformation mapping table; The step of generating feature information based on the field information includes: The corresponding feature information is determined by searching the preset feature transformation mapping table based on the field information.

4. The method for identifying dishonest customers as described in claim 1, characterized in that, Before retrieving several pieces of data to be identified from the database upon receiving the instruction to identify a dishonest customer, the method further includes: Obtain the preset initial sample set; Based on the preset field selection template, select from the preset initial sample set to obtain several sample data; The initial fusion model is trained based on the aforementioned sample data to obtain the trained target fusion model.

5. The method for identifying dishonest customers as described in claim 4, characterized in that, The step of training the initial fusion model based on the aforementioned sample data to obtain the trained target fusion model includes: Based on the aforementioned sample data, a training set and a test set are randomly generated according to a preset ratio. Based on the pre-trained target decision tree model, target Prim algorithm model, target logistic regression model and target gradient boosting model, predictions are made on the training set and the test set respectively to obtain a first prediction result and a second prediction result. The training set is adjusted based on the first prediction result to generate a target training set; The test set is adjusted based on the second prediction result to generate a target validation set; The initial fusion model is trained based on the target training set and validated based on the target validation set. When the validation is successful, the trained target fusion model is obtained.

6. The method for identifying dishonest customers as described in claim 5, characterized in that, Before predicting the training set and the test set respectively using the pre-trained target decision tree model, target Prim algorithm model, target logistic regression model, and target gradient boosting model to obtain the first prediction result and the second prediction result, the method further includes: Obtain the configuration parameters corresponding to each algorithm model; Configure the training conditions for each algorithm model according to the configuration parameters; Under the training conditions, the initial decision tree model, the initial Prim algorithm model, the initial logistic regression model, and the initial gradient boosting model are trained according to the training set to obtain the target decision tree model, the target Prim algorithm model, the target logistic regression model, and the target gradient boosting model.

7. A device for identifying dishonest customers, characterized in that, The device for identifying dishonest customers includes: The acquisition module is used to retrieve several pieces of data to be identified from the database when a command to identify dishonest customers is received. The selection module is used to select the field information corresponding to each piece of data to be identified from the plurality of data to be identified according to a preset field selection template. The feature generation module is used to generate feature information based on the field information; The identification module is used to input the feature information into a pre-trained target fusion model to obtain the bad customer concentration mining result output by the target fusion model, and to determine the corresponding dishonest customers based on the bad customer concentration mining result. The target fusion model is formed by superimposing and fusing a preset decision tree model, a preset Prim algorithm model, a preset logistic regression model, and a preset gradient boosting model. The bad customer concentration is the ratio of the number of identified bad customers to the total number of bad customers. Select the template generation module to retrieve all field information for several customers from the dataset; Perform dummy variable transformation on all the field information to obtain all feature information; Based on all the feature information, determine the prediction capability information value corresponding to each field item; Based on the predicted ability information value, each field item is filtered to obtain the remaining field items; Generate a preset field selection template based on the remaining field items.

8. A device for identifying dishonest customers, characterized in that, The device includes: a memory, a processor, and a default customer identification program stored in the memory and executable on the processor, the default customer identification program being configured to implement the default customer identification method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a dishonest customer identification program, which, when executed by a processor, implements the dishonest customer identification method as described in any one of claims 1 to 6.

Images